Advances in mobile robotics: proceedings of the Eleventh International Conference on Climbing and Walking Robots and the support technologies for mobile machines, Coimbra, Portugal, 8-10 September 2008 9789812835765, 981-283-576-8
Эта книга рассказывает о современных научных и технических результатах исследований и событиях в области мобильной робот
235 9 99MB
English Pages 1309 Year 2008
CONTENTS......Page 12
Preface......Page 6
Conference organisers......Page 7
Conference committees......Page 8
Conference sponsors and co-sponsors......Page 11
Section-I: Plenary Presentations......Page 28
Development of dance partner robot -PBDR- K. Kosuge......Page 30
From micro to nano robotics B. Nelson......Page 32
1. Introduction......Page 33
2. Classification of adhesion techniques by nature of forces......Page 34
2.1. Pneumatic adhesion......Page 35
2.2. Magnetic adhesion......Page 38
2.4. Chemical adhesion......Page 39
3. Classification on the basis of the need of energy......Page 40
4. Considerations concerning the locomotion methods......Page 41
5.1. ROBINSPEC......Page 42
5.3. SURFY......Page 43
5.4. SCID......Page 44
5.6. VENOM......Page 45
5.7. ALICIA 2......Page 46
5.9. ALICIA 3......Page 47
5.10. SPlDERBOT 2......Page 48
6. Conclusion......Page 49
References......Page 50
Section-2: Autonomous Robots......Page 56
1.1. Advantages of Two-Wheeled Mobile Robots......Page 58
2. Mathematical Modelling......Page 59
2.2. Energy Requirements......Page 60
4. Analysis and Simulation Results......Page 61
4.4. System Energy requirements......Page 64
References......Page 65
1.1. Two- Wheeled Mobile Robots......Page 66
2. System Modelling and Description......Page 67
3. Control Strategy......Page 68
4.1. Motion of the Payload and the COM......Page 69
4.2. Ejfect of D{fferent Disturbance Levels......Page 70
4.3. Effect of Disturbance Duration......Page 71
5. Conclusions......Page 72
References......Page 73
1. Introduction of the Project......Page 74
2.1 Magnet adhesion......Page 76
2.2 Two section structure......Page 77
2.3 Prototype oj the robot......Page 78
3. Control of the Robot......Page 79
5. NDT Inspection......Page 80
References......Page 81
1. Introduction......Page 82
2.1. Dual Pendulum Model......Page 83
2.2. Analysis of the Model in One Step......Page 84
3. Optimization......Page 86
4. Application in the Arrangement of Robot's Hip Yaw Movement......Page 88
References......Page 89
1. Introduction......Page 90
2. Detection of the foot placing......Page 91
3. Classification of sensed torques......Page 92
References......Page 97
1. Introduction......Page 98
2.2. Drive system......Page 99
2.3. Docking mechanism......Page 100
2.5. Ultrasonic-sensor mechanism......Page 101
4. Electrical System & Hardware Integration......Page 102
5. Implementation of Docking......Page 103
References......Page 105
1. Introduction......Page 106
3.1. Tree Construction......Page 108
3.2. Computing Distances......Page 109
3.3. Best Path Computation......Page 110
4. Simulations......Page 111
5. Conclusions......Page 112
References......Page 113
1. Introduction......Page 114
2. Building of a new modular walking robot MERO......Page 115
3. On the shift system mechanisms of the MERO modular walking robot......Page 116
4. The Control system of the MERO modular walking robot......Page 119
5. Further Work......Page 121
References......Page 123
1. Introduction......Page 125
2. Behavior Modules......Page 126
3. Reactive Layer of Behavior Network......Page 127
4.1. Predictive obstacle handling......Page 128
5.1. Obstacle Handler Group......Page 129
6. Related Work and Discrimination......Page 130
7. Conclusion......Page 131
References......Page 132
1. Introduction......Page 133
2. RFID Barriers......Page 135
3. Semantic Map......Page 136
4. Global Planning......Page 137
6. Conclusion and Outlook......Page 138
Acknowledgements......Page 139
References......Page 140
1. Introduction......Page 141
2.1. Vacuum Pad Normal Holding Force Theoretical Model......Page 142
2.2. Electromagnet Normal Holding Force Predictive Model......Page 143
3. Experimental Methodology and Results Obtained......Page 144
4. Discussion of Results......Page 146
References......Page 148
1. Introduction......Page 149
2. Simultaneously localisation in Multiple Topological Paths......Page 151
3.2. Identifying Overlap in View Sequences......Page 153
4. Experiments and Results.......Page 154
References......Page 156
Section-3: Benchmarking and Standardization......Page 158
1. Introduction......Page 160
2. Advisory Group on service robots......Page 161
3. Robots in personal care......Page 162
4. Vocabulary on robots and robotic devices......Page 163
5. Conclusions......Page 164
6. References......Page 165
1. Introduction......Page 166
2.1. Conference Tracks......Page 167
2.2. Benchmarking proposals......Page 168
2.4. Research coordination activities......Page 169
3. Novel approaches......Page 170
References......Page 171
1. Introduction......Page 173
2. Required characteristics for a benchmark......Page 175
3.1. Slalom test......Page 176
3.2. Crinkle test......Page 178
4. experimental results......Page 179
References......Page 180
1. Introduction......Page 181
3. CLAW AR Delphi study on benchmarking......Page 183
4. Example of a conceptual robot design......Page 184
5. CtA W AR benchmarking questionnaire......Page 185
6. Results from the CLAW AR benchmarking study......Page 186
8. References......Page 188
Section-4: Biologically-Inspired Systems and Solutions......Page 190
1. Introduction......Page 192
2. Modular Multi Network Architecture for Learning Grasping Tasks......Page 193
2.2. Learning the Inverse kinematics of the fingers. LMI......Page 194
3. Simulations and Results......Page 197
References......Page 199
1. Introduction......Page 201
2.1. Theory oj Ortony et al. (1988)......Page 202
2.4. MBT! of Meyers-Brigg and Meyers......Page 203
3. Related works......Page 204
4.1. GRACE - fieneric !l..ohotic drchitecture to £reate !J..motions......Page 205
4.2.4. MBTlfor personality......Page 206
References......Page 207
1. Introduction......Page 209
2.1. Studied animals......Page 210
2.2. Experiments in Vivo......Page 211
3.1. The model......Page 212
3.2. Angular trajectories......Page 213
References......Page 215
1. Introduction......Page 217
2. The parametric 3D model of the bones......Page 218
2.1. Studies of normal regime bones mechanical loads......Page 219
3. Modular adaptive implant......Page 220
References......Page 224
1. Introduction......Page 225
2. Kinematics......Page 227
3. Kinetics......Page 229
4. Conclusion......Page 231
References......Page 232
1. Motivation......Page 233
4. Techno-biological analysis and bionics of climbing......Page 234
4.1. Locomotion......Page 235
5. Technical realisation......Page 237
6. Conclusion......Page 239
References......Page 240
1. Introduction......Page 241
2. Mechanical System......Page 242
3. Up / Down Climbing Decision......Page 243
4. Experimental Results......Page 245
References......Page 246
1. Introduction......Page 248
2. Configuration of System......Page 249
3. Analog Neuron Model......Page 250
4. CPG Network......Page 253
5. Conclusion......Page 254
References......Page 255
1 Introduction......Page 256
2 Mammalian leg muscle architecture......Page 258
3.1. Actuators......Page 259
4.1. Work Transfer......Page 260
4.2. Methods......Page 261
4.3.1 Contributions of so and GA to ankle work and p~wer......Page 262
4.3.2 Variation in Timing of so Activation and GA activation......Page 263
References......Page 264
1.1. Ionic Polymer-Metal Composites (IPMC)......Page 265
2. Methodology......Page 266
2.1.1. Individual modeling......Page 267
3. Results......Page 268
3.1. [SAD performance......Page 269
References......Page 272
1. Introduction......Page 274
2. Related Work......Page 275
4. Analysis......Page 276
5. Simulation and Experiments......Page 279
6. Discussion......Page 280
References......Page 281
1. Introduction......Page 282
2. The 3D CLIMBER Robot......Page 283
2.1. Required Torque and Force......Page 284
3. Pneumatic Muscles......Page 285
3.1. Selection of the appropriate PM for the 3DCLIMBER:......Page 286
5. Serial Configuration of Pneumatic Muscles......Page 287
References......Page 288
Section-5: Biomedical Robotic Assistance......Page 290
1. Introduction......Page 292
2. Voluntary Upper Body Effort - An Overview......Page 293
3.l. Indoor Rowing Exercise Model......Page 294
4. Implementation of Control Strategies......Page 295
5. Simulation Results......Page 296
References......Page 299
1. Introduction......Page 301
2. Indoor Rowing Exercise Model......Page 302
3.2. Inference......Page 303
4.1. Genetic Algorithm Optimization Process......Page 304
5. Simulation Results......Page 305
References......Page 307
1. Introduction......Page 309
2.2. Design specification......Page 310
2.3. Muscle Stimulation......Page 312
2.4. Modelling the rowing cycle......Page 313
3. Results......Page 314
References......Page 315
1. Introduction......Page 316
2. Array design......Page 317
3. Steerability......Page 318
References......Page 319
1. Introduction......Page 321
2.1. SBO equipped humanoid model......Page 322
2.3. Simulation of the voluntary hand support......Page 323
3. Results......Page 324
References......Page 326
Section-6: Climbing, Guidance and Navigation......Page 328
1. Introduction......Page 330
2. Alicia VTX robot structure......Page 331
3. Experimental Test bed......Page 332
4. System identification......Page 334
4.1. Model selection......Page 335
5. Conclusion......Page 336
References......Page 337
1. Motivation and state of the art......Page 338
2. Climbing Robot CROMSCI......Page 339
3. Negative Pressure System......Page 340
4. Control System......Page 341
5. Experimental Results With Seven-Chamber Prototype......Page 342
5.1. Generated Forces......Page 343
5.2. Driving Experiments......Page 344
References......Page 345
1. Introduction......Page 346
2.1. Wheel design......Page 347
3. Motivation, goal and approach......Page 348
4.1. Model of a vehicle that is passing a concave corner......Page 349
4.2. Simplified calculation without the effect of gravity......Page 350
5. Mechanical design of a simple test prototype......Page 351
6.1. Tests on concave corners and comparison to calculation results......Page 352
6.2. Other tests with the prototype......Page 353
7. Conclusion and outlook to further work......Page 354
References......Page 355
1. Introduction......Page 356
3. Sliding-sock locomotion......Page 357
4. Modular rescue robots with sliding sock locomotion......Page 358
5. Robot architectures with improved climbing performance......Page 360
6. Conclusions......Page 362
References......Page 363
1. Introduction......Page 364
2. Evolution in the design and control of climbing robots......Page 366
2.1. Configuration design criteria......Page 367
3. Conclusions: Lessons learned and new directions......Page 370
References......Page 371
1. Introduction......Page 372
2. Complete Coverage in DYLEMA......Page 373
3. Experiments......Page 377
References......Page 379
1. Introduction......Page 380
2.1. Reduced-scale environment......Page 381
4. Fire Searching Algorithm......Page 383
5. Experimental Results......Page 387
Acknowledgments......Page 388
References......Page 389
Section-7: Flexible Mechanisms for Mobile Machines......Page 390
1. Introduction......Page 392
3. Rolling soft robot......Page 393
4. Experimental result of rolling soft robot......Page 394
5. Simulation of rolling soft robot......Page 395
References......Page 399
2. Principle of Jumping via Robot Body Deformation......Page 400
3.1. Flexural potential energy......Page 401
3.2. Particle-based model of circular robot......Page 402
3.4. Impulse from the ground during jumping......Page 403
4.1. Realization of a dish shape......Page 404
4.3. Prototype......Page 405
References......Page 406
1. Introduction......Page 408
3. Design and Implementation of Augmented Control Scheme......Page 409
4. Results......Page 411
5. Conclusion......Page 414
References......Page 415
1. Introduction......Page 416
2. A Robotic Catapult based on the Closed Elastica with a High Stiffness Endpoint......Page 417
2.1. Introducing High Stiffness Endpoint......Page 418
2.2. Experiment......Page 419
3. Application to Impulsive Swimming Robot......Page 420
3.1. Start Motion......Page 421
3.2. Change Direction Motion......Page 422
References......Page 423
1. Introduction......Page 424
2. Conventional Robotic Catapult based on the Closed Elastica......Page 425
3.1. Dynamics......Page 426
3.3. Simulation Result......Page 427
4. Basic Experiment......Page 428
5.1. Jump Motion......Page 429
References......Page 430
1. Introduction......Page 432
2.1. Geometry of Curves and Virtual Joints......Page 433
2.3. Closed-loop Structure represented by the External Wrench......Page 435
3. Quasi-static Energy Analysis......Page 437
4. Conclusion......Page 438
References......Page 439
Section-8: Flexible Maneuvering Systems......Page 440
1. Introduction......Page 442
2. Crane description......Page 443
3. The control system design......Page 444
4. Simulation results......Page 445
5. Conclusion......Page 448
References......Page 449
1. Introduction......Page 450
2. System description......Page 451
3. The control system design......Page 453
4. Simulation results......Page 454
References......Page 457
2. The Flexible Manipulator System......Page 458
3. Genetic Algorithms......Page 460
4. Simulation Results and Discussion......Page 461
5. Conclusion......Page 463
References......Page 464
1. Introduction......Page 466
2. Experimental Set-up......Page 468
4.1. Tuning the antecedent part......Page 469
4.2. Tuning the consequent part......Page 470
5. Result and discussion......Page 471
References......Page 472
1. Introduction......Page 474
2.1. Basic structure......Page 475
2.2. Dynamic modelling of the single - link flexible arm......Page 476
3. Feedback control of nonlinear system......Page 477
3.1. Input - state linearization......Page 478
3.2. Pole placement design......Page 479
4. Conclusions and future works......Page 480
References......Page 481
1. Introduction......Page 482
3. Design and Implementation of Augmented Control Scheme......Page 483
4. Results......Page 486
5. Conclusion......Page 488
References......Page 489
1. Introduction......Page 490
2.1. Methodology......Page 492
2.2. Unscented Kalman filter......Page 494
4. Results......Page 495
References......Page 496
Section-9: Human-Machine Interface, Tele-Presence and Virtual Reality......Page 498
2. Problem formulation......Page 500
3. Robot architecture......Page 501
4. Tactile Language......Page 502
5.1. Workflow Overview......Page 503
5.2. Action selection......Page 505
5.3. Proactive execution......Page 506
References......Page 507
1. Introduction......Page 508
3. Measurement Methods......Page 509
3.1. Performance Measurement......Page 510
3.3. Telepresence Measurement......Page 511
4. Results and Discussion......Page 512
5. Conclusions......Page 514
References......Page 515
1. INTRODUCTION......Page 516
2.3. Motor imagery paradigm......Page 518
2.4.1. P 300 - Bayesian Approach......Page 520
2.4.2. Motor imagery - Fisher Linear Discriminant Approach......Page 521
3.2. Motor imagery experiments......Page 522
References......Page 523
1 Introduction......Page 524
2.1 Camera Calibration......Page 525
2.2 Obtaining Morphological feature......Page 526
2.3 Head Segmentation......Page 527
2.4 Head Pose estimation......Page 528
4 Experimental Result......Page 530
5 Conclusion......Page 531
References......Page 532
2.1 The KHR-1 robot......Page 533
2.2 Client-server architecture......Page 534
2.3 Central virtual reality......Page 535
2.5 Feedback from physical platforms......Page 536
3.1 Locomotion......Page 537
3.3 An example fight scenario......Page 538
4. Conclusion......Page 539
References......Page 540
1. Introduction......Page 541
2.2. Exoskeleton Controller (ECO)......Page 543
2.4. The 3D Visualization Client......Page 545
3. Preliminary Tests and Results......Page 546
Acknowledgments......Page 547
References......Page 548
Section-10: Innovative Design of CLA WAR......Page 550
1. Motivation and Requirements......Page 552
2. State of the art......Page 553
3.1. Kinematics......Page 554
3.2. Control System......Page 555
4. Experimental Evaluation......Page 556
5. Discussion......Page 557
6. Conclusions......Page 558
References......Page 559
1. Introduction......Page 560
2. Mechanical Design......Page 561
3. Embedded controller design......Page 563
4. Noise Control......Page 564
5. Experimental work......Page 565
Acknowledgments......Page 566
References......Page 567
1. Motivation to the Work......Page 568
2. Formulation of the Problem......Page 569
3. Technical Solutions......Page 570
4. Process of Operation......Page 571
5. Experimental Tests......Page 572
6. Conclusions......Page 573
References......Page 574
1. Motivation to the Work......Page 576
3. Technical Solutions......Page 577
4. Process of Operation......Page 579
References......Page 580
1. Introduction......Page 582
2. Requirements and defects to be scanned......Page 583
3. Types of defects......Page 584
4. Specification of the COncEPT Prototype tomography scanner......Page 585
5. Payload feasibility considerations......Page 587
6. Climbing robot prototype......Page 588
References......Page 589
Section-11: Inspection and Non-Destructive Testing......Page 590
1. Introduction......Page 592
2. System Design......Page 593
2.1. Locomotion......Page 594
2.2. Control......Page 596
2.3. NDT Equipment......Page 598
References......Page 599
1. Introduction......Page 600
2.2 Design Probe Holder with Linear Actuator......Page 602
2.3 Onboard Couplant Supply System......Page 603
2.4 Monitoring signal stability vs speed of scan......Page 604
2.5 Laboratory experimental results on the vertical weld surface......Page 605
3. Conclusion......Page 606
Reference......Page 607
1. Introduction......Page 608
2. The Semantic Inspection Approach......Page 609
2.1. The Mission Ontology......Page 610
2.3. Inspection Planning......Page 611
3. Realization......Page 612
5. Conclusion and Future Work......Page 614
References......Page 615
1.1 Application of wall climbing robots......Page 616
1.2 The robotic system......Page 617
2.1 The control system......Page 618
2.2 Automated eddy-current inspection of turbine blades......Page 620
2.3 Automated PFM controlled inspection......Page 622
References......Page 623
1. Introduction......Page 624
2.1. Wheel Design......Page 626
2.2. The Chosen Solution......Page 627
3. Wheel Experiment......Page 628
5. Acknowledgement......Page 630
References......Page 631
1. Introduction......Page 632
2. Development of an underwater wall climbing......Page 633
2.2. First prototype climbing robot design......Page 634
2.3. Structural design of the first prototype......Page 635
2.4.1 Robot structure......Page 636
2.4.2 Adhesion force of the robot to vessel wall......Page 637
References......Page 639
1. Introduction......Page 640
2. Robot System Development......Page 641
3. Robot Trajectory for Weld Inspection......Page 644
5. NDT Results......Page 646
References......Page 647
Section-12: Locomotion......Page 648
1. Introduction......Page 650
2. Methods......Page 651
3. Results......Page 652
4.1. General Discussion......Page 653
References......Page 655
1. Introduction......Page 657
2. Design Concepts for Legged Machines......Page 658
3. Biomimetic Design Concept......Page 659
4. Biological Investigation......Page 660
6. Methods......Page 661
7. Simulation Results......Page 662
8. Discussion......Page 663
References......Page 664
1. Introduction......Page 665
2. Method......Page 667
3. Results......Page 668
4. Discussion......Page 670
Acknowledgments.......Page 671
References......Page 672
1. Introduction......Page 673
2. Related Work......Page 674
3. Controlling Dynamic Motions of Bipeds with Reflexes and Motor Patterns......Page 675
4. Initiation of Walking......Page 677
References......Page 679
1. Introduction......Page 681
2. Main Restrictions of Motion Design......Page 682
3. Sample Results......Page 683
References......Page 688
1. Introduction......Page 689
2. System Dynamics......Page 690
3. Periodic Trajectories......Page 692
4. Performance Index......Page 693
5. Results......Page 695
References......Page 696
1. Introduction......Page 698
3. Walking Planning......Page 699
4.2. Mathematical Model from the Input of the Thigh Link to the Attitude of the Body[IJ......Page 700
5.1. 3D Simulations on the Even Terrain......Page 701
5.2. 3D Simulations on the Irregular Terrain......Page 702
References......Page 705
1. Introduction......Page 706
3.1. Walking Planning......Page 707
3.3. Feedback Control Input......Page 708
References......Page 712
1. Introduction......Page 714
2. Dynamic Modeling of the Four-Link, Three-Joint Planar Biped Robot as a Free-floating Robot......Page 715
3.1. Design of the relative trajectory......Page 717
3.2. Maintaining the stable posture while walking with the RTC method......Page 718
4. Experiments......Page 719
References......Page 721
1. Introduction......Page 722
2. FUZZY Q-LEARNING CONCEPT......Page 723
3.1. Low-level control......Page 725
3.2. High-level control......Page 726
4. SIMULATION RESULTS......Page 727
5. Conclusion......Page 728
References......Page 729
1. Introduction......Page 730
2. Markov Decision Processes......Page 731
3. Related Work......Page 732
4.1. Characterization of the Sets S and A......Page 733
4.2.1. Low-level layer......Page 734
5. Results and Discussion......Page 735
References......Page 736
1. Introduction......Page 738
2. Definition of Walking Gaits......Page 739
4. Mode Change Using Methodology-l......Page 740
5. Mode Change Using Methodology-2......Page 741
6. Mode Change in Same Hip Joint Rotation Direction......Page 742
7. Experimental Results......Page 743
7.3. Mode change from W-type to L-type......Page 744
References......Page 746
1. Introduction......Page 747
2. The DLR-Crawler - Mechanical System, Sensors and Electronics......Page 748
2.1. Joint Compliance Control......Page 749
3.1. Predefined Gait Patterns......Page 750
3.2. Coordination Inspired by Cruse's Rules......Page 751
4. Results and Discussion......Page 752
References......Page 754
1. Introduction......Page 755
2.2 Leg Interface......Page 756
3.1.2 Leg shortening......Page 757
3.1.4 Leg swinging in single support phase......Page 758
3.2.1 Up righting......Page 759
3.3 Output Parameters......Page 760
5. Conclusion......Page 761
References......Page 762
1. Introduction......Page 763
2. The six legged robot LAURON IVe......Page 764
3. Behaviour networks......Page 765
4. Behaviour-based Control of LAURON......Page 766
5. Local adaptations for stable walking......Page 768
6. Experiments and real world tests......Page 769
References......Page 770
1. Introduction......Page 771
2. Particle Swarm Optimization......Page 772
3.1. Proposal description......Page 774
3.3. Stability control policy......Page 776
4. Results, discussion and further work......Page 777
References......Page 778
1. Introduction......Page 779
2. Concept of mechanism design......Page 780
3.1. Kinematics of transformable track......Page 781
3.2. Length of track......Page 782
3.3. Calculation and simulation of variable length of track......Page 784
References......Page 786
1. INTRODUCTION......Page 787
2. ANALYSIS OF PASSIVE SUSPENSION SYSTEM......Page 788
3. ANALYSIS OF LFA-V......Page 791
4. RESULTS AND DISCUSSION......Page 792
S. CONCLUSIONS......Page 793
REFERENCES......Page 794
Adaptive stair-climbing behaviour with a hybrid legged-wheeled robot.. M. Eich, F. Grimminger and F. Kirchner......Page 795
2. Adaptive Control for Hybrid Legged-Wheeled Robots......Page 796
3. Results......Page 799
4. Conclusion......Page 801
References......Page 802
1. Introduction......Page 803
2.1. Stability Margin......Page 804
2.2. Stability Potential Field......Page 805
3. Differential Kinematic Model......Page 806
4. Decoupled Control......Page 807
5. Results......Page 809
References......Page 810
1. Introduction......Page 811
2.1.2. Terrain roughness......Page 812
2.2.4. Motion resistance of the transfer leg's wheel in the rolking mode......Page 813
3.1. Indicators for changing from wheeled mode to rolking mode......Page 814
4.1. Structure of automatic locomotion mode control......Page 815
4.2.1. Changing from wheeled to rolking mode......Page 816
5. Experiment with the wheel-legged robot WorkPartner......Page 817
6. Conclusion......Page 818
References......Page 819
Section-13: Manipulation and Gripping......Page 820
1. Introduction......Page 822
2. Robot Description......Page 823
3. Trajectory prediction filter:The Kalman filter......Page 825
4.2. Trajectory prediction graphs......Page 826
5.1. Experimental set up......Page 827
6. Conclusion......Page 828
References......Page 829
1. Introduction......Page 831
2. Master and slave robot Relations......Page 832
3. Teleoperation system model......Page 833
4. Bilateral control by state convergence based in position and velocity references......Page 834
5. Experimental Results......Page 836
Bibliography......Page 838
1. Introduction......Page 839
2. Exploration of the grasp space......Page 840
2.2. Computation of the grasp space......Page 841
3. Case study......Page 842
4. Discussion......Page 845
References......Page 846
1. Introduction......Page 847
2. Problem FormulationIMotivation......Page 848
3. Neural Network Force Controller......Page 849
4. Experimental Results and Conclusions......Page 850
References......Page 854
1. Introduction......Page 855
3.1. Configuration......Page 857
3.2. Materials......Page 858
4. Control......Page 859
5. Results......Page 860
References......Page 862
1. Introduction......Page 863
2.1. Hardware description......Page 864
2.2. Software description......Page 866
3.2. Reading performance experiment......Page 867
4. Conclusions and future work......Page 868
References......Page 869
1. Introduction......Page 870
2. Handling requirements and system basic features......Page 871
3.1. Reconfigurable hanger......Page 873
3.2. Metamorphic gripper......Page 874
3.3. Picking module design......Page 875
Acknowledgement......Page 876
References......Page 877
1. Introduction......Page 878
2. Manipulation framework......Page 879
3. Visual analysis......Page 881
4. Grasp planning and execution......Page 882
5. Implementation and resnlts......Page 883
6. Discussion and conclusion......Page 884
References......Page 885
1. Introduction......Page 886
3. Bond-Graph Model and State Space Equations......Page 888
4. Simulation and Results......Page 889
5. Conclusion......Page 890
References......Page 891
1. Introduction......Page 894
2. The Grasp Mechanics......Page 895
3. The Soft Finger Contact Model......Page 896
4. Contact Size......Page 898
References......Page 900
1. Introduction......Page 902
2. Overall technology and design......Page 903
2.1. Vein localization......Page 904
3. Tactile vein localization......Page 905
4.1. SMAC actuator......Page 906
4.2. Mechanical setup......Page 907
4.3. Initial results......Page 908
5.1. Test setup......Page 909
5.2. Vein or no vein......Page 910
5.3. Practical verification......Page 912
6. Conclusion......Page 914
7. Future work......Page 915
References......Page 916
Section-I4: Modeling and Simulation of CLA WAR......Page 22
1. Introduction......Page 920
2. Hopping vibration - driven robot......Page 921
3. Numerical results of modeling......Page 923
4. Robot prototype modeling......Page 926
References......Page 928
1. Motivation of work......Page 929
2.1. Model-l......Page 931
2.2. Model-2......Page 932
3.2. Forward kinematics......Page 933
4. Results. Computational cost comparison......Page 934
4.1. Computational efficiency......Page 935
5. Discussion......Page 936
References......Page 937
1. Introduction......Page 938
2. Mechanical Design and kinematic models for the quadruped......Page 940
3.1. Hardware......Page 941
3.2. Programming environment......Page 942
4. Velocity control......Page 943
5. Dynamic Behavior Analysis......Page 945
Appendix B. Inverse Kinematic Solutions for the back legs......Page 946
References......Page 947
1. Introduction......Page 949
2. Modular six-legged robot "ANTON"......Page 951
2.1. Hardware control system......Page 952
3. Experimental results......Page 954
References......Page 956
1. Introduction......Page 957
2.2. Orin-Marhefka Model......Page 959
2.3. Proposed Model......Page 960
3. Simulation Results......Page 961
4. Discussion......Page 962
References......Page 963
1. Introduction......Page 964
2. Bipedal Locomotion Patterns......Page 965
3. Nonlinear Oscillators System......Page 967
4. Application of the System......Page 969
References......Page 971
1. Introduction......Page 972
2. LindenLabs' SecondLife......Page 973
3. Physics modeling and simulation......Page 975
4. Physics engine limits as a simulator for climhing and walking rohots......Page 976
5. Discussion......Page 978
References......Page 979
Section-15: Perception, Sensing and Sensor Fusion......Page 980
1. Introduction......Page 982
2.2.1. Gr01lnd-sensor group......Page 983
2.3. Contour following......Page 984
3.1.2. TerTain sweep......Page 985
3.2. Contour following......Page 987
4. Experiments......Page 988
References......Page 989
2. The basics of optical motion measurement......Page 990
2.1. Related work......Page 991
3.1. Basic assumptions......Page 993
3.2. Sensor parameters and the effect of texture......Page 994
3.3. Simulation results......Page 995
References......Page 997
1. Introduction......Page 998
3. Algorithm components......Page 999
4. Implementation......Page 1002
References......Page 1003
1. Introduction......Page 1004
2. Omnidirectional vision system......Page 1005
3. Structure from motion method: motion field estimation......Page 1006
4. Analysing human-robot interaction: study case six-legged robot's manipulator......Page 1008
5. Conclusions......Page 1010
References......Page 1011
1. Introduction......Page 1012
2. The Design of The Calibration Equipment for the Six-Component Force-Torque Sensors......Page 1014
3. The Control of the Walking Humanoid Stability......Page 1016
4. The Control of the Walking Leg Dynamics......Page 1017
References......Page 1018
1. INTRODUCTION......Page 1020
2. DEVELOPMENT......Page 1021
2.2. The kheNose Modules......Page 1022
3. kheNose SOFTWARE ORGANISATION......Page 1023
3.1. kheNose TIM......Page 1025
ACKNOWLEDGMENTS......Page 1026
References......Page 1027
Section-16: Personal Assistance......Page 1028
1. Introduction......Page 1030
2.1 Energy Storage Device......Page 1031
2.2 Dead Points......Page 1032
2.3 Overcoming the Dead Points......Page 1033
3. Results......Page 1035
4. Conclusion......Page 1036
References......Page 1037
2. The sit-to-stand movement......Page 1038
3. The humanoid model......Page 1039
4. Reference trajectory inpnt......Page 1040
6. Results......Page 1041
7. Seesaw model......Page 1042
9. Reference Knee Trajectory......Page 1043
11. Discussion and concluding remarks......Page 1044
References......Page 1045
1. Introduction......Page 1046
2. Proposed gravity compensation system......Page 1048
3. The experimental verification of the function of the proposed mechanism as a gravity compensation system......Page 1050
4. Simulations......Page 1051
References......Page 1053
1. Introduction......Page 1054
2.1. Problem Specification......Page 1055
3.1. Required Condition......Page 1056
3.2. Force Control of Active Walker......Page 1057
4. Seating position adjustment assistance......Page 1058
5. Experiment......Page 1059
References......Page 1061
1. Introduction......Page 1062
2. 4WD Omnidirectional Mobile System......Page 1063
3.1. Static Analysis of Wheel-Step System......Page 1064
3.2. A Chair Tilting Mechanism......Page 1066
4. Step Climbing Experiments......Page 1067
5. Conclusions......Page 1068
References......Page 1069
1. Introduction......Page 1070
2. Wheelchair Model......Page 1071
3. Steering Motion Controller......Page 1072
4. Results & Discussion......Page 1074
References......Page 1076
Section-17: Planetary Exploration and Localization......Page 1078
1. Introduction......Page 1080
2. Measurement of an Earthworm's Peristaltic Crawling......Page 1081
3. Robot......Page 1082
4. Forward Movement on the Ground and in a Tube......Page 1083
5. Turning Experiments......Page 1085
6. Experiment in Dirt......Page 1086
Reference......Page 1087
1. Introduction......Page 1088
2. Mechanical Design......Page 1089
2.2. Leg Mechanism......Page 1090
2.3. Structural Analysis......Page 1091
3. Electrical Design......Page 1092
5. Conclusion......Page 1094
References......Page 1095
1. Introduction......Page 1096
2. Sensor concept and arrangement......Page 1097
3.1. Marker identification......Page 1099
3.2. Marker tracking......Page 1100
4. Experimental setup and results......Page 1101
References......Page 1102
Section-18: Planning and Control......Page 1104
1. Introduction......Page 1106
2.2. Formation Control Problem......Page 1107
3.1. Single View Depth Estimation Scheme......Page 1108
3.2. Control Laws......Page 1110
3.3. Modification fOT Non-Holonomic Agents......Page 1111
4. Implementation......Page 1112
5. Simulation Results......Page 1114
References......Page 1116
1. Introduction......Page 1117
2.1. Snake-like Robot......Page 1118
2.2. Inch-worm Robot......Page 1120
2.3. Loop-like Robot......Page 1121
3.1. Snake-like Robot......Page 1122
4. Mechanical Design......Page 1123
References......Page 1124
1. Introduction......Page 1125
2. Problem Statement......Page 1126
3. Kinematic Model......Page 1127
5. Local Stability Issue......Page 1129
6. Simulation, Results and Discussion......Page 1130
References......Page 1132
1. Introduction......Page 1133
2. Theoretical Background......Page 1134
3. Experimental Setup and Methodology......Page 1137
4. A Discussion on the Test Rig......Page 1138
References......Page 1140
1. Introduction......Page 1141
3. The Control Architecture......Page 1142
3.1. Control Schema......Page 1143
4. Mechatronics......Page 1144
References......Page 1146
Section-19: Service Robots......Page 1150
1. Introduction......Page 1152
2.1. Analysis of Riding Motions......Page 1153
2.2. Saddle Mechanism......Page 1154
2.3. Riding Motion Control......Page 1156
3. Handle Mechanism and Bio-signal Feedback......Page 1157
4. Experiments of Riding Motion Control......Page 1158
Reference......Page 1159
1. Introduction......Page 1160
2. Climbing Robots Applications......Page 1161
3.3. Locomotion using Legs......Page 1162
4.1. Suction Force......Page 1163
4.4. Other Adhesion Principles......Page 1164
References......Page 1165
1.2. Periphery......Page 1168
2.1. Supply chain......Page 1169
2.2. Transportation of media......Page 1170
3.2. Navigation......Page 1171
References......Page 1172
1. Introduction......Page 1173
3. Energy transmission......Page 1174
3.1.1. The first simulation......Page 1175
3.1.2. The second simulation......Page 1176
3.1.6. The third practical experiment......Page 1177
4. Other methods......Page 1178
5. Conclusion......Page 1179
References......Page 1180
1. Introduction......Page 1181
2. Methodology......Page 1182
3. Fusing color and depth information......Page 1183
4. Results......Page 1185
References......Page 1186
1. Introduction......Page 1188
2. Humanoid......Page 1189
4.1. Mechanical properties identification......Page 1190
5.1. Equations of motion......Page 1191
5.4. Results......Page 1193
References......Page 1195
Section-20: Workshop on Humanoid Robotics......Page 1196
Selecting and learning multi-robot team strategies M. M. Veloso......Page 1198
Fractional calculus: Application in control and robotics 1. A. T. Machado......Page 1199
1. INTRODUCTION......Page 1200
2.2. Leg design of Stepper-Senior......Page 1201
3. ELECTRICAL AND CONTROL SYSTEMS......Page 1203
4.2.1. Movement in sagittal plane......Page 1204
4.2.2. Movement in lateral plane......Page 1205
5.1. Forward walking......Page 1206
References......Page 1207
1. Introduction......Page 1208
2. Problem Statement......Page 1210
3. Deterministic Way of Achieving Stable Walking......Page 1211
3.1. Leg Stability......Page 1212
4. Deterministic Way of Achieving Speedy Walking......Page 1213
4.2. Planning of Joint Velocity for Foot Stability......Page 1214
6. Implementation......Page 1215
7. Conclusions......Page 1216
References......Page 1217
1. Introduction......Page 1218
2. Dynamics modelling......Page 1219
3. Results......Page 1223
References......Page 1225
1. Introduction......Page 1226
2.2. Planning motion......Page 1227
2.3. The Generalized ZMP (GZMP)......Page 1228
2.4. Dynamics analysis of ZMP and GZMP......Page 1229
3. Simulation Results......Page 1231
Acknowledgments......Page 1232
References......Page 1233
1. Introduction......Page 1234
2. Problem Formulation of Walking Pattern Generation......Page 1235
3.1. Objectives......Page 1237
3.2.3. Constraint on angular velocity......Page 1238
3.3. SOGP formulation......Page 1239
5. Conclusion......Page 1240
References......Page 1242
1. Introduction......Page 1243
2. Human ankle biomechanics: a brief description......Page 1244
3.1 MIT Foot......Page 1245
3.2 SPARKy 1......Page 1247
3.3 Powered Below-Knee Prosthesis of the VUB......Page 1248
References......Page 1250
1. Introduction......Page 1252
2.1. Thick soft cover......Page 1253
3. Sensors on the surface of mechanical structure......Page 1254
6. Discussion about sensors for sensing whole body contact behavior......Page 1255
7.2. Unexpected physical interaction with human during other task......Page 1257
8. Conclusion......Page 1258
References......Page 1259
1. Introduction......Page 1260
2. Position retargeting......Page 1262
3. Angle retargeting......Page 1263
4. Combined retargeting......Page 1264
5. Experimental Results......Page 1265
6. Conclusions and Future Work......Page 1267
References......Page 1268
1. Introduction......Page 1269
2. Open Dynamics Engine Simulation......Page 1270
3.1. Servo motor and simulator models......Page 1271
3.2. Trajectory planning......Page 1272
4. Experimental Results......Page 1274
References......Page 1276
1. Introduction......Page 1277
2.1. Transformation of General Force Based on Lie Group......Page 1278
2.2. ZMP Detecting in Single Support Phase......Page 1279
2.3. ZMP Detecting in Double Support Phase......Page 1281
3. Experiment......Page 1282
4. Conclusion......Page 1283
References......Page 1284
1. Introduction......Page 1285
2. Vision System Architecture......Page 1286
2.2. Software......Page 1287
3.1. Stimuli-Based Actions......Page 1288
3.2. Image-Based Approach......Page 1289
4. Experiments and Results......Page 1290
5. Conclusions and Future Work......Page 1291
References......Page 1292
1. Introduction......Page 1293
2. Robotics setup......Page 1294
2.1. Software architecture......Page 1295
3.1. CAMSHIFT algorithm......Page 1296
3.2. 3D reconstruction......Page 1297
4. Reaching and grasping preparation......Page 1298
5. Experiments and results......Page 1299
References......Page 1300
Author index......Page 1302
Recommend Papers
![Climbing and Walking Robots: Proceedings of the 8th International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines (CLAWAR 2005) [1 ed.]
3540264132, 9783540264132, 3540264159, 9783540264156](https://ebin.pub/img/200x200/climbing-and-walking-robots-proceedings-of-the-8th-international-conference-on-climbing-and-walking-robots-and-the-support-technologies-for-mobile-machines-clawar-2005-1nbsped-3540264132-9783540264132-3540264159-9783540264156.jpg)
![Climbing and Walking Robots: Proceedings of the 7th International Conference CLAWAR 2004 [1 ed.]
3540229922, 9783540229926, 3540294619, 9783540294610](https://ebin.pub/img/200x200/climbing-and-walking-robots-proceedings-of-the-7th-international-conference-clawar-2004-1nbsped-3540229922-9783540229926-3540294619-9783540294610.jpg)
- Author / Uploaded
- L. Marques
- A. De Almeida
- M. O. Tokhi
- G. S. Virk
- Similar Topics
- Technique
- Electronics: Robotics
File loading please wait...
Citation preview
ADVANCES IN
MOBILE ROBOTICS
This page intentionally left blank
ADVANCES IN
MOBILE ROBOTICS Proceedings of the Eleventh International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines Coimbra, Portugal
8 -10 September 2008
editors
L Marques University of Coimbra, Portugal
A de Almeida University of Coimbra, Portugal
M
o Tokhi
The University of Sheffield, UK
G SVirk Massey University, New Zealand
Ｇｾｷｯｲｬ､＠
Scientific
NEW JERSEY. LONDON· SINGAPORE· BEIJING· SHANGHAI· HONG KONG· TAIPEI· CHENNAI
Published by World Scientific Publishing Co. Pte. Ltd. 5 Toh Tuck Link, Singapore 596224 USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601 UK office: 57 Shelton Street. Covent Garden, London WC2H 9HE
British Library Cataloguing-in-Publication Data A catalogue record for this book is available from the British Library.
The front cover image has been provided by: Professor Philippe Bidaud Institute des Systemes Intelligents et de Robotique, University Pierre and Marie Curie - Paris 6
ADVANCES IN MOBILE ROBOTICS Proceedings of the Eleventh International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines Copyright © 2008 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof; may not be reproduced in any jorm or by any means, electronic or mechanical, including photocopying, recording or any injiJrmation storage and retrieval system now known or to be invented, without written permission ,trom the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.
ISBN-13 978-981-283-576-5 ISBN-IO 981-283-576-8
Printed in Singapore by B & JO Enterprise
PREFACE Robotics has been an exciting field in engineering and natural sciences for many decades. It has held considerable fascination for researchers and scholars and many important contributions have been made by industrial robots in various manufacturing tasks such as assembly, welding, painting, and material handling. In recent times, we have been witnessing the emergence of new service robots which are intended to perform a variety of tasks in new environments such as search and rescue, surveillance, exploration and security missions as well as provide assistance to various users. The emergence of mobile machines for these new service missions in unstructured environments has significantly broadened research and development challenges (both technical and non-technical) that need to be considered for successfully widening their adoption. CLAW AR 2008 is the eleventh in a series of international conferences on Climbing and Walking Robots and the Support Technologies for Mobile Machines. The aim of the conference is to provide an open forum where researchers, scientists, engineers and practitioners from throughout the world can come together to present and discuss the latest achievements, future challenges and exciting applications for mobile service machines in general, and climbing and walking robots in particular. The proceedings of CLA W AR 2008 include state-of-the-art research and development findings presented during the CLAW AR 2008 conference in 153 technical presentations by authors from 32 countries covering the five continents. The editors would like to thank members of the International Programme Committee, International Advisory Committee and National Organising Committee for their efforts in reviewing the submitted papers, and the authors in addressing the comments and suggestions of the reviewers in their final submissions. It is intended that the CLA WAR 2008 proceedings will be a valuable source of reference for research and development in mobile robotics.
L. Marques A. de Almeida M. O. Tokhi G. S. Virk
v
CONFERENCE ORGANISERS Institute of Systems and Robotics Department of Electrical and Computer Engineering University of Coimbra, Portugal
CIAWAR Association A non-profit making membership based professional organization serving the robotics community www.clawar.org
vi
CONFERENCE COMMITTEES General Chairs and Programme Chairs of Eleventh International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines L. Marques (General Co-Chair) A. de Almeida (General Co-Chair)
M. O. Tokhi (Programme Chair) G. S. Virk (lAC Chair)
-
ISR, University of Coimbra, Portugal ISR, University of Coimbra, Portugal University of Sheffield, UK Massey University, New Zealand
International Advisory Committee (lAC) of Eleventh International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines
Manuel Armada Yvan Baudoin Karsten Berns Philippe Bidaud Bryan Bridge Krzysztof Kozlowski Giovanni Muscato Lakmal Seneviratne Ming Xie
- Spain -Belgium - Germany - France -UK - Poland - Italy -UK - Singapore
vii
viii
International Programme Committee (IPC) of Eleventh International Conference on Climbing and Walking Robots and the Support Technologiesfor Mobile Machines Ahmadabadi M. N. Aldbrez F. M. Althoefer K. Azad A. K. M. Arkin R. Carderfa C. Balaguer C. Billingsley J. Bonsignorio F. Bostater C. Bouazza-Marouf K. DaiJ. Dillmann R. Dodd T. J. Dubowsky S. Dutra M. S. Dutta A. Fontaine J.-G. Fu Y. Fujimoto H. Fukuda T. Garcia E. GeS. S. Gonzalez de Santos P. Gradetsky V. Halme A. Hossain M. A. HowardD. Huang L. Jarvis R. Jatsun S. F. KadarE. E. Kaynak O. Kiriazov P.
- Iran -UK -UK -USA -USA - Portugal - Spain - Australia - Italy -USA -UK -UK - Germany -UK -USA - Brazil - India - Italy - China - Japan - Japan - Spain - Singapore - Spain - Russia - Finland -UK -UK -NZ - Australia - Russia -UK - Turkey - Bulgaria
- Japan Kosuge K. - Korea Lee S. Lefeber D. - Belgium - Portugal Machado J. T. - Italy Molfino R. - Korea Moon S. Pipe T. -UK Pota H. R. - Australia - India Pratihar D. K. Preucil L. - Czech Republic Qian J. - China Quinn R. - USA - Russia Rachkov M. Ribeiro M. - Portugal - Spain Salichs M. A. - Portugal Santos V. -UK Sattar T. P. - Germany Schilling K. Shaheed M. H. -UK Shamsudin H. M. A. - Malaysia -UK Siddique N. H. Silva F. - Portugal - Belgium Steinicke L. Vajta L. - Hungary Vitko A. - Slovakia Waldron K. J. -USA WalkerR. -UK WoernH. - Germany Yigit A. - Kuwait - Finland Ylonen S. - China Zheng H. - Singapore ZhongZ. W. - Singapore Zhou C. J. - Poland Zielinska T.
ix
National Organising Committee of Eleventh International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines
Lino Marques Helder Araujo Jorge Dias J. Norberto Pires Rui Araujo Urbano Nunes
-
ISR, University of Coimbra ISR, University of Coimbra ISR, University of Coimbra DEM, University of Coimbra ISR, University of Coimbra ISR, University of Coimbra
Secretariat Office of Eleventh International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines
Cristina Luz Amanda J Virk
- University of Coimbra - CLA WAR Association
CONFERENCE SPONSORS AND CO-SPONSORS
Ie
The Knowledge Network
x
CONTENTS Preface .................................................................................................................. v Conference organisers ......................................................................................... vi Conference committees ...................................................................................... vii Conference sponsors and co-sponsors .................................................................. x
Section-I: Plenary Presentations Development of dance partner robot -PBDR- ...................................................... 3 K. Kosuge From micro to nano robotics ................................................................................ 5 B. Nelson Adhesion techniques for climbing robots: State of the art and experimental considerations ................................................................................. 6 D. Longo and G. Muscato
Section-2: Autonomous Robots Balance control of a TWRM with a static payload ............................................. 31 K. M. K. Goher and M. O. Tokhi Balance control of a TWRM with a dynamic payload ....................................... 39 K. M. K. Goher and M. O. Tokhi A cooperative climbing robot for melt weld inspection on large structures ....... 47 1. Shang, S. Monda I, A. A. Brenner, B. Bridge and T. Sattar Power analysis and structure optimization in the design of a humanoid robot .................................................................................................. 55 L. Wang, M. Xie, Z. W. Zhong, M. Wang and L. Zhang Detection and clustering of the erroneous torques developed in the femur joint of a walking roboL .......................................................................... 63 A. Vitko, L. lurisica, M. Kl'uCik, R. Murar and F. Duchon
xi
xii
A modular mobile self-reconfigurable robot ...................................................... 71 M. Zhong, P. Wang, M. Li and W. Guo Localizing from multi-hypotheses states minimizing expected path lengths for mobile robots .................................................................................... 79 K. Hemanth, S. Subhash, K. M. Krishna and A. K. Pandey Design of new modular walking robot MERO ................................................... 87 I. Ion, J. Simionescu, A. Curaj, V. Luige and A. Vasile Behavior network control for a holonomic mobile robot in realistic environments ......................................................................................... 98 M. Goller, T. Kerscher, J. M. Zollner and R. Dillmann RFID transponder barriers as artificial landmarks for the semantic navigation of autonomous robots ...................................................... 106 M. Goller, F. Steinhardt, T. Kerscher, J. M. Zollner and R. Dillmann Development of predictive models of climbing robot attachment mechanisms for an energy-efficient adaptive control scheme .......................... 114 S. A. Jacobs and A. A. Dehghani-Sanij Merging topological maps for localisation in large environments .................... 122 F. Ferreira, 1. Dias and V. Santos
Section-3: Benchmarking and Standardization ISO standards for service robots ...................................................................... 133 G. S. Virk, S. Moon and R. Gelin The replication of experiments and the performance evaluation in Clawar system research .................................................................................... 139 F. P. Bonsignorio A scalable benchmark for motion control of mobile robots ............................. 146 A. Marjovi and L. Marques Benchmarking of the robot design process ....................................................... 154 G. S. Virk
xiii
Section-4: Biologically-Inspired Systems and Solutions A neural network architecture for progressive learning of robotic grasps ................................................................................................... 165 1. Molina- Vilaplana and 1. Lopez-Coronado GRACE - Generic robotic architecture to create emotions .............................. 174 T. H. H. Dang, S. L. Zarshenas and D. Duhaut Biornimetic approach for tortoise-like robot modeling and control ................. 182 H. El Daou, l.-C Guinot, P.-A. Libourel, S. Renous and V. Bels Application of smart materials - Bionics modular adaptive implant.. .............. 190 N. G. Bfzdoaca, D. Tarn ita, D. Tarnita and E. Bfzdoaca Kinematics and kinetics analysis of rectilinear locomotion gaiL ..................... 198 A. Ghanbari, A. Rostami, S. M. R. S. Noorani and M. M. S. F akhrabadi INSPIRAT - Towards a biologically inspired climbing robot for the inspection of linear structures ..................................................................... 206 1. Maempel, E. Andrada, H. Witte, C Trammer, A. Karguth, M. Fischer, D. Voigt and S. N. Gorb Control of the multi-track type robot inspired from antennae of a centipede ........................................................................................................ 214 T. Chung, K. H. Hyun, C-K. Woo, S. Kim and Y. K. Kwak Implementation of analog controller based on biological nervous system for biomimetics walking robot.. ............................................................ 221 S. H. Kim, T. H. Kang and l.-H. Cho On the design of walking machines using biarticulate actuators ...................... 229 T. 1. Klein, T. M. Pham and M. A. Lewis Modelling and design of IPMC devices ........................................................... 238 I. Chochlidakis, G. S. Virk and A. Dehghani-Sanji Analysis, simulation and implementation of a human inspired pole climbing robot .................................................................................................. 247 A. Sadeghi, H. Moradi and M. N. Ahmadabadi
xiv
A comparison study on pneumatic muscles and electrical motors using the 3DCLIMBER as a case study ........................................................... 255 M. Tavakoli, L. Marques and A. T. de Almeida
Section-5: Biomedical Robotic Assistance Impact of upper body effort in FES-assisted indoor rowing exercise ............... 265 Z. Hussain, M. O. Tokhi and S. Gharooni GA-tuned fuzzy logic control for FES-assisted indoor rowing exercise ........... 274 Z. Hussain, M. O. Tokhi, S. Ahmad and S. F. Toha Compliant control of FES-rowing with energy store and release mechanism ............................................................................................ 282 S. Sareh, B. 1. Andrews, S. Gharooni and M. O. Tokhi An electrode array design for use with a multichannel functional electrical stimulator .......................................................................................... 289 N. Sha, L. P. Kenney, D. Howard, M. Moatamedi, B. Heller and A. T. Barker Body-weight-supported treadmill locomotion with spring brake orthosis ........ 294 M. S. Huq and M. O. Tokhi
Section-6: Climbing, Guidance and Navigation Structure and model identification of a vortex-based suction cup .................... 303 F. Bonaccorso, C. Bruno, D. Longo and G. Muscato CROMSCI - A climbing robot with multiple sucking chambers for inspection tasks ................................................................................................ 311 C. Hillenbrand, D. Schmidt and K. Berns Magnetic wheeled robot with high mobility but only 2 DOF to controL ........ 319 W. Fischer, F. Ti'iche, G. Caprari and R. Siegwart A climbing rescue robot ................................................................................... 329 L. Rimassa, M. Zoppi and R. Moifino
xv Evolution and perspectives of climbing robots at the industrial automation institute. Lessons learned and new directions ................................ 337 M. A. Armada, T. Aki'1fiev, R. Fernandez, P. Gonzalez de Santos, E. Garda, S. Nabulsi, H. Montes, 1. C. Grieco and G. Fernandez Efficient sensor-based path planning for landmine location using walking robots .................................................................................................. 345 A. Ramos, E. Garcia and P. Gonzalez de Santos Sniffing a fire: Simulated experiments in a reduced scale scenario .................. 353 P. Oliveira, L. Marques and A. T. de Almeida
Section-7: Flexible Mechanisms for Mobile Machines Rolling locomotion of a deformable soft robot with built-in power source .................................................................................................... 365 Y. Matsumoto, H. Nakanishi and S. Hirai Jumping via robot body deformation - Mechanics and mechanism for higher jumping ................................................................................................. 373 M. Miyazaki and S. Hirai Augmented control scheme for input tracking and vibration suppression of flexible manoeuvring systems: Experimental investigations ........................ 381 F. M. Aldebrez and M. 0. Tokhi A robotic catapult based on the closed elastic a with a high stiffness endpoint and its application to impulsive swimming robot .............................. 389 M. Watari, A. Yamada, H. Mochiyama and H. Fujimoto A robotic catapult based on the closed elastica with an anisotropic stiffness point and its application to compact continuous jumping robot ................................................................................................... 397 A. Yamada, M. Watari, H. Mochiyama and H. Fujimoto Quasi-static energy analysis of the robotic catapult based on the closed elastica ................................................................................................... 405 H. Mochiyama, A. Yamada and H. Fujimoto
xvi
Section-8: Flexible Maneuvering Systems Modelling and control of an overhead crane with 3DOF ................................ .415 O. A. A. Shaebi and M. O. Tokhi Impact of the hook attachment mechanism on control of an overhead crane ................................................................................................. 423 O. A. A. Shaebi and M. O. Tokhi Genetic algorithm optimization of PID controller for a flexible manipulator ......................................................................................... 431 B. A. M. Zain and M. O. Tokhi Genetic optimisation of ANFIS network for modelling of a TRMS ................. 439 S. F. Toha, M. O. Tokhi and Z. Hussain Control of a single - link flexible arm to be used as a sensing antenna ............ 447 1. C. Fernandez and V. F. Batlle Augmented control scheme for input tracking and vibration suppression of flexible manoeuvring systems: Simulation studies ....................................... 455 F. M. Aldebrez and M. O. Tokhi Output feedback nonlinear model predictive control of a twin rotor MIMO system .................................................................................................. 463 A. Rahideh and M. H. Shaheed
Section-9: Human-Machine Interface, Tele-Presence and Virtual Reality Intuitive human-robot cooperation ................................................................... 473 H. Woern and A. 1. Schmid The influence of human factors on task performance: A linear approach ........ 481 Y. Catsoulis, C. S. Virk and A. A. Dehghani Brain computer interface approaches to control mobile robotic devices .......... 489 C. Pires, U. Nunes and M. Castelo-Branco Stereo camera based head orientation estimation for real-time system ............ 497 Y.-O. Kim and S. fun
xvii Humanoid robot game: A mixture of vr and teleoperation ............................... 506 T. Juhtisz and L. Vajta EXOSTATION: 7-DOF haptic control chain featuring an arm exoskeleton and virtual reality tools ................................................................. 514 P. Letier, M. Avraam, S. Veillerette, M. Horodinca, A. Preumont, J.-P. Verschueren, E. Motard and L. Steinicke
Section-lO: Innovative Design of CLA WAR Dexterous energy-autarkic climbing robot ....................................................... 525 W. Brockmann, S. Albrecht, D. Borrmann and J. Elseberg Wall climbing robotic system and noise control for reconnaissance purpose .................................................................................... 533 P. Wang, M. Li, W. Li and M. Zhong Design of climbing cleaning robot for vertical surfaces ................................... 541 T. Akinfiev, M. Armada and S. Nabulsi Design of wheeled climbing robot with changeable structure .......................... 549 T. Akinjiev, R. Fernandez and M. Armada Climbing ring robot for inspection of offshore wind turbines .......................... 555 H. Leon-Rodriguez, B. Bridge and T. P. Sattar
Section-ll: Inspection and Non-Destructive Testing On the mechanized inspection of glass fiber plastic pipes and pipe joints ....... 565 P. Chatzakos, A. Lagonikas, D. Psarros, V. Spa is, K. Hrissagis and A. Khalid Remote automated non-destructive testing (NDT) weld inspection on vertical surfaces ........................................................................................... 573 S. C. Mondal, A. A. Brenner, J. Shang, B. Bridge and T. P. Sattar Can semantics help autonomous service robots in inspecting complex environments? .................................................................................... 581 M. Ziegenmeyer, K. Uhf, S. Sayler, B. Gassmann, 1. M. Zi5llner and R. Dillmann
xviii
Robotic system for inspection of test objects with unknown geometry using NDT methods .......................................................................... 589 A. A. Brenner and T. P. Sattar A proposed wall climbing robot for oil tank inspection ................................... 597 R. Fernandez-Rodriguez, V. Feliu and A. Gonzalez-Rodriguez Underwater wall climbing robot for nuclear pressure vessel inspection ........... 605 H. E. Leon-Rodriguez, T. Sattar and 1. Shang Amphibious inspection robot ........................................................................... 613 T. P. Sattar, H. E. Leon-Rodriguez and 1. Shang Section-12: Locomotion Stable upright walking and running using a simple pendulum based control scheme ....................................................................................... 623 H. M. Maus, 1. Rummel and A. Seyfarth From biomechanical concepts towards fast and robust robots ......................... 630 D. Renjewski, A. Seyfarth, P. Manoonpong and F. Worgotter From hopping to walking-how the biped Jena-walker can learn from the single-leg Marco-hopper .................................................................... 638 K. T. Kalveram, D. Hdufle and A. Seyfarth Initiating normal walking of a dynamic biped with a biologically motivated control. ............................................................................................. 646 T. Luksch and K. Berns Motion design for an insectomorphic robot on unstable obstacles ................... 654 Y. F. Golubev and V. V. Korianov The effect of leg segmental proportions on the energetic cost of robotic locomotion ........................................................................................... 662 P. Chatzakos and E. Papadopoulos Sliding mode attitude control of a six-legged robot in consideration of actuator dynamics ........................................................................................ 671 H. Uchida and K. Nonami
xix
Generation method of feedback control input of leg link using an attitude sensor for a six legged robot consisting of one link ........................ 679 H. Uchida, Y. Shimizu and S. Nakayama Simple intuitive method for a planar biped robot to walk ................................ 687 G. Chung Obstacle avoidance strategy for biped robot based on fuzzy q-Iearning ......................................................................................................... 695 C. Sabourin, K. Madani, W. Yu and 1. Yan Adaptive locomotion for a hexagonal hexapod robot based on a hierarchical Markov decision process .............................................................. 703 G. Cuaya-Simbro and A. Munoz-Melendez Walking gait control for making smooth locomotion mode change of a legged and wheeled robot.. ........................................................................ 711 T. Okada, W. T. Botelho and T. Shimizu The DLR-crawler: Gaits and control of an actively compliant hexapod .......... 720 M. Goerner and G. Hirzinger Developing fast bipedal locomotion method for inclined floors ...................... 728 A. Eshghinejad and M. Keshmiri Behaviour-based control of the six-legged walking machine LAURON IVc .................................................................................................. 736 T. Kerscher, A. Roennau, M. Ziegenmeyer, B. Gassmann, 1. M. Zoellner and R. Dillmann Particle swarm optimization for humanoid walking-gaits generation ............... 744 N. Rokbani, E. Ben Boussada and A. M. Alimi Mechanism for variable transformation shapes of a single-tracked mobile structure ................................................................................................ 752 1. Kim and C. Lee Optimal posture control for force actuator based articulated suspension vehicle for rough terrain mobility .................................................. 760 V. P. Eathakota, S. Kolachalama, K. M. Krishna and S. Sanan
xx
Adaptive stair-climbing behaviour with a hybrid legged-wheeled robot.. ........ 768 M. Eich, F. Grimminger and F. Kirchner Stability control of a hybrid wheel-legged robot using the potential field approach ................................................................................................... 776 G. Besseron, Ch. Grand, F. Ben Amar and Ph. Bidaud Logic-based automatic locomotion mode control of a wheel-legged robot.. .... 784 1. Leppanen, P. Virekoski and A. Halme
Section-13: Manipulation and Gripping Motion planning to catch a moving object ....................................................... 795 I. Serrano, C. Perez, R. Morales, N. Garda, I. M. Azorln and I. M. Sabater Teleoperation of a manipulator with a master robot of different kinematics: Using bilateral control by state convergence ................................. 804 C. Pena, R. Aracil, R. Saltaren, 1. Banfeld and 1. M. Sabater Influence of the sampling strategy on the incremental generation of the grasp space ............................................................................................. 812 M. A. Roa, R. Suarez and I. Rosell Robot-human cooperation holding and handling a piece of fabric ................... 820 P. N. Koustoumpardis and N. A. Aspragathos A sub € I 000 robot hand for grasping - Design, simulation and evaluation .................................................................................................. 828 1. E. Tegin, 1. Wikander and B. lliev Improving manipulation capabilities by means of radio frequency identification and computer vision ................................................................... 836 I. Sales, X. Garda, P. 1. Sanz, R. Marin, M. Prats and Z. Fa/omir A cooperative gripper for handling and hanging limp parts ............................. 843 E. Carca, M. Zoppi and R. Moifino Robust grasping of 3D objects with stereo vision and tactile feedback ............ 851 B. I. Grzyb, E. Chinellato, A. Morales and A. P. del Pobil
xxi
Bond graph modeling of soft contact for robotic grasping ............................... 859 A. Khurshid and A. Ghafoor Optimum size of a soft-finger contact in robotic grasp .................................... 867 A. Ghafoor and 1. S. Dai Tactile sensing methods for automated blood samples on humans .................. 875 A. S. S¢rensen, T. R. Savarimuthu, E. Pedersen and A. G. Buch Section-I4: Modeling and Simulation of CLA WAR Study of a vibration driven hopping robot... ..................................................... 893 S. Jatsun, V. Dyshenko and A. Yatsun Computational cost of two forward kinematic models for a S-G based climbing robot ........................................................................................ 902 M. Almonacid, R. Saltaren, R. Araci, C. Perez, N. Garcia, J. M. Azorin and J. M. Sabater RobuDOG's design, modelling and control ..................................................... 911 P. Bidaud, S. Barthelemy, P. Jarrault, D. Salle and E. Lucet Walking robot "ANTON": Design, simulation, experiments ........................... 922 M. Konyev, F. Palis, Y. Zavgorodniy, A. Melnikov, A. Rudskiy, A. Telesh, U. Schmucker and V. Rusin A nonlinear model for simulating contact and collision ................................... 930 D. A. Jacobs and K. 1. Waldron U sing nonlinear oscillators to create a pattern generator of bipedal locomotion ........................................................................................... 937 A. C. de Pin a Filho and M. S. Dutra Internet 3.0 for the simulation of networked clawar systems ............................ 945 F. P. Bonsignorio Section-IS: Perception, Sensing and Sensor Fusion Application of lateral obstacle sensor in following contours for terrain recognition tasks ................................................................................... 955 R. Ponticelli and P. Gonzalez de Santos
xx;; True ground speed measurement, a novel optical approach ............................. 963 V. Kalman and T. Takacs Simple optoelectronic exteroceptive sensor for the control of the dynamic equilibrium of a walking robot... ........................................................ 971 E. Kral Analysing human-robot interaction using omni-directional vision and structure from motion ................................................................................ 977 C. Salinas and M. A. Armada Six DOF sensory system for the force-torque control of walking humanoid ............................................................................................ 985 M. Kvasnica kheNose - A smart transducer for gas sensing .................................................. 993 1. Pascoal, P. Sousa and L. Marques
Section-16: Personal Assistance FES-assisted cycling with quadriceps stimulation and energy storage ........... 1003 B. S. K. K. Ibrahim, S. C. Gharooni, M. O. Tokhi and R. Massoud Modelling and simulation of sit-to-stand in humanoid dynamic model... ....... 1011 S. C. Gharooni, M. loghtaei and M. O. Tokhi A new gravity compensation system composed of passive mechanical elements for safe wearable rehabilitation system ........................................... 1019 T. Nakayama, T. Asahi and H. Fujimoto A robotic walker with standing, walking and seating assistance .................... 1027 D. Chugo, T. Asawa, T. Kitamura and K. Takase Step climbing of a four-wheel-drive omnidirectional wheelchair. .................. 1035 M. Wada Steering control of wheelchair on two wheels ................................................ 1043 S. Ahmad, M. O. Tokhi and K. M. K. Goher
xxiii
Section-17: Planetary Exploration and Localization Development of an underground explorer robot based on an earthworm's peristaltic crawling .................................................................... 1053 H. Omori, T. Nakamura and T. Yada Mechanical and electrical design of a two segmental eight-legged mobile robot for planetary exploration ........................................................... 1061 B. Ugurlu, C. M. Uzundere, H. Temeltas and A. Kawamura Visual odometry technique using circular marker identification for motion parameter estimation .......................................................................... 1069 S. Chhaniyara, K. Althoefer and L. D. Seneviratne
Section-IS: Planning and Control Single view depth estimation based formation control of robotic swarms: Implementation using realistic robot simulator ................................ 1079 V. Gazi, B. Fidan and S. Zhai Mechanical design and motion planning of a modular reconfigurable robot ....................................................................................... 1090 A. H. H. A. Memar, P. Z. H. Bagher and M. Keshmiri Approximation control of a differential-drive mobile robot.. ......................... 1098 H. Marin-Reyes and M. O. Tokhi Experimental study on track-terrain interaction dynamics in an integrated environment: Test rig ..................................................................... 1106 S. Al-Milli, S. Chhaniyara, E. Georgiou, K. Althoefer, 1. S. Dai and L. Seneviratne A step toward autonomous pole climbing robots ........................................... 1114 M. Tavako/i, A. Marjovi, L. Marques and A. T. de Almeida
Section-19: Service Robots Development of the riding robot like as a horse and motion control for the healthcare and entertainment.. ............................................................. 1125 M. Lim and 1. Lim
xxiv
Climbing robots: A survey of technologies and applications ......................... 1133 M. F. Silva and 1. A. T. Machado UNIFIER - Unified robotic system to service solar power plants ................. 1141 R. Azaiz A feasibility study for energy autonomy in multi robot search and rescue operations ............................................................................................ 1146 Y. Sindi, T. Pipe, S. Dogramadzi, A. Winfield and C. Melhuish Person following with a mobile robot using a modified optical flow ............. 1154 A. Handa, 1. Sivaswamy, K. M. Krishna, S. Singh and P. Menezes Development of a simulation environment of an entertainment humanoid robot doing a handstand on a high bar ........................................... 1161 P. Teodoro, M. A. Botto, C. Cardeira, 1. Martins, 1. sa da Costa and L. Schweitzer
Section-20: Workshop on Humanoid Robotics Selecting and learning multi-robot team strategies ......................................... 1171 M. M. Veloso Fractional calculus: Application in control and robotics ................................ 1172 1. A. T. Machado Development and gait generation of the biped robot stepper-senior .............. 1173 Y. Liu, M. Zhao, 1. Zhang, L. Li, X. Su and H. Dong A deterministic way of planning and controlling biped walking of LOCH humanoid robot... ................................................................................ 1181 M. Xie, Z. W. Zhong, L. Zhang, L. B. Xian, L. Wang, H. 1. Yang, C. S. Song and 1. Li Inverse dynamics modelling for humanoid robots based in Lie groups and screws .................................................................................... 1191 M. Arbulu and C. Balaguer Human-humanoid robot cooperation in collaborative transportation tasks .... 1199 M. Arbulu and C. Balaguer
xxv
A convex optimization approach for online walking pattern generation ........ 1207 R. Xiong, C. Zhou, L. Zhang and 1. Chu Energy-efficient humanoid walking with ankle actuation: Learning from biomechanics ......................................................................................... 1216 R. Versluys, B. Vanderborght, R. Van Ham, P. Beyl, P. Cherelle and D. Lefeber A fall down resistant humanoid robot with soft cover and automatically recoverable mechanical overload protection .................................................. 1225 M. Hayashi, R. Veda, T. Yoshikai and M. Inaba Retargeting system for a social robot imitation interface ............................... 1233 1. P. Bandera, R. Marfil, R. Lopez, 1. C. del Toro, A. Palomino and F. Sandoval Realistic humanoid robot simulation with an optimized controller: A power consumption minimization approach ............................................... 1242 1. L. Lima, 1. C. Gonqalves, P. 1. Costa and A. P. Moreira A Lie group formulation for realtime ZMP detection using force/torque sensor ......................................................................................... 1250 L. Zhang, C. Zhou and R. Xiong Visual tracking on an autonomous self-contained humanoid robot.. .............. 1258 M. Rodrigues, F. Silva and V. Santos Pose estimation for grasping preparation from stereo ellipses ....................... 1266 G. Saponaro and A. Bernardino Author index ................................................................................................... 1275
This page intentionally left blank
SECTION-l PLENARY PRESENTATIONS
This page intentionally left blank
DEVELOPMENT OF DANCE PARTNER ROBOT -PBDRKAZUHIRO KOSUGE Department of Bioengineering and Robotics, Tohoku University Sendai, Japan
A Dance Partner Robot, PBDR (Partner Ball Room Dance Robot), dances a waltz as a female dancer together with a human male dancer, The waltz, a ball room dance, is usually performed by a male dancer and a female dancer, The dance consists of a certain number of steps, and transition of the steps, which is lead by the male dancer based on the transition rule of the dance, The step transition rule allows the male dancer to select a step from a class of steps determined for the current step so that the line of dance is traced, The female dance partner estimates the following step through physical interactions with the male dancer, The design of the mechanism of the PBDR has been done together with a dress designer, Tastuya Oconogi, The robot mechanism, which has to be put into the robot body designed by the dress designer, is designed so as to have enough number of degrees of freedom for performing the dance, An upper body of the robot consists of two manipulators, each of which has four-degrees-of-freedom for forming a frame, and a head rotating around a neck. A lower body of the robot has an omni-directional mobile base having three degrees of freedom, which is attached to the upper body through a parallel link mechanism driven by three linear actuators. The upper body motion is realized by the parallel link mechanism. The dance robot has a database about the waltz and its transition rule which is used to estimate the following dance step and generate an appropriate step motion. The step estimation is done based on the time-series data of the force/torque applied by the male dancer to the robot upper body. The robot motion is generated for the estimated step using the step motion in the database compliantly against the interface force/moment between the human dancer and the robot in real time. The estimation of the following step, however, could not be done perfectly. We are continuing the development of the robot, and current
3
4
version could watch the human's dance step all the time during the dance and if the step is different from the estimated one, the step is corrected according to the human's step. The development of the dance partner robot suggests us a lot of important issues for robots having interaction with a human. Why we are developing the dance partner robot and how the concept will be applied to other robot systems will be discussed in the presentation.
FROM MICRO TO NANO ROBOTICS BRADLEY NELSON
ETHZurich Switzerland
Robots are currently exploring many environments that are difficult if not impossible for humans to reach, such as the edge of the solar system, the planet Mars, volcanoes on Earth, and the undersea world. The goal of these robotic explorers is to obtain knowledge about our universe and to answer fundamental questions about life and human origins. Microrobotics has entered this field by exploring life at a much smaller scale and more fundamental level. Microrobotic systems for physically exploring the structures of biological cells are being developed, and robotic motion planning strategies are being used to investigate protein folding. Microrobotic mechanisms have been used to investigate organism behaviors, such as the flight dynamics of fruit flies as well as the neurophysiology that govern many other biologically interesting behaviors. These recent research efforts and others like them illustrate how several areas of robotics research are rapidly converging to create this new discipline I refer to as BioMicroRobotics. These new directions in robotics represent only a beginning and indicate that robotics research, and biomicrorobotics in particular, has the capability of making significant contributions in the understanding of life. In moving from the micro domain to nanometric scales, completely different issues in developing nanorobotic systems and in their application arise. The second part of the talk will detail recent efforts at the Institute of Robotics and Intelligent Systems at ETH -Zurich in fabricating nanometer scale robotic components.
5
ADHESION TECHNIQUES FOR CLIMBING ROBOTS: STATE OF THE ART AND EXPERIMENTAL CONSIDERATIONS DOMENICO LONGO and GIOVANNI MUSCATO
Dipartimento di Ingegneria Elettrica Elettronica e dei Sistemi, Universita degli Studi di Catania, viale A. Doria 6, 95125 Catania, Italy Climbing robots are now widely accepted as valid options in situations where it is important to move on sloped or vertical structures in order to inspect, paint, clean or perform the required operations. Even if the first applications of climbing robots appeared more than 40 years ago, many new ideas are continuously being proposed in the scientific literature and in the market. In this work a classification of the different adhesion techniques proposed for climbing
robots is proposed and discussed. Adhesion methodologies can be classified as active when they require an external energy supply to support the robot, or passive if no energy is needed (e.g. permanent magnets or suction cups). Another classification can be done on the basis of the nature of the forces required to support the robot: pneumatic, if the adhesion force is generated by a pressure difference; magnetic if the force is magnetic; mechanical if it depends only on mechanical supports, chemical if it is due to some particular glue, or electrostatic. Moreover within each of these categories, different kind of robots have been proposed in the last years, also on the basis of the locomotion architectures: walking with legs, frame walking, with wheels, sliding, jumping, etc. Recently biologically-inspired gripping methods, trying in many cases to imitate gecko skin, appeared in several research works. However some doubts remain concerning the applicability of such systems in real applications. Some critical considerations on the different techniques and on their practical advantages and drawbacks will be exposed and an overview of the different climbing robots developed in the last 12 years at University of Catania is also presented.
1. Introduction
The main challenge of a climbing robot is to fight against gravity. Different approaches have been explored in the last years and new adhesion techniques are continuously proposed in the research literature. A complete survey on the hundreds of climbing robots developed worldwide is beyond the purpose of this paper, where we will mention only some particular examples of climbing robots, in order to better classify each adhesion technique. An interesting history of many of the original climbing robots developed by the University of Portsmouth, during the last decades is
6
7
reported in [1], while an overview of some applications of climbing robots for nondestructive testing is described in [2]. A summary of some climbing robots developed in Japan and in China is reported in [3] and [4] respectively; while in [5] a survey of commercial applications is reported. It is interesting to observe that, although the first examples of climbing robots were developed more than 40 years ago, a recent article published in the prestigious IEEE journal Spectrum, was considering "vertical surfaces and climbing as the new frontier for robotic research"! [6] Climbing robots are usually adopted in all those sectors where it could be dangerous for human to operate directly. Typical applications areas include inspections, cleaning and maintenance, repairing in industrial and civil structures, but many others have been proposed. An important step toward the dissemination of knowledge of climbing robots has been established in 1998 through the formation of the CLAW AR network, initially funded by the European Commission and then in 2005 converted to an association [7]. The network, through meetings, newsletters and the organisation of the yearly CLAW AR conference, represents a central point for the dissemination of research activities on climbing robots [8] [9]. In particular the proceedings of the CLAW AR conferences are the main source of reference for the last developments in climbing robots. Several special issues in different journals have been also organised by the network [10]. In the following sections a classification of adhesion techniques by the nature of forces will be exposed and then other classifications based on energy or locomotion typology are proposed. Finally an overview of the different climbing robots developed in the last 12 years at University of Catania is also presented with some concluding remarks.
2. Classification of adhesion techniques by nature of forces Adhesion is fundamental to climb. Adhesion forces must be able to counteract gravity force caused by the weight of the robot, when the robot is moving upside-down on an horizontal surface, or to generate a vinculum reaction, due to the presence of the adhesion and the friction, when moving on a vertical wall. Adhesion techniques for climbing robots can be classified with respect to different principles.
8
The first classification that we propose is done on the basis of the nature of the forces needed to remain attached to the wall. The following typologies have been found into the literature, with a few examples of hybrid solutions: • • • •
•
PNEUMATIC == When the adhesion force is generated from a difference of pressure. MAGNETIC == Valid only on ferromagnetic surfaces, when the force is generated by a permanent magnet or an electromagnet. MECHANICAL == When the robot is capable through the adoption of mechanical hinges or hooks to remain attached to the surface. CHEMICAL == When particular chemical substances are adopted that generate adhesive forces between the robot and the surface. In some cases the real origin of these forces at a microscopic level is mechanical (since the strong friction of the adhesive is the real nature of adhesion), or electrostatic (if Van der Waals forces are exploited), or sometime pneumatic (micro suction). However for the purpose of this classification these forces are considered at a macroscopic level as chemically generated. ELECTROSTATIC == When an electrostatic force is generated between the robot and the surface.
In the remaining part of this section these adhesion methodologies will be reviewed individually and some examples and considerations expressed.
2.1. Pneumatic adhesion This is probably the most adopted method for adhesion in climbing robots. One of the first examples of pneumatic adhesion robots are those systems using classical suction cups, usually adopted for picking parts, as robotic feet. The vacuum can be generated by using an electrical vacuum generator, on board or external to the robot and connected by a pipe, or by using a pneumatic (Venturi effect based) vacuum generator. In this last case a strong consumption of pneumatic energy is required, but neither long vacuum pipes nor heavy electrical vacuum generators on board the robot are needed. These robots moves walking or frame walking, are usually slow in motion and needs some redundancy in the feet, to increase reliability in case of the failure of a suction cup. The surface of the wall needs to be flat enough, clean and non-porous to ensure a suitable sealing of the suction cups. Moreover usually the individual suction cups should not be connected simply in parallel to a vacuum generator, to avoid that the loss
9
of negative pressure inside a single suction cup, due to a leak, is propagated to the other cups, causing a sudden loose in adhesion. On the positive side there is the possibility for such robots to pass over obstacles and the fact that really cheap systems can be built, since most of the robots can adopt COTS components from pneumatic gripping manufacturers. Particular examples of this category are the ZIG-ZAG, really simple robot built in University of Portsmouth [11], and the ROBUG II, considered as one of the first example of climbing robot capable to make a transition from floor to wall [12]. Many different other configurations have been realised using suction cups. Among these are really interesting some Japanese examples of robots using tracks of suction cups designed in such a way to be automatically connected and detached from the vacuum generator [62]. Another important pneumatic category of robots are those using a large vacuum chamber and wheels for locomotion by sliding the vacuum chamber. In this case the most important factor is related to the particular sealing that must be adopted between the chamber and the surface. In fact the sealing system cannot be completely hermetic, such in the case of classical suction cups of the previous category, otherwise the robot could not move, but not too weak, otherwise air leakage would be too high and a low pressure difference would result in an insufficient adhesion force. In this case an important advantage is the possibility to climb on not perfectly flat or clean walls (Fig. 1) and to climb over small obstacles (usually lower than lcm). Several examples have been realised adopting this principle as the BIGFOOT built by Portech Ltd [13], the CROMSCI robot developed by the University of Kaiserslautern with seven vacuum chambers [14], and many other robots developed by the Harbin Institute of Technology in China [4],[15J. The CROMSCI is rather interesting since it adopts separate chambers that can be excluded from the vacuum generator in case of air leakage. A particular combination of single modules has been also proposed in the ALICIA 3, to permit this robot to pass over obstacles [16]. A big drawback of these systems is caused by the large amount of power needed to generate the vacuum inside the chamber, since the compensation of the unavoidable loss of air in the leakages is needed. In some cases even an internal combustion engine has been adopted as the power source for the aspirator. The pressure difference to be generated is rather low, since the surface of adhesion is large, however the air flow needed is high.
10
An important improvement in pneumatic adhesion techniques has been established by Duke University with the adoption of the Vortex technology [17]. This category is similar in principlc to thc previous one; however no sealing is required between the cup and the surface, since a vortex generated inside the cup creates a pressure difference between the inner part and the outer part of the cup that does not require sealing. In this case a smaller power is required to generate the adhesion force and no friction is present between the cup and the surface.
Leakage
Leakage
Figure 1. Example of vacuum ehamber adhesion. Working principle and test on a not-perfectly-flat wall.
As a consequence faster and completely autonomous robots have been built capable to safely overcome obstacles or leaks in the wall. Moreover these robots are capable to easily transit from the ground to a 90 degrees wall. However at present the physical phenomena is not deeply understood and a precise characterisation of the relations between the power of the vortex generator and the generated adhesion forces has not been done. Only a few examples of small robots exists using vortex adhesion and several doubts concerning the scalability of these systems remain. Another application of vortex can be done for underwater cleaning robots. In this case the vortex generated by a rotating brush creates an adhesion force that can be exploited to support the robot.
11
When working underwater it must be considered that the task is simplified by the help of buoyancy forces. An example of this category is the ROBT ANK designed for inspecting petrochemical tanks [18].
2.2. Magnetic adhesion Magnetic adhesion is the most reliable form of adhesion that can be adopted when a ferromagnetic surface is available. In fact really strong forces can be generated easily by using rare-earth magnets or strong current electro-magnets. Important applications are the inspection and repairing of ships or surface vessels in general, or storage tanks. The classification can be done between robots with legs, peeling robots and robots with wheels. Robots with legs have usually electromagnets in the feet that are switched sequentially according to the walking sequence to allow the supporting legs to stay in contact with the surface, while the other legs move. An example of this category is the REST robot designed by IAI-CSIC Spain [19]. In some case to minimise energy, permanent magnets have been proposed connected to magnet circuits mechanically switched. An advantage of this method is the capability to pass over small obstacles but a drawback is the low speed that can be reached and the need to adopt redundant systems to avoid the loss of adhesion of a single foot. Recently another method has been proposed using permanent magnets named as compliant distributed magnetic adhesion [20]. In this case each foot is composed by a matrix of permanent rare-earth magnets connected to a flexible support. An interesting feature of this method is the so-called detachment procedure by peeling. Magnet robots with wheels can adopt permanent magnetic wheels with the strong advantage to minimise the gap between the magnet and the surface. However, really flat and clean surfaces are needed since the magnets are usually made by hard materials. Otherwise classical mobile robot soft wheels and a permanent magnet at a small distance to the surface can be used. A serious drawback of the adoption of permanent magnets is caused by the attraction of ferrous dust that little by little is accumulated into the magnets. If long time operations are performed in not really clean surfaces this aspect can determine serious problems to the robot. Several examples of robots adopting permanent magnets have been realised. An interesting application has been proposed by Cybernetyx with the OCTOPUS
12
robot used to clean surface of ships or by some robots developed by South Bank University London for NDT [21].
2.3. Mechanical adhesion It is the simplest way to remain attached to a surface especially in the case of
particularly structured surfaces, or when it is possible to suitably modify the structure. Many different examples exist and their structure depends from the specific application. Two examples developed by the Universidad Carlos III de Madrid are the ROMA family of robots, designed to climb on metallic structures using special grippers to move [22], and MATS an assistive personal climbing robot capable to move in three dimensional space [23]. Another interesting category is represented by those robots designed to climb over trees or over or inside cylindrical structures such as pylons, pillars or pipes. An example is the pole climbing robot proposed to climb on palm tree to harvest the palm oil fruit bunch [24]. Recently a bio inspired system has been proposed that uses micro-spines in order to adhere to the micro-irregularities in the surface of the wall. These robots have claws, hard nail-like structures with no compliances that penetrate the climbing substrate, or spines that latch onto small asperities in the surface. Spinybot [25] and RiSE [26] are two examples of this category. Many other examples of robots adopts ropes to climb on different types of surfaces generating the adhesion force as a resultant of the vinculum reaction caused by the gravity force and the rope traction. For example the ROBOCLIMBER, developed within an European project, is an interesting example of teleoperated climbing robot for slope consolidation and landslide monitoring [27].
2.4. Chemical adhesion In some case simple adhesive tape has been connected to the surface of the feet that are sequentially attached and detached from the surface. However since the adhesive force tend to decrease rapidly after a few sequences, also due to the unavoidable presence of dust, this solution is not really useful in practical applications. It has been also proposed to use automatic tape dispensing systems
13
to replace periodically the adhesive surface, but in any case a limited mobility is present. Interesting examples are the Mini-Whegs robot [28], the StickyBot [29], and the Waalbot [30]. In the future maybe particular electrically controllable adhesives will be produced, but in any case the accumulation of dust remains a strong obstacle toward their applicability.
2.5. Electrostatic adhesion It is not really completely understood. but it is now well established that
the incredible climbing capability of the Gecko lizard over a wide variety of different surfaces and slopes is due also to the presence of van der Waals electrostatic forces [31], [323. This phenomena has been tried to be reproduced in many robots, however Geckos are biological and have obviously self-cleaning and self-repairing capabilities that are actually far to be reproduced in artificial systems. In semiconductor industry electrostatic chucks are adopted to produce distributed adhesion to manipulate silicon wafers [33]. However these are adopted in vacuum and with very flat surfaces. Recently SRI International has proposed an electroadhesion system to control adhesion in climbing robots. This technology involves inducing electrostatic charges on a wall substrate using a power supply connected to 2 compliant pads situated on the robot [34]. A clamping pressure up to 1,5N/cm to conductive and non-conductive surfaces has been experimented with small power consumption and fast switching capabilities. This is a really promising approach still at a first stage and in the future we will see if this method will outperform all the others.
3. Classification on the basis of the need of energy Another important classification can be done on the basis of the need of energy to support the robot. In theory to stay attached to a wall in a fixed position no energy is required. As a consequence several methods can be classified as passive, since do not consume energy to remain fixed. For example robots that adopt permanent magnets, or passive suction cups, or mechanical gripper belongs to this
14
category. These robots can be really useful in all those situations where the system should remain for long periods in a fixed position. Otherwise other robots have been designed that needs a power source to remain attached. For example, robots that use electromagnets, or pneumatic suction cups, or even mechanical grippers connected to electric motors, dissipate energy just to remain in a fixed position. These systems can be classified as adopting Active adhesion techniques. It should be observed that we are dealing with the adhesion methods and it does not matter if energy is adopted or not for locomotion. There are also several examples of robots that are active or passive with respect to the locomotion, but in this work we are concentrating on the adhesion methods.
4. Considerations concerning the locomotion methods Most of the classical locomotion methods proposed for mobile robots have been experimented in climbing robots: Walking, Frame Walking, Sliding, Wheeled, Hybrid, Tracked, Brachiating (arms with grippers) methodologies have been proposed. Some methods are more useful for some typology of forces, some other not. Another possible classification concerning the locomotion methods of climbing robots regard their level of mobility. In particular there are 1-D mobility robot that move on flat surface or on tracks or ropes, 2-D mobility robot where full motion over a plane is possible and 3-D mobility robots when a full climbing mobility is allowed, for example they are able to perform a transition from a plane to another [10]. Finally it is important to consider a classification in autonomous climbing robots, when the power supply is on-board, and non-autonomous robots when an umbilical cable supplies the power needed. In this last case the capability of movement is limited by the cable, but these robots have greater payload capabilities and there are no limits in the duration of the operations. It must be observed that in many climbing robot applications safety reasons lead in any case to the adoption of a safety rope, to support the robot in case of undesired detachment from the surface. In this last case inserting a power supply cable is not a big problem and considerably increases the capabilities of the system.
15
S. Experimental considerations In the Robotic Laboratory of DIEES of University of Catania since 1996 we experimented many different type of robots, here quickly classified and commented in historical order.
S.1. ROBINSPEC
Figure 2. The ROBJNSPEC robot with three legs and 12 electromagnets for adhesion.
Magnetic switching, Active, Walking, Autonomous This robot has been designed for the inspection of petrochemical tanks. An interesting feature was its capability to move by using three legs. During the movement of the legs each motor was responsible for the motion of a single leg, while when the body was moving, all three motors acted concurrently to generate the motion [35], [36], [37 J.
16
5.2. WALLY
Cilinders
Figure 3. The WALLY robot and its kinematic architecture.
Pneumatic Suction cups, Active, Walking, Non-Autonomous It represents a very simple and cheap solution for climbing. The main drawbacks were the low speed in movement, the need of an umbilical cable for pressure and vacuum supply and the low reliability in general [38].
5.3. SURFY
Figure 4. The SURFY robot.
17
Pneumatic suction cups, Active, Frame Walking, Non-Autonomous
In this classical configuration the chassis was built by using Plexiglas. This fact allowed to reduce cost and weight and also to simplify manufacturing. Accurate design by means of finite element method was performed [39 J, [40].
5.4. SCID
y
x
Figure 5. The SCID robot and an example of trajectory.
Magnetic switching, Active, Sliding. Autonomous.
This robot was capable to move passively on a wall. It can go only downwards but no energy is required for the motion. As a consequence it is completely autonomous and very small and lightweight [41], [42].
18
5.5. ALICIA 1
Figure 6. The ALICIA 1 during a climbing robot competition.
Pneumatic Vacuum Chamber, Active, Wheeled, Non-Autonomous Simple demonstration of climbing robot with a vacuum chamber, designed for the CLAW AR climbing robot competition (Winner in Paris CLAW AR 2002). It has several infrared sensors that permit it to move autonomoLlsly avoiding obstacles in the path [601,[61].
5.6. VENOM
Figure 7. The VENOM robot.
19
Magnetic Constant Force, Passive, Wheeled, Autonomous. Another robot designed for the CLAW AR competition (Madrid, 2004). In this case an optical mouse was adopted as odometer in order to localise the robot on the surface. A small rare-earth magnet inserted between the wheels guarantees the needed support force [60], [61].
5.7. ALICIA 2
Figure 8. Alicia 2 testing and drawing of an internal view.
Pneumatic Vacuum Chamber, Active, Wheeled, Non-Autonomous. It is the basic module of vaeuum chamber for pneumatic adhesion system. It has
been designed as a demonstrator and to perform research activity on the control of the pressure inside the chamber [43], [44], [45], [46], [47], [48].
20 5.8. SPIDERBOT 1
Figure 9. The SPIDERBOT 1 and a view of its eight suction cups.
Pneumatic suction cups, Active, Frame Walking, Non autonomous. It is an evolution of the SURFY robot with a stronger mechanical chassis and improved control system.
5.9. ALICIA 3
Figure 10. ALICIA 3 workiug principle and test on a concrete wall.
21
Pneumatic Vacuum Chamber, Active, Hybrid Wheeled/Walking, Non-autonomous. Three identical modules connected together by means of two arms. In this way the robot can pass over medium height obstacles [16], [49], [50], [51], [521, [53], [54], [551.
5.10. SPlDERBOT 2
Figure II. The SPfDERBOT 2 robot is an evolution of SPfDERBOT 1 with a higher payload.
Pneumatic suction cups, Active, Frame Walking, Non-Autonomous It is an evolution of the SPIDERBOT 1, more reliable with larger payload capabilities and stronger mechanical construction.
22
5.11. ALICIA VTX
Figure 12. The ALIC1A VTX robot during a horizontal/vertical transition in cooperation with a mobile rohot.
Pneumatic Vortex, Active, Wheeled, Autonomous.
This robot is capabJe also to move from the floor to a wall, has been adopted also to show the cooperation between a mobile robot and a climbing robot [56],[57].[581·
6. Conclusion If we consider a matrix when we put on one axis the nature of the adhesion force
and on another the locomotion typology and inside the cells an example of a robot, the following classification can be obtained: Table I. Adhesion force type versus locomotion architecture. Pnenmatic
Mechanical
Chemical
Electrostatic
Wheeled
[43]
[21J
[64]
??
??
Walking
[12J
[19]
[25]
[29]
[34]
Frame Walking
[39]
[37J
[27]
??
??
Tracked
r62]
[66]
??
??
??
Hyhrid
[16]
??
??
[28]
??
Sliding
??
[41]
'!?
??
??
[22]
??
??
23 This table could be furthermore divided into several categories for each force typology and with respect to the active or passive nature of the adhesion force. Many empty cells exist in this table, some of them are impossible to build, some are useless, some are unknown to the authors, but probably some other if experimented could give innovative solutions or suggestions for new typologies. From this overview it results that climbing robots are a real interesting topic of research in robotics. At present all commercial applications are based on magnetic and on pneumatic adhesion forces; in the future some new and interesting robots will be discovered.
7. Acknowledgments The authors would like to acknowledge the contribution given to this research by the many students of the Engineering Faculty of the University of Catania that in the last 12 years helped us with their work and their fantasy to the development of all our robots.
References 1. J. Billingsley, "Climbing up the wall", in "Advances in Climbing and Walking Robots", Proceedings of the 10th International conference CLAWAR 2007, Singapore 16-18 July 2007, pp. 5-14. 2. B. Bridge, "Climbing Robots for Nondestructive Testing: Historical th Perspective and Future Trends", Proceedings of the 10 International conference CLAWAR 2007, Singapore 16-18 July 2007, pp. 25-32. 3. A. Nishi, "Development of wall-climbing robots", Computers Elect. Eng. ,Vo!. 22, No.2, pp 123- 149, 1996. 4. Wang Yan, Liu Shuliang, Xu Dianguo, Zhao Yanzheng, Shao Hao & Gao Xueshan, "Development & Application of Wall-Climbing Robots", Proceedings of the 1999 IEEE Int. Conf. on Robotics & Automation, Detroit Michingan, May 1999, pp 1207-1212. 5. K. Berns, C. Hillenbrand, T. Luksch, "Climbing Robots for Commercial Applications - a survey" 6th International Conference on Climbing and Walking Robots (CLAWAR) - 771-776, September 1719,2003 - Catania, Italy. 6. P.Patel-Predd, "Wall-Climbing Robot Spies", IEEE Spectrum, May 2008. 7 . CLAW AR Association Web site: http://www.clawar.org.
24
8. G.S. Virk, "The CLAW AR project: developments in the oldest robotics thematic network", IEEE Robotics & Automation Magazine, Vol. 12, N.2, June 2005, pp.14 - 20. 9. G.S. Virk, G. Muscato, A. Semerano, M. Armada and H.A. Warren, "The CLAWAR project on mobile robotic", Industrial Robot: An International Journal Vol.3 1, • N. 2, pp. 130-138,2004. 10. C. Balaguer, G.Virk, M. Armada, "Robot Applications Against Gravity", IEEE Robotic and Application Magazine, March 2006, pp. 56. 11. J. Billingsley, A.A. Collie, B.L. Luk, "A Climbing Robot with Minimal Structure", Proc. lEE conf Control 91, March 1991, Edinburgh, pp. 813-815. 12. B.L. Luk, A.A. Collie, J. Billingsley, "Robug II: an intelligent wall climbing robot", IEEE Conference on Robotics and Automation, April 1991, Sacrtamento USA, pp. 2342-2349. 13. C. Hillenbrand, K. Berns, F. Weise, J. Koehnen, "Development of a Climbing Robot System for Non-destructive Testing of Bridges", Proceeding 8th IEEE Conference on Mechatronics and Machine Vision in Practice, 2001, Hong Kong. 14. C. Hillenbrand, D. Schmidt, K. Berns, T. Leichner, T. Gastauer, B. Sauer, "Development of a sealing system for a climbing robot with negative pressure adhesion", 10th International Conference on Climbing and Walking Robots (CLAWAR) pp.115-124, July 16-18, 2007. 15. Pan Peilin,Wang Yan , "The Sealing Property of the Sucking Disc for the Wall-climbing Robot", Robot. Vol.18, No.4, 1996, pp. 217-220. 16. D. Longo, G. Muscato, "The Alicia3 climbing robot for automatic wall inspection", IEEE Robotics and Automation Magazine, Vol. 13, N.l, pp. 42-50, March 2006. 17. hUp:llsecure.pratt.duke.edu/pratcpress/web.php?sid= 186 18. ROBT ANK INSPEC web site, http://www.robtank.com 19. J.C. Grieco, M. Prieto, M. Armada, P. Gonzales de Santos, "A SixLegged Climbing Robot for High Payloads", Proc. Of the 1998 IEEE Int. Conf. on Control Applications, Trieste, Italy, 1-4 Sep 1998, pp. 446-450. 20. Berengueres, J.; Tadakuma, K.; Kamoi, T.; Kratz, R., "Compliant Distributed Magnetic Adhesion Device for Wall Climbing", IEEE International Conference on Robotics and Automation, 2007 10-14 April 2007 Page(s): 1256 - 1261. 21. J. Shang, B. Bridge, T. Sattar, S. Mondal, A. Brenner, "Development of a climbing robot for inspection of long weld lines", Industrial Robot, Vol. 35, N.3, 2008, pp. 217-213.
25 22. C. Balaguer, A. Gimenez, CM. Abderrahim, "ROM A robots for inspection of steel based infrastructures", Industrial Robot, Vol.29, N.3, pp. 246-251, 2002. 23. C. Balaguer, A. Gimenez, AJ. Huete,A.M. Sabatini, M. Topping, G. Bolmsjo, "The MATS robot: service climbing robot for personal assistance", IEEE Robotics & Automation Magazine, Vol. 13, N. 1, March 2006, pp. 51 - 5S. 24. Ripin, Z.M. Tan Beng Soon Abdullah, A.B. Samad, Z. "Development of a low-cost modular pole climbing robot", Proceedings TENCON 2000, Volume: 1, pp. 196-200, Sep 2000, Kuala Lumpur, Malaysia. 25. A. T. Asbeck, S. Kim, M. R. Cutkosky, W. R. Provancher, and M. Lanzetta, "Scaling hard vertical surfaces with compliant microspine array", Int. 1. Rob. Res., 25(12): 1165-1179,2006. 26. M.1. Spenko, G.c. Haynes, J.A. Saunders, M.R. Cutkosky, A.A. Rizzi, R.1. Full, D.E. Koditschek, "Biologically Inspired Climbing with a Hexapodal Robot", Journal of Field Robotics, 200S. 27. Cepolina, F. Maronti, M. Sanguinet, M. Zoppi, M. Molfino, R.M., "Roboclimber versus landslides: design and realization of a heavy-duty robot for teleoperated consolidation of rocky walls", IEEE Robotics and Automation Magazine, Vol. 13, N. I, pp.23-31, March 2006. 2S. K.A. Daltorio, S. Garb, A. Peressadko, A. D. Harchler, R. E. Ritzmann, R.D. Quinn, "A Robot that Climbs Walls using Micro-structured Polymer Feet", Proceedings of the Sth International Conference on Climbing and Walking Robots CLAWAR 2005, London, Springer, Sep. 2005. 29. D. Santos, B. Hayneman, S. Kim, N. Esparza, M.R. Cutkosky, "GeckoInspired Climbing Behaviors on Vertical and Overhanging Surfaces", Proceedings of the IEEE International Conference on Robotics and Automation ICRA 200S, Pasadena, April 200S. 30. M. P. Murphy and M. Sitti. Waalbot, "An agile small-scale wall climbing robot utiiizing dry elastomer adhesives", IEEE/ASME Transactions on Mechatronics, 12(3), pp. 330-33S, 2007. 31. K. Autumn and A. Peattie, "Mechanisms of adhesion in geckos", Integrative and Comparative Biology, 42(6), pp.lOS1-1090, 2002. 32. K.Autumn, "BiologicalAdhesives", vol. XVII. Berlin,Germany: Springer-Verlag, 2006. 33. K. Asano, S. Aonuma and F. Hatakeyama, "Fundamental Study of an Electrostatic Chuck for Silicon Wafer Handling", IEEE Trans. on Industrial Applications, Vol. 3S, No.3, May/June 2002. 34. SRI Web site: http://www.sri.com/rd/electroadhesion.html 35. G.Muscato, "Soft-Computing Techniques for the Control of Walking Robots", IEE COMPUTING & CONTROL Engineering Journal, (IEE London, U.K.) Vol. 9, N. 4, pp193-200, August 1995.
26 36. L. Fortuna, A. Gallo, G. Giudice, G. Muscato, "Sensor Fusion to Improve the Control of a Mobile Walking Robot for Automatic Inspection: The ROBINSPEC System", Proceedings of the 6th IMEKO International Symposium on Measurement and Control in Robotics ISMCR96, Brussels (Belgium), 9-11 Maggio 1996, pp.376-380. 37. L. Fortuna, A. Gallo, G. Giudice, G. Muscato, "ROBINSPEC: A Mobile Walking Robot for the Semi-Autonomous Inspection of Industrial Plants", in Robotics and Manifacturing: recent trends III research and applications, Vol. 6, AS ME PRESS New York (USA), pp. 223-228, Maggio 1996. 38. G. Muscato, G. Trovato, "Motion control of a pneumatic climbing robot by means of a fuzzy processor", First international symposium CLAW AR '98 Climbing and Walking Robots, Brussels, 26-28 Novembre 1998. 39. G. La Rosa, M. Messina, G. Muscato, R. Sinatra, "A low-cost lightweight climbing robot for the inspection of vertical surfaces", Mechatronics, (Pergamon Press, Exeter U.K.),Vo1.l2, N.l, pp.71-96, Jan. 2002. 40. G. La Rosa, M. Messina, G. Muscato, "SURFY: A low weight surface climbing robot", 16th IAARC/IFAC/IEEE International Symposium on automation and Robotics in Constructions, Madrid, Spain, 22-24 September 1999. 41. D. Longo, G. Muscato, "A Small Low-Cost Low-Weight Inspection Robot with Passive-Type Locomotion", Integrated Computer-Aided Engineering, Vol. 11 N. 4, pp. 339-348, 2004. 42. D. Longo, G. Muscato, "SCID: A non actuated robot for walls exploration", Proceedings of the 2001 IEEE/ASME International Conference on Advanced Intelligent Mechatronics, pp.874-879, Como, Italy July 2001. 43. D. Longo, G. Muscato, "Design and Control Methodologies for the Climbing Robot Alicia II", 5th International conference on Climbing and Walking Robots, CLA WAR 2002, Paris France, September 25-27, 2002. 44. D. Longo, G. Muscato, "Neural control of the climbing robot Alicia", 1st International IEEE/EURON Workshop on Advances in Service Robotics ASER 2003 , Bardolino (Verona), March 2003. 45. D. Longo, G. Nunnari, G. Muscato, "Neural network system identification for a low pressure non-linear dynamical subsystem onboard the Alicia 2 climbing robot", 13th IFAC Symposium on system Identification, Rotterdam (The Netherland), 27-29 August 2003. 46. D. Longo, G. Muscato, "Design of a Single Sliding Suction Cup Robot for Inspection of non Porous Vertical Wall", Proceedings of the 35th
27
47.
4S.
49.
50.
51.
52.
53.
54.
55.
56.
57.
International symposium on Robotics ISR 2004, Paris France, March 23-26, 2004. D. Longo, G. Muscato, "A modular approach for the design of the Alicia3 climbing robot for industrial inspection", Industrial Robot: An International Journal, Vol. 31, N.2 ,pp. 14S-15S, 2004. G. Cacopardo, D. Longo, G. Muscato, "Design of the Alicia3 robot - A modular approach", 6th International Conference on Climbing and Walking Robot, Catania (Italy), pp. SOI-SOS, 17-19 September 2003. D. Longo, G. Muscato, "Control Architecture for the Alici aJ\3 Climbing Robot", Proceedings of the World Automation Congress 2004, Seville Spain, June 2004. D. Longo, G. Muscato, "Adhesion control for the Alicia3 climbing robot", 7th International Conference on Climbing and Walking Robot, Madrid (Spain), September 2004. D. Longo, G. Muscato, "A climbing robot for risk reduction in petrochemical industrial inspection", CISAP-1. 1st Italian Convention on Safety & Environment in Process Industry, Palermo, Italy, November 2S-30, pp.261-266, 2004. D. Longo, G. Muscato, S. Sessa, "Simulator for locomotion control of the Alicia3 climbing robot", Proceedings of the Eight International Conference on Climbing and Walking Robots and the support Technologies for Mobile Machines CLAWAR 2005, London U.K., 1315 September 2005. D. Longo, G. Muscato, S. Sessa, "Simulation and locomotion control for the Alicia3 climbing robot", 22nd International Symposium on Automation and Robotics in Construction ISARC 2005 - September 1 114,2005, Ferrara (Italy). S. De Francisci, D. Longo, G. Muscato, "A Direction Dependent Parametric Model for the Vacuum Adhesion System of the Alicia II Robot", Proceedings of the IEEE 14th Mediterranean Conference on Control and Automation, Ancona, Italy, June 28-30 2006. D. Longo, G. Muscato, S. De Francisci, "A State-Space Direction Dependent Parametric Model for the Suction Cup of the Alicia II Robot", European Control conference 2007, Kos, Greece, 2-5 July 2007. D. Longo, D. Melita, G. Muscato and S. Sessa, "A Mixed Terrestrial Aerial Robotic Platform for Volcanic and Industrial Surveillance", IEEE International Conference on Safety, Security and Rescue Robotics 2007, Rome (Italy), 27-29 September 2007 C.Bruno, D. Longo, D. Melita, G. Muscato, S. Sessa, G. Spampinato, "Heterogeneous robot cooperation for intervention in risky environments", Proceedings of the 7th IARP Workshop Robotics and
28
58.
59.
60.
61.
62.
63.
64.
65.
66.
Mechanical assistance in humanitarian demining and similar risky interventions HUDEM 2008, Cairo, Egypt, 28-30 March 2008. F. Bonaccorso, C. Bruno, D. Longo, G. Muscato, "Structure and model identification of a vortex-based suction cup", Proceedings of the I Jth Conference on Climbing and Walking Robots, CLAW AR 2008, 8-10 September 2008, Coimbra Portugal. G. Muscato, D. Longo (Editors), Proceedings of the Sixth International conference on Climbing and Walking robots and their Supporting Technologies CLAW AR2003, Professional Engineering Publishing, Bury St Edmunds and London, UK, September 2003 (ISBN I 86058 4098). D. Longo, G. Muscato, "Climbing Robot Competition experience at University of Catania", First International Conference on Dextrous and Autonomous Robots and Humanoids, Yverdon-Ies-Bains, Switzerland, 19-22 May 2005. D. Caltabiano, D. Longo, C.D. Melita, G. Muscato, S. Sessa, "CLAW AR Competitions", EUROBOT Workshop on Educational Robotics, Acireale, Italy, I June 2006. Zhu J.; Sun D.; Tso S-K, "Development of a tracked Climbing Robot", Journal of Intelligent and Robotic Systems, Vol. 35, NA, Dec. 2002, ppA27 -443. V. Gradetsky, "Wall Climbing Robot: Evolution to Intelligent Autonomous Vehicle", Proceedings of the first International Symposium CLAW AR 98, Brussels ｾＶＭＲＸ＠ Nov. 1998, pp. 53-60. T. Okada, T. Sanemori, "MOGRER: A Vehicle Study and Realization for In-Pipe Inspection Tasks", IEEE Journal of Robotics and Automation, Vol. RA-3, N.6, Dec. 1987. Kotay, K. and Rus, D. "Navigating 3d steel web structures with an Inchworm robot", Proceedings of the 1996 International Conference on Intelligent Robots and Systems, Osaka, 1996. W.Shen, J. Gu, Y. Shen, "proposed Wall Climbing Robot with Permanent Magnetic Tracks for Inspecting Oil Tanks", Proceedings of the IEEE Int. Conference on Mechatronics & Automation, Niagara Falls, Canada, July 2005.
SECTION-2 AUTONOMOUS ROBOTS
This page intentionally left blank
BALANCE CONTROL OF A TWRM WITH A STATIC PAYLOAD KHALED M K GOHER and M 0 TOKHI Department of Automatic Control and Systems Engineering, The University of Sheffield, UK. k,mourad@shefac,uk This work presents investigations into controlling a two-wheeled robotic machine (TWRM) with a payload positioned at different locations along its intermediate body (ill), Two types of control techniques are developed and implemented on the system modeL The traditional proportional-derivative (PD) control and fuzzy logic (FL) controL An external disturbance force is applied to the rod that constitutes the ill in order to test the robustness of the developed controllers. The simulation results of both control algorithms are analyzed on a comparative basis for the developed two control techniques.
1. Introduction The type of intelligent robot proposed in this work is a mobile robot with a twowheeled inverted pendulum. This design was chosen because its mechanism has an innately clumsy motion for stabilizing the robot's body posture. The robot has a body with two wheels for moving in a plane and a head similar to a human head for controlling the motion. Two independent driving wheels are used for position control, for fast motion in a plane without casters, and for some actions that help the robot appear to be intelligent. The mechanical structure of a robot with only two driving wheels is similar to an inverted pendulum. Compared to common inverted pendulums, the two wheeled inverted pendulum robot has different, complicated problems, [8]. The inverted pendulum (IP) is one of the most popular examples used for illustrating various control techniques. The goal of controlling the IP is to balance the pendulum in the upright position when it initially starts with some nonzero angle off the vertical position. The dynamics of balancing a pendulum at the unstable position can be employed in the applications of controlling walking robots, rocket thrusters, etc, [8]. 1.1. Advantages of Two- Wheeled Mobile Robots Two-wheeled machines have different applications due to their own advantages which arise from the special design. For example, a two wheeled vehicle may be
31
32
safer for the occupants while simultaneously being more agile to navigate narrow city streets. Furthermore, the reduced volume and lower mass of this configuration would increase fuel efficiency and overall functionality. However, because such a vehicle would be inherently unstable it would require an intelligent control mechanism to provide dynamic balancing. 2. Mathematical Modelling
The dynamic characterisation of the robot is described in this section by introducing the governing main equations describing the mathematical model of the two-wheeled robotic machine (TWRM). The model is derived based on the Newton-Euler equations of motion. Consider Figures I and 2 which represent a schematic diagram of the vehicle and free body diagram of the intermediate body (IB) with the external applied disturbance force F.
z
V,.
Fig. 1 Schematic diagram of the vehicle
Fig. 2 Schematic diagram of the IE
Manipulating the basic dynamic equations of the DC motor powering the vehicle, the vehicle wheels and the IB of the machine yields the following two first order non-linear differential equations describing the motion of the system under the effect of an applied payload and impact disturbance force on the rod:
8"
= (
(1
)
2 )
Ig + M +M" Lg
(1)
(21;---
.
-
!
Leg! Leg2 ... ,Leg6 Ｍ｟ｾ＠
.---
f--- - -
I
-'-0
ｾＭｯ＠
-,--
i
---' , '
"
Ｍｾ＠
lnteiface 2 ,Source;
i
.----r-- - , I i
.--
Walking robot's Stance wlatfonnmeasuring -'- - -signal
I
I
I
Position measuring - i- Stength measuring - ---signal signal - --.-- j C1l3Tll ｾ＠ Ｍｾ＠ (:t==== , , , HC!2'\T12 Leg I'
ｈｃｉＳｾｔＱ＠
-c.c
.
Figure 4 Data acquiring system block scheme
- calculate and check the movement control, regulate the posItIon, its force sensors detect the steps, automatically generates the movement laws (space, speed, acceleration), real time interpolation, automatic estimation of the offset servo-valves, convert direct and indirect coordinates systems, attached to the walking robot's elements [5]. The mechanical, hydraulic, electrical and conducting systems are functionally inter-connected and they cooperate in fulfilling the duties the operator claims. Nevertheless, a precise and stable functioning requires that the axis position be automatically maintained according to the internal position reference in the computer's memory, so that the robot's axes could precisely and repeatedly carry out the movement laws generated by the robot's guidance program. As it is shown, in the computer's memory records two automatically run functions, namely those of the off set adjustment, estimation and compensation.
5.
Further Work
In order to achieve the complex mISSIons, the walking robot, planned to be autonomous, needs a hierarchically intelligent control due to its specific natural working conditions, where inaccurate descriptive information and data concerning its own movements as well as the environment are received. The main components of the walking robot control system are shown in Fig. 5. The goal and complexity implied by the realization of such a technical machine make necessary the existence of an integrated system for the development of mechatronic equipment. The mechanical system will be modular. A module of
95 the mobile system will be made by a central body to which two legs are joined and will received within its structure the necessary elements of the actuator system. The actuation will be performed by means of a complex device that is mounted on the central body [4]. The multi-sensor system consists of (fig 5): sensors to survey the system components (parameters of the actuating, supplying and distributing system); Testinf!, and Development Monitorin"D device
-Positioninf!, and con trol./ocl/ssinf!, ,Specific contraltoI' , data acquisition
Kerboard ,' HDD l'Epro, m/R,-AM- monitor, Testing MOllse: FDD Disk Unvolatile _device
I,'
--t =-c
Ｍｾ＠
-r=
HDD+ FDD SVGA COM interface AT Bus
Ｌｾ＠
-
-
-
, TV servo camera- - - --TV TV controller - ｾ＠ (on-line) .
Ｍｾ＠
IBM compatible PC _ _ ._ with stand-by workink - - - - - - ' - - - Peljornllnf!, ｾ＠ ,panel _ Interface moduls DUO. AUO i L L___ - '/ +interlocs AT - Bus ｾ＠ Controller for ｾ＠ Direct controls--r--r "'" radio -confirmations- - - ｾ＠ - - -, --------.l...- ＬＭｾ＠ emission/reception Interface extension Robot interfd't:e ". - , f - -.. ＾ｾ＠ -------.- - .....,- ,- 'SU;nd-by,:'
-----.J"-
r-- ,---
r'System sensors,
rr- -----T
I
ｾ｟ＯＭ］
-=--
Robot sensors"
Servo robot comand
Robot
ｩＭ｟Ｋｾ＠
-1-ｾ＠
Motorl,------
I
I
ｇ･ｮｲ｡ｴｯｾＧ＠
_ Tii)!(Jraulic - - , - ,
ｾｹｳｴ･ｭ＠
I
,-'
-,--220Vac/50Hz NOTE" Three axes controller for each leg
TV picture : V 'Controller forradio
'-::--"
emlSSlon
--
--,-' reception
Power with , accumulators
,
1
-
-Start system control -P rocessed picture , Programs
- -
I Master
ＭｾＬ＠
station (PC)
I
Figure 5 The proposed scheme for the control system of the modular mechatronic system
- sensors to survey and control forces, positions, speeds, accelerations and targets; - sensors to scan route, proximity, to video control in order to acquire images, as well as the computerized system needed for image analysis, parameter extraction, coordination of the navigation systems etc. The system of the information transfer and processing will consist of the module for data transmission and the controller of the data and images transmission and reception.A proper work and functioning of the autonomous motion systems implies the existence of a careful coordination between motion planning, environment perception and control performance, in order to get a suitable behavior within loosely structured environments. For each operation
96
there is hierarchically architecture, which integrated the sensorial information, covering the whole perception function from the lowest to the highest programming level. The integral system works within a stage sequence that repeats itself cyclically, namely: it senses - it decides - it operates. The operating and control system works by interacting with the environment and operates taking into account its changes in real time. The functions of this system are: - to control in real time the multi-articulated mechanical system of the walking robot; - to continuously supervise of the sensorial information; - to process the primary data at the sensorial level and to transform them into coded information about the environment conditions; - to take the necessary decisions as a response to the changes that have appeared within the robot-environment interaction; - to control the hydraulic servo-distributors within the actuating system.
6.
Conclusions
The new modular walking robot MERO made by the authors is a multi-functional mechatronic system designed to carry out planned movements aimed at accomplishing several scheduled tasks. The walking robot operates and completes tasks by permanently inter-acting with the environment where there are known or unknown physical objects and obstacles. Its environmental interactions may be technological (by mechanical effort or contact) or contextual ones (route identification, obstacle avoidance, etc) The successful fulfillment of the mission depends both on the knowledge the robot, through its control system has on the initial configuration of the working place, and by those obtained during its movement.
References 1.
2.
3.
4.
A.P. Bessonov, N.V. Umnov, The Analysis of Gaits in six-legged Robots according to their Static Stability, Proc. Symp. Theory and Practice of Robots and Manipulators, Udine, Italy, 1973 Denavit J., Hartenberg R.S., A Kinematic Notation for Lower Pair Mechanisms Based on Matrices, Journal of Applied Mechanics, Tr. ASME, 1955, Vol. 77 1. lon, 1. Simionescu,. A. Curaj, (2005) MERO Modular Walking Robot Support of Technological equipments, The 8 th International Conference on Climbing and Walking Robots, September 12-14, 2005 London, UK Ion, 1., Stefanescu, D.M., (1999) Force Distribution in the MERO FourLegged Walking Robot, ISMCR'99 - Topical Workshop on Virtual Reality and Advanced Human-Robot Systems, vol. X, Tokyo, Japan
97 5.
6.
7.
Hiller M., Muller J., Roll U., Schneider M., Scroter D., Torlo M., Ward D. (1999). Design and Realization of the Anthropomorphically Legged and Wheeled Duisburg Robot Alduro, The Tenth World Congress on the Theory of Machines and Mechanisms, Oulu, Finland, June 20-24 S.M. Song, and J.K. Waldron, An Analytical Approach for Gait Study and its Application on Wave Gait, International Journal of Robotics Research, Vol. 6, No.2, 1987, pp. 60-71 J.K. Waldron, Modeling and Simulation of Human and Walking Robots Locomation , Advanced Scholl Udine, Italy 1996
Behavior Network Control for a Holonomic Mobile Robot in Realistic Environments Michael Goller, Thilo Kerscher, 1.Marius Zollner and RUdiger Dillmann Interaktive Diagnose- und Servicesysteme (IDS), Forschungszentrwn Informatik an der Universitdt Karlsruhe (FZI), www.Jzi.de. Karlsruhe, 76131, Germany £-mail: (goellerlkerscherlzoellnerldilimann)@Jzi.de
A main challenge in the field of service robotics is the navigation of robots in human everyday environments. Supermarkets, which are exemplary chosen here, pose a challenging scenario because they often have a cluttered and nested character. They are full of dynamic objects. Especially the presence of large numbers of people is a special challenge to cope with. It can often be difficult to map the locality because of a frequently changing environment. Therefore a broad approach is needed to cover three main tasks: the reactive local navigation, the interaction with dynamic objects and the flexible global navigation and task planning. A three layered navigation concept was developed where each of these fields is dealt with in a dedicated layer. This paper presents the bottom layer of this three-layered approach. The focus lies on providing an reactive local navigation system that is able to accomplish movement tasks in a complex scenario. The control is based on behavior networks. To enhance the manoeuvrability of the robot (Fig. I) it uses a holonomic drive system with mecanum wheels. This paper is associated with the CommRob* project. Keywords: Behaviour Network, Behaviour Based Control, Holonomic Drive
1. Introduction
Among the most important and most basic features of an autonomous robot is the ability to move to a certain location without colliding with any obstacle in the path. A navigation capable of this is often based on a metrically map. But what happens if the environment changes too often to measure it exactly? Here the robot has to find its way on its own in close touch with the real environment. At this point the idea of a combination of topological navigation and a behavior-based control for local navigation gets involved. A topological map can be built based on the definitively static structures of the environment like rooms and corridors without the need of exact measurement. When moving inside such a node the robot receives *www.commrob.eu, contract number IST-04544l under 6 th framework programme
98
99
Fig, I,
The holonomic robot InBot
Fig, 2, The layered structure of the navigation concept. The focus in this paper lics on the bOltom layer, the Reactive Layer.
instructions on how to reach the desired target. Three main fields of duty can be identified here, In bottom up order these are navigation inside a local topological area, the handling of dynamic objects and the interaction with humans and on the highest level the task and topological route planning. As fourth task the mapping can be mentioned, but this can not be covered here. In a work from B.Kuipers2 a detailed hierarchical approach on learning cognitive or topological maps in such scenarios can be found. A three layered navigation concept (Fig.2) was developed to cover the mentioned three identified tasks. This paper focuses on the bottom layer of this three layered approach which is called Reactive Layer. The goal of this layer is to fulfill movement tasks in a local area. The second layer, that is like the third one not described in detail in this paper, contains a behavior based control to deal with dynamic objects and with the interaction with nearby humans. The top layer contains a global topologic navigation and is explained in detail in a corresponding work. I The last two layers arc mentioned for the general idea only. A sketch of the navigation system is displayed in Fig. 3.
2. Behavior Modules A behaviour, which is the basic unit of this behavior-based control, is a small software-module which is dedicated to a single task (see Fig. 4). All behaviors of such a control work independently and parallel at all the time. They range from reactive behaviors like turning away from the wall to deliberative behaviors like freeing the way for another robot. Typically the lowest level behaviours, the reflexes, generate the motor commands while the higher behaviours orchestrate the
100
rig. 3. The navigation conccpt: !\ global navigation based on a topological map instructs a local navigation with modules for predictive and reactive obstacle handling as well as safety oriented modules.
lower ones. Besides calculating the dedicated output (n) based on given input (e), which represents the data flow, each behavior has the possibility to interact with other behaviors directly, which represents the control flow. A behavior can be motivated (m) or inhibited (i) by other behaviors, and each one can inform others about how active (a) it is and how satisfied it is with the actual situation (r). This information can be used by other behaviors to estimate how efficiently the first one is working and accordingly motivate or inhibit it. Using these capabilities, individual behaviors can be woven into behavior networks. In such a network a whole set of behaviors combine their strictly dedicated work to fulfill a higher task. Additionally we use a special kind of very simple behavior modules, the fusion behaviors. These unify several inputs of the same type to one output of the same type. This is necessary when several modules provide input for a one single other module. The most prominent types are the maximum fusion, where on input is selected based on the activity of the source module and the weighted fusion where all inputs are weighted by the activity and then merged for the output.
3. Reactive Layer of Behavior Network The Reactive layer is split into two levels itself (see Fig. 5). The task-oriented level deals with the selection or generation of (sub-)targets, so each module generates a target vector it wishes the lower modules to execute. One of the targets is chosen by a maximum fusion behavior. The behaviors in the object-oriented level are based on a dynamic potential field method using repelling or attracting vectors as
101
Control
e Data e . data input vector u dala output vector
Pig. 4. A single behaviour module with the interface (e) and (u) for the data to be processed, (r),(a) for the control output and (m),(i) for the control input for the module.
Fig. 5. The behavior network is divided into two levels: the top onc is for the task generation and to bottom one for the handling of local objects. Each level has a dedicated fusion behavior.
common language. Each module contributes one vector and all vectors are merged by a weighted fusion.
4. Task-oriented Level The top level of the behavior network is instructed to fulfill a certain task and the appropriate data is provided. In the case of a target location that is to reach, first a predictive obstacle handling takes place. Other possible goal oriented data sources are the velocity vectors generated for example by a module responsible for person following or the set-point vector acquired by the input of the robot trolley's haptic handle. The vectors are merged in a maximum fusion what means that one data source is selected. This vector is handed down to lower potential field-based level as data source.
4.1. Predictive obstacle handling If a target location location is known, therefore the task is a point-to-point movement instead of e.g. the haptic control of the robot, predictive obstacle handling
102
Fig. 6. A sequence of sub-targets is generated on the Fig. 7. The robot automatically avoids way to the target location, forming some kind of dy- U-shaped obstacles or moves out of namic visbility graph. them by choosing the best suited corner as sub-target.
takes place to overcome the local minima problem of the potential field approaches 34 and to generate more efficient and shorter paths. If the direct path to the target is blocked the obstacles are scanned for corners. The corners are then ordered by their quality that depends on the corner's distanee and direction compared to the robot and to the target as well as the free room around it. The best corner is chosen as new sub-target and temporarily replaces the original target. This way the robot is driving on some kind of dynamic visibility graph. A brief example is given in Fig. 6 and 7. 5. Object-oriented Level In this level the task-oriented attracting vector is mcrged with attracting and repelling vectors added by several special behaviors. These are grouped in two subnetworks or groups. The group for avoidance of static obstacles and the group for the reactive handling of dynamic objects. The latter one highly depends on information acquired in the Tactical Layer (Fig.2) so it will not be explained in this paper. The data basis for this level are occupancy maps (A.Elfes5 , S.Thrun6 ). This way a highly dynamic potential ficld is created that depends for example on thc robot's velocity vector an thercfore eannot be caleulatcd preliminary.
5.1. Obstacle Handler Group The most important modules are for the obstacle avoidance (Fig. 8). Instead of the commonly used two instanees for left- and right-hand obstacles here are used two for eaeh of them to take the holonomic drive into account. Here the actual driving direction and the scheduled driving direction in that is accelerated can differ remarkably. So for both of them the obstaclc avoidance has to bc active. The
103
Obstacle Handler Group
Fig. 8. The sub-network for the avoidance of static obstacles. The modules for right- and left-handed avoidance have to he instantiated twice with respect to the ho]onomic drive of the robot.
repelling vector 1lAO consists of the unified vector CP Fl - CoG ) from the obstacles weighted Center of Gravity (Pcoc) to the robot, the weight of the CoG itself (VV(Pcoc ), the velocity of the robot (/I), a factor j, the motivation 7n and the inhibition i. The activity a is proportional to ilAO and the rating T is proportional to the sum of weights representing the unhappiness of the behavior. The weight WO for a point or a cell of the occupancy map depends on their distance (dO) (0 the robot and their angular position (00) compared to the rclevant direction, e.g. the actual movement vector or the direction in which the robot will accelerate. Both values are checked if they arc located within a certain active area (. .. A). An extract of a Matlab simulation of this behavior is illnstrated in Fig.9. 1'.1
. W(2:= Oi . W(Oi))
'Ivl'
j·m· (Ii)
i=l 2
d(Px,PH);O) ·ma:x: (OA
O(Px,PT);O) M
M
L Oi' W(Oi),
(]
IluAOII,
T
I LW(Oi)11 Ｇｩｾｬ＠
These modules arc assisted by special modules that are only active in special situations. If the Avoid Obstacle modules have an equally strong activity two behaviors are motivated. The first one looks for narrow gaps roughly in the desired direction and leeds the robot through, the other one forces a decision for one direction if the robot stands in front of a wall. The attracting and repelling vectors are merged by a weighted fusion and so a resulting vector is generated that defines the new acceleration for the robot.
6. Related Work and Discrimination 1997 M. Mataric 7 discriminated control architectures between deliberative, reactive, behaviour-based and hybrid. The modular character of the behaviour-based
104
Fig. 9. Entry and passage of a small con'idor in a Matlab simulation. Red and green: repelling vectors from obstacles, light blue: attracting vector to target, dark blue: resulting vector after fusion
Fig. 10. The robot is avoiding two obstacles and Fig. II. Avoidance of aU-shaped obthen passing a narrow passage without oscillating stacie due to the predictive obstacle handling in the task-oriented level
approach simplifies the system's design but new problems arise, like the behaviour coordination and fusion. These problems are analyzed in various works like from P.Althaus 8 . The most proposals are based on special behaviour selection mechanism or arbilers. 9 The approach presented in this document uses activity, target rating, motivation and inhibition values for each behaviour to discriminate the control flow from the data flow as well as very simple fusion behaviours to merge different data flows. This approach is based on works from J.Albiez lo for the behavior-based control of walking machines. Other derivations, besides the one presented in this document, are actually used to control walking machines II or wheel driven robots like RAVON.12
7. Conclusion A hierarchical behavior-based control for a holonomic robot was developed. In all tests the reactive component was able do avoid collisions with static obstacles (e.g.Fig.l 0). The predictive obstaclc handler on the other hand generates efficient paths that are comparable to those generated by visibility graph methods.This way the most prominent shortcoming of potential field methods 34 , the local minima problem, is negated as well (Fig. I I ). It should be kept in mind that the robot has no global map knowledge and therefore is only able to plan the path in visibility range of the sensors or within a local memorized area. The network character of the control enables us to easily extend the control with new functionalities. Either
105
by hooking in new behavior modules which is done straight forward due to the use of fusion behaviors or by recombining present functionalities by activating the corresponding behavior modules. This way it is possible to use the obstacle avoidance functionality to augment the steering functionality of the haptic handle so that the intelligent trolley moves around obstacles while it is being pushed by it's user. Another augmentation can be achieved by slightly activating the target attracting functionality while the robot is pushed so that the robot pulls the user slightly when a direction change or a turn off is necessary. Acknowledgements This research has been carried out in the CommRob project (www.comrnrob.eu) and is partially funded by the EU (contract number IST-04544I under 6th framework programme). References 1. M.GOLLER, F.STEINHARDT, T.KERSCHER, 1.M.ZOLLNER, R.DILLMANN: RFID Transponder Barriers as Artificial Landmarks for the Semantic Navigation of Autonomous Robots, Submitted to CLAWAR 2008 2. BENJAMIN KUIPERS: An Intellectual History of the Spatial Semantic Hierarchy, Computer Sciences Department, University of Texas at Austin, Austin, Texas 78712 USA 3. JAVIER ANTICH: Extending the Potential Fields Approach to Avoid Trapping Situations, Department of Mathematics and Computer Science, University of the Balearic Islands, Palma de Mallorca, Spain 4. J. BORENSTEIN: Potential field methods and their inherent limitations for mobile robot navigation, in Proc. of ICRA, vol.2, 1991, pp.1398-1404. 5. A.ELFES: Using occupancy grids for mobile robot perception and navigation, Computer, 22(6), 46-57 (1989). 6. S.THRUN: Learning Occupancy Grids with Forward Sensor Models, Carnegie Mellon University, 2002 7. M. MATARIC: Behavior-Based Control: Examples from Navigation, Learning, and Group Behavior. Journal of Experimental and Theoretical Artificial Intelligence, Special issue on Software Architectures for Physical Agents, 9(2-3):323-336, 1997. 8. P. ALTHAUS, H.I. CHRISTENSEN: Behaviour Coordination for Navigation in Office Environments, Proceeding of the 2002 IEEE/RSJ Int. Conference on Intelligent Robots and Systems, EPFL, Lausanne, Switzerland, 2002 9. D. LANGER, 1. ROSENBLATT, M. HEBERT: A Behavior-Based System for Off-Road Navigation IEEE Journal of Robotics and Automation, 1994 10. J. ALBIEZ, T. LUKSCH, K. BERNS, R. DILLMANN: An Activation-Based Behavior Control Architecture for Walking Machines. The International Journal on Robotics Research, Sage Publications,vol. 22:pp. 203-211, 2003. II. J. ALBIEZ: Verhaltensnetzwerke zur adaptiven Steuerung biologisch motivierter Laufmaschinen. GCA Verlag, 2007. 12. H. SCHAFER, K. BERNS: RAVON - An autonomous Vehicle for Risky Intervention and Surveillance, University of Kaiserslautern (Germany), 2006
RFID Transponder Barriers as Artificial Landmarks for the Semantic Navigation of Autonomous Robots Michael Goller, Florian Steinhardt, Thilo Kerscher, 1.Marius Zollner and RUdiger Dillmann Interaktive Diagnose- und Servicesysteme (IDS), Forschungszentrum InJormatik an del' Ulliversitdt Karlsruhe (FZI), www.fzi.de. Karlsruhe, 76131, Germany E-mail: (goeller/steil1hardtlkerscherizoelinerldillmann)@Jzi.de
A main challenge in the field of service robotics is the navigation of robots in human everyday environments. Supermarkets, which are exemplary chosen here, pose a challenging scenario because they often have a cluttered character. They are full of dynamic objects. Especially the presence of large numbers of people is a special challenge to cope with. It can often be difficult to map the locality because of a frequently changing environment. Therefore a broad approach is needed to cover three main tasks: the reactive local navigation, the interaction with dynamic objects and a the global navigation and task planning. A three layered navigation concept was developed where each of these fields is dealt with in a dedicated layer. This paper describes the top layer with a semantic navigation using RFID tags as landmarks. It is based on a topological map with semantic attributes. Barriers of RFID tags are used to discriminate the environment into topological areas. Combining a topological navigation with a behavior-based control makes the navigation independent of a global metrical map. This paper is associated with the CommRob* project. Keywords: RFID, landmarks, semantic navigation, topologic map, topologic navigation
1. Introduction
Among the most important and most basic features of an autonomous robot, in this case the intelligent shopping cart InBOT(Fig.l), is the ability to move to a certain location without colliding with any obstacle in the path. A navigation capable of this is often based on a metrical map. But what happens if the environment changes too often to measure it exactly? In supermarkets special offers or advertisements are deployed or removed all the time. Here the robot has to find its way on its own in close touch with the real environment. At this point the idea of a combination of topological navigation and a behavior-based control for local navigation gets involved. A topological map can be built based on the definitively static structures of the environment like rooms and corridors without the 'www.commrob.eu, contract number IST-04544I under 6 th framework programme
106
107
Fig. 1.
The holonomic robot InBot
Fig. 2. The layered structure of the navigation concept. The focus in this paper I ies on the top layer, the Strategic Layer.
need of exact measurement. When moving inside such a node the robot reccives instructions on how to reach the desired target. Three main fields of duty can be identified here. In bottom up order these are navigation inside a local topological area, the handling of dynamic objects and the interaction with humans and on the highest level the task and route planning on the topological level. As fourth task the mapping can be mentioned, but this can not be covered here. In a work from B.Kuipers7 a detailed hierarchical approach on learning cognitive or topological maps in such scenarios can be found. A thrce layered navigation concept was developed to cover the mentioned three identified tasks (Fig.2). The bottom laycr's goal is to fulfill movement tasks inside a local topological area using a behavior network as control. The content of this layer is not explained in detail here. It can be found in a conesponding work l instead. The second layer, that is not described in detail in this paper as well, contains a behavior-based control to deal with dynamic objects in a local area and with the interaction with nearby humans. The focus of this paper lies on the top layer that is called Strategic Layer. It contains a global navigation that is based on the discrimination of the area into several topological nodes and applying semantic attributes to them. The global navigation generates a plan based on the topological map and instructs the behavior-based control in the bottom layer with movement tasks and parameterizes it based on the semantic attributes. A sketch of the navigation system is displayed in Fig. 3. At this point the crucial question arises. How does the robot know when it is entering a new topological area? This is commonly done by letting the robot look for certain landmarks. Because natural landmarks are often hard to detect or to
108
Fig. 3. The navigation concept: A global navigation based on a topological map instructs a local navigation with modules for predictive and reactive obstacle handling.
identify, artificial landmarks are commonly used (Ch.5).
2. RFID Barriers Many landmark-based systems for localization share a common disadvantage compared to a short-range Radio Frequency Identification (RFID)12 system. The landmarks can be occluded. Wall- or ceiling-mounted long-range RFID-based systems can be blocked as well, for example if too many humans are located bet ween transponder and reader. To avoid these occlusions and to get more precise position information this paper proposes to use ground-mounted RFID barriers (Fig A). These are placed for example across corridors (see Fig.5) to divide the corridor into several topological areas. The robot knows all IDs of the tags contained in a barrier. By driving over a barrier the robot recognizes that it just entered a new area. By doubling the tags, as depicted in FigA or 5, it even knows in which direction it has driven past the barrier. To identify a proper height for assembling the RFID reader" on the robot the reader's coupling area was measured with the transponder parallel to the reader. As shown in Fig.6 a distance of 2cm between reader and barrier has proven optimaL Here the coupling area, consisting of a combination of the main and a side area, has approx. 35cm in diameter. The larger the coupling area in diameter the faster the robot can move past the barrier without missing it. Additionally the low distancc of 2cm makcs disturbances between reader and barrier very unlikely. Using this setup the robot was able to detect all barriers in the tests up to velocities "liD ISC.PRlOl·A Proximity Reader, PEIG ELECTROKIC GmbH
109
Fig. 4. The barrier consists of several RFID Fig. 5. A system of corridors is divided into sevtransponders (small white rectangles) that are eral areas by special RFID barriers. A RFID reader is placed inside the robot (white box), so the robot placed inside or under the barrier sparsely. is able to detect a barrier by moving over it.
Z[cmJ 20
140 120
15
100
10
80
60
5 0
ｾ＠
40 20 2.5
Fig. 6. The Coupling area in 3D view. This figure shows the balloon like shape of the main area and the circular shape of the side area. The distance with the largest diameter is 2cm where both of the areas meet.
Fig. 7. Reading performance of a double RFID tag depending on the velocity of the reader. Up to a velocity of 2m!s at least one of the two tags of a double tag could be detected in all tcsts.
of approx. 2m/s when double tags have been used. The reading performance with a single tag is significantly inferior (see Fig.7).
3. Semantic Map The shop (Fig.S) is split up into several areas by using the described RFID barriers. Figures 5 and 9 show examples for such a partitioning. The areas and their connections are mapped to a topological map without metrical local maps (Fig. 10). The only information needed for a certain area are the relative positions of the contained entry points. The semantic map is a semantic extension of the topological map. Each node is enhanced with semantic information (Fig. I I ) about the node. The semantic information can be used as an indication of whether a node should be used or avoided in a global plan. Additionally these information can include directions how to reach the next node from the current entrance of the actual node, restrictions to velocity or acceleration, certain dangers which wait in this area like a grate in the floor which would render an optical motion sensor useless or a milTor that could trick a camera system. Additionally actual information
110
Fig. 8.
Map of the shop.
Fig. 9. Map divided into a1'- Fig. 10. eas by RFID-Barriers (red cal map lines)
Resulting topologi-
Type: [Corridor. Mom, etc.] Primary behaviour: (Follow corridor, straight through mom]
drive
Restrictions: [Velocity, Acceleration, etc.] Special 10catio11s
Possible sensor failures (Ground is not suited flY1' opticaimotllJll sensors, mirror might trick camera, etc.} Actua! information: [corridor blocked., etc.1 COllOecticltls to neighbour nodes and landl!k1l'ks indicaling this node Fig. 1l.
Topological map with semantic annotations
like a blockage of a corridor or the prohibition to enter a room or other warnings might be of intercst. The information can be updated online by the robot to inform other robots about knowledgc it has acquired.
4. Global Planning First of all the robot plans a route from its current topological node to thc nodc that contains the target. This planning is done by a standard A * algorithm. This algorithm is enhanccd in a way that it is able to utilize the semantic information provided, for example to generate constrains or weights out of them to influence the plan. Because the semantic map is a very high level of discretization the navigation takes place on a small state space. Therefore the processing costs are very low. Based on this plan a list is generated that contains the relative coordinates of the barriers to each other. In the end the relative coordinates of the target in the last
111
area are added to the list. This list is then executed by the behavior-based control. It is described in a corresponding work 1 but taking a further look here would go
beyond the scope of this paper.
5. Related Work In this section works are summarized that make use of comparable technologies or methods like the presented ones. The main focus lies on artificial landmarks like optical ones or RFID tags as well as on the usage of topological maps. 5.1. Landmarks In works by Gu0 4 and Yoon 11 colored visual landmarks are used. They are designed in such a way that they are robust versus perspective transformations and easy to identify. Beside these camera-based approaches there are several others like infrared based landmarks 10 or long range RFID systems 53 . By using these RFID concepts only rough position estimations are possible. Therefore these systems are only used as advancement for other systems. For example Hahnel 5 uses the rough information to reduce the space of possible solutions for a vision-based method. Another advantage of the RFID-based system is the possibility of depositing additional information on the tags. This can for example be the geometry of a local crossroad 6 , local maps9 or even information to control the robot like in a work by Kulyukin 8 where behaviors of the robot are controlled. 5.2. Topological Maps These maps are very compact because they contain hardly any metric information. They represent large structures like corridors or rooms so they are as far as possible time-invariant. 15 Additionally they are easy to construct even for large places. When considering human robot interaction the facts stands out that topological maps are much easier understood by humans then large metrical ones. Often the topological maps are combined with local metrical maps.7 They can be constructed manually, like done for Robox l6 but often the robot learns them automatically13 or they are used to support a mapping process. 14
6. Conclusion and Outlook In this work a navigation system for an mobile robot was introduced that was developed especially for highly dynamic environments. A main aspect to deal with was the lack of metrical maps in such environments. Therefore the navigation was
112
Fig. 12. Successful path of the robot. The Fig. l3. Robustness: while moving the robot was higiven task was to move to several locations jacked (dotted line) and moved past a RFlD-barrier it inside different topological areas. did not expect on its rOLite. It was always able to COfrect this' mistake'.
divided into two parts. The upper one, focused on in this document, works based on very few static structures that arc assumed to be invariant. The resulting areas are discriminated by ground-mounted barriers of RFLD transponders and build up a topological map of the location that is annotated with semantic attributes afterwards. The global navigation takes place on this semantic map. A focus was on the development of the reliable barrier detection. Because of the low number of tags needed and the sparse distribution of barriers in the shop this system is cheap and its installation is low invasive. The robot was in all tests able to perceive the barriers at velocities of up to 2m/s in a stable and reliable manner. In human environments it is hardly imaginable to let autonomous robots maneuver with such a speed or even more. Velocities of up to 1.5m/s arc more realistic. In several tests the reliability and robustness of the system has been proven. Fig.12 depicts a test where the robot was instructed to move to severalloeations in different topological areas and Fig.13 shows the hijacking of the robot where the robot was able to recover from. In future work the semantic aspect of this kind of navigation can be enhanced. It is still necessary to determine the relative coordinates of neighboring barriers. Instead, the navigation could be performed on a purely semantic basis. For example the robot could be given 'the third door' as waypoint instead of the coordinates. Thus the last remaining mctrical information of the location could be dismissed. Acknowledgements This research has been carried out in the CommRob project (www.commrob.eu) and is partially funded by the EU (contract number IST-045441 under 6 th framework programme).
113
References 1. M.GOLLER, THILO KERSCHER, J.MARWS ZOLLNER, RUDIGER DILLMANN: Behaviour Network Control for a Holonomic Mobile Robot in Realistic Environments, SUBMITTED to CLAWAR 2008 2. J.ALBIEZ, T.LUKscH, K.BERNS, R.DILLMANN: A Behaviour Network Concept for Controlling Walking Machines, Proceedings of the 2nd International Symposium on Adaptive Motion of Animals and Machines, Kyoto, MarchA-8, 2003 3. HEESUNG CHAE, KYUSEO HAN: Combination of RFJD and Vision for Mobile Robot Localization, Proceedings of the 2005 International Conference on Intelligent Sensors, Sensor Networks and Information Processing, S. 75 - 80, 2005 4. YANG Guo, XINHE Xu: Color Landmark Design for Mobile Robot Localization, IMACS Multiconference on Computational Engineering in Systems Applications, Volume 2, S. 1868 - 1874,2006 5. D. HAHNEL, W. BURGARD, D. Fox, K. FISHKIN, M. PHILIPOSE: Mapping and Localization with RFJD Technology, Proceedings of the 2004 IEEE International Conference on Robotics & Automation, S. 1015 - 1020,2004 6. OLAF KUBITZ, MATTHIAS O. BERGER, MARCUS PERLICK, RENE: DUMOULIN: Application of Radio Frequency Identification Devices to Support Navigation of Autonomous Mobile Robots, Vehicular Technology Conference, 1997 IEEE 47th Volume I, S. 126 - 130, 1997 7. BENJAMIN KUIPERS: An Intellectual History of the Spatial Semantic Hierarchy, Computer Sciences Department, University of Texas at Austin, Austin, Texas 78712 USA 8. VLADIMIR KULYUKIN, CHAITANYA GHARPURE, JOHN NICHOLSON, SACHIN PAVITHRAN: RFID in Robot-Assisted Indoor Navigation for the Visually Impaired, Proceedings of 2004 IEEE/RSJ International Conference on Intelligent Robots and Systems, S. 1979 - 1984, 2004 9. WEIGUO LiN, CHANGLI Mu, KUNIKATSU TAKASE: Path Planning with Topological Map built with JD Tag and WEB Camera, Proceedings of the 2006 IEEE International Conference on Mechatronics and Automation, S. 1739 - 1744, 2006 10. T. TOMIZAWA, Y.M. SAIKI, A. OHYA, S. YUYA: Property Modifiable Discreet Active Landmarks, 2007 IEEE International Conference on Robotics and Automation, S. 3420 - 3426, 2007 11. KUK-JIN YOON, IN-SO KWEON: Artificial landmark tracking based on the color histogram, IEEE/RSJ International Conference on Intelligent Robots and Systems 2001, Proceedings. 2001 12. BILL GLOVER, HIMANSHU BHATT: RFID Essentials (Theory in Practice), O'Reilly, 2006 13. S. THRUN: Learning metric-topological maps for indoor mobile robot navigation, Artif. Intel!. 99(1), 21-71 (1998). 14. J. MODAYIL, P. BEESON AND B. KUIPERS: Using the Topological Skeleton for Scalable Global Metrical Map-Building, Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), (2004). 15. P.ALTHAUS: PhD-Thesis, Indoor Navigation for Mobile Robots: Control and Representations, Royal Institute of Technology, Stockholm, 2003 16. R.SIEGWART: Robox at expo.02: A Large Scale Installation of Personal Robots, Swiss Federal Institute of Technology Lausanne (EPFL) , 2003
DEVELOPMENT OF PREDICTIVE MODELS OF CLIMBING ROBOT ATTACHMENT MECHANISMS FOR AN ENERGYEFFICIENT ADAPTIVE CONTROL SCHEME S. A. JACOBS
School of Mechanical Engineering. University of Leeds, Clarendon Road, Leeds, LS2 9fT, UK A. A. DEHGHANI-SANU
School of Mechanical Engineering, University of Leeds, Clarendon Road, Leeds, LS2 9fT, UK This paper presents the investigation of two climbing robot attachment mechanisms. The two mechanisms are tested to find their maximum holding forces under a range of conditions, and models to predict these forces are developed. We describe the experimental methodology and the results of testing three types of vacuum pads and an electromagnet are presented, which are then compared to the predictive models. The outcome of this work is to be used to develop a fully autonomous and efficient method of adaptive control for climbing robots.
1. Introduction
Climbing robots are commonly used when steep surfaces need to be climbed and there are dangers, impracticalities or economic disadvantages in human workers gaining access to the environment. This includes inspection, cleaning and maintenance tasks on the exteriors of tall buildings and structures such as bridges and dams, ship hulls and aeronautical structures, and the interiors of petrochemical tanks and pipe/duct systems. An important aspect of the climbing robot is the mechanism by which it can securely and reliably attach itself to the climbing surface. The four main types of attachment mechanism are magnetic (electromagnets, permanent magnets), adhesive (artificial gecko material, dry adhesives), suction/vacuum (vacuum pads, inverted hovercraft) and mechanical (micro-spines, clamps). Whilst many climbing robots make use of the above attachment mechanisms, and there are examples of robots that can perform a search to obtain a secure placement of end effector [1, 2], there are currently none that use the attachment mechanism in an energy-efficient manner, i.e. by supplying the effector with energy sufficient to support only the weight of the
114
115
robot, or performing a search to determine the location of an energy-optimised placement. We have identified two attachment mechanisms as being suitable for implementation in a proposed adaptive control scheme: electromagnets and vacuum pads; there are examples of climbing robots using vacuum pads that take a vacuum pressure directly from a small, on-board vacuum pump which could be used to supply a variable vacuum pressure [3], and whilst switched-bypass is one possible way of implementing a magnetic attachment mechanism there are still robots that use only electromagnets as effectors for reasons of size and weight [4]. Whilst others have already carried out experiments on electromagnets [5] and vacuum pads [6, 7], the results presented are not comprehensive and do not investigate the many additional factors that affect the holding force such as steel thickness or air gap width.
2. Theoretical and Predictive Models Developed 2.1. Vacuum Pad Normal Holding Force Theoretical Model A simple theoretical model for the normal holding force of a vacuum pad is the product of pad area and vacuum pressure. We have tested pads from three manufacturers and observed the pads to be compliant, the diameter reducing as the pull-off force increased. We took this into account with our theoretical model, which assumes that the pad is a thin-walled conical section and that we can model a segment of the pad as a simple cantilever beam. The displacement is proportional to the cube of the beam length; for small angles, horizontal displacement of the pad rim is proportional to vertical displacement. A model based upon the above reasoning is shown in Eq. (1), where Ro is the overall pad radius and R, is the length of the cantilever section, determined by taking measurements of the pads (see Table 1). P is the vacuum pressure measured relative to atmosphere, and a is the constant of pad deformation.
F110r111a/ -- - 7r (R0 - aR I 3p )2 P
(1)
Table 1. Lengths of cantilever sections for the three vacuum pad ranges. Vacuum Pad Diameter (mm)
Length of Cantilever Section RJ (mm) SMC Pads Festa Pads Pisco Pads
10
N/A
3
3
J3
5 8
N/A
N/A
5 7
5 7
N/A
N/A
20 30 32 40
50
N/A 13 17 22
7
7
8
8
116
Theoretical holding forces are given in Figs. 1-3, with the experimental results. Note that the Festo and Pisco pads did not deform as much as the SMC pads. This was due to the rigid central section of the vacuum pad reducing the size of the cantilever section R J • a for the SMC pads was derived by examination of the results from all five pads, for which a = -0.0055. For the Festo pads a = -0.066, determined from results for the 10, 30 and sOmm pads. For the Pisco vacuum pads, only the 30mm pad was tested experimentally, for which a = -0.093.
2.2. Electromagnet Normal Holding Force Predictive Model Our initial attempts to create a purely theoretical model were unsuccessful. Part of the problem was the lack of data on the relationship between magnetic field strength H and flux density B of the electromagnet and steel sheets, so we used our experimental results to derive this data. Note that the 3.19mm and 9Asmm curves in Fig. 4 are near-identical; this is because above a certain thickness of steel sheet, the holding force is limited by the construction of the electromagnet and not the steel sheet (due to the flux reaching its saturation level in the electromagnet before the steel). Calculation of the maximum flux that can flow through the electromagnet (which is commonly given as IAT to 1.6T [8]; we assume that the saturation flux density is l.sT) shows that this threshold steel sheet thickness is around 3.2mm. For sheets thinner than 3.2mm, the holding force will decrease according to a function of the thickness. The first step in creating this new model is to recognise that the force-current curve is made up of two parts - the non-saturated region, where the field strength H is such that the flux density B has not begun to approach its saturation level, and the saturation region, where H continues to increase such that B approaches and eventually reaches its saturation value of l.sT. The two parts of the curve were modelled separately, and are given in Eq. (2). The transition point of I = O.IIA was found by inspection; it is the point where B has the largest gradient, and hence must be the point where B is about to enter its saturation region.
o< 1 < O.II}
F= { 1>0.11
25
16000/ 21177/3-15334/2+3718.8/-184.53
(2)
We cannot expect the force-current curve to be directly proportional to the relative thickness of the steel sheet, because for thin steel sheets the flux density B may already have approached its saturation level before the transition between the two parts is reached. It was therefore necessary to establish a piecewise
117
relationship f(t) between the steel thickness t (in m) and the relative holding forces at I = O.IIA (i.e. the transition current), and is shown in Eq. (3). 2
0 < t < 0.001 } 384400t + 43.5t f(t)= 0.001Rectus Femoris-> Gastrocnemius-> Ankle
Power transmission and shock absorption via biarticulate muscles has been documented in humans [2]. While these properties are remarkable, biarticulate muscles are not used in walking machines, with at least one exception [3, 4]. Many attempts have been made to produce human-like walking in robots, the most famous example being Honda's 'Asimo' [5]. However, the speed and flexibility of which humans and animals are capable (i.e. running, jumping, and balancing) surpasses anything produced artificially thus far. Asimo, remarkable as it is, can demonstrate only a small fraction of human locomotor capability. Most walking robots use rotational motors at the joints to control movement, while the concept of biarticulate muscles has been neglected in the robotics literature (a notable exception being [6]). Undoubtedly, this is partially due to need to minimize the number of actuators in a robot (to reduce weight and increase system reliability). However, with widely available smart actuators, for example Animatics (http://www.animatics.com) and Robotis (http://www.robotis.com). highly reliable modular actuators are readily available. We anticipate the movement toward high DOF systems will accelerate.
231
In this article, we describe theory behind biarticulate robots, and an implementation in a robotic leg. We then report on development of methodology for measuring work transference and demonstrate the use of this methodology in a real robot. We show energy flowing via the GA, and show that the timing of SO versus GA activation seems to be important for achieving maximum power output at the ankle.
2
Mammalian leg muscle architecture
The human leg can be modeled as a system of three parallel joints (hip, knee, and ankle) and nine muscle actuators (See Figure 1.) The muscles include three biarticulate muscles: the GA, which spans the knee and ankle, the RF, and the HA, which both span the hip and knee. The leg uses an agonist/antagonist, or flexor/extensor design with regard to monoarticulate muscles. Extensor muscles are used to support the body weight of the robot against gravity. Flexor muscles are used to lift the limb. The flexor muscles are generally much smaller than the extensor muscles. Monoarticulate muscles on the ankle include the T A, which flexes the foot, and the SO, which in conjunction with the GA, extends the foot. The knee is extended by the VA, while the BFS helps flex the knee. The GM holds the hip upright, while the Illiacus (IL) flexes the hip. In human beings it has been demonstrated that knee extensors generated significantly more force than the knee flexor [7]. An even more dramatic example is of the ankle flexor versus extensor. In humans, the cross sectional area (CSA) of various muscles have been measured. While the particular configuration of the muscle (e.g. Pennation angle) can affect the force generating capability of muscle, we note that muscles such as the VL and SO have a much greater CSA than muscles such as Figure J. Model of the human leg. TA is tibialis anterior, SO is soleus, GA is gastrocnemius, VA is Yatus lateral us, RF is rectus femoris, BFS is short head of biceps femoris, HA is two-joint hamstrings, GM is gluteus maximus, and IL is iliacus. Redrawn from [l J.
232 biarticulate GA and RF f8j. This implies that the monoarticulate extensor muscles must produce more force. Moreover it suggests that implementation of a robot based on human leg muscle architecture can be done using smaller motors for the flexor and biarticulate muscles, thus reducing the weight penalty for using multiple motors for each joint.
3
Implementation of Design Concepts
We implemented these ideas in an human-like leg, with pin joints at the hips, knees and ankles. Each joint is Figure 2. Actuator architecture of robot actuated by a combination of actuators leg, Shown above is a cut away view of designed to mimic the mechanics of the rahat limb, High performance modular motors pull Oil Kevlar straps to muscles. The following muscles were activate the joints. Bimiiculate actuators modeled: GA, T A, SO, V A, RF, IL and are: Gastrocnemius and Rectus Femoris, GM (the HA and BFS were not Hamstrings are not implemented here. modeled). The toe was modeled as a passive joint. An elastic cord is used to straighten the toe. The usc of a toe proves the robot with the ability to 'stand on tip-toes.' The distances and proportions of the limb segments were based on human anthropometric data [8].
3.1.
Actuators
The actuators were composed of stiff Kevlar strap connected to a motor. The straps were affixed to a mounting bracket that pulls and releases the strap in one direction. We have selected Robotis RX-28 motors for the combination of force and compactness for the GA, SO and VA muscles and a Futaba S3150 for the TA muscle. Robotis RX-64 are used for the HS, RF and IL. We used an agonist/antagonist muscle configuration. This type of actuation allows the motor to pull but not push, similar to muscle action. As an example, we modeled the T A with one motor placed in the calf and connected by a Kevlar flat strap from it to the front of the foot. The SO was modeled by connecting a flat strap between the rear of the foot and the calf. The dual strap also permit
233
joints to be "stiffened" by applying force to both sides, which assists in leg stabilization during foot touch down. Figure 2 shows a 3D CAD drawing of the leg design, including motor positions and Kevlar straps running down the front and back of the leg. Figure 3 shows photos of the completed leg as constructed.
3.2.
Sensors
Angle sensing pots (Murata SVOlAI03) are used to measure joint angles of the foot, knee and hip. Each sensor was calibrated in radians by comparing the voltage output versus a known angular reference. We found the pots to be highly linear. At the attachment point of the straps with the motors, we designed a custom made force sensor. This sensor is based on a Futek FSH0l463 Force gage. The assembly was mounted between the motor bracket and the strap. As force is exerted by the strap, this force is measured by the gage. Finally, we used a load cell to measure tension at the Achilles' tendon (both the SO and GA act on the ankle via the Achilles' tendon). The model number of the Achilles' tendon load cell was unknown.
4 4.1.
Measuring work and peak power at the ankle of the robot Work Transfer
Work produced by a rotational torque can be written in time-discrete form as:
"v
W = L., T*I1Bv
The torque at the ankle is given by: 7 TA , 7 GN , 750 are torques produced by the T A,GN and SO respectively. Multiplying by the angular displacement, in radians, of the joint we have the net work at the ankle:
ｾＱｫｉ･＠
Wso,
=MMder,"lkIe =-M""lkIeITA +M;.lkIetav+M;.iJet )() ｾｮｫｬ･＠ =-Rj)\ + Wm + WCN
where ltjA' WGN is the work contribution by the T A,SO and GN respectively. Likewise, for the knee: W Knee = W\-L
+ W RF -
WCN - WSfS - W HA
In the simplified case of activating the extensors alone: WAnkle WCN =
= WCN + WSO
WvL
+ W RF -
W/V1ee
234 Substituting:
WAnkle
= Wso + WVL + WRF -
WKllee
Because muscles pull, but cannot push we note ｗｭｾｏＬ＠ and W GN ｾｏ＠ thus Hence WVL + W RF - W Knee ｾ＠ 0 . the work contribution of the upper leg and the lower limb sum together when the OA is active; the work done on the ankle is greater than just the SO alone. We conclude that the work done at the ankle on the environment IS Figure 3. Bent knee squat. This configuration was assisted by the RF and the VL. used to measure the contribution of SO and GA to the ankle power as well as to analyze the effect of Note, it is critical that an SO and GA activation timing on peak power actuator be cable of pulling and production. not pushing. If the OA could push, it would be possible for the upper leg to take energy away from the ankle. Thus, a stiff rod connection between actuator and joint cannot be used, for example, if we are to adhere to biological principles. 4.2.
Methods
During the following experiments, the robot was commanded to do a weight lifter's 'squat.' The timing of the SO versus ON was varied as was the activation of the SO or ON. Simultaneous measurements were made of the potentiometer and the force gage using a PICO Scope (Model 5203). The sampling rate was WOK Hz. To improve differentiation, data was decimated to sampling rate of 100 Hz. The data was smoothed using a alpha tracker (a = 0.2). The raw angle and force signals were converted to radians and Newtons using a linear calibration equation. Time and angle difference were computed. Next the change in distance at the ankle attachment point was computed using the following formula: & = r ｾｂｷｨ･ｲ＠ r is the moment arm from the ankle rotation axis to the ankle attachment point of the Achilles' tendon in units meters (0.0023 meters) and Bis the angular measurement of the ankle in radians.
*
235
1.R
Ankle Work Contribution by SO, GN and SO & GN
ＱｾＳＭ ＱＺｾ｟＠
1:3---+-
ｯＮｾ＠
"4-0- - - - ' - - _ . N : -_ _- ' -_ _ｾ＠
_ _ｾ＠
_
___1
TIME (50100
x 20 x 20
Spring steel thickness (mm) 0.10 0.10
0.34 0.34
the order of SMA-contraction was arranged as in Fig. 2-(c), which we call an AE pattern. The switching time for each SMA is 5.0 s, making the total driving time of robot II 180 s for this AE pattern. 4. Experimental result of rolling soft robot
In this section, we experimentally show the rolling of robots I and II on a horizontal surface and on a surface inclined 10 degrees. Each robot was driven with an AE pattern, and alternately switched at 5.0 s. Initially the shape of the rolling robot was close to a perfect circle. The material of the floor was polyvinylchloride (PVC), with static coefficient of friction of 0.45. The ambient temperature during each experiment was 25°C. Figure 3 shows sequential snapshots of the rolling robot. Robots I and II were compared on horizontal ground and inclined ground while the robots had an oval shape. The oval shape was maintained constantly so that each robot would not slope down during the experiments. Figure 4 shows the distance moved in these experiments. Stair-shaped locomotion is a characteristic of a rolling robot. The sharp rising response indicates the moment generated by the robot while transforming and rolling. To compare the two robots, we performed experiments testing the robots on a horizontal and an inclined surface. vVe found that robot
368
(a-I) 08
(a-2) lOs (a-3) 208 (a) Robot I on flat ground
308
(b-l) Os
(b-2) 10 s (b-3) 208 (b) Robot II on flat ground
(b-4) 30 s
(c-l) Os
(c-2) 10 8 (c-3) 20 s (c) Robot I on slope (10 deg)
::lOs
(d-l) Os
(d-3) 20s (d-2) lOs (d) Robot II on slope (lOdeg)
(d-4) 308
Fig. 3.
Sequential snapshots of the rolling robot.
II was 4 % slower on horizontal ground and 1 % slower on inclined ground when compared with robot I. In contrast, robot II had a more stable and robust locomotion than robot I.
5. Simulation of rolling soft robot
The dynamic simulation of robots I and II using particle-based modeling was performed for two conditions, rolling motion on a horizontal surface and rolling motion on a slope. The main objective was to analyze the effect of robot weight on rolling. The particle mode1 2 shown in Fig. 5 was employed, with a total of 81 mass points distributed throughout the body (64 points) and SMAs (17 points). In robot II, the weight of the battery and the circuit was included.
369
( a) flat ground Fig. 4.
(b) slope (1O deg)
Distance moved, as assessed experimentally.
We used the penalty method to express the reaction force and frictional force on the ground. As expressed, the generated force was related to the martensite and austenite transformations on SMA 4 . 5 In this simulation, the static coefficient of friction was 0.5 and the ambient temperature was set at 25°e. Potential energy was estimated as follows: The deformed circular shell , f being the coordinates of the i-th was divided into n particles, with particle. If Ai is the mass of the robot, and if the entire mass is distribnted on the shell, the mass of each particle is given by rn = Min. Thus, the gravitational potential energy Ugrav is
vl
n
Ugrav
L
Tn
g Yi
(1)
i=l
where g is the acceleration of gravity. Figure 6 shows the simulated configurations of the rolling robot and the gravitational potential energy changes of each using a quasi-static method. The configuration shown in Fig. 6 closely agreed with the configuration shown in Fig. 3. In Fig. 6, the local minimum point shown with the blue dotted line indicates the rotation angle that is measured from the origin at the initial state (O.Os). Note that the local minimum point is shifted as the robot rolls; thus, this local minimum point refers to the angle of rotation. These results indicate that rolling locomotion uses the gradient of a gravitational potential energy. By comparing the results shown in Fig. 6, we found that the potential energy graphs were similar to each other and that the rolling speeds were almost the same. In addition, robot weight had no effect on rolling locomotion on horizontal ground. Thus, it is easier to roll the robot leftward than rightward down the slope because the peak
370
(a) Robot I Fig. 5.
(b) Robot II Particle-based model of a soft robot.
potential energy is on the right half, not on the left half. If the weight is increased, the valley of the potential energy is deepened, and the robot thereby becomes more stable. The simulated distance moved during rolling locomotion is shown in Fig. 7. We found that the robot behaves in a manner similar to stair-shaped locomotion. Simulation and experimental results were in good agreement, with a margin of error of 2-6 % on fiat ground and of 4-6 % on a slope. The distances traversed by robots I and II coincides with one another, indicating that weight did not significantly affect locomotion. The locomotion distance on a slope was less than that on fiat ground due to the slip between robot and ground (Fig. 7-(a) and (b)).
371
Os
(a-2) lOs (a-3) 208 (a) Robot I on flat ground
(a-4) 30s
(b-l) Os
lOs (b-3) 20 s (b) Robot II on flat ground
(b-4) 30 s
(e-l) Os
(e-2) lOs (c-3) 20s (e) Robot Ion slope (lOdeg)
30s
(d-I) Os
( d- 2) 10 s (d-3) 20 s (d) Robot II on slope (lOdeg)
(d-4) 308
Fig. 6.
Simulation results of a rolling soft robot.
372
(a) flat ground Fig. 7.
(b) slope (lOdeg)
Simulated distance moved by a rolliug soft robot.
6. Conclusion
In this paper, we have described the development of a soft robot with a built-in power source. By applying dynamic simulation with particle-based modeling, we analyzed the rolling motion of this robot. This simulation indicated that rolling locomotion uses a gradient of gravitational potential energy. We found that inereasing robot weight had almost no influence on its kinematie performance on flat ground, whereas inereased robot weight enhanced its stability on inclined ground. These results indicate advantages of attaching sensors to a rolling soft robot. In future experiments, we will apply sensors to the rolling soft robot and try to develop a jumping soft robot with a built-in power souree. References 1. S. Hirose, Biologically Inspired Robots
2.
3.
4.
5.
Snake-like Locomotor's and Manipulator's, Oxford Science Publications, 1993. Y. Sugiyama, S. Hirai, "Crawling and Jumping of Deformable Soft Robot", Fmc. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, pp.3276-3281, Sendai, September, 2004. H. Nakanishi, S. Hirai, "Passive Crawling of a soft robot", 2007 IEEE/ASME Int. Conj. on Advanced Intelligent Mechatmnics (A IM2007) , pp.1-6, ETH Zurich, Switzerland, September. 4-7, 2007 . K. Ikuta, M. Tsukamoto, and S. Hirose, "Mathematical Model and Experimental Verification of Shape Memory Alloy for Designing Micro Actuator", Pmc. of IEEE Micm Electm Mechanical Systems, pp.103-108, 1991. D. R. Madill, D. Wang, "Modeling and L2-Stability of a Shape Memory Alloy Position Control System", IEEE Tmnsactions on contml systems technology, Vo1.6, No.4, pp.473-481, 1998.
JUMPING VIA ROBOT BODY DEFORMATION - MECHANICS AND MECHANISM FOR HIGHER JUMPING M. MIYAZAKI* and S. HIRAI
Dept. of Robotics, Ritsumeikan University, Kusatsu, Shiga 525-8577, Japan * E-mail: [email protected] www.ritsumei.ac.jp/se/- hirai/
As jumping is an effective method of moving over rough terrain, there is much interest in building robots that can jump. Deformation of a soft robot's body is an effective method to induce jumping. Our aim was to develop a jumping robot by deformation of a circular shell made of spring steel to result in the highest jump. Higher jumping requires enlargement of the contact area between the robot body and the floor. We developed a jumping mechanism that utilized a dish shape to maximize contact area.
Keywords: jumping, deformation, shape, mechanism
1. Introduction
Because jumping is a very effective method of maneuvering, especially over obstacles, jumping is used ubiquitously by animals and insects as a means of locomotion. For this reason, there are a variety of jumping robots, most of which have rigid bodies .1-3 A fiexible body exhibits high stability when moving in rough terrain. The aim of the present study is to improve the jumping capability of a soft robot.
2. Principle of Jumping via Robot Body Deformation Figure 1 shows the principle of jumping. Deformation of the robot body results in the accumulation of energy. Instantaneous release of this energy causes the robot to jump. The principles of jumping include 1) bouncing off the ground (Fig. l-(a) and 2) stretching upward with a fiat bottom (Fig. 1(b». A shape with a fiat bottom has been reported have increased impulse during jumping, resulting in higher jumps .4 We investigated which of the
373
374
(a) bouncing off the ground Fig. 1.
(b) stretching upward
Principle of jumping via deformation
four initial deformation shapes, shown in Fig. 2, of a deformable robot that is circular at rest results in the highest jump. The robot's body is made of spring steel, weighs 4.6 g, and is 100 mm in diameter, 12 mm in width, and 0.15 mm in thickness. All four shapes have the saIne flexural potential energy, Uflcx 16.0 x 10-- 2 Nm. Fig. 3 shows how each shape jumps. The arrows indicate the bottom of the objects in their highest positions and the value indicates the height of the bottom point. As shown in Fig. 3, the dish jumps the highest, followed in decreasing order by the peanut, cup, and cap. Clearly, jumping ability depends on the initial shape of deformation.
(a) cap Fig. 2.
(b) cup
(c) peanut
(d) dish
Initial deformation shapes tested for jumping capability
3. Mechanics of Jumping
3.1. Flexural potential energy We next calculated the flexural potential energy of a circular body. If L is the length of the circumference, P(09) is a point on the body at distance 8 from the point of origin along the circumference, and O( s) is the angle subtended by the tangent to P (s ), the flexural potential energy of a circular robot can be calculated as:
jo
'L
1
2Rflex
(dO)2 ds ds,
(1)
375
(a) cap
(b) cup Fig. 3.
Fig. 4.
(c) peanut
(d) dish
Effect of initial deformation shapes on jumping
Flexural Voigt model around ,1 shell particle
where Rflcx is the ftexural rigidity at point P(s). liVe used this equation to evaluate the ftexural potential energy of a circular body during dynamic simulation. 3.2. Particle-based model of circular robot
Figure 4 shows the ftexural Voigt model around a shell particle. Pi and Ph are the particles adjacent to P j , separated by angle OJ around particle P j. The torque Ij around particle Pi is then given by
(2) where is the ftexural elastic constant, bbcnd is the ftexural viscous constant, ei,j is the unit vector along the edge from P·i to Pi and ni,) is the unit vector perpendicular to vector ei,j' vVe assume that vectors ei,.i and ni,) form a right-handed coordinate system. Distance l is fixed between two neighboring particles. Torque Ij can be equivalently converted into three forces, ('j/l)ui,j on Pi, h/l)nj,h on Ph and /l)ni"i ('j/l)nj,h on
Pj
.
376
Fig. 5.
Model of the ground
3.3. Model of the ground Collision with the ground makes a circular robot jump. We therefore modeled the ground to simulate the collision between a robot and the ground. Fig. 5 shows a Voigt model of the ground, where kground is the coefficient of elasticity and bground is the coefficient of viscosity. Particle P j is assumed to be beneath the surface of the ground, and dj is the depth to which particle P j has penetrated the ground. A repulsive force !ground is then applied to the particle. This repulsive force can be expressed as:
(3) If f..L is the coefficient of friction of the ground, then, the force of friction on particle P j is f..L!ground .
3.4. Impulse from the ground during jumping We next simulated the reaction force and the impulse from the ground during the jumping process. Fig. 6 shows the reaction force from the ground during jumping on the shapes shown in Fig. 2. Fig. 6-(d) shows reaction forces on the dish shapes in Fig. 2-(d). No impulsive force is generated. A relatively small force is applied for a longer time, almost 15 ms. The result indicates that the maximum reaction force from the ground does not directly influence the height of the jump. The impulses of the particles from the ground can be computed by integrating the force with respect to time. Fig. 7 shows the calculated results. When the stored flexural potential energy of the initial shapes is identical, the maximum reaction force of the dish is the lower than for the other three shapes. Surprisingly, the maximum impulse of the dish is optimal. Comparison of the Fig. 2, jump height is related to the maximum force of the impulse.
377
(a) cap shape
(b) cup shape Fig. 6.
(a) cap shape
(d) dish shape
Reaction force from the ground
(b) cup shape
Fig. 7.
(c) peanut shape
(c) peanut shape
(d) dish shape
Impulse from the ground during jumping
4. Jumping Mechanism
4.1. Realization of a dish shape We developed a special mechanism to realize the dish shape, as shown in Fig. 8. The mechanism includes a frame and a wire (Fig. 8-(a)). By pulling the wire, the frame is pushed toward the robot body, making it flat (Fig. 8(b) ).
(a) natural shape Fig. 8.
(b) deformed shape
Mechanism for the dish shape
378
4.2. Wire pulling and releasing The wire must be pulled to deform the robot body from its natural shape into the dish shape. By releasing the wire, the dish shape turns into the natural shape, making the robot jump. The wire should be released quickly for the robot to jump. We therefore used a quick return mechanism (Fig. 9) to pull and release the wire. The end of the wire is fixed to the rigid bar. When the rigid bar moves to the opposite side, the wire is wound around the curved body of the mechanism. This mechanism works repeatedly in order of Fig. 9-(a) to (d).
(a) wire released
(b) slow rotation
(c) wire pulled
(d) quick rotation
Fig. 9.
Quick return mechanism
4.3. Prototype
We made a prototype of a jumping robot (Fig. 10-(a)), using a dish shape mechanism weighing 10.7 g, a quick return mechanism weighing 51.6 g, and spring steels weighing 31.9 g, yielding a total robot weight of 94.2 g. The
379
spring steels were 200mm in diameter, 15mm in width, and O.20mm in thickness. By pulling the wire, the robot body was converted from its natural shape (Fig. lO-(b)) into the dish shape (Fig. 10-(c)). Jumping with the dish shape was achieved by this prototype. When the voltage of 6.0 V was applied to the robot, it could jump up to 35mm in height (Fig. 11). In designing the wire pulling system, the robot body deformation was constrained to a maximum of 130mm. However, because the motor torque is insufficient, the attainable deformation was only 115 mm. Because of this, the maximum jumping height of the robot was about :35 mm.
(a) prototype Fig. 10.
(b) natural shape
(c) deformed shape
Prototype of a jumping robot
5. Conclusion
\Ve have demonstrated here the principle of jumping via robot body deformation. vVe found that, of the four initial shapes of deformation, the dish shape was able to jump the highest. The results of our simulation showed that impulse was important for jumping. An experimental prototype was constructed, its ability to attain a dish shape was evaluated, and its jumping performance was determined. Higher jumping requires increasing the motor torque by installing a gear in the motor and increasing the degree of deformatioIl.
References 1. H. Tsukagoshi, Y. Mori, M. Sasaki, T. Tanaka, and A. Kitagawa, Development of Jumping and Rolling Inspector to Improve the Debris Traverse AbilityCJournal of Robotics and MechatTOnics, Vo1.l5, No.5, pp.482-490, 2003.
380
0.008
0.168
0.338
0.538
0.608
0.668
0.738
0.80 8
0.868
Fig. 11.
Jump of a robot
2. S.B. Kesner, J.S. Plante, P.J.Boston, T. Fabian, and S. Dubowsky, Mobility and Power Feasibility of a Microbot Team System for Extraterrestrial Cave Exploration, Pmc. IEEE Int. Conf. on Robotics and A'utomation, pp.48934898, Rome, April, 2007. 3. Y. Sugiyama, A. Shiotsu, M. Yamanaka, and S. Hirai, Circular/Spherical Robots for Crawling and Jumping, Pmc. IEEE Int. Conf. on Robotics and A'utomation, pp.3595-3600, Barcelona, April, 2005. 4. Y. Matsuyama and S. Hirai, Analysis of Circular Robot Jumping by Body Deformation,?roc. IEEE Int. Conf. on Robotics and Automai'ion, pp.1968197:3, Rome, April, 2007.
AUGMENTED CONTROL SCHEME FOR INPUT TRACKING AND VIBRATION SUPPRESSION OF FLEXIBLE MANOEUVRING SYSTEMS: EXPERIMENTAL INVESTIGATIONS F. M. ALDEBREZ and M. O. TOKHI Department of Automatic Control & Systems Engineering, The University of Sheffield, UK This paper presents an investigation into the development of an augmented control scheme for vibration suppression and rigid body motion control of a twin rotor multiinput multi-output system (TRMS) in hovering mode. The augmented control scheme comprises feedforward and feedback control methods. A 4-impulse input shaper is designed and employed as a feedforward control method to pre-process the command signal applied to the system, based on the identified modes of vibration. It is then combined a hybrid feedback controller to form an augmented control scheme referred to as hybrid PD-type FLC and PID with 4-impulse input shaper (FPDPID+IS). The developed control strategy is designed and implemented within the real-time environment of the TRMS rig. Results of the response of the TRMS with the developed controller are presented in time and frequency domains. The performance of the proposed control scheme is assessed in terms of input tracking and level of vibration reduction. This is accomplished by comparing the system response with the developed controller to the one without control action in both open and closed loop configurations. Keywords: Command shaping, flexible manoeuvring systems, augmented control scheme and vibration control.
1. Introduction Flexible manoeuvring systems exhibit many advantages over their rigid counterparts due to their lighter weight and lower power consumption. They are easily transportable and have higher manipulation speed [2]. However, the control task of such systems is a challenging problem due to nonlinearities and flexible dynamics. Moreover, motion induced vibration is exhibited by such systems. The system vibration leads to additional settling time before the new manoeuvre can be initiated [1]. Therefore, in order to achieve a fast system response to command input signals, it is imperative that this vibration is reduced. Several different approaches have been proposed to reduce residual vibration in flexible systems. Both feedforward and feedback control structures have been utilised for the control of flexible manoeuvring systems [5, 8 and 9].
381
382
An acceptable system tracking performance with reduced vibration that accounts for system changes can be achieved by developing an augmented control scheme that caters for rigid body motion and vibration of the system independently. This can be realised by utilising strategies consisting of combined feedforward and feedback control methods, where the former can be used for vibration suppression and the latter for input tracking of the system. The current investigation is confined to the development of augmented control scheme that combines feedforward control component with feedback control method for input tracking performance and vibration suppression of the vertical movement in the TRMS rig. The proposed control structure is referred to as augmented feedforward and feedback control scheme (AFFCS).
2. The TRMS experimental set-up The TRMS is a laboratory platform designed for control experiments by Feedback Instruments Ltd [4]. The schematic diagram of the TRMS is shown in Figure I. Its behaviour in certain aspects resembles that of a helicopter. For instance, like a helicopter there is a strong cross-coupling between the main rotor and the tail rotor. As such the TRMS can be considered a static test rig for an air vehicle. These platforms are often employed to test the suitability of different control methods for such systems.
Beam
u
M O. The parameters 10 and ao are initial leg length and landing angle, respectively. We analyze dynamic stability of the system using a Poincare map F for the system variables (y, x, B, e) at the instant of upper apex. The system is stable if (1) a periodic solution exists (subsequent apices are equal) and (2) all eigenvalues of the Jacobian of F at the periodic solution have magnitudes less than one. We use a Newton-Raphson algorithm to find periodic solutions.
3. Results Periodic solutions exist for both gaits, walking and running (Fig. 1 (b) and 1 (c)), with small pitching motions. The strategy for generating the hip torque is equal in both gaits. Fig. l(b) shows that in walking the hip torque is almost zero during midstance. For the selected solutions, three out of four eigenvalues of the Jacobian of F lie within the unit circle and one eigenvalue is one. Once these patterns are perturbed the system tends to another periodic solution. Such periodic patterns are called partially asymptotically stable. 6 For every velocity within a certain range exists at least one periodic solution. Fig. 2 shows the pitch angle of perturbed gait patterns. A body oscillation around the vertical axis occurs with a frequency similar to a pendulum (with moment of inertia J) which is mounted at the point P. The oscillation is slightly damped. Table 2 shows that the model can handle relatively large disturbances in running and comparatively small disturbances in walking. aThe body axis is defined as the connecting line between hip and COM.
626 94
biJ
"
ｾ＠
92
Q)
on"
90
§
,g "" ""
88 86 0
timet (8)
4
20
21
22
4
20
21
22
(a) 94 biJ ｾ＠
"
92
Q)
on"
'" ""Ｚｾ＠ '" ""
90 88 86 0
2
time t (8)
(b) Fig. 2. Pitch angle in perturbed walking (a), and running (b). The periodic patterns of Table 2 were perturbed by an increased initial pitch angle: 2 deg in (a) and 1 deg (b). Gray regions indicate stance phase in running and donble support phase in walking.
4. Discussion and Conclusion
4.1. General Discussion In this paper, an intuitive strategy for stabilizing the trunk in bipedal walking and running is presented. Except for the landing angle, this controller requires only internal sensors. The model does not preserve energy, however, it reveals periodic solutions which do not affect the system energy. Because the body oscillates like a pendulum mounted at the point P, we call P a Virtual Pivot Point (VPP). The emergence of stability can be understood as results of the underlying systems which are asymptotically stable (SLIP)7 or indifferently stable (pendulum), respectively. Apparently the combination is such that the indifferent stable pendulum becomes asymtotically stable. The VPP-strategy could be a basic framework for investigating postural stability and the hip function in dynamic bipedal locomotion. Interestingly, the model predicts hip torque patterns for walking which are similar in shape and magnitude to corresponding data observed in human overground walking. s
627 Table 2. Initial conditions of the selected periodic solutions (middle column) and the range of initial conditions that lead to stable gaits. initial conditions Yo (m)
xo
(m/s) 00 (deg) 00 (deg/s)
min 1.059 1.021 87.3 -5.83
walking periodic 1.082 1.038 90.0 6.47
max 1.100 1.076 93.3
min 1.021 4.81 46.6
13.84
-202.2
running periodic 1.065 5.00 90.0 -5.5
max 1.538 5.93 132.1 75.8
The proposed VPP-strategy offers an alternative concept to the ZMp9,10 for stabilizing upright gait. In contrast to state-of-the-art humanoid robots, this control offers the unique opportunity to stabilize walking and running without relying on a specific foot shape. Here, gait stability does not depend on the size of the foot but largely on the maximum stiction force allowed by the ground-foot-contact. Assuming a stiction coefficient of 0.8, to achieve the same effect of a VPP in a ZMP-controlled biped the foot length would need to be over 1.6 times leg length. Therefore, both strategies could he used dependent on external conditions and geometry of the leg. For example, on slippery surfaces the VPP torque is limited whereas on uneven ground a ZMP-based robot might fail.
30
25
ｾ＠
20
6
"'" " w w
'"
15
4-< Ｎｾ＠
OJ)
.£
10
o L __ _- L____
45
50
____L __ _-L____ ｾ＠
55
60 landing angle
65
"0
__ ｾ＠
70
__ ｾ＠
75
ｾ＠
80
(dcg)
Fig. 3. Regions of a) for stable running for a given speed (x = 5 positions of P. Only the left borders differ dependent on rp.
and three
628
4.2. Comparison with the SLIP Stable running patterns exist within a region of parameter combinations (Fig. 3). This is a subset of the region of stable running in the SLIP model. 7 With increasing rp, the size of the stable region is reduced. The model is indifferently stable for r p = 0 as the total torque vanishes and so any rotation will persist. For rp = rH = 0, the model is identical to the SLIP. There exists a minimum rp for stable solutions which is approximately 0.01 m. By shifting the VPP forward or backward with respect to the body axis, the hip torque pattern is changed such that the system energy is continuously decreasing or increasing, respectively. This additional breaking or thrusting force can be used to cope with external loads (e.g. carrying a cart or uphill locomotion). The corresponding adaptation of trunk posture is in line with the observations in human locomotion (e.g. trunk lent forward for acceleration) as well as to the concept of the Segway two-wheeled mobile systems ll or similar robots. 12
4.3. Future Work We plan to implement the proposed VPP-strategy in various bipedal robots. Also, we plan to extend the VPP-strategy for three-dimensional locomotion.
Acknowledgement This work is supported by the German Research Foundation (DFG, SEI042/4).
References l. H. Geyer, A. Seyfarth and R. Blickhan, Proc. Roy. Soc. B 273, 2861 (2006). 2. R. J. Full and D. E. Koditschek, J. Exp. Biol. 202, 3325 (1999). 3. M. H. Raibert, Legged robots that balance (MIT Press, 1986).
4. N. Neville, M. Buehler and 1. Sharf, A bipedal running robot with one actuator per leg, in Proc. IEEE Int. Con/. Robotics and Automation, 2006. 5. 1. Poulakakis and J. W. Grizzle, Monopedal running control: SLIP embedding and virtual constraint controllers, in Proc. IEEE/RSJ Int. Con/. Intelligent Robots and Systems, 2007.
6. P. Holmes, R. J. Full, D. Koditschek and J. Guckenheimer, SIAM Review 48, 207 (2006). 7. A. Seyfarth, H. Geyer, M. Guenther and R. Blickhan, J. Biomech. 35, 649 (2002). 8. S. J. Lee and J. Hidler, J. Appl. Physiol. 104, 747 (2008).
629 9. M. Vukobratovic and J. Stepanenko, Math. Biosci. 15, 1 (1972). 10. S. Kajita, T. Nagasaki, K. Kaneko and H. Hirukawa, IEEE Robot. Autom. Mag. 14, 63 (2007). 11. Segway Inc. (2008), http://www.segway.com. 12. F. Grasser, A. D'Arrigo, S. Colombi and A. C. Rufer, IEEE Trans. Ind. Electron. 49, 107 (2002).
FROM BIOMECHANICAL CONCEPTS TOWARDS FAST AND ROBUST ROBOTS D. RENJEWSKI* and A. SEYFARTH
Locomotion Laboratory, Friedrich-Schiller- Universitiit Jena, Dornburger StrajJe 23, 07743 Jena, Germany * E-mail: [email protected] 711711711.laufiabor. de P. MANOONPONG+ and F. WORGOTTER
Bernstein Center for Computational Neuroscience, Georg-A ugust- Un'iversitiit Gottingen BunsenstrajJe 10, 37073 Gottingen, Germany + E-mail: [email protected]
Robots of any kind, highly integrated mechatronic systems, are smart combinations of mechanics, electronics and information technology. The development of bipedal robots in particular, which perform human-like locomotion, challenges scientists on even higher levels. Facing this challenge, this article presents a biomimetic bottom-up approach to use knowledge of biomechanical experiments on human walking and running, computer simulation and neuronal control concepts to sequentially design highly adaptable and compliant walking machines.
Keywords: biped walking, compliant actuator, neuronal control, biomimetic design
1. Introduction
Although human technology advances rapidly and demonstrates impressive power in special applications a short look into our environment shows a lot of more flexible, robust and advanced properties and behaviors in natural beings. In nature almost nothing is developed for high performance and specialised tasks, but technically seen animals are versatile, robust and adaptive, highly integrated systems. Locomotion is a major challenge in autonomous robotics as well as in animals. As the amount of energy is limited in mobile systems, energy efficiency is of high importance. Humans
630
631
invented an efficient and high performance solution that cannot be found in nature - the wheel. Nevertheless it is limited to locomotion on even ground, across its boundaries, i.e. on unstructured terrain wheel-based systems will fast knock its limits. The alternative natural concept for fast and versatile movement on solid ground is legged locomotion. 2. Design Concepts for Legged Machines Even while the concept of legged locomotion is inspired by nature the technical systems often did not exceed the stage of morphological biomimetics. Simple walking robots have been built already in the middle of the 20 th century and advancing permanently. In the beginning machines with 4 and more legs were build to assure static stability. Drives in all joints ensure full controllability.l The biological inspiration was limited to the morphology of the leg, design and functional elements evolved from a purely technical toolbox to say stiff rotational drives, rigid mechanical chains and inflexible joints. The development of biped walking machines was strongly motivated from prostethics and service robotics. New challenges for stability, mass distribution and light-weight elements appeared. An early biped walker was WAP-1 of Ichiro Kato,2 that already used artificial rubber muscles and so was one of few elastic exceptions. On the other hand a lot of modern advanced biped robots which in tradition of mechanical engineering are built as stiff as possible. This leads to complex control tasks to avoid impacts that are typical for natural biped locomotion and may damage the structure and the joint drives. Static stability was an early and quite simple control paradigm that limited biped robots to square-cut movements and low speed. The more advanced control concept ZMP, that is still used by many up to date robots, was introduced in late 1970s. 3 This method requires permanent knowledge of system states namely precise joint-angle control, but is powerful in controlling biped machines to execute different tasks. The design and realisation process of Johnnie exemplarily shows, that ZMP-robot performance increases with computational power and battery capacity4 and is still under-achieving in terms of efficiency, disturbance handling, and natural appearance compared to human walking. 5 Similar robots of this kind are ASIMO,6 HRP-2 7 or REEM-2a. A promising, especially in energy efficiency, but mostly also stiff mechanics approach is the passive dynamic walker and its bipedal robot offsprings. The aim of passive dynamic walkers is to generate human-like movement with pure mechanics. The lack of control and energy supply does not per-
632
mit them to walk on level ground, but on shallow slopes they perform impressive dynamic gait. 8 The design once more was driven by engineering mechanics. The stability paradigm of this robot realises "limit cycle gait" for finding more efficient, natural, fast and robust walking motions. 9 On this basis actively driven robots were developed which consume remarkably low energy.10 These concepts are not advanced enough today to fullfil complex motion task. The presented approach will use functional biomimetics in addition to morphological biomimetics to bring technology closer to human running and walking skills.
3. Biomimetic Design Concept Functional biomimetics as a scientific discipline systematically deals with a technical implementation of structures, methods and development processes of biological systems.u In a biomimetic bottom-up approach the technical development is inspired by biological findings. In a sequential process specific biological functionality is translated into functional components by means of a technical design process like simulation, CADband iteration. Most modern biped walking machines use stiff kinematic chains, a large number of different sensors and powerful computers for effective locomotion. 12 This is to keep themselves in balance, to avoid impacts and to react on external disturbance, e.g. obstacles. A biped Fig. 1. Structural diagram of a biped robot as a mechatronic walking robot is a complex mechatronic system embodiment system that consists of sensors, actuators and data processing (Fig. 1). Compliant mechanisms are still difficult to handle and therefore not often used by engineers, but adaptable compliance like observed in human walking is just about to enter technical applications. To build a robust biped machine able to dynamically walk and even to run on different surfaces, it requires adaptive compliant mechanisms to handle impacts and thus reduce control effort.
ahttp://www.pal-robotics.com/index.php bComputer-Aided-Design
633
4. Biological Investigation Biomechanics of human gait was investigated in the Locomotion Laboratory. Probands walked and ran at different speeds on an instrumented treadmill and their joint motion was captured with a high speed optical system. Ground-reaction force (GRF) and center of pressure (CoP) were measured, center of mass (CoM) motion (Fig. 2), angular motion in joints Human Walking at 1.2 m/s
I
J
10il 0.99 0.97 0.95 0
2
4
3
8
567 x [m]
9
10
11
12
CoM
ｾｏｬ＠
ｾ＠ Ｎｾ＠ 0
o
'0
lJ..'"
-250 0
2
3
4
6
7
8
9
10
456 time [s]
7
8
9
10
5
time [s]
1000 ｾ＠ 1::
lJ..2
500
•
o1 : }
0
2
3
Fig. 2. Gait experiment on instrumented treadmill, CoM-motion, horizontal and vertical GRF
and joint torques were calculated from recorded motion capturing. These experiments on locomotion proofed that joint function in human legs does not correspond to any traditional technical ·actuator. Force-displacement and torque-angIe-relation suggest spring-like properties with switchable or adaptable stiffness in joints. Besides impact avoidance elastic elements can store and release energy and increase energy efficiency. Biological experiments furthermore discovered, that cyclic movements like walking or running may be driven by neural pattern generators 13 and only major disturbances are controlled on a higher level. This leads to the approach of designing mechanical parts and actuation robust enough to generate biped motion from simple patterns driving the actuators and to negoti-
634
ate obstacles until higher level control intervenes. It becomes obvious, that joints behave like springs (displacement-foree-relation) where compliance changes over time.
5. Problem Formulation The aim is to build mechanical devices that can reproduce joint behavior observed in experiments. Starting on single joint level, mechanical characteristics of human joints in motion are implemented. Extending this to all joints and considering biarticular elastic connections the complete leg behavior will be mimiced. The compliant design should reduce impacts, energy consumption and high-level control effort. The mechanical structure will serve as an explanatory model to confirm biomechanical theories and form the basis of robust walking robots that in ease of disturbances can be controlled by adaptive neural networks 14 that will actively adjust the compliance properties. This will enable these biped robots to adapt the gait to new situations. A further question in this project is asymmetry in mechanical properties. This question addresses the requirement of mechanical precision of biped walking systems and will also arise new impulses for prostetics.
6. Methods The design process for the envisioned biped robot will consist of several iterations. The aim of the first iteration stage is to build a knee joint with a clutching mechanism that can engage an extension spring in stance and disengage in swing. Biological data of the knee joint 15 show spring behaviour in stance phase and almost no internal force in swing phase. Different existing technical approaches to adjust compliance like fluidic muscles 16 or MACCEPA 17 were considered but these compliant systems may not reproduce the observed behaviour. Simultaneously a computer model is established based on the springmass model for walking and runninglS (Fig. 3). This model will guide the mechanical design and serve for defining mechanical paFig. 3. basic spring mass model rameters as well as for designing controllers and for testing control
635
stragegies. The simulation results will be validated on robot test beds derived from existing robotic platforms like JenaWaiker 19 or RunBot. 14
7. Simulation Results The first iteration addresses the interaction between the mechanical setup and the environment (Fig. 4). The corresponding model consists of a simple point mass with two spring-like, mass-less legs programmed in Matlab and Simulink. Its di4 P . d . · aSSlVe ynamlc mensions are equal to average humans (mass 80 F Ig. . setup in first iteration kg). The model is conservative and runs without actuation on flat ground. Spring stiffness, angle of attack and leg length are adjustable parameters. In first simulations self-stability was approved and experimental data matched (compare Figs. 2 and 5). To demonstrate the ability for self-stabilisation of this passive compliant walker model, the parameters were set to leg length
VW\MNWVWWMMMN
I
10
12
xCoM [mJ
,; u.
5 time [s1
10
5 time [s1
10
,; t U.
ｾ＠
-500
0
Fig. 5. Simulation results of a walking biped spring mass model (xo=O m, Yo=O.981 m, v x o=l.l m/s, Vyo=O m/s, C! =16 kN/m, c2=16 kN /m) show displacement (YCoM (XCoM), above) and GRF (Fhoriz(t) center, Fvert(t) below). Touch-down (triangle down) and take-off (triangle up) of left (empty markers) and right (filled markers) legs as well as footpoints (circles) arc depicted.
1 m, angle of attack 69"and spring stiffness
Cl =16
kN 1m. Simulation re-
636 suIts are shown in Fig. 5. In a second simulation the spring constants were changed asymmetrically to Cl =16 kN 1m and c2=18 kN 1m for left and right leg respectively. The results are shown in Fig. 6. These simulations will be validated in experiments later and help to explain biomechanical theories of human walking and running. 8. Discussion
As a result, even with asymmetric springs the model is able to stabilize passively in a very short time with appropriate initial conditions. This first result leads to some important conclusions: (1) compliant walkers are able to generate stable gait pattern, (2) certain asymmetry in design may be compensated, (3) control strategies could be derived to minimize asymmetry by tuning stiffness. As asymmetry is a common feature in simple robots and also in locomotor dysfunctions, e.g. due to amputation, further investigations may help to understand human gait far more than today. The next steps are to set up modules for a new biped robot using adaptable compliant mechanisms and to test their ability to reproduce the required force and torque characteristics for dynamic walking and running, to compare the robot results with experimental data and to introduce obstacles into the simulation for stability testing.
10
12
xCoM [m] 400
ｾ＠ ｾ＠
ｌｾ＠
5 time [51
ｾ＠ LL
ｾ＠
10
500
-500
0
10 time [5]
Fig. 6. Simulation results of a walking asymmetric biped spring mass model with similar initial conditions (compare Fig. 5) cr=18 kN/m, cl=16 kN/m
637 Acknowledgments
This work is supported by the German Research Foundation (DFG, SEI042/05). The authors would like to appreciate Susanne Lipfert and our other collaborators for providing us the experimental results of human locomotion. References 1. D. C. Kar, Journal of Robotic Systems 20, 671 (2003). 2. M. F. Silva and J. Tenreiro Machado, Journal of Vibration and Control 13, 1447 (2007). 3. M. VukobratoviC and B. Borovac, International Journal of Humanoid Robotics 1, p. 157173 (2004). 4. M. Gienger, K. Loffier and F. Pfeiffer, Towards the design of a biped jogging robot, in Robotics and Automation, 2001. Proceedings 2001 ICRA. IEEE International Conference on, 2001. 5. S. Collins and A. Ruina, Robotics and Automation, 2005. ICRA 2005. Proceedings of the 2005 IEEE International Conference on 1, 1983(April 2005). 6. K. Hirai, M. Hirose, Y. Haikawa and T. Takenaka, Robotics and Automation, 1998. Proceedings. 1998 IEEE International Conference on 2, 1321(May 1998). 7. K. Kaneko, F. Kanehiro, S. Kajita, H. Hirukawa, T. Kawasaki, M. Hirata, K. Akachi and T. Isozumi, Robotics and Automation, 2004. Proceedings. ICRA '04. 2004 IEEE International Conference on 2, 1083(26-May 1, 2004). 8. T. McGeer, The International Journal of Robotics Research 9, 62 (1990). 9. D. G. Hobbelen and M. Wisse, Limit Cycle Walking (I-Tech Education and Publishing, Vienna, Austria, 2007), ch. 14, pp. 277-294. 10. S. Collins, A. Ruina, R. Tedrake and M. Wisse, Science 307, 1082 (2005). 11. D. Neumann, D. Bechert et al., Technologieanalyse bionik., in Analyse und Bewertung zukiinftiger Technologien, ed. V.-T. im Auftrag des BMFT (VDITechnologiezentrum, Dusseldorf, 1993) p. 123. 12. J. Angeles, Fundamentals of Robotic Mechanical Systems - Theory, Methods, and Algorithms, second edn. (Springer-Verlag, June 2002). 13. K. Pearson, Brain Research Reviews 57, 222(January 2008). 14. P. Manoonpong, T. Geng, T. Kulvicius, B. Porr and F. Worgotter, PLoS Comput Bioi 3, p. e134(Jul 2007). 35, 1459(November 15. M. Gunther and R. Blickhan, Journal of ｂｩｯｭ･｣ｾ｡ｮｳ＠ 2002). 16. B. Tondu and P. Lopez, Control Systems Magazine, IEEE 20, 15 (Apr 2000). 17. R. V. Ham, B. Vanderborght, M. V. Damme, B. Verrelst and D. Lefeber, Maccepa: the actuator with adaptable compliance for dynamic walking bipeds, in 8th International Conference on Climbing and Walking Robots, 2006. 18. H. Geyer, A. Seyfarth and R. Blickhan, Proc. R. Soc. London B 273, 2861 (2006). 19. J. A. Smith and A. Seyfarth, J. Biomech. 40, p. S306 (2007).
FROM HOPPING TO WALKINGHOW THE BIPED JENA-WALKER CAN LEARN FROM THE SINGLE-LEG MARCO-HOPPER KARL THEODOR KALVERAM
Cybernetical Psychology, University of Dusseldorf, 40225 Dusseldorf Germany. E-mail: [email protected] DANIEL HAuFLE and ANDRE SEYFARTH
Locomotion Laboratory, University of lena 07743 lena, Dornburgerstr. 23, Germany. E-mail: ｯ｡ｳ｀ｵｮｩｾｪ･Ｎ､＠ Fast dynamic biped walking also includes a vertical "hopping" component which demands provisions for take-off as well as for touchdown. To explore the conditions for stable hopping with minimal risk of damage, we designed - inspired by the previously described Marco robot - a hopper model with a leg consisting of two cascaded compliant elements. Optimization runs resulted in positive damping coefficients (i.e. negative velocity feedback) to be applied from tOllch down until midstance, but negative damping (i.e. positive velocity feedback) from mid stance to take off, while stiffness barely differed in both stance phases. Thereby, the energy management by positive and negative velocity feedback turned the hopper model functionally into a mass-spring assemble capable of both stable hopping and avoidance of hard impacts. Keywords: cascaded compliance, stable hopping, positive velocity feedback, landing impact
1. Introduction
Legged locomotion as targeted by the mass-spring model I includes, besides of the horizontal propulsion, also a vertical component with alternating stance and flight phases of each leg. This "hopping" movement is the topic of the present paper. The Marco hopper robot 2.3 is a testbed designed and build for the investigation of the conditions leading to stable hopping. The robot (see Fig.l) comprises a 1.4 kg sledge (the body), and a 0.4 kg rod (the leg) which is actuated by a DC motor fixed to the body and driven by software. Upright bars force both the body and the leg to move in the vertical direction only. First experiments 2 revealed that stable hopping could be achieved using software mimicking a mechanical spring, whereby the energy lost by damping and friction was supplied either by
638
639 augmenting the stiffness in the seeond half of the stance phase, or by an appropriately shaped positive force feedback during the whole stance phase. IShank
Y
tooth belt --+ motor body ----)
Ybody
Figure 1: Schematic of the Marco hopping robot2 Using Simulink, arbitrary forcing functions can related to the body - be applied to the shank via current, motor and tooth belt, The ground reaction force (grt) is measured by a strain gauge, the body position Ybody related to ground by a Posimag system, and the shank position Y,bank related to Ybody by integration of the output of a tachometer fixed to the motor shaft.
However, hard ground impacts occurred which subjected the robot to damage. We softened these impacts intuitively by attaching a passive damper (ball of Adipren®), d=O.035m) to the leg's lower end. The added compliance, however, made the flow of energy within the now two-segmented leg intransparent, and also stability was difficult to maintain. Here we present an analysis of a hopper model with a leg comprising two compliant elements in series, and propose how a soft landing self-stable hopper can be achieved applying such a cascaded compliance design.
640 2. Method Fig.2 depicts snapshots of the behavior of the model taken for the analysis. The stance phase ranges from t, (touch down) to t3 (take off), with midstance at t2 where the COM reaches the minimum height. V=O
body
shank
leg
foot
ｾ＠
ｾ＠ ｾＡ＠
ｾ＠
ｾ＠
J
ｾ＠
touch downJ stance1 stan'ce2 Ltake off
Figure 2: Hopper model consisting of three point masses mo. ml, m2 represented as spheres. Between the masses spring-damper elements with lengths b, h are placed, which have the rest lengths 1\0, ho, the stiffness coefficients kl' k2, and the coefficients of viscous damping bl, b2. Mass mo can be interpreted as body, the masses ml and m2 as belonging to segments of a leg, here called shank and foot. yo, YI, Y2 denote the actual vertical positions of the masses. h indicates the initial value of Y2 when the hopper is released, and YCOM (not shown) the corresponding position of the center of mass (COM). Displayed are five snapshots at points of time to, tl, t2, t3, 14, which represent a full stancelflight cycle. The character v on top indicates the velocity of the COM in the respective snapshot.
We split the stance phase into the phases stance 1 responsible for breaking, which runs from touch down to mid stance, and stance 2 responsible for the subsequent acceleration, which runs from midstance to take off. The scenario followed Marco's mechanics and style of operation as close as possible, and was
641
given by the masses mo=1.4kg, ml=OAkg, m2=Okg, rest lengths 110 = O.lm, l2o=0.03Sm, and start height h=0.20m, which placed Ybody at 0.33Sm. The coefficients of stiffness kI, k2 and viscous damping bI, b2 were subject to a twostage evolutionary optimization procedure. In the first optimization run, the algorithm started with k l =IS00N/m, k 2=7000N/m, b l=b 2=ONs/m, and aimed at minimizing the peak ground reaction force during stance 1. The algorithm comprised 100 trials, and in each trial the coefficients of stiffness and damping where randomly changed once. The second optimization run started with the coefficients resulting from the last trial of the first run, and enclosed, too, 100 trials. Here, the goal was to get a parameter combination applied during stance 2, which drove the COM's apex during flight as close as possible to the start height YCOM' But now only kl and b l were allowed to vary, according to the fact that also in the Marco hopper the spring properties only of the first segment of the leg could be submitted to variation. So, after finishing the optimization task, during stance 2 the parameters resulting from run 2 were applied, while during stance 1 and flight the parameters resulting from run 1 remained valid. Computations were made by MATLAB and SIMULINK (fixed step sample time 1 ms, solver odeS). 3. Results Fig.3a reveals that continuous hopping based on undamped springs is possible, but leads to inter-mass oscillations and hard multiple ground contacts unseen in human or animal hopping. As outlined in Fig.3b, via optimization, however, parameter sets can be found by which stable hopping is achieved, while mUltiple ground contacts and hard ground impacts are avoided. Table 1 pinpoints that this demands different spring parameters for breaking in stance 1 and acceleration in stance 2. The greatest and most important change undergoes spring 1: Here, the optimization routine turns the positive coefficient of damping bI into a negative one, while the stiffness kl is left unchanged in essential. FigA illustrates how the spring parameters and the optimization criteria approach their final values in the two-stage optimization procedure. The first optimization run (trials 1 to 100) diminishes the peak ground reaction forces by about one third, but lessens also the apices considerably. In the second run (trials 101 to 200) the apices completely recover, while the peak ground reaction forces retain the low level reached in the first run. The drop of the damping coefficient bl from positive to negative values in the second run seems most remarkable, while the remaining coefficients do not essentially change.
642 ground
a
contact
0.5
1.5
1
b E ...... c:
o
:;:::;
·iii
2.5
2
-
ground
_ L - -_ _'---------',
::ntact
02
ﾷｯｾ＠
o
c.. -0.2: _ _ __
o
0.5
1
1.5
2
2.5
time [5]
Figure 3: Hopping a with undamped springs, b with springs whose stiffness and damping coefficients are optimized for minimal ground reaction forces (grf) and stability. Ground contacts are indicated by black bars on top of the positional trajectories.
Table 1: Parameters (absolute values) applied to the leg according to the optimization procedure. PO: Start values. PI: Results of a typical run I (to be applied in stance I and flight). P2: Results of a typical run 2 (to be applied in stance 2, while in stance I and flight the results of run I remain valid).
PO PI P2
stance 1, flight bl kl k2 b2 [kN/m] [Ns/m] 0 0 1.5 7.0 1.2 6.7 11.0 33.5
kl k2 [kN/m] 1.5 7.0 1.3 6.7
stance 2 bl b2 [Ns/m] 0 0
-18.1 33.5
643
2
....... _ ＬｾＭ
........ --.
ＱＮＵｾ［Ｚ｟＠ 1· Q)
Ｎｾ＠
)1··
ＢｾｫＲ＠
= ｾ＠
apex
0.5.
fI)
Q)
:::-
;: ccs (j) b.
Oil
grf
-0.5· ｾ｢Ｑ＠
I
-H
-1.5
_21 __-
o
I
ＭＢＮｾ
50
100 trial number
150
200
Figure 4: Relative sizes of the ground reaction force (grf), the apex, and the coefficients of stiffness (kl, k2) and damping (bl, b2) during optimization. Trials I tolOO mirror a typical run aiming at minimization of peak gronnd reaction force during stance I. Trials 101 to 200 reflect a typical run aiming at keeping the apex onto the start height in the flight phase. Parameters of stiffness and damping are divided by the values reached in the first run, while grf and apex are divided by their start values.
4. Discussion The Marco robot is designed to perform real world hopping: On touch down, the foot when reaching the ground breaks the shank, and the shank when coming to rest breaks the body. For take off, the body leans on the leg standing on the ground and gets accelerated for leaving the ground. In order to understand the flow of mechanical energy within this machine, we modeled it in a first approximation by software using virtual damped springs to actuate shank
644
and foot. Main results of the investigation of such a cascaded compliance design are: - Gentle hopping inevitably requires damping, a fact often ignored in hopping models. Damping, however, deprives the system of energy, so hopping must cease after a while. To regain stability, the lost energy must be replenished. - Optimization with respect to ground reaction force provides coefficients of stiffness and damping which, when applied from touch down until midstance (during stance 1), will lead to minimal peak ground reaction forces. - Optimization with respect to stability gets just that negative damping coefficient which, when applied from midstance to take off (during stance 2), will lead to constant apices. Considered physically, a positive damping coefficient means leakage of energy, while a negative damping coefficient means supply of energy. Referring to the differential equations describing a damped mass-spring system, respectively using control-theoretical terms, positive damping can be described as "negative velocity feedback", and negative damping as "positive velocity feedback". In a biological system, sensors in the muscle spindles are available which code the velocity of a muscle's shortening/lengthening. This signal, when fed back to the motor-neurons innervating the muscle, can - dependent on the absolute value and the sign of the gain in the feedback loop - provide a scaleable energy management during the hopping cycle. Evidence for such a type of neuromuscular compliance control comes from experiments in human goal directed forearm movements 5 : Here, externally applied positive damping was completely compensated through internally generated negative damping, and vice versa - a finding which is explainable in the framework of adaptable inverse controI 6,7. Regarding the present paper, energy management by negative velocity feedback during stance 1 and positive velocity feedback during stance 2 turns the hopper model functionally into a mass-spring assemble capable of both stable hopping and avoidance of hard impacts. The next step will be to compare the reported results with data produced by the real Marco hopper.
Acknowledgments. This study was supported by grants KA417124 and SE104212 of the German Research Foundation (DFG).
645 References 1. Blickhan, R. The spring-mass model for running and hopping. 1. Biomech. 22,1217-1227(1989) 2. Seyfarth, A., Kalveram, K.T., Geyer, H. (2007) Simulating muscle dynamics in a simple hopping robot. In: K. Berns, Luksch,T.(eds) Autonome mobile systeme. 20. Fachgesprach, Technische Universitat Kaiserslautern, Springer, 2007 (pp.294-300) 3. J. Smith, K. Kalveram, F. Ida, R. Rummel, Y. Blum, S. Lipfert, A. Karguth, and A. Seyfarth, Biologically inspired compliance strategies in robotic legged locomotion (2008, submitted) 4. Geyer H, Seyfarth A, Blickhan R. Compliant leg behaviour explains basic dynamics of walking and running. Proc. Roy. Soc. Lond. B, 273: 28612867,2006. 5. Kalenscher, T., Kalveram, K. Th., & Konczak, J. Effects of two different dynamic environments on force adaptation: Exposure to a new force but not the preceding force experience accounts for transition- and after-effects. Motor Control, 2003, Vol.7, 242-263. 6. Kalveram, K. T., Seyfarth, A. Learning the inverse model of the dynamics of a robot leg by auto-imitation. In: K. Berns, T. Luksch eds) Autonome mobile systeme. 20. Fachgesprach, Technische Universitat Kaiserslautern, Springer, 2007 (pp.308-314) 7. Kalveram, K. Th. A neural network model rapidly learning gains of reflexes necessary to adapt to an arm's dynamics. Biological Cybernetics, 68, 183191, 1992
INITIATING NORMAL WALKING OF A DYNAMIC BIPED WITH A BIOLOGICALLY MOTIVATED CONTROL T. L UKSCH* and K. BERNS
Robotics Research Lab, University of Kaiserslautern Kaiserslautern, Germany * E-mail: [email protected] www.rrlab.cs.de
Two-legged locomotion is a much reseached topic in the robotics community since many decades. Nevertheless human walking and running is still unequaled. This paper introduces a biologically motivated approach of controlling bipeds that is based on recent results from neurological research on human walking. It features a hierarchical network of skills, motor patterns and reflexes that works locally and distributed and tries to exploit the natural dynamics of the system. The control concept is illustrated by the process of walking initiation.
Keywords: Biped Locomotion, Reflexes, Reactive Control, Behaviour-based Control, Passive Dynamics
1. Introduction
When controlling the locomotion of two-legged robots, two different approaches can be observed. The technical approach relies on concepts developed for industriel robotics and sound mathematical calculations, but several shortcomings can be observed: • • • • •
Mostly no exploitation of natural dynamics or elasticities No natural looking motions High energy costs and computational demands Low robustness and adaptability Dynamic model is necessary, which can never be exhaustive
On the other hand, the biological approach tries to transfer results from neurological or biomechanical reseach to technical systems. As nature's solution to biped walking still outclasses any technical solution of today, the authors suggest to follow the second, biologically motivated approach. Un-
646
647
fortunately the knowledge on nature's neuromechanical control concepts is far from being completely understood, so only part of them can be used as inspiration to a biped control system.
2. Related Work Research results of the last years seem to support the assumption that neural motor control is of a hierarchical layout. Bizzi et al. found a spatial connection of stimulation of regions in the spinal cord of frogs and the kinematic reaction of its legs. 1 They suggested the existence of motor programs or modules creating activities of whole groups of muscles. Later results on the combination of such modules for movement show that the same modules are even used for different modes of locomotion. 2 Analysis of human motor control lead to similar finding. Ivanenko et al. used statistical methods like factor analysis or peA to show that muscle activity patterns recorded using EMG during walking can be explained by the sum of only five basic temporal components. This holds true even for different walking speeds and on supporting body weight to various degrees. 3 A spatial mapping to the spinal cord could also be shown. 4 Further results imply that the same five motor patterns ean account for both walking and running with only a phase shift of one of the temporal components. 5 Still there remains the question of a semantic interpretation of these basic patterns, and if a interpretation explains how the patterns evolved. While the observed motor patterns seem to be mostly of a feed-forward nature, one must ask how sensory feedback is incorporated. The reaction of reflexes must be combined with the muscle activation of motor patterns that are themselves modulated by various stimuli. Rossignol et al. discuss the dynamic sensorimotor interactions in the spinal cord and at supraspinal levels. 6 Zehr and Stein review research on the modulation of reflex responses during static tasks and locomotion. 7 During the initiation of normal walking, the human body rotates about the ankles like a flexible inverted pendulum. This motion is created by stereotypical activation of the lower extremities' muscles. s The main action results from activity of the hip muscles. 9 The following key aspects of natural motion control can be identificd: • Mechanics is optimized for the task by evolution ("intelligent mcchanics") • Heavily exploitation of natural dynamics and energy storage in elastic components • Self-stabilizing properties of elastic elements
648
• Hierarchical control from brain via spinal cord to motor neurons • Proprioceptive feedback triggers reexes and modulates motor programs and CPOs • Distributed subsystems reduce signal density and parameters • High performance despite relatively slow signal transfer and computation units There have been various attempts to control biped robots using methods inspired by biological insights as those just mentioned. Oeng et al. implement a reflexive neural network for a small planar walker. 10 They show that fast walking is possible without planing of trajectories but rather by using local reflexes and by exploiting the passive dynamics of the mechanical system. l l In can be demonstrated that a purely reactive sensorimotor neural network can produce a walking gait in a 8 DoF simulated biped. 12 Only a few works can be found on controlling fully articulated bipeds, let alone experiments on real hardware. Endo and his colleagues propose the use of a neural oscillator and feedback pathways similar to Kimura's work on quadrupeds. 13 They tested the approach on the QRIO robot, but used inverse kinematics of the legs to generate trajectories.
3. Controlling Dynamic Motions of Bipeds with Reflexes and Motor Patterns The approach described in this paper tries to incorporate features of natural locomotion control as those described above into a robotics control archi tect ure: (1) The system is structured as a hierarchical network of control modules. This way it is possible to represent different levels of neural motor control like reflexes or motor patterns. The layout of the control system is shown in figure lao (2) The control components are local and distributed. No elaborate models of the robot or its environment are necessary and no explicit trajectories are includes. The complexitys is reduced. (3) Reflexes introduce a tight sensor/actor coupling for fast responses to stimuli. Reflexes can be inhibited or react different depending of the phase or mode of locomotion as it is the case in biological control. (4) Motor patterns allow for temporal synergies of a few cooperating joints. The patterns can be modulated by descending pathways or proprioceptive inputs, i.e. by high-level modules, sensors like inertial system or
649
load cells, or measurement of joint torques and angles. Torque impulses instead of trajectories do not force robot into unsuitable motions. (5) The passive dynamics of the mechanical system and its interaction with the environment are allowed to contribute to the overall motions. This leads to low energy consumption and natural motions. (6) Easy and transparent fusion of different control unit's outputs for similar actuators is possible, so no additional work on arbitration is necessary. The system is based on a behavior-based control framework that was successfully used before on various robots by the authors and others (e.g. Ref. 14) and allows to implement the characteristics just mentioned 3 • Brain
Ｑ Ｇ｟［ｌｇＮｾｯｭＬｴｩｮ＠
Behaviors
(a)
I
(b)
Fig. 1: Layout of the control units and the simulated biped Designing an architecture supporting the features just mentioned is not sufficient. There still remains the difficult part of finding the proper reflexes and motor patterns for the control network to do the aspired job. One of the ways proposed here is the analysis of muscle activities and temporal basic components appearing in human motor control. For parts of this data there already exists a semantic interpretation by biological research, e.g. there exists common agreement on the existence of several reflexes involved aThe behavior-based control framework iB2C can be downloaded at http://rrlib . CS. uni-kl.de
650
in locomotion and posture regulation. Other reflexes or motor patterns are designed to match certain muscle activities or to handle remaining control issues. One of the common design guidelines for control units is to prefer torque control over position control to incorporate the passive dynamics of the robot and the environment. Instead of biological data, results from numerical optimization calculations can also give similar insights that are closely fitted to the technical system. This technique has also been used by the authors. 15 In contrast to most approaches using a more biologically inspired control, the proposed method is applied to a highly complex biped robot. The fully articulated humanoid features six degrees of freedom (DoF) per leg, a three DoF spine and three DoF arms, 21 DoF in total (Fig. 1b). The robot is dynamically simulated and includes mechanical properties like elasticity as those found in the biological example. A control network for the simulated biped is developed using the proposed methodology. It enables the robot to walk and to keep balance on moving ground and against other disturbances like external forces.
4. Initiation of Walking To illustrate the features and the design procedure of the proposed approach, the initiation of normal walking is presented. The process of initializing the first step is examined in biomechnical research,8,9 but is seldom in the focus of robot control.
Fig. 2: Interaction of locomotion behaviours in the brain group
Figure 2 gives an overview on the high level locomotion behaviours. When stimulating the walking skill, it will first enable the walking initiation.
651
When this behaviour is content, the cyclic walking behaviour will become active and inhibit the standing skill (inhibiting connection with the dotted end).
500 ms ,-----,
L Gluteus
0.2 mV
]
medius
L Adductor
R Gluteus
::J
medius
R Adductor
lateral torque 10 Nm
J
-----t-'
Fig. 3: EMG measurements during human walking initiation 9
The walking initiation process consists of a forward motion of the whole body by adding torque to the ankle joints, taking the load of the first swing leg and starting the first step. For moving the body's center of gravity in direction of the stance leg, a motor pattern is stimulated. This pattern is derived from EMG measurements during human walking initiation (Fig. 3). It can be seen that the adductor muscle of the stance leg and the glueteus medius muscle of the swing leg is active. This is translated into a motor pattern producing torques in the hip joints. It must be noticed that the rest of the body joints remain passive except a certain elasticity, so the whole body movement results from passive dynamics. The foot contact forces in simulation (Fig. 4b) can be compared to the ground reaction forces measured in human initiation of normal walking (Fig. 4a). It can be seen that the characteristics of force progression are the same. Most noticable the load of the designated swing foot first increases as reaction to the hip movement, but than decreases to zero as the natural body dynamics moves the centre of gravity over the stance leg. The swing leg is than free to take the first step.
652
ｎＷＰｾＧｅｩ＠
aJ
2 ｾ＠
ｾ＠
ｾ＠
ｾｐ｜＠
400
_
ｾＧ｜＠
I
\,
-----
"
200
\, \\
" ＧｾＭ
o 0.00
0.20
0.40
0.80
D.IO
1.00
Tlm,,{uc)
(a) Human contact forces 8
ｾ＠
OJ
40
600 500
300 200 100
ｉｾｐ］ＭｃＺｪ＠
ｯｾＭ＠
o
0.2
0.4
0.6
0.8
1
t [sec]
(b) Simulation results
Fig. 4: Comparing foot forces during walking initiation
5. Conclusion and Outlook A control concept for dynamical biped motions has been suggested. Based on the findings in biomechanical and neural research, a hierarchical network of skills, reflexes and motor patterns is designed. Those control units are derived e.g. from motion and muscle activity analysis, but can also be found by mathematical optimization. A network for stable standing, walking initiation and walking has been created. Future work will focus on walking robustness by adding further reflexes modulating the walking motion. Standing stability will be increased by adding a stepping strategy besides the ankle and hip strategies. The construction of a biped prototype will continue to test the control concept on a real robot.
References 1. M. A. Lemay, J. E. Galagan, N. Hogan and E. Bizzi, IEEE Tmnsactions on Neuml Systems and Rehabilitation Engineering 9(March 2001). 2. E. Bizzi, V. Cheung, A. d'Avella, P. Saltiel and M. Tresch, Bmin Research Reviews 57, 125 (2007). 3. Y. P. Ivanenko, R. E. Poppele and F. Lacquaniti, Journal of Physiology 556 (2004). 4. Y. P. Ivanenko, R. E. Poppele and F. Lacquaniti, The Neuroscientist 12,339 (2006). 5. G. Cappellini, Y. P. Ivanenko, R. E. Poppele and F. Lacquaniti, Journal of Neurophysiology 95 (2006). 6. S. Rossignol, R. Dubuc and J.-P. Gossard, Physiological Reviews 86, 89 (2006). 7. E. Zehr and R. Stein, Progess in Neurobiology 58, 185 (1999).
653 8. R. J. Elble, C. Moody, K. Leffler and R. Sinha, Movement Disorders, 139 (1994). 9. S. Kirker, D. Simpson, J. Jenner and A. Wing, J. Neural. Neurosurg. Psychiatry 68, 458 (2000). 10. T. Geng, B. Porr and F. Worgotter, Neural Computation 18, 1156 (2006). 11. P. Manoonpong, T. Geng, B. Porr and F. Worgotter, IEEE Symp. on Circuits and Systems (ISCAS) (2007). 12. C. Paul, Adaptive Behavior - Animals, Animats, Software Agents, Robots, Adaptive Systems 13, 67 (2005). 13. G. Endo, J. Nakanishi, J. Morimoto and G. Cheng, Experimental studies of a neural oscillator for biped locomotion with qrio, in Proceedings of the IEEE International Conference on Robotics and Automation, (Barcelona, Spain, 2005). 14. J. Albiez, T. Luksch, K. Berns and R. Dillmann, A behaviour network concept for controlling walking machines, in 2nd International Symposium on Adaptive Motion of Animals and Machines (AMAM) , (Kyoto, Japan, 2003). 15. T. Luksch, K. Berns, K. Mombaur and G. Schultz, Using optimization techniques for the design and control of fast bipeds, in 10th International Conference on Climbing and Walking Robots (CLA WAR), (Singapore, 2007).
MOTION DESIGN FOR AN INSECTOMORPHIC ROBOT ON UNSTABLE OBSTACLES YU. F. GOLUBEY' Mech.-Math. dept., Lomonosov Moscow State University, Moscow, 119899, Russia • E-mail: [email protected]
Y. Y. KORIANOY" Keldysh Institute for Applied Mathematics, Moscow, 125047, Russia ** E-mail: [email protected]
This paper develops the results of works Refs. 1-4. Methods for designing of insectomorphic six-legged robot motion to overcome complicated obstacles by means of Coulomb friction are presented. Those obstacles are two identical lofty horizontal shelves connected by a narrow horizontal beam, a ladder leaned against a vertical wall of the shelf and a ball which can roll along a horizontal plane. The 3D computer simulation was fulfilled to demonstrate effectiveness and robustness of elaborated methods for obstacles overcoming.
Keywords: insectomorphic robot; control design; obstacles; simulation.
1. Introduction
The ability of a walker to overcome a terrain with a conglomeration of high obstacles can be formed by teaching step by step the robot to overcome isolated obstacles as well as reasonable their combinations. Some examples of overcoming a terrain with small roughness are given in Ref. 5. Also machines which uses vacuum suckers or principle of an electrostatic ad hesion 6 are developed intensively as they allow moving along vertical walls. It seems that in some particular cases walking machines could move along vertical constructions of significant height simply using Coulomb friction forces as animals does it. Methods for overcoming of vertical column by means of friction forces are presented in Ref. 1. The design of insectomorphic walker motion for surmount combination of two obstacles was presented in Refs. 2,3. The robot motion on the step ladder leaned along the vertical
654
655 wall was investigated in Ref. 4. In this paper the new results for overcoming pair combinations of obstacles are discussed. In particularly, the problem of saving equilibrium on the massive ball which can roll along the horizontal plane without sliding is considered.
2. Main Restrictions of Motion Design The robot consist of parallelepiped-shaped rigid body of mass Tn with dimensions a - - side of the body (length), b - front or rear edge (width), c - height. Six identical insectomorphic legs arc symmetrically attached to the sides of the body. Points in which the legs are attached (legs attachment points) are located uniformly along the sides. Each leg consists of two links: hip, length II, mass mi and shank, length 12 , mass m2. Dimensions of the robot meet the following condition: a: b : c:
h : 12
=
1 : 0.5 : 0.1 : 0.5 : 0.33.
Let us link the right Cartesian coordinate system Bxyz with the body. B is the center of the body, axis Bx is directed to its front edge parallel to the lateral facets, Bz is the structural vertical. The position of a leg is determined by three joint angles, two of which (Oi' Pi) defines the position of a hip relatively to the body, and third h'i) - of a shank relatively to the hip. Thus, the total degrees of freedom of the robot is 24. The joint angles of the leg numbered 1 :s; i :s; 6 can be found unambiguously of vector ri, directed from the attachment point to the foot on the assumption of an orientation of the knee. By default the knee is oriented so that if the foot moves forward the knee moves forward. Legs are numbered so that leg 1 is the rear right leg relative to the positive direction of Bx axis, rear legs have numbers 1 and 2, middle legs - 3, 4. Thus, all right legs have odd numbers. Joint angles are determined by the inversion of the following correlations: I Txi = Xi - PJi = [( -l)i- h sin(3i + 12 sin((3i + Ii)] sinoi, I Tyi = Yi - Pyi = -[( -1)i- l sin (:Ii + 12 sin(pi + Ii)] cos Oi, 1
T zi
= Zi -
Pzi
= (-1) i-I h cos Pi + 12 cos (Pi + Ii)'
where (Xi, 7J;, Zi) - coordinates of the 'i-th foot, (P'£i, ,Pyi, Pzi) - coordinates of the i-th attachment point in the body reference frame. The interaction of a foot with the surface is supposed as the viscouselastic Coulomb model of friction forces i . The motion of the robot should be comfortable. It means that condition of static stability must be satisfied as possible. When static stability is
656
impossible, the algorithm should use methods of dynamic stability 3. Robot can touch supporting surface only by feet, and legs of robot doesn't have to intersect each other during all time of the motion. We assume that robot is equipped with the electromechanical drives in joints and has full access to the following information: the geometry of the obstacles, own position relatively to the obstacles, joint angles and velocities. The programmed values of the joint angles are generated by the algorithm of control. The algorithm is not strictly fixed, the information about the actual robot configuration during the motion essentially used. Realization of programmed values is accomplished as in Ref. 1. Legs transfers are realized on the base of the plane step cycles 1 . Step cycles are modified in dependence on surfaces of obstacles, velocity of motion of the robot, prescribed footsteps. Motions of feet along the step cycles are smoothed to save the continuity of motions and their velocities. 3. Sample Results
Algorithms of robot's motion design was worked out by means of computer simulation of full dynamics of robot and its environment. For that purpose the software complex "Universal Mechanism" 7 was used. Example 3.1. Motion on a Horizontal Beam Consider an obstacle composed by two parallel shelves of the same height with a horizontal top area. These top areas are connected at the same level by a narrow beam that is perpendicular to the vertical walls of the shelves. Suppose that, at the initial moment, the robot has a symmetric pose on one of the shelves before its edge, and the planes of the legs are perpendicular to the longitudinal axis of the body. The robot has to go from one shelf to another along the beam. The assumption that the beam is narrow means that the transversal size is approximately equal to the margin of static stability. Therefore the front and back legs are still applied to provide the support of the robot on the surface, and the middle legs work as a flywheel in order to provide the robot stability when the conditions of static stability are violated. The tracks are to be under the robot body in the course of motion. For a small distance between the tracks, because of the danger of mutual intersection of symmetric legs, it is advisable to admit that the knees move in the direction opposite to the direction of movement of the feet. The motion is executed according to the following stages3 : (1) the robot configuration is changed for the motion with a narrow track; (2) the robot goes on the beam with a four-legged diagonal gait to the another shelf; (3)
657 the robot configuration is changed in the reverse way in order to go with a wide track. When the diagonal pair of legs is supporting, the body is a physical pendulum fixed on the axis that passes through the supporting points and located at the upper unstable equilibrium position. To stabilize this position, the straightened middle legs execute a coordinated rotation in the plane perpendicular to the longitudinal body axis. As the measure of deflection of the body from the vertical axis, we can take the projection ｾ｣＠ of the robot center of mass on the axis ｏｾＮ＠ The control torque Mf for angle f33 of middle third leg is performed by the formula (3
M3 = ＭＨｃｬｾ･＠
+ ｣ＲｾＩＯ＠
.
- M g,
Mg = (ml
+ m2)gle sinf33,
where Ie is the distance from attachment point to the center of mass of the straightened leg; Ci, i = 1, ... ,6 are the feedback gains. The motion of the fourth leg relative to the body is skew-symmetric to the motion of the third leg (Fig. 1). The angular velocity of the middle legs accumulating in the loss of static stability is eliminated in the course of joint standing of the front and back legs.
ｉｾ＠
Fig. 1.
Balance on the beam
Example 3.2. Climbing Along the Step Ladder. Ladder of length I leans on the vertical wall and is on the horizontal plane and forms an angle 'P with it. It can be proved that the equilibrium of the ladder and robot on it can be achieved only when tan'P 2': 1/ j, where j is the friction coefficient on horizontal plane. The robot, locating at the bottom on the horizontal supporting plane, should move to the upper horizontal ground of the shelf of height h using a step ladder (Fig. 2a). For definiteness, we accept that I r:::; h. The pitch between the top edges of ladder footsteps is O.4a. The absolute right coordinate system ｏｾＷ｝Ｈ＠ is connected with the shelf. The origin 0 is put at its base. The axis O( is upward directed and lies in the plane of the side wall of the shelf on which
658
(
ｾ＠
(a) Fig. 2.
ｾ＠
"
...
ｾ＠
71
(b) The initial position for climbing the ladder (a), the scheme of body motion (b).
the ladder leans, and the axis 071 is oriented along the outer normal of the specified wall (Fig. 2a). The ladder plane is parallel to the axis ｏｾＮ＠ The climbing onto the ladder is executed by the gallop gait in the course of three maneuvers. 1. The maneuvers of rotation of the body from a horizontal position to an inclined one. 2. Regular motion along the ladder in the upward direction. 3. Maneuvers of inverse body turn to the horizontal position and the passage to the standard position for the triples gait. Irregular carryings are determined by a list in which the number of the carried pair of legs is explicitly specified 4 . The maneuver of body turn in the passage on the ladder is arranged similarly to that in Ref. 1. The lateral segment of the body is inscribed in the angle between two guidelines £ and M (Fig. 2b). The distance (J" between the front edge of the body and point A is given by the function of time 4 . The main idea of regular motion along the ladder repeats the modification of the gait gallop in climbing a corner 3 - the middle legs are carried twice as often so that they turn out to be close to the pair of legs that has to be carried next. To revolve to a horizontal position, inverted guidelines for the motion of the front and rear body edges are used. The guideline parallel to the ladder is in charge for the motion of the rear leg, and the horizontal guideline provides the motion of the front edge. Thus the triangle in Fig. 2b is turned over. Figure 3 contains the plots of vertical coordinates of robot's body and feet in dependence of time. Symbols (i) i = 1,3,5 means (-coordinates of corresponding feet, (f corresponds to the front edge of the body. Feet
659 with even numbers have as a rule identical coordinates except small time segment at 36.48 s when 1 and 2 legs are transferred one after another. Time to is initial moment for climbing to the ladder, t2 is the end of the gait adaptation, t12 is the end of the body rotation for climbing, at t17 all legs are on the ladder, t33 is the ending of the body rotation to the horizontal position. Figure 4a gives some position of the robot in course of the motion.
,
·_-_··--··1-
2.34
__________ L
.. __ ..... ___ L
1.95
-----f-
--- ---
1.56
----j
------- ---r
3.12
..
2.73
1.17
------
ｾＭ
I
I
0.78
------f
---------4-
0.39
----I
- - -------r I
0
to t2
12.16
24. 32 t 17
iI2 Fig. 3.
36.48
Climbing the ladder, plots
(b)
(a) Fig. 4.
Robot on obstacles
t33 48.64
660 Example 3.3. Balance on the Moving Ball. The robot is staying on the massive ball and should provide for own stability at this position (Fig. 4b). Initially the vertical projection of robot's center of masses isn't coincide with the point where ball touch the horizontal supporting plane and ball may have some initial velocity. Robot have to move its center of masses so that mentioned projection coincide with the supporting point of the ball and ball doesn't move. Firstly let us consider the case when supporting points don't change on the surface of the ball. Denote Ks angular momentum of the system relative the contact point S of the ball and the horizontal plane. Let to and tk are the initial and the final time moments of stabilization step. Suppose the body of robot moves parallel to the velocity of ball center. Then we have h
Ks(tk) - Ks(to) =
J
M[g x rc(t)]dt,
to
where M is the full robot's mass, g -- vector of the weight acceleration, r c - the radius-vector of robot's center of mass from point S. It is supposed that Ks(to) -.l g. Between the radius vector rb of body center and rc in the neighborhood of equilibrium position we should have the differential equation: drb = Xdrc(t). At the end of the stabilization step we should have Ks(tk) = 0, rc(tk) II g. When rc(to) isn't parallel to g the body motion is defined by formula rb(t) = rb(td'\'(7) + (1 - '\'(7))rb(tO), where 7 = t - to, T = tk - to and
0::; 7 < 71, 72/2' '\'(7) -,8(7-7d2/2+71(7-7d+7f!2, 71::;7 YUpright
］ｾＮ＠
NxiTxi
ｾ＠
is always perpendicular to
N;
.We have
+ NviTvi + NziTzi =0
From (3) the maximum magnitude of
ｉｾ＠
under no slip condition is fLi
(4)
IN I. i.e i
(5)
762
Case I: If
IN;
I= 0 then ｉｾ＠
components N x;, Ny;, N zi
= O. Case 2:
If
IN;
I"* 0 then anyone of the
* 0, let N zi * 0, then from (4) and (5) we get
a"* 0 equation (6) is quadratic in nature and will have real roots if
.Since
b 2 - 4ac 2 0 ｾ＠-'-" JLilNil JNziJ > - Txi which is always true by (2). Therefore we get T. = - ,l1i N ,i!Ni!IN,il
and
" A
t;
= ｉｾ＠
T
= [t,i tvi
Now
ｎＧｩｾＫ＠
ANT = IN:I = [n x; n v; n zi ] .Now
T
tz;J and
let
quasi-static
n;
analysis is done on the system assuming the masses of legs and wheels are net negligible when compared to the mass of the chassis. Net force and moment acting on the system is given by F
= [Fx
the
M
r=I (I, 4
F, Fz
+ N;) + Mg . M is the mass of the chassis and g is
acceleration
= ｉ｛ｾｬＱｩ＠
4
X
Ｈｾ＠
+N
due
j )]
= [M x My M z
to
Y·
gravity
ｾｉｬｩ＠
the CG of the chassis to the point of contact of the the chassis, 2 W ｾ＠
r
jeli
where,
r;
{R.[-W(-IY+la
= =
R.[
W
is the radius vector from ith
wheel.
2a ｾ＠
Length of
Width of the chassis
ｾｉｬｩ］ｪ｡Ｋｲ［＠ _
and
(_ly+l
a
r;=R.[OO-l;f, r;=-r;.n i ·
of ViE {1,2} of ViE {3,4)
R.[O - rsin(rJ - rcos(r) f, ri "'" arctan ( ｎｾｺｊ＠
Where
I; is the length of the
ith
of the torodial cross section.
leg, r is the radius of the wheel, Let[m tx ; m ly ; m lzi
f ,
r; is the radius
[m nx; m llyi mnzi
f'
be the
763 unit moment vectors due to T; and
N;
respectively.Now
A.C=D
(7)
Where
C ］｛ｉｾ＠
IINzllt; ｉｎｾ＠
t,1
nxl
tyl
I1Yl
Intx1
nx2
tx2
ny2
I1z1
tzl
A==
[F, Fy Fz M
D
nz2
tz2
In nx ]
II
N311 ｾ＠
t x3
I1x3
t 1'3
ny4
t z3
My M z]T,
nz3
tx4
t,4
mn>:2
mlx3
mnx3
mtx4
m!y3
mny3
m ty4
nttz3
m llyl
m'1'2
mny2
lntzl
1nll21
In/::;2
mnz2
mnd
4
n y4
mtx2
m ty1
IIN 1Y nx4
mtz4
n zA Innx4
,
m lly4
mnz4
4
Let rnin(S),S
= IIr:1
(8)
;=1
!Ni !2::0
(9) (10)
r/
nax
,
r
mill j
are the maximum and minimum torques that the motor can generate.
For the vehicle to move forward and remain in equilibrium we set Fy and
1M I= O.
;::: 0 ,
To depict the regions of infeasibility we solve (8) subject to (7),
(3), (9) and (10) to minimize motor toques. For the simplicity of analysis we assume the front wheels are at contact angle y] and rear wheels are at contact angle r2 which are varied from 0 to 7[/2 and IjI is varied from 0 to (rl - r2) .
Fig. I Plot of min(TI + T2 +73+ 74) Vs contact angles showing regions of infeasibility
764 4
Fig 1 shows the plot of
min(IIf;I) as a function of the contact angles rl' r2 . ｩｾｬ＠
The discontinuities show the regions of infeasibility. 3. ANALYSIS OF LFA-V For controlling the contact forces we develop an actively articulated suspension system and exploit its internal degree of mobility by an actuated prismatic joint through a linear force actuator mounted on the chassis. To develop our model we use a generic platform consisting of a chassis, prismatic joints and four toroidal wheels each pinned to an outer slide link of a prismatic joint, where as the inner slide link is fixed to the chassis. The forces that act at the the vehicle are normal force N i
,
traction force ｾ＠
i'" wheel of
F;
and actuator force
A
which is always perpendicular to the chassis. Now a similar analysis is done to frame the Quasi Static analysis for LFA- V.
IN I
=
The actuator force is given by F.I A
i
A
dat(n i , Ii
A
fA.A
= R.rO L'
fAA
I'
)
I
Remark: Along the sliding direction the only force to be considered is
0 l]T.
F;
A •
The
property of a prismatic joint with linear force actuator is that only the components of ｾ＠
and Ni perpendicular to
To find the components of ｾ＠ resulted
AA
i;
when
is
F;
A
get transmitted to the chassis.
perpendicular to
F;
/2
rotated by
by Tneli
= dat(!;, R,ai )1'F1.R,ui· is
given
and
Ni
about
F; A is given
perpendicular to
A
nnel;
=
INN
nel;
.1 =
[ nnetx; nne!)';
nett
net
force
and
net
moment
4
4
F = I(pA
towards 'i;
Let
netl
the
which is
by
= lielll =[t nelx; t nety; t netz; f
Hence
R,a;
perpendicular to
Similarly component of
AT. t nel;
we find
radians
1l
[9] .Hence the component ｾ＠
A
+fleti
+ NnelJ,
M
=I
[(rlai X
PA) +
is r,ni
given
X (fneti
by
+ N neli )]
ｩｾｬ＠
Now a similar quasi static analysis is done for LFA-V. From the Fig.2 it is easy to see that all infeasibility regions are eliminated. To attain a desired posture, the vehicle parameters that need to be controlled while traversing a terrain are the
765 height h of the chassis, velocity V in Y direction, If, ro and yo .To achieve desired velocity Vd ｾ＠
Fy
for the vehicle we command a suitable value of
ｾ＠
kl'e v + ki:1' , ev
= Vd - V
. Similarly we have
-
-
F z =kpe/z +kJ:/z +Mg , Mx =kl'e lf +k)!1f ,My =kpe ro +k):ro
ro d e yo
= yOd
-
ro, e"
= hd -
h, e
- yo .Where e v ' e",e lf , ero , e yo and eyoare
IV
the
differences
between the desired values and the instantaneous values of the parameters.
k p' k p, k1" k p' k p and kv' kv' (, (, k1' are the proportional and derivative gains respectively. Control equations are developed to overcome the differences in the dynamics of thc system which arise due to the assumption of negligible wheel and link mass when compared with the chassis.
Fig. 2 Plot of min(T,+ T2 + T3+ T4 ) vs Contact angles showing regions of infeasibility are eliminated
4. RESULTS AND DISCUSSION Fig.3 and FigA shows simulations for the terrains 1,2 which are modeled such that every point on the surface has a finite and unique gradient in any three orthogonal directions. These simulations were performed using MATLAB, Simulink and MSC Visual Nastran. We obtain N;,Zi' V, If, ro and yo from MSC Visual Nastran. The controllers were applied to maintain
= rOd = yOd = 0 and hd =OA2m by assuming w=O.2m, ｾ＠ = 0.0125 m and r =O.OSm. Fig.S and
If/d
Y:l =O.Smls
M =9.31 Kg, a =O.3m, Fig.6 show the plots of
Euler angles for terrain 1, 2.It is easy to see that the deviations of these angles are well within acceptable limits.
766
Fig. 3 LFA-V negotiating terrain 1
Fig. 4 LFA-V negotiating terrain 2
lime
Fig. 5 Plot of Pitch, Roll and Yaw terrain-l
Fig. 6 Plot of Pitch, Roll and Yaw - terrain-2
S. CONCLUSIONS In our work we present LFA-V whieh negotiates uneven terrain by modifying the eontaet forees while maintaining a desired posture. Such an approach for rough terrain mobility does not seem to have reported in the literature. The paper presents a quasi static analysis for the system and also a motivation for using
767 LFA-V over and above a passive suspension system by depicting enhanced feasibility regions. The efficacy of this method is confirmed by the plots of Euler angles which are well within the acceptable limits ensuring desired posture. From the noise analysis the system is stable for reasonable values of sensor noise at the points where contact forces are measured.
REFERENCES [1] T. Estier, Y. Crausaz, B. Merminod, M. Lauria, R. Piguet, and R. Siegwart, "An innovative space rover with extended climbing abilities," in Proc. Int. Con! Robotics in Challenging Environments, Albuquerque, USA, 2000. [2] R. Volpe, J. Balaram, T. Ohm and R. Ivlev, " A next generation mars rover prototype," 1. Advanced Robotics, vol. 11, no. 4, pp. 341-358, Dec. 1997. [3] K. Iagnemma, A. Rzepniewski, S. Dubowsky and P. Schenker, "Control of robotic vehicles with actively articulated suspensions in rough terrain," Autonomous Robots, vol. 14, no. 1, pp. 5-16, 2003. [4] S. V. Sreenivasan and K. J. Waldron, "Displacement analysis of an actively articulated wheeled vehicule configuration with extensions to motion planning on uneven terrain," ASME 1. Mechanical Design, vol. 118, no. 6, pp.312-317,1996. [5] Ch. Grand, F. BenAmar, F. Plumet and Ph. Bidaud, "Stability and traction optimization of a reconfigurable wheel-legged robot," Int. 1. Robotics Research, Oct. 2004. [6] K. Iagnemma and S. Dubowsky, "Traction control of wheeled robotic vehicles in rough terrain with application to planetary rovers," Int. 1. Robotics Research, vol. 23, no. 10-11, pp. 1029-1040, Oct.-Nov. 2004. [7] John J. Craig, "Introduction to Robotics - Mechanics and Control," Third Edition, Prentice Hall. [8] Ch. Grand, F. BenAmar, F. Plumet and Ph. Bidaud, "Decoupled control of posture and trajectory of the hybrid wheel-legged robot hylos," in Proc. IEEE Int. Can! Robotics and Automation, New Orleans, LA, 2004, vol. 5, pp.5111-5116.
ADAPTIVE STAIR-CLIMBING BEHAVIOUR WITH A HYBRID LEGGED-WHEELED ROBOT MARKUS EICH, FELIX GRIMMINGER, FRANK KIRCHNER Robotics Group, German Resear'Clt Center for Artificial Intelligence, Bremen, 28359, GeT'many E-mail: [email protected] www.dfki.de/robotics Inspired by quadruped animals we developed the hybrid legged-wheeled robot ASGUARD. We showed already that this robot is able to cope with a variety of stairs, very rough terrain, and is able to move very fast on flat ground. We will describe a versatile adaptive control approach for such a system which is based only on proprioceptive data. An additional inclination and roll feedback is used to make the same controller more robust in terms of stair-climbing capabilities. At the same time, high velocities can be reached on flat ground without changing the configuration of the system. By using twenty compliant legs, which are mounted around four individually rotating hip-shafts, we abstract from the biological system. For the locomotion control we use an abstract model of bioinspired Central Pattern Generators (CPG) which can be found in biological systems frOID humans to insects. In contrast to existing work, ASGUARD uses the sensed feedback of the environment to adapt the walking pattern in real time.
Fig. 1.
The hybrid legged-wheeled robot ASGUARD
Keywords: robot locomotion; legged wheel; pattern generator; stair climbing
768
769 1. Platform Description
ASGlJARD a is a hybrid quadruped outdoor robot which is driven by four directly actuated legs with one rotational degree of freedom. 1 It has an additional rotational degree of freedom along the body axis, serving as an elastic spinal column. Related hybrid systems like the robots RHEx can be found in. 2- 5 In contrast to the mentioned RHEX system, ASGUARD uses no fixed trajectory for the locomotion, but uses direct proprioceptive data (i.e. motor torque, leg position and inertial sensors) to adapt the motion pattern on-line, without any predefined pattern. WHEGS is another state-of-theart system regarding hybrid legged-wheeled locomotion and is described in 6 and. 7 WHEGS uses, in contrast to ASGUARD, a mechanical design to adapt the locomotion pattern. Similar to ASGUARD, no predefined pattern is needed but is defined implicitly by the force-feedback of the ground. While the intelligence of WHEGS concerning the adaptation is hidden in the mechanical design, ASGUARD uses an FPGA-based controller, making on-the-fly changes in the control concept possible without changing the hardware configuration. The locomotion of ASGlJARD is performed by an abstract model of Central Pattern Generators (CPG). The compliant legs of the robot are arranged around four hip shafts, with an angular distance of . Because of the symmetry of the legs, we have only to consider the phase in between [-in, in] (cf. Figure 2).
Fig. 2.
The
ASGUARD
wheel. Five compliant legs are mounted around each hip shaft.
2. Adaptive Control for Hybrid Legged-Wheeled Robots The concept of using CPGs (Central Pattern Generators) is well-known and utilized in the area of ambulating robots. CPGs are the major mechanisms a Advanced
Security Guard
770
in animals to control and to produce rhythmic motion. CPGs are characterized by the ability to produce rhythmic motion patterns via oscillation of neuronal activity without the need of sensory feedback. 8 From the CPG control methods, used in a variety of walking machines,9-14 we developed an efficient approach to control such systems by using trajectories in the time-angle space. For the directional and speed control of ASGUARD a high level controller, which receives its input via a joystick, sends the parameters for phase, frequency, and direction to the CPGs (d. Figure 3). From our high-level controller we can modify the pattern parameters by changing the pattern frequency, the direction of the generated pattern as well as a phase offset. By this phase offset we can change the synchronization of each of the legs. In our case, a simple sawtooth pattern is used. During the run, the environment generates a direct feedback on the system,
Fig. 3. Our architecture of the behaviour-based control using proprioceptive sensor feedback
which is in our case the feedback of inclination of the robot as well as the torque feedback of the actuators. An advantage of our design is that we can modify the amplification factor of the proportional control term of the inner control loop at runtime. By changing those parameters on-line, we allow a larger error between the current leg position and the target trajectory. This is an important factor because we are using the proprioceptive information of the applied force to change those parameters.
771
Changing the proportional part of the position control parameters on-line has an effect like a mechanical spring. The stiffness at each leg is directly adjustable by the proportional part of the position controller. We release the spring for each actuator if the measured torque is higher with respect to the average torque. This does not directly effect the generation of the CPGs, only the actual trajectory of each leg. This is comparable to an injury, like a sprained ankle, where humans get a direct negative feedback on the nerves which will result in a more "'elastic'" way of walking without changing the motion pattern directly. The error of the actual trajectory and the target trajectory is then fed back to the motion pattern generator. If the error gets larger than ｾ＠ of a step phase (410 7f), the pattern of the specific actuator is synchronized with the actual position of the leg. By this we close the loop between the CPG and the position controller, which, to our best knowledge, has not been done with other hybrid legged-wheeled robots. This is an important feature because we do not have to synchronize the left and right side of the legs while climbing a stair manually or by a fixed stairs motion pattern. This task is performed directly by the adaptive controller. When As GUARD climbs a stair, the first front leg which has contact to the first step will have a higher torque on the actuator. The controller will release the leg and therefore allow a larger position error. When the second leg makes contact with the first step, the torques are distributed between the front legs. The position controller reacts with a higher stiffness, making the first two legs able to lift the robot up the stairs. In order to distribute the torque of each actuator, which is directly related to the measured motor current, we use an approach to modify the proportional part of the position controller, which is responsible for following the trajectory of the generated pattern (d. Equation 1).
2:= cur (1) (curi - - - ) * ii) n Pi refers to the proportional part of the position controller in leg i and curi to the measured motor current in A for each motor. The constants fi:i and Li are used to map the proportional factor of the controller within the bounds of the controller. In l we showed already increases the climbing behaviour on a flight of stairs. The reader should note that this approach does not directly change the motion pattern in its phase, frequency or direction, but changes the way the internal position controller changes the actual trajectory by allowing a larger error between the target and the actual trajectory. The difference Pi
= (fi:i -
772
between the torque of a specific motor (which is proportional to the motor current) with respect to the average torque gives a negative feedback on the controller. This results in a higher elasticity, similar to an adaptive spring. In our approach, a high positive discrepancy in the motor torque results in a higher elasticity in the leg. In contrast to a stair climbing behaviour, the robot has to bring as much power as possible to the ground, especially while accelerating. This is best achieved by a strict position controller within our control architecture (cf. Figure 3). For this we use a simple proportional controller (cf. Equation 2) with maximal error amplification. Due to the inertia of the system, no differential term within the position control loop is needed. Oi = Pmax
* (errori)
(2)
To make our first control approach (cf. Equation 1) versatile for flat as well as for steep terrain and stairs, we take into account the inclination of the robot which is measured by an inertial-based tilt sensor. We assume that the robot should adapt its control parameters if the inclination is positive, i.e. pitch 0,0011 0.(1)1))
(>,0007
O,0005 {,-(lJ03
i),Clon1 0
wu
Figure 9. Diagram of different regimes of hopping of robot body dependence on frequency and relation m, rim.
899 ＲＵｾＭＮＬ＠
hmax,MM
ＲＰｴＭＫｾ＠ ＱＵＫＭｾｌＴ＠ ＱＰＫＭｫｌｾ
_ _ 1,8
--2
o ｾｃｩＧｬＺＮＭＫ｟＾ｉｊＬ＠
l/c
100
-150
200
250
300
Figure 10. Diagram of maximal height of hopping robot dependence on frequency and parameter
mlr/m.
Diagram on fig.lO shows dependence on maximal height of hopping robot dependence on frequency. We defined that height of hopping robot increase when frequency goes up. Maximum average velocity of robot body in horizontal direction corresponds to frequency equal about 260 lis (see fig. 11). Vx, we
4.6 4,4 4.2
3.8 3.6
3A 3,2 ＳＫＭｾﾷ ｾ＠
m
_____ ｾＭＫＮ＠
ｾ＠
ｾ＠
____ _ ｾＭＫｌ＠
_
m ___-+-+ _ __-L _ __m W, pa,q/c
Figure II. Diagram of average velocity of robot body dependence on frequency for (mi T)/m=2,34*l0-4.
4.
Robot prototype modeling
Fig. 12 shows the scheme of prototype of hopping robot with rotating masses. The robot consists of five main part two DC -motors maintained in a frame, IRsensors for remote control, vibration absorber and internal control system. The weight of prototype is equal to 120 grams.
900
7 Figure 12. The scheme of prototype of hopping robot with rotating masses. I -robot body, 2 -frame of electrical DC-motor, 3 --rotating internal mass, 4 -·IR - sensor, 5 viscous - elasticity elements, 6 - video camera, 7-USB - connector.
It is very important to reduce the vibration level transmitted to the equipment of the robot from the exciters it is reasonable to placc this equipment on a special platform, isolated from the vibration-driven robot body by a springdashpot vibration absorber. Besides equipment maintained on hopping robot has additional acceleration during landing of robot body on supporting surface. For reducing of dynamical forces acting on the equipment we have considered viscous elasticity supporting elements. That provides soft landing of robot. Parameters of these elements are calculated by assistance of finite element method in SolidWorks software.
901
5.
Conclusions
In this paper original scheme of mobile robot for hoping motion with rotating internal mass were developed. Mathematical models of 2-D vibration-driven robot for flight and landing processes were presented. By numerical method average velocity of robot body in different motion regimes were calculated. Numerical algorithm for calculation of dynamical parameters of motion allowed investigating periodical regimes of motion. Acknowledgment
This investigation is supported by RBRF project 08-08-00438. References 1. Aoshima, S.; Tsujimura, T.; Yabuta, T.: A miniature mobile robot using piezo vibration for mobility in a thin tube. Journal of Dynamic Systems, Measurement, and Control Vol. 115 (1993), pp.270-278 2. Yeh, R.; Hollar, S.; Pister, K.S.J.: Design oflow-power silicon articulated microrobots. Journal of Micromechatronics, Vol. I, Num. 3,2001, pp. 191-203 3. Gradetsky, V.G.; Knyazkov, M.M.; Kravchuk, L.N.; Solovtsov, V.N.: Microsensor control of motion of compact robots inside tubes (in Russian), Mikrosisternnaya Tekhnika [Micro system Engineering], No.8, 2002, pp. 11-19 4. Ma, J.; Lo, M.; Bao, Z.; Wang, A.: Micro peristaltic robot simulating earthwonn and its control system. Journal of Shanghai Jiaotong University Vol. 33 No.7, 1999 5. Bolotnik N.N., Chernousko F.L., Kostin G.V., and Pfeiffer F. Regular motion of a tubecrawling robot in a curved tube II Mechanics of Structures and Machines. 2002. Vol. 30. No.4. P. 431-462. 6. Panagiotis Vartholomeos, Evagelos Papadopolos. Dynamics, Design and Simulation of Novel Microrobotic Platfonn Employing Vibration Microactuators. Journal of Dynamics System, Measurement and Control. Vol. 128, March 2006 pp.122-133. 7. Bolotnik N.N., Zeidis LM., Zimmennann K. and Yatsun S.F. Dynamics of Controlled Motion of Vibration-Driven systems. Journal of Computer and Systems Science International,2006 Vol45 ｎｾ＠ 5pp.834-840. 8. Bolotnik N.N., Zeidis l.M., Zimmennann K. and Yatsun S.F Mobile vibrating robots. Proceedings of the 9 th International Conference on climbing and walking robots(CLA WAR2006). Brussels, Belgium. 2006. p.558-563. 9. Bolotnik N.N., Yatsun S.F., Cherepanov A.A. Automatically controlled vibration-driven. Proc. Intern conf. on mechatronics ICM2006. Budapest, 2006. p.438-441 10. Samuel Kesner, Jean - Sebastien Plante, Steven Dubovsky A Hopping mobility concept for a rough terrain search and rescue robot. Advances in climbing and walking robots.
COMPUTATIONAL COST OF TWO .FORWARD KINEMATIC MODELS FOR A S-G BASED CLIMBING ROBOT MIGUEL ALMONACID*, ROQUE SALTAREN**, RAFAEL ARACIL**
*Polytechnic University of Cartagena, **Polytechnic University of Madrid *Muralla del Mar sIn Cartagena, - **C/ Jose Gutierrez Abascal,2, Madrid. CARLOS PEREZ, NICOLAS GARCIA, JOSE M. AZORIN, JOSE M. SABATER
Virtual Reality and Robotics Lab., Miguel Hernandez University A vda. Universidad sin, Elche, 03202, Spain An interesting novel application of the Gough-Stewart platform as a climbing robot and its kinematics control has been proposed to climb autonomously through long structures describing unknown spatial trajectories, such as palm trunks, tubes, etc. For planning the motion of the parallel robot, inverse and direct kinematics problems have to be solved continuously in the path planning algorithm in a minimum time. Computation efficiency of the model is very important. This paper presents a comparison between two models of the 6-UPS parallel mechanism. Inverse and direct kinematic problems have been numerically solved with classic methods and compare for four different configurations for the two models. The analysis and simulation of the kinematics problems show the computational efficiency of the proposed model for the path planning of the climbing parallel robot.
1. Motivation of work
The 6-DOF parallel mechanism [1] configuration in this paper consists of regular hexagonal lower and upper platforms and six identical linear actuators. The legs are six identical UPS kinematics chains (where U is a universal joint connecting the base to the linear actuators, P is a prismatic actuated joint and S is the spherical joint connecting the linear actuator to the end-effector). These parallel mechanisms are well known as Gough-Stewart platforms since Gough [2] first introduced it in 1962 as a tire testing machine and Stewart [3] popularized it in an aircraft simulator three years later. The main features of parallel mechanism [4] are accuracy, high velocity and high load capacity. One of the drawbacks of this mechanism is its limited workspace. Our research group is improving that limitation using parallel mechanisms as mobile robots [5]. For example Aracil and Saltaren [6] presented several morphologies of the parallel mechanism for different climbing applications.
902
903
Computational tools using multibody dynamics have been developed in [7 J to solve the kinelbatic andaynatruc equations of 6-UPS parallel mechanism. The contributions to solve the complex direct kinematic problem in parallel mechanisms ate,'vel'Y w,ide [1,4, 8-10]. The large systems of equations using multibody dynamics in 6-DOF parallel mechanisms [11] are balanced with the new processors capacity. In this paper the forward kinematic problem has been accomplished using mechanical constraint formulation and numerical methods. The problem of the multiple solutions of this parallel mechanism is solved. This method provides a unique solution for the direct kinematic problem. The climbing procedure consists on fixing one of the rings to the structure using a clamping system, as long as the other ring can be move and orientate to the planned point. Planning the motion for climbing non-straight long structures requires computing inverse and forward kinematics several times (depending on the curvature) during a complete advancement and retrieval of the linear actuators. A computationally efficient model of the 6-UPS parallel mechanism is presented using universal-translational composite joint. The contribution presented in this paper has been implemented in a real parallel robot used as a climbing robot for palm tree maintenance Fig. lea).
End-
Uy
(a)
(b)
Fig. 1. General architecture of a 6-UeS parallel climbing robot.
904
2. Structure of a parallel climbing robot The spatial model of the 6-UJ:S parallel mechanism is shown in Fig. l(b), where X-Y-Z is the reference frame fixed at the base, and x'ry'rZ'J is the coordinate frame fixed at the centre of gravity of the end-effector with the Z'raxis and Zaxis pointing vertically upward. The position vectors SoAn (n=1 ... 6) locate the universal joints at point An with respect to the base reference, and the position vectors S' /n (n=l ... 6) locate the spherical joints at point Bn with respect to the mobile coordinate system. rABn denotes the vector connecting An to Bn.
2.1. Model-l The 6-UPS parallel mechanism is modelled using 13 bodies. One of them corresponds to the end-ejector, and the other ones twelve correspond to the couple of links that form each one of the six lineal actuators. The base is considered as the Ground. The model is defined in Table 1. If Euler parameters are used for orientation, each body has seven generalized coordinates. There are, therefore, 91 number of coordinates (nc) for the 13 bodies. The spherical, prismatic and universal joints have tree, five and four constraint equations, respectively, yielding 12 constraint equations for each out of six kinematic chains. In addition, the 13 Euler parameter normalization constraints yield a total for 85 kinematic constraint equations (nh). The difference provides the degrees of freedom of the system. Table 1 Model-l of the 6-UPS parallel mechanism Bodies 13 Constraints Spherical (end-effector-piston) Prismatic (piston-cylinder) Universal (cylinder-base) Euler parameter normalization constraint
nc= 13x7=91
6x3=18 6x5=30 6x4=24 13xl=13 nh = 85
DOF=91-85=6
Taking the base as a reference, the composite set of generalized coordinates for the entire system is (1)
905 where ql denotes the generalized coordinates system for the end-effector and q2, q3, ... ,qI3 the generalized coordinates for the couple of links of the six linear actuators. Kinematic modelling of a 6-UfS parallel mechanism involves the selection of the bodies that make up the mechanism, kinematic constraints that act between pair of bodies and time-dependent kinematic drivers. So, next the constraint equations vector of the parallel robot ( q,t) is derived as
(2)
where K(q)nXI is a vector of 72 holonomic kinematic constraints imposed by joints of the mechanism. D(q,t)6xl is a vector of 6 constraints imposed by the time-dependent linear actuators. And P(q)13xl is a vector of 13 constraints for the normalization of Euler parameters.
2.2. Model-2 A computationally efficient model-2 of the parallel robot can be obtained using composite constraints between a pair of bodies [12]. In this case, the composite joint is formed by the universal joint and the translational joint of the kinematic chain of the robot. This means to consider the cylinder of each leg as an intermediate body that serves only to define the kinematics constraints that are connected, without introducing it as a separate body, with its associated cartesian generalized coordinates. The new model of the 6-UfS parallel robot is defined in Table 2. The robot is modelled using 7 bodies. Therefore, the number of generalized coordinates (nc) becomes 49. The spherical and universal-prismatic composite joints have tree plus tree constraint equations, respectively, yielding 6 constraint equations for each out of six kinematic chains. In addition, the 7 Euler parameter normalization constraints yield a total for 43 holonomic constraint equations (nh) Table 2 Model-2 of the 6-UPS parallel mechanism
Bodies 7 Constraints Spherical (end-effector-piston) Universal-prismatic composite (piston-base) Euler parameter normalization constraint DOF = 49-43 = 6
nc= 7x7=49 6x3=18 6x3=18 7xl=7 nh =43
906
3. Kinematic analysis of the parallel climbing robot Geometric parameters of Table 3 have been used for the numeric solution of the kinematic problem of the 6- UfS parallel mechanism. Table 3 Geometric parameters of the 6-UPS climbing parallel robot
Parameter Description Ie Cylinder length (j) c"
0
ID
08
'"'0
075
C
ｓｾｴ＠
."ｾＰＶＵ＠ "
"
()
Co
o
o
CCOL--:--c--3o----C,--o-c-Oc6 -c,--o----;----c'.
"OL--:----'c-3O----C,--O-S-Oc6 ｾＬ｣ＭＡＢ＠
Penetration Velocity (m/s)
Penetration Velocity (m/s)
Figure 3. Results of Hunt-Crossley method
Figure 4. Results of Orin-Marhefka method
Equations of Motion were generated via the symbolic manipulator, Autolev [6], and numerical integration was performed with a fourth-order, variable step, Kutta-Merson integrator with a time step of 0.1 ms and a relative error tolerance of 1.0xlO-7 • .; -
I
c
ｾＰＶＵ＠
.""
-
o
Croquet Data Set ｌ･｡ｳｴｓｾｵｲｆ＠
SmulaloOnRBsults
"
o
()
Do 0.1
CｃｏＡＭＺｾＳ｣ＬＵ＠
ｾＶｯＭＢＬ［｣ＧＮ＠
0,
10
2'0
30 ＧｾＴＰ＠
50·' ＧｾＢＶＰ＠
Penetration Velocity (m/s)
Penetration Velocity (m/s)
Figure 5. Results of Proposed Method
Figure 6. Results of High Velocity Test
In many cases, alternative measures of the coefficient of restitution are used such as the Stronge's or Poisson's formulations to better illustrate the relationship between energy and velocity. For the simple case of a ball bouncing in the normal direction with no frictional energy losses, Stronge's coefficient of restitution, which is the square root of the ratio of the restitutive work to the compressive work, produces the same result as Newton's coefficient of restitution.
935 Table I. Simulation Results of the Coefficient of Restitution and the compressive and restitutive work done for several collision velocities.
Newton CoR: Proposed Method Compressive Work (J) Restitutive Work (J)
Impact Velocity (m/s) 3.670 4.240
2.190
2.780
0.798
0.792
0.782
0.240
0.387
-0.153
-0.243
4.990
5.500
0.775
0.766
0.760
0.674
0.900
1.246
1.513
-0.412
-0.541
-0.731
-0.874
4. Discussion Given a real set of coefficient of restitution data, the proposed method is capable of more realistic simulations than previous methods. The results show that the new method is able to match the given data set well over the entire range. As shown in Fig. 6, the method is also capable of stably simulating very high velocities with small errors between the desired curve and the simulation results. However, it is clear that at such high velocities a method that deals explicitly with the other collision effects such as plastic deformation is necessary. Furthermore, the proposed technique decouples the selection of the stiffness and damping and allows for selective tuning of parameters to match data. The contact time and the resultant of the ground reaction force are easily measured experimentally and are largely a factor of the stiffness of the colliding materials. To tune the model, the stiffness, K , can be adjusted first to match the model with these two measurements. For determining the damping properties, the stiffness is not used in determining alpha and is only found in Eq. (7). As a result, changing the stiffness of the material does not require resolving Eq. (8) for new coefficients. Thus, the stiffness and damping can be adjusted independently to reproduce the given collision behavior of the experimental system. Future work on this project will be to experimentally measure the performance of KOLT' s feet and verify the contact model's ability to predict the collision performance of the leg design. We will also focus on expanding the model into a contact framework for rigid bodies to allow for oblique collisions and collisions with angular velocity to be simulated realistically.
936
References 1. J.G. Nichol, S.P.N. Singh, K. Waldron, L. Palmer, D. Orin., System Design
2. 3. 4.
5. 6.
of a Quadrupedal Galloping Machine. Int. J. Robotics Research 23, 1013 (2004) K. Hunt, F. Crossley, Coefficient of Restitution Interpreted as Damping in Vibroimpact. J App. Mech. 440 (1975) W. Goldsmith, Impact: The Theory and Behavior of Contacting Solids. London, UK: Edward Arnold (1960) D. Marhefka, D.Orin, A Compliant Contact Model with Nonlinear Damping for Simulation of Robotic Systems. IEEE Trans Systems, Man and Cybernetics 29, 566 (1999) D. Gugan, Inelastic Collision and the Hertz Theory of Impact. Am. J. Physics 68, 920 (2000) AutoLev, Online Dynamics (http://www.autolev.com!)
USING NONLINEAR OSCILLATORS TO CREATE A PATTERN GENERA TOR OF BIPEDAL LOCOMOTION ARMANDO CARLOS DE PINA FILHO Graphical Engineering Department, Polytechnic School, Federal University of Rio de Janeiro, Technology Center, 21949-900, Ilha do Fundiio, Rio de Janeiro - RJ, Brazil, e-mail: [email protected] MAX SUELL DUTRA Mechanical Engineering Program, COPPE, UFRJ, Federal University ofRio de Janeiro, Technology Center, 21945-970, Ilha do Fundiio, Rio de Janeiro - RJ, Brazil, e-mail: [email protected]·ufrj.br
The bipedal locomotion is perfonned by means of rhythmic and synchronized motions. These movements have a pattern produced by nervous networks in the spinal marrow. These specialized systems are known as central pattern generators (CPGs). Oscillators can be used to create similar systems to the human CPG, providing approximate patterns of locomotion. The objective of this work is to present a nonlinear oscillators system used to generate patterns of bipedal locomotion, taking into consideration a 2D model, with three detenninants of human gait, which perfonns parallel movements to the sagittal plane. Using this system, the behavior of the hip and knee angles was detennined. Modification of the step length and gait frequency can be obtained from change of few parameters in the oscillators. An analysis of the system provides excellent results when compared to experimental analyses. Based on these results, we conclude that the use of nonlinear oscillators can represent an excellent method for creation of pattern generators, allowing their application to simulate a CPG of bipedal locomotion.
1. Introduction
A large number of degrees of freedom is involved in bipedal locomotion and well-coordinated movements of these degrees of freedom is essential. The main part of this coordination occurs in the central nervous system, which generates signals according to the desired gait. According to Mackay-Lyons [1], the nervous networks in the spinal marrow are capable to produce rhythmic movements when isolated from the brain and sensorial inputs. These specialized nervous systems are known as nervous oscillators or central pattern generators (CPOs). We have some interesting works about the locomotion of vertebrates controlled by central pattern generators in [2-4].
937
938
The human locomotion is controlled, in a way, by a CPG. Nonlinear oscillators can be used to create similar systems to the human CPG, providing approximate patterns of locomotion. In these systems, each oscillator generates angular signals of reference for the movement of the legs, and they have amplitude, frequency, parameters, and coupling terms. We have some works about oscillators systems applied to the locomotion in [5-7]. This work shows a CPG simulated by Rayleigh oscillators, which was little explored in previous research, mainly the application in locomotion. Besides, a mutually coupled oscillators system was proposed, which has an advantage in relation to other systems, since the mutual influence between the elements provides a mutual reaction, when one of them is submitted to a disturbance. Then, the objective of this work is to present a nonlinear oscillators system used to generate patterns of bipedal locomotion, taking into consideration a 2D model, with three determinants of human gait, which performs parallel movements to the sagittal plane. Using oscillators with integer relation of frequency, the behavior of the hip and knee angles were determined. Modification of the step length and gait frequency can be obtained from change of few parameters in the oscillators. The study of this oscillators system has great application in the project of autonomous robots and rehabilitation technology.
2. Bipedal Locomotion Patterns The choice of an locomotion pattern depends on the combination of a central programming and sensorial data, as well as the instruction for a specific motor condition. This information determines the organization of muscular synergy, which is planned for adequate multiple conditions of posture and gait [8]. SCltsorlnl infonnntion
Central nervous S)'£Ielll
Figure 1. Control system of human locomotion.
939
Figure 1 presents the control system of human locomotion, controlled by the central nervous system, which the CPG supplies a series of curves for each part of the locomotor. This information is transmitted to the muscles by a network of motoneurons, and the conjoined muscular activity perfOlms the locomotion. Sensorial information about the conditions of the environment or some disturbance are supplied as feedback of the system, providing a fast action proceeding from the CPG, which adapts the gait to the new situation. Despite of the people not walk in completely identical way, some characteristics in the gait can be considered universal, and these similar points serve as base for description of patterns of the kinematics, dynamics and muscular activity in the locomotion. In the study presented here, the greater interest is related to the patterns of the kinematics, in particular, of the hip and knee angles. Taking into consideration the movements in the sagittal plane, which divides the body in left and right side, the hip and knees perform movements of flexion and extension. From the use of an optic-electronic system of three-dimensional kinematical analysis, Raptopoulos [9] define the angular behavior of the hip and knee in the course of the locomotion cycle. Twenty-four healthy young male volunteers participated in this work. They had no previous history of surgery or musculoskeletal problems that could affect their walking pattern. They were asked to walk at their normal cadence. Figure 2 presents the graphs of angular displacement and phase space of the hip related to the movements of flexion and extension. Figure 3 presents the graphs of angular displacement and phase space of the knee. The results of the work performed by Raptopoulos [13] are used in this research only as reference results for comparison with the results from nonlinear oscillators system. 290 ｲＭｾＢＧ＠ ISO
-189
-20
Figure 2. Angular displacement and phase space of the hip (mean ± deviation)[9l
940
ｾ＠ ｾｩ＠
20
"L
40
ｾ＠
c
60
i ｾｏ＠
100
I
60
lyd,·!",,!
Figure
Ｓｾ＠
Angular displacement and phase space of the knee (mean ± deviation)[9].
3. Nonlinear Oscillators System The nonlinear oscillators system in this work uses Rayleigh oscillators, similar to the system proposed by Pina Filho and Dutra [10]. Thus, we have:
0-
5(1- qil )0 + n2(0 - 0
0
)-
coupling term =
0
(1)
where: 0 represents the angles in study; 5, q and n are the parameters of the oscillator. The coupling term represents the way each oscillator interacts with the others. In this work, the term
C
k,;[ 0i (Oi - 0iJJ is responsible for the
coupling between oscillators with different frequencies, while the term Ci,j
(Oi - OJ)
makes the coupling between oscillators with the same frequencies.
Applying the Eq. (1) to the proposed problem, we have:
Olk - 5lk
(1- ｱｬｫｏｉｾ＠
)Olk + nTk (Olk - OlkJ
(2)
-Clk,h[Oh(Oh -OhJJ-Clk,rk(Olk -Ork)=O Oh - 5h(1- qh O; )Oh + ｮｾＨｏｨ＠
- OhJ
- ch,lk [Olk (Olk - OlkJJ - ch,rdOrk(Ork - OrkJJ = 0
(3)
(4) where: h is related to hip, lk is the left knee and rk is the right knee. The synchronized harmonic functions corresponding to the desired movements are:
941
(5) (6) (7) where: (]) is the gait frequency, and a is the phase value. Considering ark = ah = ark = 0 and deriving the Eqs. (5-7), inserting the solution into the differential Eqs. (2-4), the necessary parameters of the oscillators (qi and Qi, i EO {lk, h, rk}) can be determined. Then: (8)
(9)
(10)
(11) (12)
(13) 00 61l
,
eve 1
,',
Ell
.11
fO
'"
il, ["I.l
0
.,., .fO ｾ＠
ｾＭＧｏｉＤ＠
2!\
limer-l
./Ij,j
aJ ｾ＠
,.,
.,., 6,(']
Figure 4. Behavior of the hip and knees as function of the time and stable limit cycles.
fO
60
00
942
From Eqs. (2-4) and (8-l3), and using the MATLAB software, the graphs shown in Fig. 4 were generated, and represent, respectively, the behavior of the angles as function of the time and the stable limit cycles of the oscillators. These results were obtained by using the parameters showed in Table 1, as well as the initial values provided by Table 2. All values for the model were experimentally identified through tests performing in the MA TLAB. Table I. Parameters of Rayleigh oscillators.
Table 2. Experimental initial values.
Comparing the graphs of Fig. 4 with the graphs presented in Figs. 2 and 3, despite of the different amplitudes demonstrated by the experimental tests and by nonlinear oscillators system, taking into consideration only the cycle of locomotion, where 100% of cycle in the Figs. 2 and 3 is equivalent to 6.28 s in Fig. 4, the graph generated by oscillators system represent the pattern of behavior of the elements (hip and knees), being able to be used to generate the approximate patterns for locomotion. This confirms the possibility of the use of nonlinear oscillators in the modeling of a pattern generator. 4. Application of the System In this work, we consider a 2D model of bipedal locomotor with movements in the sagittal plane, capable to represent three of the six most important determinants of human gait: the compass gait, the knee flexion of the stance leg, and the plantar flexion of the stance ankle. The model has three pairs of elements: femur, tibia, and foot, and these elements have identical lengths. The locomotion cycle can be divided in two intervals: single support phase (one foot in the ground) and double support phase (two feet in the ground). Applying the nonlinear oscillators system to generate the degrees of freedom of the locomotor, a simulation was performed taking into consideration the Eqs. (2-4) and (8-13), with the values showing in Tables 1 and 2. The simulations presented here were performed only through a kinematical analysis. Figure 5 presents the bipedal gait for angular amplitude of the hip equal to 50°.
943 0.8
0.6
y[mj 04
0.2
/ [)
x leu]
Figure 5. Gait for A" = 50°.
In this case, the adopted value for amplitude of the hip angle provide a step length equal to 0.63 m. A alteration of this amplitude value cause a change of step length, consequently modifying the gait. Thus, the pattern generator system makes possible the change of step length by means of amplitude alteration. In the case of alteration of the hip angle, with initial amplitude 500 changing to 30°, the step length will be equal to 0.38 m. The gait simulation with these conditions is presented in Fig. 6. ＰＮＸｾＭ＠
0.6
y Em] 0.4
0.2
o a
0.5
1
1.5
2
X [Ill]
Figure 6. Gait for Ah = 30°.
Besides the change of step length, the system makes possible too the change in the gait frequency, which can be modify by means of the Eqs. (5-7), choosing a new value for OJ. Then, modifying the value OJ from 1 to 2, we have logically a duplicated gait frequency (Fig. 7).
944 ＶＱＩｲＭｾ｟］Ｌ＠
---- 0:·=2 -@:::ol
-40
Locomotion cycle
Locomotion cycle
Figure 7. Behavior of the hip and knee with modification of the gait frequency.
5. Conclusion From presented results and their analysis and discussion, we come to the following conclusions: the use of nonlinear oscillators can represent an excellent method for creation of pattern generators, allowing their application to simulate a CPG of bipedal locomotion; the adopted model is capable to characterize three of the six most important determinants of human gait; and by change of some parameters in the oscillators, modification of the step length and the gait frequency can be obtained. The model presented in this work is based only on kinematics, and dynamics will be studied in future research. Acknowledgments The author Armando Carlos de Pina Filho would like to express your gratitude to FAPERJ (Fundayao Carlos Chagas Filho de Amparo aPesquisa do Estado do Rio de Janeiro), for the financial support for purchase of the equipment used in the course of this present research. References
1. 2. 3. 4. 5. 6. 7. 8. 9. 10.
M. Mackay-Lyons, Physical Therapy 82-1(2002). S. Grillner, Science 228, 143-149 (1985). K. G. Pearson, Annu. Rev. Neurosci. 16,265-297 (1993). J.1. Collins and S. A. Richmond, Bio!. Cybernetics 71, 375-385 (1994). 1. S. Bay and H. Hemami, IEEE Trans. Biomed. Eng. 34,297-306 (1987). M. S. Dutra, Shaker Verlag, Germany (1995). A. C. de Pina Filho, D.Sc. Thesis, COPPE/UFRJ, Brazil (2005). F. B. Horak and L. M. Nashner, Journal Neurophy. 55, 1369-1381 (1986). L. S. C. Raptopoulos, D.Sc. Thesis, COPPE/UFRJ, Brazil (2003). A. C. de Pina Filho and M. S. Dutra, CONEM, Recife, Brazil (2006).
INTERNET 3.0 FOR THE SIMULATION OF NETWORKED CLAWAR SYSTEMS FABIO P. BONSIGNORIO Heron Robots s.r.l., V.R. Ceccardi, 1118 Genova, /-16121, Italy The 3.0 or 3D internet, exemplified by environments such as LindenLab's SecondLife, supplies a massively multi agent simulation environment, whose capability of simulating, for example the multi body dynamics, is so far quite limited. Nethertheless it is possible to do some preliminary simulations of the collective behaviours of multirobot systems, in a mixed environment where some of the agents are directly controlled by a human being. In this paper we discuss the limits and future opportunities of Massive MUltiplayer Game technologies applied to Internet 3.0 for the simulation of massively multi robot systems. In particular in the cooperative behaviour analysis of bioinspired locomotion systems. Some reasoned guesses on the time frame of this developments are shared and a few examples are given.
1.
Introduction
Massively multiplayer online role-playing games (MMORPG),[1,2], are a kind of computer role-playing games in which a large number of players interact in a virtual 'world', they usually, like Everquest, World of Warcraft, Ultima online, Star Trek online, are based on traditional fantasy on science fiction themes. Here we will not discuss the growing cultural and economical impact of these systems on our societies, but we will discuss their features as possible enviroment for the simulation of multi clawar and mobile robot systems. As such we will see them as massive multiagent object oriented simulation sytsems. The virtual 'world' is usually a 3D rendering of a phisical environment where objects behave according to, simplified, newtonian physical laws. Actually the game's persistent 'world', usually hosted on a dedicated server farm by the game's publisher, is modified concurrently by the action of all of the user, and as consequence, continue to 'exist' and evolve also whm a specific player is not using it.
945
946
Most MMORPGs are deployed using a client-server system architecture. The system that generates and persists the 'world' runs continuously on a server, and players connect to it via client software. A 'metaverse' is an envinroment technologically similar to those of a MMORPG, but with no scores, competition rules etc, and oriented toward social networking, community building and user' s creative expression, the word 'metaverse' comes from a 1981 science fiction novel, 'Snow Crash' by Neal Sthepenson, which first described such a system although assuming video realistic rendering of the virtual world. Since then several examples of such systems have been deployed evolving with information technologies technical progress: a text-based, low-bandwidth system called The Metaverse as part of their BBS, Illuminati Online, a similar one from SenseMedia called SnowMOO, also based on Snow Crash, Active Worlds and others. The most popular is so far Second Life, launched in 2003 by LindenLabs, a private subscritpion based initiative with free access. There are a couple of open source free project, Solipsis and The Croquet Project, which aims to 'creating and deploying deeply collaborative multi-user online applications on multiple operating systems and devices',[solipsis,croquet] 'more extensible than the proprietary technologies behind collaborative worlds such as Second Life. MMORPGs and metaverses are a growing and remarkable economic fact as their 'Gross National Product' can be compared to that of a small country with a per capita income comparable to that of countries like China, see [3].
2.
LindenLabs' SecondLife
Second Life might be considered as an example of a massive multi agent object oriented real time simulation environment providing a technical support for a sophisticated social network environment. Other examples are IMVU, There, Active Worlds. By March 2008, about 13 million accounts were registered, and more notably, about 38,000 residents are concurrently operating within the environment, this allows to consider SecondLife as a quite successful example of multiuser simulation system. The user is represented within the system by an 'avatar' a cartoon like puppet that moves into the simulated 'world', the 3D rendering of the simulation scenario is comparable to a 3-4 years ago pc or gaming station video game, see Figure 1.
947
The programming environment allows with a simple interface to create objects,statical or dynamical, associating to them methods written in a pascal like language and also to modify with simple commands the terrain where the objects interact, see Figure 2.
Figure 1. A Second Life simulated landscape
It is possible to upload graphical objects or animations created outside. One of the reason of the market success of the envinroment is the Intelectual Property Rights policy which leaves the copyright of objects to the creator (or owner). The script language is an object oriented pascal like language called Linden Scripting Language, or LSL [6]. LSL is used to add autonomous behavior to many of the objects in Second Life, such as doors that open when approached, or vehicles driven or autonomous. LSL has been used to create relatively advanced systems, such as a comparatively simple artificial life experiments on the virtual 'island' , a specific region of the simulated environment shaped as an island, of Svarga, where a simplified ecology run autonomously (with clouds, storms, plants, etc.) . An example of script is given in Figure 3.
948
Figure 2. Example of objeet parameters setting screen
3.
Physics modeling and simulation
The 'world' is built over a tlat (as gravity vectors are costant everywhere and orthogonal to same plane) euclidean earthlike surface. This is seen as a collection of 'regions', of 256x256 m, each usually running on a specific server, typically a dual core processor. Each server instance simulates the collisions and interactions of all objects, including the users' avatars, in that region. Each object is made by a limited (255) number of 'prim', primitive objects connected together. There is also also a, growing limit, to the number of prims which can be created in each region and consequentlly in all the system. This limit allows to statistically guarantee the response times, another limit is that when a region is 'full', i.e. The limitis of the server is reached, is not possible for an avatar or other object to access it. Of course care is taken to limit the frequency of this kind of situations. The 3D rendering and animation is performed by a thick client software that the user has to install on his/her computer. From the beginning of April the Havok 4 physics engine is used to simulate the syntethic world dynamics. This allow to simulate thousands of physical objects at once. The limit to the number of interactions, 'collissions', between different objects is of about 400-500, at present.
949
Figure 3. An example ofLSL script to control a kart
Each atomic entity, called an asset, primitives, textures, LSL scripts etc.,is referenced with a universally unique identifier or UUID. These entities are stored on a dedicated relational database farm based on MySQL with an estimated total storage of about 100 Terabytes (end of 2007 data). The dynamical simulation is based on a 'physics engine' , Havok 4 [7] from April 2008, Havok 1 from the beginning in 2003. A 'physics engine' is a middle ware which maps the object dynamics to a parallelized set of operations to be performed on a processor usually a graphic card, and sometimes a dedicated one [8].The usage of phisics engines is widespread in game industry as it hides the details of low level distribution of computations to the developers. In the current, from April 2008, system the Havok middleware acts on a virtual computing architecture which allows perspective to use differen physics middleware, this was not possible on previously custom built system architecture. Several uselful LSL programming methods, including sensor methods, are provided by the platform.
4.
Physics engine limits as a simulator for climhing and walking rohots
It might seem, from what was summarised above, that metaverses in general, and
SecondLife in particular, provide an ideal environment for outdoor simulation of
950
clawars , for example in search and rescue, survelliance, mine clearing, and other kind of robots, this will be probably true in the perhaps close future, but there are presently some limitations which prevent from performing very realistic simulations. A physics engine is a program that simulates Newtonian object collisions and interactions in a mathematically and computationally simulated virtual world representaion, by simulating gravity, elasticity, and the conservation of momentum between interacting, i.e. colliding, objects, but these features are not the core purpose in such social networking or gaming facilities, [10,7,8]. Although this will probably change in the next future when the system will more and more integrate the new Havok 4 physics engine, the physics engine is mainly used in Second Life to discriminate between empty space and filled space and to avoid interpenetration of object, including avatars and user created buildings not falling under the ground, to guarantee some basic motions and manage contacts between objects. In most if not all computer games, and in particular in online games wher the response time for a huge number of users must be limited enough to carry an illusion of real interaction in a real world, speed of simulation is more important than accuracy of simulation. In a game an object is quite often represented by two separate meshes. One of these has a quite complex and detailed shape which the player sees in the game, for example a greek temple or a sport car. The other invisible one is the one actually used to calculate the motion, in both the above case we may have different dimension parellepipeds including the other more complex model shape, this simplified mesh is usually called the 'collision geometry'. Even when, like in SecondLife, the collision geometry, the simplified mesh, is complex enough, the object body moves as a single quasi-rigid body. Wheels of a mobile differential or car like wheeled robot or legs in a climbing and walking robots are not moved autonomously, for example in Havok 1 physical engine, it will be possible easyly in the next future as the new phisics engine will be progressively fully integrate. A particularly clear example is given by sphere rolling in Havok 1, as they are represented internally by a many faceted polyedra in some conditions, due to the coarse precision of calculations they stop rolling on an inclined plane. Another limitation, due to the need of sparing computing resources, if an object is located on the ground and the object does not move beyond a certain threshold distance in about two seconds, then the motion simulation for that object is disabled and it remains frozen until a collision occurs with some other object. To can lead to unrealistic trajectory simulation like an high mass, slow enough, pendulum sticking on the upswing, where it should reverse direction.
951
Other issues come from rounding errors in positions and forces, for example when two independent objects are assembled with a precision that is greater than the thershold allowed by the physics engine. This can lead to an unnatural buildup energy in the object due to the rounding errors, that begins to violently shake and eventually blow the objects apart. This kind of problems may affect chain links under high tension, and wheeled objects with actively physical bearing surfaces. This limits the complexity of clawar and mobile robots that can be correctly simulated. The framerate is comparatively low. Small fast-moving objects do not move smoothly through space. This limist the maximum velocity of moving objects, but it is not a problem with clawar simulation. If there is a need for rocket or projectile simulation it must be currently managed with specific 'tricks'. It must be noticed that with the move to the the new physics engine it will possible to perform multi rigid body simulation with some limitation on the kind of joints Crag doll' simulation).
5.
Discussion
There is a remarkable number of simulation environments which can be used for simulation of clawar systems and robotics, Webots, USARSim,so that it claimed that in many if not most cases investinng resources in an ad hoc simulator may be a waste of resources for robotocs reasearchers [5]. All of them show, at present, clear advantages over Second Life and similar environments as regards the level of details of a single or small group of robots dynamics in mechanical terms. In particular the so called High Fidelity simulators like USARSim, [9], used in the virtual Urban Search and Rescue Robocup league and based by the way on the Unreal Tournament, a shooter game physics engine, allow an almost perfect representation of a multi body system dynamics, namely a clawar or mobile robot, moving and interacting in a small essentially deterministic environment. Where a Second Life based, or in general one developed into a'metaverse', simulator can give clear advantage is when you need to simulate in an open ended geographical environment, with winds, floods, clouds, crowds of real people, and you are interested in the collective emerging behavior of network of many robots, more than in the, say, the stability of walking of the single robot. There are of course at present tight limits to the control system complexities, though. Another, practical, advantage is the possibility of the researcher to share the same simulation environmen from different location in the, real, world, and even to literally enter and interact with his/her avatar, into the simulation. In future perspective, a few years as most recent physics engine, thanks to Moore law and
952
distributed sytems technologies progress will be deployable, it will be possible to simulate more realistical collective behaviour of huge networked robot system with more complex control systems. 6.
Future work and conclusions
We plan to develop, with an open source approach on top of LSL scripts and native objects, a number of basic robotic template which may be used as unit in the swarms and some basic robotics behaviour taken from the swarm intelligence literature to test in a simulated environment. There are already a few early initiatives for robotics in SecondLife which perhaps might be merged. On a different respect, simulating intelligent behaviours in synthetic world where the informational complexity of the simulated physical systems is under control and well measurable, might help shed light on the relations between information metrics and dynamics in cognitive automous systems. This will be also investigated. We have seen that, so far, Second Life is not fit for detailed dynamical simulation of single clawar and mobile robot systems. This may change in a few years. It is thought that the bigger benefits coming from performing clawar simulations in such environments will be in network and swarm robotics research, in particular for Search and Rescue, Survelliance and Mine Clearing in geographical size settings. References 1.
G. Armitage, M. Claypool and P.Branch, Networking and Online Games, Wiley, (2006). 2. T. Alexander (ed.), Massively Multiplayer Game Development, Charles River Media, (2005). 3. E. Castronova, Synthetic Worlds: the Business and Culture of Online Games, Univesrity of Chicago Press, (2005). 4. M. Rymaszewski, W.J. Au, M. Wallace, C. Winters, C. Ondrejka and B. Batstone-Cunningham, Second Life: The Official Guide, Wiley, (2007) 5. 1. Craighead, R. Murphy, 1. Burke, B. Goldiez, A Survey of Commercial & Open Source Unmanned Vehicle Simulators, ICRA2007, (2007) 6. http://www.lslwiki.net/ 7. http://www.havok.com/ 8. http://www.ageia.com/ 9. http://usarsim.sourceforge.net/ 10. D. Featherstone and D.E. and Orin, D. E., Robot Dynamics: Equations and Algorithms, ICRA2000, (2000)
SECTION-1S PERCEPTION, SENSING AND SENSOR FUSION
This page intentionally left blank
APPLICATION OF LATERAL OBSTACLE SENSOR IN FOLLOWING CONTOURS FOR TERRAIN RECOGNITION TASKS R. PONTICELLI* and P. GONZALEZ DE SANTOS Department of Automatic Control, Institute of Industrial Automation - CSIC, Madrid, CP 28500, Spain E-maiL *[email protected] [email protected] www.iai.csic.es/dca
One application of the sensor head designed for terrain scanning in humanitarian demining tasks in the DYLEMA project is presented, where the lateral distance to obstacles measured by a network of lateral range sensors is converted into a virtual contact force, which in turn is feed as input for a contact force control loop. The sensor head sweep movement is modified when an obstacle is detected (or "touched") also helping to detect the position of the obstacle's contour.
Keywords: contact force; range sensor; contour following; obstacle detection; terrain recognition.
1. Introduction
The humanitarian antipersonnel land mines removal requires the use of robotic systems putting the operator at a distance away from the infested area increasing both operator safety and the mine removal rate. 1 ,2 In the DYLEMA project, which is devoted to the configuration of a humanitarian de-mining system that consists of a sensor head, a scanning manipulator and a mobile platform based on a hexapod walking robot 3 (see Fig. 1), the sensor head consists of a commercial mine-detecting set customized with a network of range sensors helping to adapt the head to ground surface's profile and to detect lateral obstacles. The data provided by the sensor head are also used by the system controller to steer the mobile robot during mine detection missions. 4 The sensor head prototype relies on a metal detector. For plastic or wooden landmines, other mine detecting technologies can be used. 5- 7 The strategies for landmine scanning and location remain the same regardless
955
956
Fig. 1.
DYLEMA configuration
of the mine-detecting set technology.S
2. System break-down 2.1. Manipulator The mine-detector set senses one point at a time, requiring to be swept over the whole terrain. DYLEMA employs a 5 DOFs robotic manipulator for both moving the sensor head and adapting it to terrain irregularities. 9 ,lo Figure 2 shows the scanning manipulator developed. Some manipulator's dimensions, such as manipulator-link lengths, depends on the hexapod's body height and leg span. s
2.2. Sensor head The sensor head subsystem consists of a network of range sensors mounted around the mine-detecting set helping the scanning manipulator to control the sensor head height and orientation, at about 5 cm above and parallel to the ground, detecting objects in the sensor head's path as wel1. 9
2.2.1. Gr01lnd-sensor group The ground-sensor group employs range sensors to cover an area under the sensor head, enabling relative ground elevation measurement without
957
Fig. 2.
Scanning manipulator and sensor head
mechanical contact. The sensors are accommodated at the outer edge of the sensor head at regular intervals, pointing downwards at the ground, and provides data of the plane under the sensor head (angles and distance) and information for building an elevation map, combining range sensor readings and spatial coordinates obtained from the robot localization system and manipulator kinematics. 4
2.2.2. Lateral-sensor group These range sensors are accommodated at the outer edge of the sensor-head support platform at regular intervals, with their beams oriented approximately tangent to the support platform. The orientation of each range sensor is adjusted such that, seen radially, each area overlaps the next (see Fig. 3). An object approaching the sensor head in a radial direction will be detected by at least one range sensor.
2.3. Contour following The contour following algorithm presented in this paper is based on the traditional contour following techniques employed in machine tools, which employs force control loops adjusting the final position of a manipulator's
958
Fig. 3. Sensor head Infrared beams representation. (A) Ground beam, (B) Support platform, (C) Ground sensor, (D) Lateral sensor, (E) Lateral beam.
end-effector accordingly to the applied force to the target. In the DYLEMA project, the contour following helps determining the area pertaining to a terrain feature (rocks, trees, holes, steps in terrain, etc.) that can be considered as an obstacle in the path of the mobile robot.
3. Algorithms 3.1. Compliance control The compliance control implemented employs two procedures: 3.1.1. rarce sensing The most common type of force sensing is based on indirect determination of the force through the measurement of the deformation exerted by the force on a known stiff material. The application presented in this paper uses the idea of force sensing through the deformation measurement of a spring (Hooke's law). The distance measured to a potential obstacle in the robot's path is converted into a virtual contact force. 3.1.2. TerTain sweep The mobile robot employs a complete-coverage algorithm, where an exact cellular decomposition is performed dividing the field into regions or cells
959
o (1)
(2)
Fig. 4. Terrain sweep (A) Mobile robot, (B) Sensor head, (C) Scanning manipulator sweep trajectory, (D) Mobile robot sweep trajectory, (E) Scanned terrain, (F) Unscanned terrain.
free of obstacles, so that the sum of resulting cells equals the total space free of obstacles. l l The legged robot performs a coarse sweep motion whilst the scanning manipulator performs a fine sweep motion over the surface in front of the legged robot. The resulting sweep movement of the sensor head is a composition of the two main sweep motions (robot plus manipulator) working synchronously, as depicted in Fig. 4. The nominal sweep trajectory of the manipulator's end-effector presents a cross-trapezoidal loop; i.e. this movement alone, without the lineal displacement of the mobile robot, sweeps the sensor head in a cross-loop with trapezoidal-like shape. This cross-loop trajectory gets straightened (see Fig. 5) when added to the lineal displacement of the mobile robot, resulting in a succession of trapezoidal-like sweeps.
Fig. 5. Composite sensor head sweep movement. (A) Mobile robot with a constant forward velocity, (B) Sensor head, (C) Cross trapezoidal sweep trajectory W.R.T. the mobile robot, (D) Trapezoidal sweep trajectory W.R.T. the absolute reference frame (terrain).
960
B
A
Fig. 6. Sweep with compliance behavior. (A) Sweep trajectory, (B) Sweep trajectory modified by a small obstacle, (0) Sweep trajectory modified by a big obstacle.
3.2. Contour following To obtain an obstacle's contour an algorithm has been implemented based on two steps: 1) sweep movement with compliant contact, and 2) aggregate information from previous sweep passes. The compliant behavior of the manipulator end-effector allows partially following obstacle's contour whilst the manipulator is performing the nominal sweep movement of the sensor head (see Fig. 6). The distance to an obstacle in the sensor head's path is converted into a virtual contact force and feed as input to a force control loop, modifying the end-effector's position command sent to the manipulator, as depicted in Fig. 7. The manipulator sweeps the sensor head over the terrain following the surface profile. The compliant behavior is activated when an obstacle is detected closer than a predefined distance (a virtual contact is detected),
Pout
Manipulator
Fig. 7.
Force-position control loop.
961
modifying the manipulator target position with a damped-spring-like response. If the sensor head "touches" the obstacle (i.e. with a low contact force), then the sweep is modified so the sensor head follows the segment of the obstacle's perimeter that intersects the original sweep trajectory. The determination of the full contour shape is done after subsequent compliant sweeps has been performed and a map is built. 4. Experiments
The main validation experiment consists of the mobile robot settled at a fixed position whilst the manipulator performs the sensor head's sweep movement. The terrain under the sensor head is uneven and the manipulator has to adapt the position and orientation of the sensor head continuously to follow the contour. To test the performance of the manipulator's compliant behavior, different obstacles are introduced in the sweep path to test three particular situations: 1) sensor head touches a small obstacle causing a small deviation from the nominal sweep trajectory; 2) sensor head touches a big obstacle causing a mayor deviation from the nominal sweep trajectory; and 3) sensor head touches a big obstacle preventing the nominal sweep trajectory from completing (see Fig. 8). The first two situations are demonstrated in the video sequence b, and the last situation is demonstrated in the last
a
b
c Fig. 8. Video frames sequences of the experiments. a) Nominal sweep trajectory, b) T'rajectory modified because the "smooth contact" with an obstacle, c) Trajectory stalled because the "hard contact" with a big obstacle.
962 sequence c. 5. Summary and Conclusions A head to be carried by a mobile robotic platform for terrain recognition tasks has been developed, being the obstacles contour following capability the main result presented in this paper. The contour detecting is performed by a range sensors array disposed at the perimeter of the sensor head assembly. Based on Hooke's law, a force control algorithm has been implemented using the cartesian distance to obstacles readings as inputs. The qualitative system performance has been validated though the experiments videos, but quantitative results are not yet available to the date of writing this paper. Further results and analysis will be exposed at the CLAWAR'08 conference meeting. References 1. Y. Baudoin, M. Acheroy, M. Piette and J. Salmon, Mine Action Information Center Journal (1999), Vol. 3. 2. J. D. Nicoud, Industrial Robot: An International Journal, 164 (1997), Vol. 24. 3. Dylema (2008), The SILO-6 and DYLEMA Projects' Home Page. Available at http) /www.iai.csic.es/users/silo6/. 4. R. Ponticelli, E. Garcia, P. Gonzalez de Santos and M. Armada, Industrial Robot: An International Journal, 133 (2008), Vol. 35. 5. P. D. Gader, B. Nelson, H. Frigui, G. Vaillette and J. Keller, Signal Processing, Special Issue on Fuzzy Logic in Signal Processing, 1069 (2000), Vol. 80, Invited Paper. 6. K. Albert, M. Myrick, S. Brown, F. Milanovich and W. D.R., SPIE , 308 (1999). 7. A. Rouhi, Chemical €3 Engineering news, 14 (1997), Vol. 75. 8. R. Cobano, J.A. Ponticelli and P. Gonzalez de Santos, Industrial Robot: An International Journal (2008), in press. 9. R. Ponticelli, P. Gonzalez de Santos and M. Armada, Lateral sensor for ground tracking systems in humanitarian-demining tasks, in Proceedings 9th International Conference on Climbing and Walking Robots, (Brussels, Belgium, 2006). 10. P. Gonzalez de Santos, E. Garcia, J. Estremera and M. Armada, Internacional Journal of Systems Science, 545 (2005), Vol. 36. 11. E. Garcia and P. Gonzalez de Santos, Robotics and Autonomous Systems, 195 (2004).
TRUE GROUND SPEED MEASUREMENT, A NOVEL OPTICAL APPROACH VIKTOR KALMAN, TIBOR TAKAcs Department of Control Engineering and Information Technology, Budapest University of Technology and Economics, Magyar Tud6sok krt. 2, Budapest, 1117, Hungary Optical sensors supply by far the most information and as greater and greater processing capabilities become readily available their use becomes more widespread. Many researchers and companies have made more or less successful attempts at creating optical sensors for speed measurement, however there are still many questions open for research. The aim of this article is to introduce a novel method for optical speed measurement and put it into perspective by summarizing and reviewing recent related work. Practical considerations on texture analysis and sensor parameters are discussed backed up with simulation results.
1. Introduction Mobile robot navigation is a well researched discipline, looking back to a relatively long history however it is still a rich, active area for research and development. The ultimate goal for robots and intelligent vehicles seems to be autonomous navigation in complex real life scenarios[l]. In order to achieve higher levels of autonomy sophisticated sensors and a sound understanding of the robot and its interaction with the environment is needed. The development of navigation and dynamic sensors has always had a prominent place in mobile robotics research, as the key to accurate trajectory tracking and precise movements is the exact knowledge of the dynamic parameters of the mobile platform. Our work focuses on an optical speed sensor that is mounted on a moving platform and uses a camera facing the ground. It uses digital image processing to calculate displacement periodically. It has several advantages over traditional wheel based measurement methods, thus it could be used as a standalone sensor or could complement a sensor fusion system as well. 2. The basics of optical motion measurement Visual movements are caused by the relative displacement of the observer (eye, camera) and the objects of the world. Most techniques of visual motion
963
964 measurement are based on the well researched discipline called "optical flow". The basic idea is to compare consecutive images of a scene produced by camera and calculate a vector field for each image which shows the displacements of the pixels to get the next image of the scene. This vector field is often called optical flow or optical flow field [2].
2.1. Related work Optical speed measurement has existing commercial solutions and considerable research activity in the academic sector. As for commercial technologies the most widely known example is the optical mouse, which on the other hand has generated a fair amount of academic research. The optical mouse uses two distinct but essentially similar techniques for displacement calculation. The classical method uses sideways LED illumination and relies on the micro texture and cast shadows of the surface. The more advanced texture independent method is laser speckle pattern technology. Laser speckle patterns can be observed when a rough surface (rough, relative to the wavelength) is illuminated with a coherent light and the interference of the reflected light waves creates a surface dependent random intensity map on the detector, which changes according to the movement of the light source. Frequency analysis is a less frequently used method. The light reflected from the surface travels through an optical grating, and is focused on a pair of photo-detectors. The surface elements, passing in front of the grating generate a certain signal frequency in the detectors depending on the sampling frequency, ground speed, grid graduation, ratio of the image, size of the surface elements and the size of the picture on the grating. The difference of the two signals is computed and its frequency corresponds to the true ground speed. Indoor dead reckoning solutions for small mobile robots using optical mice were suggested by several authors [8,9] T.W. Ng investigated the usability and accuracy of optical mice for scientific measurements in several articles [5] with good results. They found that the readings possessed low levels of error and high degrees of linearity. However the measurements depended strongly on the distance between the surface and the detector, due to the fixed focal length and the illumination direction of the mouse. Several researchers proposed the use of optical mice as a dead reckoning sensor for small indoor mobile robots in one and two sensor configurations. By using one sensor and kinematical constraints from the model of the platform, a slip free dead reckoning system can be realized. The kinematic constraint originates from the sensors inability to calculate rotation. By using two sensors
965
the constraint can be removed and the measurements become independent of the platforms kinematics. In general all researchers found that mice were extremely height dependent, to the extent that they could not be reliably used on carpet. Also due to their limited field of view and resolution they were only usable for a few mls. Some of their problems can be eliminated with the use of homogeneous lighting and telecentric optics, with additional cost. Mouse sensors are cheap and readily available and with certain modifications they can be used for low speed mobile robot dead reckoning. Horn et al. aimed at developing a sensor system for automobiles. They used a fusion approach with two cameras and a Kalman filter. One of the cameras is a forward looking stereo camera to estimate yaw rate and forward velocity, the other camera is facing the ground and used to estimate two dimensional velocity. It was found that the camera facing the ground gave better results for lateral and longitudinal velocity than the stereo camera. The fusion approach provided good results even when one of the sensors was failing. The system was tested at slow « I mls) speeds on a towed cart in a lab [11]. [3] followed a somewhat similar approach and used a matrix camera facing the ground to estimate speed over ground. They used a mechanism that moved the camera over sand and compared optical flow speed estimates with measurements from an encoder attached to the mechanism. They used Matlab and the Lukas and Kanade algorithm to compute optical flow. They obtained good results at low speeds (0-50 mmls), however the suitability of the algorithm they used is questionable. This technology has already found its way to the transportation industry as well. [4] has a one-of-a-kind optical speed sensor used for testing the dynamics of passenger vehicles before mass production. The sensor is claimed to be working on any surface, including water and snow, but it is priced for the big automotive manufacturers. It uses the frequency analysis method. [6] describes an optical speed measurement system for automated trains. It uses the principle of laser speckle interferometry mentioned above, and "looks" directly on the rails to measure the trains speed. It is clear that much work has been done in the field of optical navigation but several issues remain open for research. Current industrial solutions are somewhat bulky and definitely not priced for the average mobile robot. Solutions by academic researchers have not matured to the level of really useful applications. Mouse chips are mostly the sensors of choice. With some modifications their problems of ground distance, lighting and calibration can be
966
helped, but their current speed and resolution is simply not enough for high speed (the order of ten m/s) applications.
3. Our approach In this section we outline the basics of the motion measurement system proposed by the authors. Our goal was to lay the foundations of an optical sensor that can measure speed of a moving platform relative to the ground in two dimensions, up to speeds of several ten meters per second, at a relatively low cost.
3.1. Basic assumptions We claim that accurate two dimensional measurements can be made with linescan cameras. The most important advantages of this type of camera in respect of displacement measurement are relatively high - several mega pixels - resolution in one dimension, frame rates at the order of 10 to 100 kHz and relatively low prices. Inherently the motion component orthogonal to the main axis causes errors in the calculation of parallel displacement. This error can not be totally eliminated but it is possible to decrease this effect with high frame rate and larger field of view of the camera. If the sampling frequency is high (which is easy to reach with line-scan cameras) then the perpendicular displacement between two consecutive images can be small enough that they will be taken of essentially the same texture element, making correlation in the parallel direction possible.
Figure 1. Multiple sensor displacement model
Some assumptions on which we based our investigations: the sensor is facing the ground, which is relatively flat, the field of view is constant because we operate in the constant magnification range of telecentric optics, there is only one velocity vector for the whole image and our sensor can only measure movements along a straight line. The last assumption makes a multiple sensor approach (Fig. 1) necessary to make the measurements independent from
967 platform kinematics. The arcs traveled by the sensor are approximated by straight lines between sampling intervals. X=d 1/lX2+ d2/l X l . Y ' d 1 +d 2
(1)
Displacement of any other point of the platform can be calculated with a simple geometrical transformation. If the reference point is in the origin of sensor #1 (namely d 1 = 0), then the equations in (1) became simpler: X = ｾｸ｝Ｌ＠ Y and a = arcsin((Llx2 - LlX2) / d 2 ). Note that the system is overdetermined = ｾｹｬ＠ the y component of the second sensor is not needed.
3.2. Sensor parameters and the effect of texture For purely texture image based systems the importance of texture can not be overlooked as it affects sensor qualities and forms the basis to define sensor parameters like sampling frequency, magnification, resolution, pixel size and shape. Texture size might be the most important feature of a given texture as it determines the size of the area the sensor needs to look at i.e. the magnification. Texture size can be hard to define as it depends on how closely we look at a given surface. Looking at micro texture might be a better option as it is usually available on otherwise homogeneous surfaces - laser speckle correlation takes advantage of this - but if we use a small image with great magnification, we limit the maximal speed measurable as for a given frame rate and resolution we might not get overlapping images. Several methods exist in the literature to determine texture size. For example [10] used difference histograms. [7] used segmentation on a binarized image to determine average particle size. The theoretical limit of geometrical precision of movement calculation also depends on texture; only the presence of sufficient high frequency components will guarantee precise correlation [12]. The highest frequency of interest can be determined from the energy spectrum of the image. Optical mice illuminate the surface at a low angle creating long shadows of miniature surface irregularities, making measurement possible on surfaces of homogeneous colour. Laser speckle interferometry offers another alternative. A serious drawback of both the above mentioned illumination methods is that both patterns change with the distance between the light source and the object. This effect makes displacement measurement hard, if not impossible.
968
3.3. Simulation results The simulator - written in Matlab - works the following way: the ground is represented by a high resolution image (Figure 2.), and it is projected onto an arbitrary rectangular detector (Figure 3.).
a.) Concrete
b.) Cork
c.) Stones
d.) Dust
Figure 2. Some of the ground textures used in the experiment
Figure 3. Projection of matrix and line-scan camera (illustration)
Magnification, ground speed and sampling rate are chosen and between sequential images Gaussian noise is added (simulating the noise of the camera). Three motion directions can be chosen, zero, 45, and 90 degrees. The two neighbouring images are then compared according to a distance measure of choice such as correlation, least squares, Manhattan and cosine distances. As the exact distance in pixels is known the error of the measurement can be obtained easily. Movements that are non-parallel to the detectors main axis introduce errors. These can be minimised by widening the field of view, at the expense of loosing contrast (Figure 4.).
a) Consecutive images with wider field of view
b) Consecutive images with nan-ow field of view
Figure 4. The effect of field of view shape factor
The only difference between a) and b) is that the width of the detector was bigger on a). It is important to note here that increasing image width much further leads to total loss of contrast making measurements impossible. Figure 5. shows the measurement error versus the frame rate (other parameters fixed). The simulated velocity of the platform is an ambitious 100 m/s and the direction of movement is 45 degrees. This frame rate range is usual for common line-scan cameras.
969
a.) Stone
b.) Cork
Figure S. Error versus frequency for different textures
From the figure the tendency can be seen that for "bigger" texture size the errors converge to zero at smaller frequencies, however more experiments are needed with different textures to verify this assumption. Figure 6. shows the error surface as a function of the two dimensions of the field of view. The main axis of the line detector is length; width of the sensor is scaled in percentage of the length, 100% meaning a square field of vision.
a.) Stone
b.) Cork
Figure 6. Error sUlfaces as a function of field of view ratio (widthllength) JSkfps
It is clear from the images that increasing the length alone does not decrease the error, image ratios of 40% or larger are needed to obtain acceptable measurements.
Figure 7. The effect of increased frame rate Cork
@
30kfps
However increasing frame rate allows us to choose ratios around 20% which is demonstrated on figure 7. These results seem logical as an increase in frame rate means smaller displacements between frames making correlation possible for narrower images too.
970
4. Summary In this paper a novel method of optical speed over ground measurement was presented. The proposed sensor setup uses line-scan cameras to measure displacement in two dimensions. The feasibility of the setup is verified by software simulation in which the effects of perpendicular movement image size and frame rate were tested, at simulated speeds in the order of several ten m/s. Future work will include tests on real hardware and extended research on applications. References
1. Bishop, R. Intelligent Vehicle Technology and Trends Artech house, ISBN 1-58053-911-4, Norwood, MA, USA. (2005) 2. Beauchemin, S. S. et al. The computation of optical flow. ACM Computing Survey, Vo1.27, 1.3, (Sept.'95), 433-466, ISSN 0360-0300 3. Chhaniyara, S.; et al. Optical Flow Algorithm for Velocity Estimation of Ground Vehicles: A Feasibility Study, Int. Journal on Smart Sensing and Intelligent Systems, Vol. 1, No.1, (March 2008.) 246-268 4. Correvit(R)-SL (2001) Non-Contact Optical Sensor for slip free measurement of longitudinal and transversal dynamics, Corrsys-Datron Sensorsysteme GmbH, 2001. [Online]. Available: www.corrsys-datron.com 5. Ng, T.W. (2003) The optical mouse as a two-dimensional displacement sensor. Sensors and Actuators, A 107, (2003.) 21-25 6. OSMES OSMES the optical speed measurement system Siemens Transportation Systems 2004. 7. Pi, M. H. & Zhang, H. Measurement of fine particle size with wavelet signature, Proc. of Int. Can! on Image Processing, ICIP 2005. IEEE Volume 3,11-14 Sept. pp.:III - 165-8 8. Sorensen, D. K. On-line Optical Flow Feedback for Mobile Robot LocalizationlNavigation MSc Thesis, 2003, A&M University, Texas, USA 9. Palacin, J. et al. The optical mouse for indoor mobile robot odometry measurement, Sensors and Actuators A 126 (2006.) 141-147 10. LepistO, L. et al. Grain Size Measurement of Crystalline Products Using Maximum Difference Method, In: Image Analysis, pp. 403-410, Vol. 452212007, Springer Berlin 11. Horn, J.et al. A Fusion Approach for Image-Based Measurement of Speed Over Ground, Proc. of Int. Can! on Multisensor Fusion and Integration for Intelligent Systems, pp. 261-266, Sept. 3-6,2006, Heidelberg, Germany 12. Forstner, W. On the geometric precision of digital correlation, Int. Archives of Photogrammetry and Remote Sensing, vol. 24, no. 3, pp. 176-189, 1982.
SIMPLE OPTOELECTRONIC EXTEROCEPTIVE SENSOR FOR THE CONTROL OF THE DYNAMIC EQUILIBRIUM OF A WALKING ROBOT ERIKKRAL
Tomas Bata University in Zlin, Faculty ofApplied Informatics Nad Stranemi 4511, CZ-76005 Zlin, Czech Republic Simple optoelectronic exteroceptive sensor for controlling and learning the dynamical equilibrium of a walking robot is presented. Sensor consists of the digital camera and structured light source, for example laser diode module with Diffractive Optical Element. Digital camera captures structured configuration of light spots projected on a surface in front of a robot. There are two variants, the light source position is fixed to the digital camera and there is no reference object at scene. In these case only medial and lateral tilt to plane in front of the robot is estimated and second variant when only light source of the robot is fixed to any part of robot and camera is placed somewhere else. In these case multi-degrees-of-freedom information can be estimated. The image information from the digital camera is input for the control of a dynamical equilibrium of a walking robot.
1.
Introduction
This article proposes simple optoelectronic exteroceptive sensor for the control of the dynamical equilibrium of a walking robot mainly for the indoor environments. It is not supposed as a main control device, but it is means as an extension for optimal motion on a planar surface. There are two versions, first the camera is fixed to the light source and placed to the robot oriented toward, then the medial and lateral tilt is estimated as the input information for the control of a dynamic equilibrium of a walking robot depicted in Figure 1. Additionally, the second camera, placed apart from the robot can be used ifthere is a reference object with two pair of parallel lines [1]. In this case, multidegrees-of-freedom information can be used for the control of a dynamical equilibrium of a walking robot, depicted in Figure 2.
971
972
I.Walking robot with mounted optoelectronic exteroceptive sensor.
2.
Figure 2. Second camera and reference rectangle
Methods
The main part of the image processing is done using the algorithm for the analysis of the image information for the accuracy of the positioning of the laser diode's light spots [3], depicted in Figure 3 which creates the input infonnation for the control of the walking robot. The laser light spots positions are determined using basic matching algorithm. Using 2-D Gauss distribution precise position of the centcr of the lasers light spot is estimated. The camera system distortion has to be considered and is eliminated by the application of the calibration camera model [4] and [5]. Perspective distortion is removed using two parallcllines in a reference object [6]. A projective transfonnation can map ideal points to finite points. Under an affine transfonnation the ideal points remain at infinity [6].
3.
Algorithm components
In Figure 3 you can see the example of the pattern consist of four points A, B, C, D projected with four laser beams from one center with the uniform inter-beam angle G. These laser beams creates pyramid. In case, that there is any reference object with know pair of parallel lines, perspective distortion could be removed using the two parallel lines in reference object [3]. It means that there should be
973 another reference pattern, square or rectangle, with known parallel lines. In this study, only the pattern with four points is described. The main reason is that it could be easily built with four low-cost laser diodes mounted in desired configuration. This algorithm is designed to be used in low cost embedded hardware like FPGA or microcontroller.
Algorithm has these basic steps 1) 2) 3) 4) 5)
Estimation of the suspicious areas Finding the suitable areas combination Camera distortion elimination Estimation ofthe centre of the light spot using 2-D Gauss distribution Estimation of the medial and lateral tilt or of the multi-DoF information.
The matching algorithm uses the pattern with the discrete Gauss distribution. Only areas with a sudden change of brightness are tested. The suspicious areas are the areas which match the pattern and are small enough to be considered as
974 light spots. First step is the affine rectification. A projective transfonnation can map ideal points to finite points. Under an affine transfonnation the ideal points remain at infinity [6]. Angles P}' P2 of the rotations of the pyramid axis around pyramid's base diagonals are computed from ratio of length of U/, U2 respective U3 ,U4 :
(1)
Where (J is the light source inter-beam angle and u/ ｾ＠ U3 ; U2 ｾ＠ U4. As the affine transfonnation preserve ratio of lengths on collinear lines, the ratio of k/ and k2 should be equal. Then, the expected positions of light's beams with known angles are constructed. These positions are compared with measured position and parameters of the inverse affine transfonnation are computed. As the affine transfonnation preserve ratio of lengths on collinear lines, the ratio of k/ and k2 should be the same. Then, the expected positions of light's beams with known angles are constructed. These positions are compared with measured position and parameters of inverse affine transformation are computed. Combinations with low affine transformation error and with the low difference of the ratio k/ and k2 are chosen. The best combination is finally estimated according to its brightness and level of matching to the pattern. By solving system of equations (2) for 3 and more corresponding points, we can find matrix T, which can be easily transfonned to the affine transfonnation matrix TA .
(2)
ax+ by + C - gx'x- hx'y - ix' = 0 dx + ey + f
- gy'x - hy'y - iy' =
0
(3)
975
XI
YI
1
0
0
0
0
0
XI
YI
x"
Y"
1
0
0
0
0
0
x"
Yn
,
0
-XIX\
- Y;XI
, -XnXn ,
0
-YnXn
-X;YI
-x;
a
-Y;YI
-Y;
b
,
- y:xn
-XII
- Y:Yn
-YII
c
(4)
,
The radial distortion can be approximated using the expression (6) and tangential distortion is often describes as expression (7)[5]
(5)
(6)
(7)
A camera model for distortion calibration is derived by combining the CCD camera projection model (5) with the correction for the radial and tangential distortion components.
Ui] v
l
J
=
K
[(u, + Ou i.I') + Ou,.U)Jj (v. + Ov I
1
I'·) J
+ Ov U) J
(8)
1
Precise positions of the centers of the laser light spots are estimated using 2-D Gauss distribution approximation [3]. And we considered the centre of 2-D Gauss distribution approximation as the centre of the laser light spot. In case, there is reference object, positions of the centers of the laser light spots can be transformed into the three rotations and three translations [7]. As a final step, the information about the medial and lateral tilt or of the multiDoF information can be used as input information to the neural network that provides control of the dynamic equilibrium of a walking robot.
4.
Implementation
Evaluating algorithms are implemented in MA TLAB and executive algorithms are implemented in the Field Programmable Gate Array (FPGA) with the Very
976 High Density Hardware Description Language (VHDL). 5.
Conclusion
This article propose simple optoelectronic sensor for the control of the dynamical equilibrium of a walking robot mainly in indoor environments. The medial and lateral tilt or multi-degrees-of-freedom information can be used for the control of the walking robot or for the learning of walking robot. Acknowledgments
The financial support from the grant MSM7088352102 "Modeling and Control of Processing Procedures of Natural and Synthetic Polymers" is gratefully acknowledged. References
1.
2.
3.
4. 5.
6.
7.
E. Knil, Preliminary Study of Construction of Optoelectronic Control Device for Object Modeling in 3D Virtual Environment, Ostrava: International Workshop Control and Information Technology (2007) ISBN 978-80-8073-805-1 M. Kvasnica, Six DoF Sensory system for the force-torque control.of walking humanoid. Proceedings of the 11th International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines, Coimbra, Portugal, 08 - 10 September 2008. E. Knil, The analysis of the image information for the accuracy of the positioning of the multicomponent sensory system's laser diode's light spots. Strbshi Pleso: 8th International Carpathia Control Conference, (2007), ISBN 978-80-8073-805-1 C. Mei, 2006. Camera Calibration Toolbox for Matlab
R. Tsai, An Efficient and Accurate Camera Calibration Technique for 3D Machine Vision, Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, Miami Beach, FL, pp. 364-374 (1986) R. Hartley and A. Zisserman, Multiple View Geometry in Computer Vision. Second Edition, Cambridge University Press (2004), ISBN 0521540518 M. Kvasnica. Algorithm for Computing of Information about Six-DOF Motion in 3-D Space Sampled by 2-D CCD Array, The paper from Invited Session by Kvasnica M.: "Robotics, Sensory Systems for Robotics and for Human-Machine Interface", Proceedings of the World Multi-Conference SCI '2001-ISAS, VoI.XV, Industrial Systems, Part II, Orlando, Florida, July 2001, USA.
ANALYSING HUMAN-ROBOT INTERACTION USING OMNIDIRI!CTIONAL VISION AND STRUCTURE FROM MOTION CARLOTA SALINAS, MANUEL A. ARMADA Automatic Control Department, Industrial Automation Institute ｾ＠ CSIC, Carretera de Campo Real, Km. 0.2, 28500 Arganda del Rey, Madrid, Spain Service Robots are intended to support people in daily activities (complex environments), and, for this reason, human-robot communication is an important subject. This work is based on the fact that humans noticeably exploit motion information from other humans' actions. This paper presents a new approach of analyzing Human Robot Interaction based on robust optical flow techniques to calculate motion information using a catadioptric system implemented to enhance the Field of Vision. This allows capturing much more information for the interpretation procedure that employs a hierarchical fuzzy making decision technique designed for the learning model.
1. Introduction
Autonomous mobile robots are intended to be used in several application areas, such as exploration activities, search and rescue, humanitarian dernining [1], and in a broad range of relevant fields as service robots for assisting people both in domestic and/or in professional environments. These robots should support people in their daily activities; hence an important part of its performance relies in the interaction between humans and robots. Since those activities are accomplished either in environments containing various uncertainties or the scenarios are non-structured or change dynamically, it should be necessary to provide robots with some kind of intelligence. Traditional studies on robotics have been mainly focused on the autonomous behaviour or service of robots rather than in their interaction performance. Even a robot performing excellent behaviour might be not-smart and carry out operations erratically. Communication is very important for robots to understand the user's intentions and to perform tasks correctly. An effective approach is to construct models by which a robot emulates cognitive functions of humans. The knowledge that humans make use of motion information of others' action [2J is where this work is focused on. The purpose of this work is to provide a new approach for analysing human-robot interaction using omni-directional vision [3] to enlarge the field of view and structure from motion techniques to calculate motion information,
977
978
aimed to allow performing smart operations and making decisions. Proposed vision system is based on omni-directional catadioptric system [4, 5], which is not limited in its FOV as conventional cameras does; omnivision systems are able to observe full 360 view around the robot. The paper is organized as follows. In Section 2, the highlights of catadioptric systems are described and the geometry of the employed camera model for this work is presented. In order to compute multiple objects in movement, the next section explains the studied methods for Structure From Motion (SFM) [8, 10 and 13]. Section 4 illustrates the methodology for humanrobot interaction analysis, based on hierarchical fuzzy (making-decision) technique. Last sections introduced the case of study of Sil06 manipulator [1] and experimental results are shown. Finally, conclusions and future work are presented. 0
2. Omnidirectional vision system In order to acquire enlarged images (omnidirectional images - ODls), researchers have had to draw on using swivelling or multiple cameras to capture the entire scene. However the mechanical control and high cost in computation time to obtain panoramic images are the major drawbacks for real-time application. The single effective view catadioptric vision system is an effective technique to enhance the field of view. It affords a 360' view angle of the environment in a single image [3]. The system is composed of a mirror in combination with a conventional imaging sensor. For manifold applications, the images reconstructed from ODls must be geometrically accurate and have the same viewpoint. The study presented in [4, 5] underlines that few mirror shapes can overcome this constraint.
Figure 1: Omni-directional and panoramic cylindrical image.
979 A single viewpoint catadioptric camera maps a single point projected on the image to each point in the 3D space (visible). Since the circular image directly acquired by the omnidirectional camera is naturally distorted, to view an entire image it is common to project the image on a cylinder surrounding the rotating mirror; the resulting image is easy to interpret. The system presented in this work employs a hyperbolical omnidirectional camera (hyperbolical mirror and CCD camera). As it was mentioned in Section 1, the study case involves a 6-leged mobile robot SIL06 [1] and the vision system placed on board it. The circular omnidirectional image and its cylindrical panoramic image acquired from the catadioptric camera are shown in Figure 1. The extraction of metric information from 2D frames is required for imaging techniques and it is an important task since the image formation model is determincd by the shape of the mirror [6]. By using a t1exible technique to calibrate omnidirectional cameras introduced in [7], it is possible to calculate the mirror profile errors, and then to compute the model of the catadioptric system, see Figure 2.
Figure 2: Illustration of omni-directional system pre-calibration.
3. Structure from motion method: motion field estimation The estimation of the motion field (optical t1ow) is generally measured from the analysis of instantaneous changes in the brightness values at any image pixel point. For these calculation it is usual assumed that image contrast is conserved throughout the image and the brightness changes are caused by the relative motion of the camera and the 3D scene. Given a brightness function J(x, y,t)
980
vJ
at a pixel position (x, t), time t and displacement vector x' conservation condition can be written as [8, 13]: this ｢ｲｩｾｨｴｮ･ｳ＠ l(x, y, t) = l{x + V xt, y + Vyt,O) (eq.(l». The methods to solve a differential equation of (I) (with respect tot) are based on the local smoothness assumption; they fit the tlow field to a constant velocity model V over a small neighborhood. Differential method assumes that the global smoothness of the brightncss changes in the images. Nevertheless, in catadioptric sensors the nonlinearities of the retlective surfaces cause affine transformation on the image contents. Consequently, the estimation neighborhood must be limited to prevent the violation of the global smoothness assumptions caused by deformations.
v(v
Figure 3: Indoor scene: segmentation of the manipulator and two objects (clusters: small & large movement) approaching the robot in radial direction.
Figure 4: Outdoor scene: segmentation of the manipulator and two objects (clusters) that move around the robot (circular motion). In order to "beat" the welI-know aperture problem, in this work a Robust Optical Flow method must be implemented [9, 10]. The studied method is presented in [11, 12], where the authors introduced an effective technique to find a robust solution based on solving a system of over-determined linear equations with all the data matrices containing both noise and outliers. The proposed
981
algorithm uses a robust regression method called the Least Median of Squares Orthogonal Distances (LMSOD) in conjunction with the Weighted Total Least Squares method (WTLS). The algorithms were modified for omnidirectional images, the experimentation results of outdoor and indoor images sequences are respectivcly shown in Figure 3 and Figure 4.
4.
Analysing human-robot interaction: study case six-legged robot's manipulator
In the natural human behavior normally the motion information from others' actions is utilized. This information is received, classified and interpreted, and later it infers in their desires and intentions [15]. Focusing in this knowledge, this paper analyzes human-robot interaction.
/I ::> Critical Angular / l/'" Region
Critical Radial Region
Critical Region (CR) = {Critical Radial Region, Critical Angular Region} Region
Not-Manipulator Region
Figure 5: Linguistic variables representation. The motion information is the detected optical flow when the manipulator/robot or any object is in movement, within the robot field of view. Because the fact that catadioptric system is able to acquire 360 around the robot, larger object surfaces can be perceived and the duration of object appearance is also superior. 0
982
The communication of human beings is described through linguistic terms and not in metric values. Fuzzy logic allows defining behavior decision rules in these terms, which simplify expert knowledge encoding of a generic mechanism [14]. A hierarchical fuzzy decision making policy is adopted. Whereas the higher level corresponds to the decision of what manipulator behavior must be accomplished, and the lower levels correspond to decisions about the world state (robot' environment) for each moment. The linguistic variable and image features are represented in Figure 5. Since the central region includes the body of the 6-legged robot it will be seen as the Prohibited-Region. To build the Critical Radial Region for manipulator movement, both the position in image coordinates of the fixed manipulator joint's (JI) and the maximum radius of manipulator motion field Rmfm are considered. The segmented motions are characterized by the centroid C; (r, e) of the distance between RL and RH . (Lower and higher radius from ｾ｝＠ which are image center 0). The Critical Angular Region is limited for ｛ＭｾＬ＠ collected from manipulator mechanical constraints. The first level in fuzzy making-decision identifies the localization of the manipulator and the n-moving clusters. The inputs or linguistic variables are the b
Euclidean Distance (polar coordinates)
L; = ｦｾ＠
dr 2 + r2 (da)2 between
JI
a
and RH, and the position of Ci. The corresponding term sets are Tl = {lower than Rmfm, similar to Rmfm, higher than Rmfm} and T2 = {inside CR, outside CR}, the outputs 01 = {Manipulator, i-cluster}.Then a process of characterization and classification of the 01 motion is performed. The n-clusters Ci positions are grouped in left and right hemispheres (respect to 0° horizontal line). In order to obtain a representative motion of each hemisphere (ML, MR) in images discs, i-clusters are weighted Wi according its values Li, and the magnitude
(circular and radial components) and direction
Vi
rpi
of its velocity.
Finally, the high level selects the optimal decision (action) that better satisfies the goals and constraints of the system. In this case the purpose invol ves a natural interaction with humans. As inputs of this level the properties of M L, M R,
ｻｗｩＧｾ［＠
rpJ and C
j
are utilized. And its term sets T3={lower,
medium, higher}, T4={go-slow, go-normal, go-fast}, T5={close-distance, medium-distance, far-distance}. The possible decision studied in this research allows the robot the following actions 02= {go-left, go-right, go-ahead, stop}, and the velocity response is a result of representative weight W; .
983 In the experimentation period an 8 images sequence was studied (Section 3), the results in the initial level reach a 95% of successes, and the 5% of failures coincides with non-common but possible situations, where humans are placed next to the manipulator (CR) or when an occlusion is presented. The obtained results in the high level shows the ability to deal with large and small movements, because of the fact that the classification process weights all cluster according it dimensions velocity vector and it centroid position. This ability also can be used to deal with non-rigid object, as it is with human beings behaviour, because human beings can move their entire body or only a part of them, see the illustration in Figure 6.
(a)
(b)
(c)
(d)
Figure 6: Correct detection of motion dimcnsion and optimal decision (a) goright and (b) Stop. Incorrect making-decision, due to the analysis of occlusions (b) and non-rigid objects has not been studied (d) go right, ERROR-human is violating CR constraint.
5. Conclusions This paper has underlined the importance of motion information in human-robot interactions. The learning model using motion information provides a robot with efficient mechanism to obtain properties of moving objects (humans, obstacles) and to establish natural interactions with a human.
984 The approach presented in this paper is a powerful methodology to analyze human-robot interaction to establish a natural communication. These prior results indicates that it is possible to provide the robot knowledge concerning to objects motion information of its environment. Furthermore, the omnidirectional system permits capturing of much more information for interpretation tread; thereby a promising cognitive model could be developed, where robot has the capability to predict or to be aware of the moving-objects close-far to it, and to launch a friendly and natural communication with humans. References 1. P. Gonzalez de Santos, J. A. Cobano, E. Garcia, J. Estremera and M.A. Armada, "A six-legged robot-based system for humanitarian demining missions", Mechatronics,Vol. 17, pp. 417-430, 2007 2. A. Agah. Human interactions with intelligent systems: research taxonomy. Computers and Electrical Engineering, Vol. 27 (1):71-107, 2001 3. Y. Yagi. Omnidirectional sensing and its applications. IEICE Trans. In! Syst. 82(3):568-579,1999. 4. S. Baker, S.K Nayar. A theory of single-viewpoint catadioptric image formation. Int. 1. of Computer Vision 35(2): 175-196, 1999. 5. T. Svodoba, T. Pajdla. Epipolar geometry for central catadioptric camera. Int. J. of Computer Vision, 49(1): 23-37,2002. 6. J. Gaspar, C. Decca, J. Okamoto, J. Santos-Victor. Constant resolution Omnidirectional Cameras. OMNIVIS'02, 2002. 7. D. Scaramuzza, A. Martinelli, R. Siegwart. A flexible Technique for Accurate Omnidirectional Camera Calibration and Structure from Motion. Procc. of 4th IEEE Int. Can! on Computer Vision Systems (ICVS 2006), 2006. 8. B. Hom, B. Schunck. Determining optical flow. Artificial Intelligence Vol. 17: 185-203, 1998. 9. M. Black. The robust estimation of multiple motions: Parametric and piecewise-smooth flow fields. Computer Vision and Image Understanding, Vol. 63 (1):75-104,1996. 10. K. Schilndler, D. Suter. Multibody structure and motion with outliers. In Proc. CVPR, 2005. 11. A. Bab-Hadiashar, D. Suter. Robust optical flow computation, HCV Vol. 29 (1): 59-77,1998. 12. F. de la Torre, M. Black. Robust principal component analysis for computer Vision. ICCV 2001. 13. J.L. Barron, D.J. Fleet, S.S Beauchemin. Performance of optical flow techniques, HCV, Vol. 12 (1):43-77,1994. 14. L.A. Zadeh. Fuzzy sets a basis for a theory of possibility. Fuzzy Sets and Systems, Vol. (I): 3-28,1978.
SIX DOF SENSORY SYSTEM FOR THE FORCE-TORQUE CONTROL OF WALKING HUMANOID* MILAN KV ASNICA
Tomas Bata University in Zlin, Faculty of Applied Informatics Nad Stranemi 4511, CZ-76005 Zlfn, Czech Republic E-mail: [email protected]
This paper focuses on current state of the art in sensory equipment for walking robotic systems and walking humanoids oriented on assistive technologies, military, security and rescue robotic systems. Described sensory systems are based on modular design that measures axial shiftings and angular displacements is a part of many sensory systems in robotics and human-machine interface applications. This is done by means of a square or annular CCD, or a set of four linear PSD elements, and four laser diode rays or planes, creating the shape of a pyramid. The positions of four light spots on the CCD or eight light spots on the PSDs are processed to produce three axial shiftings and three angular displacement values.
1. Introduction
The substance of six-component sensory system with the utilization for walking robotic systems is based on the shape of the 3D pyramid consisting of four edges created by light rays (respectively walls of light planes from structured light). The intersection of the laser light rays is forming the apex of the pyramid. The (square) basis of the pyramid is created by means of the 2-D Charge Coupled Device (or Position Sensitive Device, or CMOS) array and is used like a floating coordinate frame x,y. Intersection of four light rays with the 2-D CCD array in the basic parallel position between both two flanges results in a square shape with four light spots in the corners. Intersection of four light rays against the 2-D CCD array under acting of axial shiftings and angular displacements generally results in a trapezoidal position of the light spots. The light rays pyramid shape is coupled with the outer flange and the 2-D CCD array is coupled with the inner flange.
The financial support from the grant "Vyzkumny zamer MSMT 7088352102" in the part "Sestislozkova informace pri mefenf statickych a dynamickych vlastnostf anizotropnfch materialu" is gratefully acknowledged.
985
986 The six degrees-of-Freedom (DoF) motion (three axial shiftings and three angular displacements) between two flanges is sampled by means of the unambiguous trapezoidal light spots position, see [2], [3], [4]. Simple modular construction enables low cost customization, according to the demanded properties: A -stiff module of two flanges connected by means of microelastic deformable medium; B -compliant module of two flanges connected by means of macroelastic deformable medium; C -the module of the 2-D CCD array; D -the module of insertion flange with basic light sources configuration and focusing optics; F -the module of the plane-focusing screen; H -the module of the optical member for the magnifying or reduction of the light spots configuration. The problem of the customization of six-DoF sensory systems according to the enhanced accuracy and operating frequency of the scanning of the six-DoF information is possible to improve by means of the module of insertion flange with the configuration of light sources with strip diaphragms, creating the light planes with strip light spots and by means of the module of the single or segmented linear or annular CCD or PSD elements with higher operating frequency, respectively using the module of two, parallel working, concentric CCD annulars with higher reliability.
Figure 1. Six-Component Force-Torque Transducer.
The explanation of the activity of the force-torque transducer: Laser diodes 1 emit the light rays 2 creating the edges of a pyramid intersecting the plane of the 2-D CCD array, here alternatively the focusing screen 8 with light spots 3, see Figure 1. The unique light spots configuration changes under axial shifting and angular displacements between the inner flange 5 and the outer flange 6 connected by means of elastic deformable medium 7. An alternatively inserted optical member 9 (for the magnification of micro-movement, or the reduction of macro-movement) projects the light spots configuration from the focusing
987 screen onto the 2-D CCD array 4. Four light rays simplify and enhance the accuracy of the algorithms for the evaluation of the six-DoF information. The algorithms for the computation of three axial shiftings and three radial displacements values is based on the inverse transformation, see [11] of the final trapezoidal position of four light spots related to the original square light spots position in the plane coordinate system XCCD, YCCD on the 2-D CCD array. This algorithm determines the relative location and orientation of a floating 2-D coordinate system against a fixed 3-D coordinate system corresponding to the apex of the pyramid shape, or contrary. The information about three axial shiftings and three angular displacements is sampled and converted according to a calibration matrix to acting forces Fx, Fy, Fz and torques Mx, My, Mz, creating the vector F = [Fx,Fy,Fz, Mx,My,Mz]T.
2.
The Design of The Calibration Equipment for the Six-Component Force-Torque Sensors
The stability of control system of walking robots is under the influence not only static but mostly dynamic properties of acting forces F = [Fx,Fy,Fz, Mx,My,Mz]T from the six component force-torque sensors. For the calibration of the six-component force-torque sensory systems are used two different types of calibration equipments. The first, known variant is destined for static calibration of force-torque sensors. This calibration equipment consists of six independent loading corridors for every force-torque component separately, see Figure 2. Here the acting forces Fx, Fy, Fz and torques Mx, My, Mz, create the vector F = [Fx,Fy,Fz, Mx,My,Mz]T. The vector D = K.F of axial shiftings and angular displacements on the output of the six-component sensor enables the evaluation of the vector F = K(-I) .D, where the K is the characteristic matrix and the K(-') is the compensation matrix. The calibration matrix describes the properties of the elastic deformable medium in the force-torque sensor. The aim of the calibration equipment is to determine the interference of every forcetorque component to remaining components and to determine the linear part of the working characteristic. The first variant of the calibration equipment, see [7] is destined for the static calibration which is sufficient mostly for force-torque sensory system with longer response time. Some elastic deformable medium in the force-torque sensory system are suitable only for the sampling of force-torque process with low frequency, because some materials of the elastic deformable medium interfere significant to remaining components.
988
Figure 2. Calibration Equipment for Static Loading (from STU Bratislava).
The second new variant of the calibration equipment, see [10] is destined for the sampling of the six-component force-torque sensors at dynamic loading. In this calibration equipment are used the one-axis resistance strain gauges for the measurements of the force-torque components at both static and dynamic loading. The dynamic foree-torque loading is acting by the use of the pressing rods moved from the mechatronic, hydraulic or pneumatic actuators against the basic points of the squirrel cage of the calibration equipment, see [5]. This enable the measurement not only the static, but as well the dynamic characteristics of the six-component force-torque sensor for more components simultaneously. The measurement frequency of dynamic calibration equipment is restricted by the maximal working frequency of strain gauges of 2KHz, but this is sufficient for the activity of walking robots. -Fz
Figure 3. Calibration Equipment for Dynamic Loading (from UTB Zlin).
989 In Figure 3 is depicted the design of mechanical parts of the squirrel cage between actuators and the strain gauges: 1. Six-component sensor, 2. Axial lever arm, 3. The stiff cradle, 4. The grapple, 5. Contact surface of the force-torque acting, 6. Contact surface for the axial x,y,z torque acting, 7. One-axis strain gauges unit for the axial x,y,z force acting, 8. Strain gauges units for the angular torque acting.
Figure 4. The Six-Component Force-Torque Transducer mserted between the Shoulder and the Shaft of Humanoid Body.
3. The Control of the Walking Humanoid Stability The six-component force-torque transducer inserted between the shoulder and the shaft of humanoid body, see [11] for the control of the stability is depicted in Figure 4. The control of the walking humanoid's stability is parallel checked by the simple optoelectronic exteroceptive sensor for the control of a dynamic equilibrium, see [8]. This simple optoelectronic exteroceptive sensor located in
990 the position of the walking robot's head is connected with the shoulder by two DoF joint servos.
4. The Control of the Walking Leg Dynamics Universal, low cost, intelligent modular sensory systems enable to evaluate a humanoid's hand or leg dynamics while in motion, see [6] is depicted in Figure 5. A part of the artificial leg consists of the joint 10 connecting a shin with a foot 11.
Figure 5. Six -Component Force-Torque Sensors Mounted in Humanoid's Leg and the Range-Incline Finder Built iu the Heel.
The motion of the joint 11 is controlled by means of the six DoF information gained from two six-component sensors. The joint's 10 drive transmission is switched by means of the coupling muff 9 in order to control the dynamics of the motion. The six-component information about the leg's dynamics processed from two force-torque sensors enables us to use the drive power intelligently, even to convert the damping of the joint 10 motion for energy recuperation into the battery. The joint 13 connects the foot 11 with the toes part 14. The angular
991
displacement, (here for example a, p), of the joint 13 is used for the accommodation to the ground's incline 12a, 12b, according to the information from the range-incline finder. The ground's incline under the artificial leg is scanned by means of the range-incline finder mounted in a heel, see Figure 5, consisting of modules A, C, D, H. The light spots 3 from the light beams 2 on the ground 12a, 12b create the configuration scanned by the 2-D CCD alTay. The processing of this information enables us to evaluate the incline of the ground in two mutual perpendicular planes. Real-time algorithms are suitable for the single cheap microprocessor. A microprocessor based signal processing as an indicator of the ground's incline helps the humanoid to keep the stability.
5. Conclusion The modular design for six-component sensory system presented here enables easy customizing for wide variety applications in walking humanoids. Various combinations of the modular components enable tailoring of the sensory system properties including the use of the haptic interface for applications such as: detection of micro elastic or ｭｾ｣ｲｯ･ｬ｡ｳｴｩ＠ deformation, active compliant links, multi-DoF hand controllers, signature scanners, keyboards for blind people, tactile sensors, and range finders-positioners. In general, this modular design concept allows the maximization of service life because of ease of repair and the use of modular components for various types of sensors and the customization for a wide variety of design requirements. For example various levels of resolution and operating frequency, enhanced demands for safety and reliability in space robotics and medical use, and low cost design for manufacturing. In conclusion, introduced six component sensory system enables to built walking robots appropriate for experiments in the field of assistive technologies, human-machine interface, walking platforms for military, security, antiterrorist and rescue robotics and for the swarm robotic systems.
References 1.
2.
Dobrocka E., Geometrical Principles of the Monolithic X-Ray Magnifier, Journal Appl. Cryst., 24, 212-221, (1991) Hirzinger G., Dietrich J., Gombert J., Heindl J.,Landzettel K.,Schott J., The Sensory and Te1erobotic Aspects of Space Robot Technology Experiment ROTEX, Proceedings of the International Symposium on Artificial Intelligence, Robotics and Automation in Space, Toulouse, Labege, France, (1992).
992 3.
Kvasnica M., Intelligent Sensors for the Control of Autonomous Vehicles, Proceedings of the 6th International Conference and Exposition on Engineering, Construction and Operation in Space and on Robotics for Challenging Environments - Space and Robotics 98, Albuquerque, New Mexico, USA, (1998). 4. Kvasnica M., Improvement of Positioning Accuracy in Multi-Pod Parallel Structures", ASCE Multi-Conference on Engineering, Construction, Operations, and Business in Space and on Robotics for Challenging Situations and Environment "Space and Robotics 2002, Albuquerque, New Mexico, USA, (2002). 5. Palka M., The Design of the Calibration of the Six-Component ForceTorque Sensor, Diploma Thesis, Tomas Bata University in Zlin, Czech Republic, (2007). 6. Kvasnica M., Modular Force-Torque Transducers for Rehabilitation Robotics, Proceedings of the 6th International Conference on Rehabilitation Robotics ICORR'1999, Stanford University, (1999). 7. Sasik J., Multi-component Force-Torque Sensors Calibration Methods for Robotics Application, Strojnicky casopis, 38, No.6, Bratislava, Slovakia, (1987). 8. Kral E., Simple optoelectronic exteroceptive sensor for the control of a dynamic equilibrium of a walking robot. Proceedings of the 11th International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines, Coimbra, Portugal, (2008). 9. Kvasnica M., Algorithmic Approach for the Sampling of Six Degrees of Freedom Information Using Floating 2-D Coordinate Frame. Proceedings of the International Workshop on Robotics and Mathematics ROBOMAT 2007, Coimbra, Portugal, (2007). 10. Kvasnica M., The Design of the Calibration Equipment for the Six Component Force-Torque Sensor. Proceedings of the International Congress Mechatronics and Robotics MiR 2007, Izdatelstvo Poligraficeskij Komplex Lenexpo, Saint Petersburg, Russia, (2007). 11. hup:llwww.megarobot.net
kheNose - A SMART TRANSDUCER FOR GAS SENSING JOSE PASCOAL, PEDRO SOUSA and LINO MARQUES
Institute Jor Systems and Robotics University oj Coimbra 3030-290 Coimbra, Portugal * E-mail:{zeJranc.pvsousa.lino}@isr.uc.pt www.isr.uc.pt
This paper describes a Smart Transducer, inspired by the IEEE 1451 Standards, for olfactory sensing. The device is composed by five different types of gas sensors, three anemometers and one temperature and humidity sensor. This system senses the environmental conditions, process the acquired information and send it to a Khepera III mobile robot through an 12 C bus. An overview on the implemented real time software and data structures associated with each sensing node will be made.
Keywords: smart transducer; smart sensor; gas sensing; mobile robot
1. INTRODUCTION
This paper presents a smart sensor for measurement of the gas concentration and airflow, inspired by the IEEE 1451 Standards. A smart sensor is a sensor version of a smart transducer, which will convert a physical, biological or chemical parameter into an electrical signal! and then preprocess this signal into standardised data before sending it to a controller or other system component. While IEEE 1451 came about to serve the needs of the sensor and measurement industries, the Transducer Electronic Data Sheet (TEDS) concept has profound implications for other industries. 2 The TEDS stores manufacture-related information for the transducer(s), such as manufacturer identification, measurement range, accuracy, and calibration data, similar to the information contained in the transducer data sheets normally provided by the manufacturer. 3 The information, belonging to the TIM (Transducer Interface Module) will be sent through a serial bus to the Network Capable Application Processor (NCAP). The basic TEDS contains the minimum required information that characterises a transducer. It
993
994
Fig. 1.
System's organisation of the kheNose Smart Transducer.
shall be comprised of the Manufacturer ID, Model number, Version letter, Version number and Serial number in a total of 64 bits is included for each node of the Smart Transducer. 4 Previous works from Postolache and Ramos,5,6 used gas, temperature and humidity sensors for air quality monitoring, Pard0 7 developed a gas measurement system based on IEEE1451.2 standard and Wobscha1l8 a wireless gas monitor with IEEE 1451 Protocol. The work presented here expands the previous works since it will be used in mobile platforms, using thermal anemometers for odour tracking and specially for human injuries prevention. The developed system will be used in hazardous environments, specially in firefighting applications where the air is contaminated, monitoring and acquiring information about that environment and send it to an external control unit who will inform the firefighters about the conditions they will face. To overcome the fact that the sensors will be easily poisoned, a module, with self-identification, is the solution to this problem and to prevent a higher cost of the overall system and a easier and faster way to keep the system permanently working. Fig. 1 depicts the implemented system.
2. DEVELOPMENT This paper presents the development of a Smart Transducer which includes five gas sensors, three anemometers for air flow estimation and one sensor capable of measuring the temperature and the humidity of the environment. It is able to actuate four output ports for general purpose actions. Previous research in this area, like RoboNose from University of Coimbra,9 worked with gas sensors to gather olfactive information, specially gas concentration measurements. The goal was to upgrade it and develop a Smart Transducer, following the IEEE 1451 Standards Smart Transducer Interface for Sensors and Actuators guidelines, modular, plug and play and capable of including
995
(a) kheNose main board and modules location. Fig. 2.
(b) kheNose sensing modules.
kheNose overall system.
calibration information, to firstly interact with a Khepera III mobile platform lO and in the future, to be compliant with other devices that will be used by the GUARDIANS projectY 2.1. The kheNose
The kheNose (Fig. 2(a)) is a device developed to sense olfactive information through the use of gas sensors (Gl-G5), anemometers (Wl-W3) and a temperature and humidity sensor (TH1). A Microchip dsPIC33F controller acquires all the analog and digital information from the sensors, processes that data and sends it to the Khepera III KoreBot Extension board. lO This extension board supports several communication protocols, like 12 C, used to physically connect the kheNose to the Khepera III. The system is composed by six TIMs: An eCO, three thermal anemometers, and two eNostrils. The eCO and the anemometers are single channel IEEE1451.4 compliant boards and the eNostrils are double channel boards. All the functions related with the transducers, namely signal conditioning, data acquisition and processing and calibration management are performed by the kbeNose board. The calibration data for each sensing module is stored in a local EEPROM located in the module. 2.2. The kheNose Modules
kheNose is made of sensing modules. This approach was made since gas sensors get poisoned very easily and in this way, the maintenance, the modification and the upgrade will be carried without major difficulties. The following sections will describe kheNose modules (see Fig. 2(b)).
996
2.2.1. eNostril A setup with two different gas sensors (see Fig. 2(b), (A)) from Figaro Engineering is used to make the eNostril. It is modular, since we are able to have different sets of different sensors for a broader sensing capability. Since the module has two sensors, the TIM has two channels that were configured in the EEPROM. This I 2 C EEPROM holds the Basic, the Channel and the Calibration TEDS's from the sensors that composes the eNostril.
2.2.2. thS cale Thermal anemometers (see Fig. 2(b), (B)) are used in the system for measuring the air flow and to track odour plumes, with NTC thermistors from EPCOS. Since the hardware is previously calibrated, this board has an EEPROM, as the previous modules, with the corresponding TEDS's.
2.2.3. eTempHum eTempHum is another module (it can be seen in Fig. 2(b), (C) ) installed on the board which measures the temperature and humidity, using a Sensirion Inc SHTll sensor, that will help to compensate the data from the eNostril. This bus was driven using 2-wire interface, using two digital Input/Output (I/O) ports from the kheNose controller it is possible to measure the humidity with an absolute RH accuracy of ±3% RH and the temperature with an accuracy of ±0.4°C@25°C.
2.2.4. eGO The eCO board can be seen in Fig. 2(b), marked with (D). This board is composed by a Carbon Monoxide sensor from Figaro Engineering (TGS2442) and an EEPROM with the respective TEDS. The goal for this board is to sense the monoxide carbon in the test area.
2.3. Actuators Four digital output ports can be switched through an I 2 C register, to control the state of a pump, an electronic valve or other actuators for other purposes.
3. kheNose SOFTWARE ORGANISATION The developed software for the Smart Sensor, as referred before, runs under a RTOS, a real time kernel operating system. This is a pre-emptive real time operating system that allows software to perform the required actions in
997
real time. Since kheNose needs to process several actions at the same time, this type of operating system is an important advantage. Each action is performed by a specific task that has it is own running period, when it is a periodic task, or it is own processor time, in interrupt based tasks. The main function initialises all the important modules (ADC, I 2 C and RealTime Scheduler). Fig. 3 shows the RTOS existing tasks, as well as the data flow between sensor modules, tasks, buses and actuators. The blocks representing the tasks have their own function, as described below: Process Task: this is the main task, responsible for configuring and initialising the hardware modules, the TEDS and the calibration data, as well as ensuring the program control and stability. Acquisition Task: this task performs data collection of ADC channels from gas sensors and anemometers, dealing with timing constrains and channel selection. Acquired data is send to a queue, waiting for some other task to read them from it. SHTxx Task: this task acquires and stores in a queue the temperature and humidity values from a proprietary 2-wired serial bus. Sensor Detection Task: this task detects the presence of a sensing node, reads the TEDS data and send this information to the Process Task that will call the respective function, according to the sensor module connected. This is one of the most important tasks concerning the Smart Transducer concept. 12 C Task: the communication task, responsible for deal with data requests and sends. Whenever a request arrives, it accepts it, sends the information about the required data to the Process Task, waits for data ready and finally sends the response. In each task block corner can be seen the priority of the respective task. The priority of the tasks is proportional to the corner number, so the tasks with bigger importance like the Acquisition or the I 2 C communication gets a bigger value. Looking at the priorities, it is shown that when an I 2 C event occurs, the task should be executed as soon as possible. In other hand, the SHTxx Task (task responsible by data collection from the temperature and humidity sensor) has the lowest priority due to slow changes of these physical characteristics. Another important block for the smart transducer is the Data Container which holds all the information that characterises the kheNose, including the TEDS from all the sensors.
998
Fig. 3. kheNose's software organisation. Representation of the tasks block, the tasks priorities (at lower right corner of each task block), the data flow between functions and the main bus.
3.1. kheNose TIM As mention before and shown in Fig. 1, kheNose has a specific TIM's organisation. TIM is responsible for analog conversion and signal conditioning, as well as communication and data transfer functionalities. What concerning with smart sensors, TIM also storage the TEDS information, mostly Basic and Channel TEDS. They were organised as presented in Table 1 and 2. Referring to the IEEE 1451.4, the Basic TEDS information shall be comprised of 64 bits and the Manufacturer ID was set to 16382 since this is the identification number normally used for academic purposes.1 2
Table 1.
kheNose implemented Basic TEDS.
Basic TEDS
Bits
Type
Data
Basic TEDS length Model Number Manufacturer ID Version Letter Version Number Serial Number Total
16 15 14 5 6 24 80
int int int char int int
80 3 16382 A 3
062008
999
For instance, for the Figaro TGS2600 sensor Channel TEDS, the physical units are set to 1, since voltage is being measured instead of concentration, the ranges depend on the implemented circuitry used in the kheNose and the extra data belongs to the kheNose and to the TGS2600 sensor. 13 Table 2.
kheNose implemented Channel TEDS.
Channel TEDS Channel TEDS length Lower range limit Upper range limit Physical units Unit warm up time Uncertainly Channel model significant bits Channel sampling period Manufacturer's identification Model number Total
I
Bits 32 32 32 16 32 32 16 32 128 128 480
I
Type
Data
unsigned long int float float unsigned int float float unsigned int float string string -
480 (60 bytes) 0.660 3.143 1 600 8.07e-4 12 1.0e-4 Figaro TGS2600 -
4. CONCLUSION A kheNose prototype was implemented and basic tests like plug and play detection and data acquisition were performed. This system is being employed in Khepera III mobile robots to develop and test navigation algorithms involving chemical detection in small scale scenarios. IEEE 1451.4 compliant sensing modules were implemented and the respective TEDS for gas sensors and thermal anemometers were designed and successfully implemented (see Fig. 4). In the future, the capabilities of the kheNose system will be improved in order to implement highly advanced smart sensing systems that can be used in the normal size robots used in the framework of the EU GUARDIANS project (e.g. Era-Mobi 14 and Rescuer 15 ).a
ACKNOWLEDGMENTS This work was partially supported by the Portuguese Science and Technology Foundation (FCT/MCTES) by project RoboNose, contract POSI/SRI/48075/2002 and by project GUARDIANS contract FP6-IST045269. ahttp:www.guardians-project.eu
1000
Fig. 4.
Khepera III and kheNose with sensing modules.
References 1. IEEE Instrumentation and Measurement Society, IEEE standard for a smart
2. 3. 4.
5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15.
transducer interface for sensors and actuators - digital communication and transducer electronic data sheet (TEDS) formats for distributed multidrop systems The Institute of Electrical Engineers.(March, 2004). F. Rubinstein, S. Treado and P. PettIer, Industry Applications Conf., 2003. 38th lAS Annual Meeting. Conf. Record of the 2, 805 (12-16 Oct. 2003). E. Y. Sonag and K. Lee, IEEE Instrumentation and Measurement Magazine 11, l1(April 2008). IEEE Instrumentation and Measurement Society, IEEE standard for a smart transducer interface for sensors and actuators mixed-mode communication protocols and transducer electronic data sheet (TEDS) formats The Institute of Electrical Engineers. (March, 2004). O. Postolache, M. Pereira and P. Girao, InstT1tmentation and Measurement Technology Conf., 2005. IMTC 2005. Proc. of the IEEE 1 (2005). H. G. Ramos, O. Postolache, M. Pereira and P. S. Girao, MWSCAS '06. 49th IEEE Int. Midwest Symp. on Circuits and Systems 1, 177(Aug. 2006). A. Pardo, L. Camara, J. Cabre, A. Perera, X. Cano, S. Marco and J. Bosch, SensoTs and Actuators B: Chemical 116, l1(July 2006). D. Wobschall, Pmc. of IEEE Sensors Applications Symposium, 162 (2006). L. Marques, N. Almeida and A. de Almeida, Pmc. of IEEE Int. Conf. on Sensors 1 (20m). K-TEAM, K-TEAM Corporation Web reference, (2008). GUARDIANS Project, Group of unmanned assistant robots deployed in aggregative navigation supported by scent detection Web reference, (2008). J. Kim, D. Kim, H. Byun, Y. Ham, W. Jung, D. Han, J. Park and H. Lee, Talanta 71, 1642 (2007). Figaro, TGS 2600 Product InfoTrnation, tech. rep., Figaro. Videre Design, Era-mobi: Mobile robot platform Web reference, (2008). Robotnik Automation S.L.L., Rescuer Web reference, (2008).
SECTION-16 PERSONAL ASSISTANCE
This page intentionally left blank
FES-ASSISTED CYCLING WITH QUADRICEPS STIMULATION AND ENERGY STORAGE B.S. KSM KADER IBRAHIM, S.c. GHAROONI, M.O. TOKHI and R. MASSOUD
Department of Automatic Control & System Engineering, The University of Sheffield. United Kingdom Using quadriceps only to obtain smooth FES-cycling reduces the number of stimulated electrodes applied to a paraplegic. Therefore, the preparation process for FES-cycling become more comfortable and less time consuming. This paper discusses how elastic cables can be used to assist quadriceps FES-cycling to eliminate the dead points of the pedal cycle. Implementation of an energy storage device (elastic cable) has many desirable influences on FES-cycling, such as decreasing the number of stimulated muscles. This in tum reduces the number of FES surface electrodes, thereby saving excess energy that is released during the cycling. This subsequently reduces stimulation intensity and duration and minimizes muscle fatigue.
1. Introduction Spinal cord injuries prevent the affected person from walking, standing and moving freely. Functional electrical stimulation (FES) is the best way to restore movement of the working limb muscles. Cycling by means of functional electrical stimulation (FES-cycling) makes exercising easier, safer and more enjoyable for subjects with spinal cord injuries. It has been proved that FEScycling enhances the paraplegic's health to a large extent such as improvement in the cardiovascular function and muscle forces [3,4,5]. Neuroprosthesis cycling machines may differ in terms of their balance strategies (tricycles, stationary bicycles, recumbent bicycles, etc.), their function (indoor, outdoor), the number muscles stimulated, the type of FES electrodes (surface, percutaneus, implanted), and the control strategies used to control the exercise. To obtain steady FES-cycling most researchers have stimulated at least two muscle groups. A novel approach involving the stimulation of only the quadriceps to obtain steady FES-cycling stimulating knee has been proposed by Massoud,2007. The most challenging problem facing paraplegics during the preparation for FES-cycling is placing the surface stimulation electrodes in their proper positions. Since this plays an important role in the stimulation performance, the electrodes should be as close as possible to the corresponding nerves.
1003
1004
The more muscles involved in the stimulation process the more electrodes are required to be correctly positioned and the more time consuming this will be. To overcome the problem of repositioning a large number of electrodes at the beginning of each FES stimulation session, some researchers have used implanted electrodes placed very close to the corresponding nerve [6]. Due to the associated risks with surgical operations, a large number of patients do not choose the implanted electrode method. In this study elastic cables are used to assist quadriceps-FES-cycling to eliminate the dead centres of the pedal cycle. The two elastic cables are stretched and compressed as the crank revolves. With the correct positioning this may allow the legs to store elastic energy that can subsequently help overcome the dead centre of the pedal cycle. 2. Method 2.1 Energy Storage Device
An energy storage device can assist FES-cycling, and could possibility replace conventional auxiliary electric motor with the energy storage actuator, in order to increase cycling performance. Pons et al. (1989) demonstrated that it is possible to have the motor assist the pedalling during part of the cycle and retard it during another part of the same cycle. The motor resistance can be replaced with an energy storage device such as an elastic cable which releases its energy when assistance is needed. Gharooni et. al (2000) used a spring brake orthosis (SBO) as the gait restoration system in which stored spring elastic energy and potential energy of limb segments are utilised to aid gait. This energy is released as kinetic energy at the optimal time to provide the desired limb motion. Rasmussen et al. (2005) tried to eliminate the dead centres of the pedal cycle by providing the bicycle frame with a spring in their ergonomic bicycle. In this research elastic cables are used for this purpose. The elastic cable slack length, fixation points on the frame, and fixation points on the crank are optimized. A recumbent cycle with the elastic cables is shown in Figure 1.
1005
Figure 1: A recumbent cycle with elastic cables
2.2 Dead Points Each quadriceps for each leg plays two roles during the pedalling cycle: (1) pushing to speed up cycling and (2) resisting the speed when it exceeds the desired range. The main hindrance in cycling are the dead points, which are defined as thc two pedal positions at which the net moment referred to crank rotation centre is zero [2]. Figure 2 shows the knee angles and crank angle. A pedalling cycle of 360° has two so-called dead points, which on an ordinary bicycle fall when the pedal arms are close to vertical. These two points (dead spots) of crank angle, 0 occur at 0° and 180° as shown in Figure 3. The moment arms in these points will become zero.
1006
Right Ankle Figure 2: Knee angles and crank angle
Figure 3: Dead point
In these positions it is difficult for the human body to produce much crank torque because the tangential pedal force direction is perpendicular to the preferred force direction of the legs. Much effort has been invested into mechanisms that reduce or eliminate these dead points. One of the best known initiatives was the oval chain wheel marketed by Shimano in the 1980's. Rasmussen et al. [7] demonstrated an intricate mechanism bicycles for paraplegics that helps overcome the dead point by making sure that it does not occur simultaneously for the two legs. In this study another possible solution would be to allow the mechanism to store energy when the leg has maximum leverage and subsequently release the energy when the dead point is approached.
2.3 Overcoming the Dead Points In normal conditions, to overcome the dead points, the corresponding leg must push to move the crank in the cycling direction while the other leg is flexing to give a sufficient driving torque to overcome the situation. After passing the dead point the cycling speed begins to increase, and the corresponding leg starts to
1007
flex to overeome this inerease and to negate the influenee of its earlier extension. These important tasks of the knee flexors during the cycling process make them absolutely essential in cycling and researchers have used them as important elements in investigating FES-cycling for paraplegic. The knee flexors are not stimulated in this study, and the only way to reduce the crank cadence after the dead point is to apply another moment in the opposite direction. This can be achieved by arranging elastic cables as shown in Figure 4. The two elastics are stretched and compressed, as the crank revolves, and this may allow the legs to store elastic energy that can subsequently help overeome the dead points of the pedal cycle. The moment of force or torque equals the product of magnitude of force, F and the moment arm, d : Torque, M
= F.d
(1)
The torque depends proportionally on the moment arm or the distance between the pedal and the location on the elastic cable under the hip. The higher the moment arm, the higher the torque produced.
F Figure 4: Moment produced with elastic cables
1008
3. Results Basically there are four major phases in the cyclical legs motion to be considered. In each of the phase, the legs will be 90° away from their previous position. The presence of elastic cables helps to overcome the dead points by providing moments in the opposite direction for each phase. a) Phase One In this phase, the left foot would have moved to the furthest point in the cycle path while the right foot is now at nearest point. In this phase the moment arms for both legs are equal. The dead point can be overcome by adding more force on the left leg than on the right leg, refer to Figure 5a. b) Phase Two In the second phase the feet are separated in a top-down position, where the right foot is at the upper position while the left foot is at the bottom position. In this phase the moment arms for right the leg reaches high level while for the left leg it reaches lower level. If some force is given to right leg, this will help moment arm back to equal position as in Phase 3, refer to Figure 5b. c) Phase Three In this phase, the left foot is at the nearest point to the humanoid while the right foot is at the furthest point away. The dead point can be overcome with adding more force on the right leg than the left leg. The moment arms for both legs are equal as illustrated in Figure 5c. d) Phase Four The final phase will be the position where the humanoid has cycled for one full cycle. The left foot is now in the upper position while the right foot is at the bottom. In this phase the moment arms for the left leg reaches high level while for the right leg it reaches lower level. If some force is given to the right leg, this will help bring the moment arm back to equal position as in Phase 1 then it will repeat the same process for the next cycle refer to Figure 5d. From this cyclical legs motion it is seen that the two elastics are stretched and compressed which allow the legs to store elastic energy that can subsequently help overcome the dead points of the pedal cycle.
1009
Figure 5: Four major phases in the cyC\icallegs motion
4. Conclusion Exercises for rehabilitation of spinal cord injury patients include standing, walking and cycling induced by electrical stimulation. This study has focused on pedalling aspect of FES-assisted cycling. The paper has highlighted how elastic cables can be used to eliminate the dead points of the pedal cycle during a cycling exercise. The implementation of elastic cables shows that the dead point can be eliminated by providing moments of force in the opposite direction for each cycling phase. Future work will investigate the development of appropriate control strategy to achieve optimal stimulated pulse width with combination of energy storage device (elastic cables) for smooth cycling motion.
1010
Acknowledgement The financial supports of Higher Education Ministry of Malaysia and University Technology Tun Hussein ann, Malaysia are greatly acknowledged.
References 1. D. J. Pons, C. L. Vaughan, and G. G. Jaros, Cycling device powered by the electrically stimulated muscles of paraplegics, Med Biology Eng Computer, 27 (1),1-7, (1989). 2. J. J Chen, N. Y Yu, D. G Huang, B. T. Ann, and G. C. Chang, Applying fuzzy logic to control cycling movement induced by functional electrical stimulation, IEEE Trans Rehabilitation Eng, 5 (2), 158-69, (1997) 3. J. S. Petrofsky, and M. Laymon, The effect of previous weight training and concurrent weight training on endurance for functional electrical stimulation cycle ergometry, Eur J Applied Physiology, 91 (4),392-8, (2004). 4. M. Kjaer, S. F. Pollack, T. Mohr, H.Weiss, G. W. Gleim, , F. W. Bach, T. Nicolaisen, H.Galbo, and K. T. Ragnarsson, Regulation of glucose turnover and hormonal responses during electrical cycling in tetraplegic humans, Am J Physiology, 271 (1), 191-199,(1996). 5. P. D. Chilibeck, G.Bell, J.Jeon, C. B.Weiss, G. Murdoch, I. MacLean, E. Ryan, and R. Burnham, Functional electrical stimulation exercise increases GLUT-J and GLUT-4 in paralyzed skeletal muscle, Metabolism, 48 (11), 1409-13, (1999). 6. R. Massoud, Intelligent control techniques for spring assisted FES-cycling, PhD Thesis. The University of Sheffield, Sheffield, UK(2007). 7. S.T Rasmussen, M. Christensen,. M. Damsgaard, Zee, Ergonomic optimization of a spring-loaded bicycle crank,. 6 th World Congresses of Structural and Multidisciplinary Optimization, Rio de Janeiro, Brazil (2005). 8. S. Gharooni, M.O.Tokhi, B. Heller, The use of elastic element in hybrid orthosis for swing phase generation in orthotic gait. The 5th Annual Conference of the International Functional Electrical Stimulation Society, Alborg, Denmark, pp. 486-489, (2000). 9. T. A. Perkins, N. D. N. de, N. A. Hatcher, I. D. Swain, and D. E. Wood, Control of leg-powered paraplegic cycling using stimulation of the lumbosacral anterior spinal nerve roots, IEEE Trans Neural System Rehabilitation Eng, 10 (3), 158-64, (2002).
MODELLING AND SIMULATION OF SIT-TO-STAND IN HUMANOID DYNAMIC MODEL S. C. GHAROONI, M. JOGHTAEI, M. O. TOKHI Department of Automatic Control &System Engineering, The University of Sheffield, United Kingdom The goal of this study is to investigate torque profiles of each active joint in sit-to- stand (SiSt) mode arm free and with seesaw exercise machine. This paper considers the development of a dynamic humanoid model for SiSt demonstration. It consists of two main parts, the development of human model using Visual Nastran and the development of a controller for controlled movement of the system. A closed loop feedback finite-state PID control strategy is used. In order to reduce torque amplitude profile on each joint, a seesaw mechanism is realised to reduce the upper body load on lower extremity. In this method only knee joint torque in the range of stimulation muscle torque (50Nm) was adequate to perform sit to stand without relying on hip and ankle joints.
1.
Introduction
Rising from a seated posItIOn IS one of the basic and necessary actIvItIes performed by every human being. This movement is performed by a human being many times in a single day. For a person to lead an independent life it is imperative that he/she has the ability to rise from a sitting position. But there are severe limitations for certain individuals due to injury, old age, or certain diseases. Standing up to an upright position is a prerequisite for many activities of daily life such as initiation of walking, reaching high objects, making transfers and face to face interactions with others. In addition to functional benefits, regular standing and weight bearing may also provide important therapeutic benefits to individuals with spinal cord injury (SCI) [IJ. This research investigates to develop a mechanism for sit to stand such as seesaw to allow disable people with limited muscle power to achieve sit to stand. 2.
The sit-to-stand movement
Sit-to-stand (SiSt) seems to be a very simple task as it is performed by everyone. However, studying this in detail reveals that SiSt is one of the most mechanically demanding tasks. It is considered to be the prerequisite for gait. The SiSt movement has varied between different studies. Two separate phases
1011
1012
were identified in some previous studies. The phases are the initial forward trunk lean and the upward extension of the trunk. Most of the studies have described three separate phases of the SiSt movement, namely the initiation phase, the seat unloading phase and the ascending phase [2]-[3]. The rising phase is described by Riley et al [3] as the most mechanically demanding functional activity in daily activities. Rising phase is when the subject rises from seated position to standing position. Figure 1 shows components of the rising phase where the first diagram on the left is the initial position while the subsequent ones show each of the components Figure 1: Rising phase components listed in Table 1. Table 1: Sit-to-stand phases
Phase T
••
Activity I Forward Momentum
Cycle
0%-27%
Seat Unloading
Vertical Acceleration
27%-45%
Ascending
Deceleration
45%-73%
Balancing Body
73%-100%
Kralj et al [4] defined four phases of the SiSt movement. These are the initiation phase, seat unloading phase, ascending phase and the body balancing phase. The phases of SiSt movement may vary with age. The SiSt movement described in here is for a healthy person. In this study the SiSt trajectory for ac humanoid model is developed based on SiSt phases of a healthy person.
3.
The humanoid model
A realistic bipedal humanoid model was designed based on anthropometric data of human in full scale with human locomotion pattern to evaluate joints' power and torque profiles. Standard measurements of human body height and weight are given by Winter [5]. The model (Figure 2), based on simple geometrical
1013
shapes, as developed in the Visual Nastran (VN) environment, allow its motion control from VN as well as from MATLAB-SIMULINK. Figure 2 depicts a humanoid dynamic model with location of centre of mass of shank, hip and trunk. Each segment in the VN model has its own dynamic properties such as, inertia, centre of mass, angular and linear velocities in three dimensions. These parameters would be defined and fed in the model development stage. The main characteristics of the developed humanoid model are: total mass M=70 kg and leg length 42 cm, giving height L=42/0.24=175 cm [5]. Thus, considering the model height to be 175cm, each segment height was calculated. Table 2 shows the segments' weights for a total body weight of 70kg. The centre of mass and inertia are also Figure 2: Humanoid model considered in the model as further specifications of body segments. Table 2: Body segments dimentions
Segment
Foot length
Shank
Thigh
Head & Neck
Trunk
Weight (kg) Height (cm)
1.015 26.60
3.255 43.00
7.000 42.90
5.810 31.85
34.790 50.40
4.
Reference trajectory inpnt
The essential part of this study is to design realisable trajectory plan for the SiSt movement. It is important that in all the phases in the study the centre of gravity is projected over the foot. A desired angular trajectory for each joint was developed according to the defined joint angle of the model. Some commands from fuzzy logic toolbox of MATLAB were used to build the trajectories. Figure 3 shows the rising trajectories thus
Figure 3: Rising phase trajectories for knee, hip and ankle
1014
obtained for each joint. 5.
Controller
Finite state PID controller was developed in MATLAB-SIMULINK so as to track a set of reference trajectories for each joint, and to balance the trunk position throughout the rising up. The control strategy for the humanoid SiSt switches from one mode to another mode based on transition on reference trajectory in order to keep the trunk balanced during SiSt transition. Two PID controllcrs operate for each joint trajectory with different parameters. The controller gains for both PID controllers were tuned using the Ziegler-Nichols method. 6.
Results
Results obtained from the simulation of the dynamic human model are shown in Figure 4. From the results obtained, the model achieved its' final position in the rising phase after 1.5 seconds. Figure 4 also gives the desired trajectory and the measured output for the ankle, hip and knee joints. The ankle torque profile varies in the range of -90 Nm to 170 Nm. The largest torque value (163 Nm) occurs during the initiation of ankle dorsiflexion. The smallest value (-90 Nm) occurs during the initiation of ankle plantar flexion. The hip torque profile varies in the range of -150 Nm to 225 Figure 4: Measured data for ankle, hip and knee joint during SiSt Nm. The largest torque value (225 Nm) occurs during the initiation of lean forward component where the hip starts to flex forward. The smallcst value (-150 Nm) occurs during the initiation of recovery component where the hip starts to flex back to its' final position.
1015
The knee torque profile varies in the range of -100 Nm to 50 Nm. Peak knee absolute value torque (100 Nm) occurs during the initiation of knee extension. There are also some abrupt changes in the joints' torque, when the control system switches from the first controller to the second controller at time t = 1.15s. This value of torque is behind the ability of a disabled person muscle torque which has been produced by the use of electrical stimulation. The maximum torque produced by means of electrical stimulation on knee joint is around 50 Nm. In addition to that for this type of sit to stand using hip and ankle joint torques are needed which are difficult to produce via electrical stimulation in a disabled person. In the following section a seesaw technique is developed in order to eliminate hip and ankle joint torques. The only torque required for performing sit to stand in seesaw technique is the knee joint.
7.
Seesaw model
The function of seesaw in this study is to support humanoid body in the process of standing up and sitting down for SiSt and stand-to-sit (StSi) exercise. Moreover, it acts as a mechanical construction that constrains the position of the pelvis in such a way that it moves in the sagittal plane on a circular path about the seesaw axis. Figure 5 shows the complete model design, where the seated humanoid model and load are placed on the seesaw mechanism.
Figure 5: Seesaw model
The load will exert an anti clockwise turning moment of magnitude W LxML and the humanoid model will exert a clockwise turning moment of magnitude WHxM H. When WLxM L is greater than WHxM1h there will be resultant anticlockwise moment acting on the seesaw, such that the humanoid model will be lifted upwards as the load descends. (ML= distance from pivot joint to load's
1016
COM, M H= distance from pivot joint to humanoid's COM, WL= load's weight, WH= humanoid's weight). The load is placed at 1.2m from centre of the seesaw which is the same distance as humanoid model's centre of mass (COM) position, from centre of the seesaw. Rigid joint is used between the bar and load to ensure that the load is always fixed on the seesaw. Moreover, the feet of the humanoid model are fixed to the ground to avoid slippage and instability during its motion.
8.
SiSt and StSi Kinematics
Figure 6 shows sequence frames of SiSt seesaw movement. Initial phase of SiSt mode is the starting point before the model moves in which, the seesaw is at 0", while the knee angle is at _120" and heels are off ground. The humanoid model and load are in equilibrium on the seesaw. As the kuee joint goes to extend the seesaw rotates counter clockwise. Final phase for SiSt mode is the end of the model movement, where the humanoid model is able to stand when one end of the seesaw touches the ground. The seesaw angle is at 14.2", and the knee angle is at _26". The feet are flat on the ground.
Figure 6: Seesaws motion from initial to final phase
Conversely, this phase will be the initial phase for StSi mode.
9. Reference Knee Trajectory The humanoid model has been moved passively by the seesaw in order to achieve either SiSt or StSi mode as shown in Figure 7. The period of time needed to complete both phases is Figure 7: A complete cycle of SiSt and StSi vary but takes, on average, 1-3 seconds [6]. Here, the time taken to complete each mode is set to 2 seconds. 85kg load have been applied at one end
1017
of the seesaw, which indirectly lifts the other end where the humanoid model is seated. The 85kg load is chosen, due to its capability to lift the humanoid model within practical time. In StSi mode, 50kg load has been applied on the seesaw. This weight is chosen, since it is lighter which indirectly helps the humanoid model to sit down easily with the help of gravitational force.
10. Testing and Verification The performance of this control scheme has been tested by running the controller continuously in 10 cycles. The overall results have shown that the system is in stable condition where the actual output still follows the desired input Figure 8: SiSt and StSi sequential motions Dashed line= closely. The performance Desired input, Solid line= Actual output result of the continuous motion is presented in Figure 8. However, the range of stability is only valid for the applied load in between 45-75kg. Outside this limit, the knee of humanoid model will begin to oscillate which lead to system instability. Furthermore, this control system is only applicable to the already setting reference trajectories; for other trajectories the system will be unstable. In the real application, 50Nm is the maximum torque value that can be applied with electrical stimulation, and higher values might be harmful to patient's muscle. Thus, a saturator has been implemented in the control scheme.
11. Discussion and concluding remarks From the graph of sit-to-stand, (SiSt) knee trajectory as shown in Figure 9, the knee torque is increased from zero to 1l0Nm in first 0.4 seconds, and then dropped to -60Nm at 1.05 seconds. The torque magnitude is changing due to the signal sent from PD controller output, which comes from the error between the desired input and the actual output. During the standing position, the knee torque is maintained within 25Nm in order to keep it in that position. In the complete cycle trajectory with saturation limit, the maximum torque value is 50Nm, while the minimum value is -50Nm. This is illustrated in Figure 10. Even though, the knee orientation angle does not smoothly follow the desired input. This is because of the implementation of the saturation limit which limits the actual torque needed by the knee.
1018
lWr----------------------,
, 0:2
OA
0.5
i.O
L2
1.4
'1 $
111
;2,0
22
Zll'
22
Tfme,sec
Figure 10: Knee torque profile and knee orientation angle in SiSt
Figure 9: Knee torque profile with torque limiter and knee orientation angle in a complete cycle
Generally, other joints such hip, ankle and toe, are moved freely during both modes. The movements of these joints are generated from the knee joints movement. Thus, it can be said that the knee joint is the only active joint, while others joints are passive.
References J. S. Petrofsky, and M. Laymon, The effect of previous weight training and concurrent weight training on endurance for functional electrical stimulation cycle ergometry, Eur J Appl Physiol, 91 (4),392-8, (2004). 2. Millington, P. J., B. M. Myklebust, and G. M.Shambes.: "Biomechanical Analvsis of the Sit-to-Stand Motion in Elderly Persons". Archives of Physical MedIcine and Rehabilitation, volume 73, pp 609- 617,1992 3. Riley, P.O., M. L. Schenkman, R. W. Mann, and W. A.Hodge.: "Mechanics of a Constrained Chair-Rise". Journal of Biomechanics, Volume 24, issue ], pp 77-85. 1991. 4. Kralj, A., Jaeger, R.J. & Munih, M.: "Analysis of standing up and sitting down in humans: definitions and normative data presentation." Journal of Biomechanics, volume 23, Issue 11, pp 1123-1138. 1990. 5. Winter D A, "Biomechanics and Motor Control of Human Movement", 2nd Edition. Wiley-Interscience, New York, 1990. 6. Baer GD, Ashburn AM (1995), Trunk Movements in Older SUbjects during Sitto-Stand, Archives of Physical Medicine and rehabilitation, Pages: 844-849. 1.
A NEW GRAVITY COMPENSATION SYSTEM COMPOSED OF PASSIVE MECHANICAL ELEMENTS FOR SAFE WEARABLE REHABILITATION SYSTEM T. NAKAYAMA*, T. ASAHI and H. FUJIMOTO'
Omohi College, Graduate School of Engineering, Nagoya Institute of Technology, Gokiso, Showa-ku, Nagoya, Aichi 466-8555, Japan * E-mail: [email protected]
This paper proposed a new gravity compensation system which is suitable for the lower extremity rehabilitation therapy. The proposed system can compensate the gravitational moments exerting on the leg joints perfectly using the passive mechanical elements. In contrast with the previous systems, the proposed gravity compensation mechanisms are wholly embedded in the link body, so that the system is safer and well-suited to the wearable equipment. The feasibility of the gravity proposed mechanisms is confirmed through the experiments using an actual equipment. The effectiveness of the system as a lower extremity rehabilitation equipment is examined through computer simulations for the bending and stretching exercises.
Keywords: Rehabilitation system; Power Assist system; Gravity Compensation; Mechanism Design.
1. Introduction
In step with the aging of the population, many people suffer from the heavy arthralgia on the legs. To alleviate their suffering, various kinds of prosthetic joints and replacement arthroplasty have been developed by now. Owing to these technical developments in medicine, many of the patients who could not walk by themselves become possible to recover the walking ability, nowadays. However, many of the patients who had the heavy leg disease accompany the muscle weakness frequently, so that they are impossible to walk by themselves after the operation in many cases. Therefore, the support devices which can compensate the patient's body weight arc highly required for the patients. To this end, many types of the gravity support devices and rehabilitation equipments have been developed. 1- 3 One type of them is the devices to lift
1019
1020
the patient's body from the ceiling, which is used in the hospital for the patients with heavy leg disease. 3 Another one is the power assist devices which make up for the poor athletic performance of the patients. 1 Those devices already work well in the clinical practices. However, there are some problems such that the former ones are considerably restricted the freedom of the movement, and the later ones are unsuitable for continuous use by the limit of the battery life whereas has a risk to harm the patients by some unexpected malfunctions. For the rehabilitation equipments, safety and the long term availability are essential specification. Therefore, it is preferable that the gravity compensation system is composed of the passive mechanical elements entirely.
. hI 1 link system Fig. 1.
multi link system Previous gravity compensation systems 4 ,5
As to the passive gravity compensation systems, there has been a long research history.6 At first, they were developed in the robotics to reduce the load exerting on joints and realize the smooth manipulation of the heavy robot. Shirata et al. proposed a simple power assist mechanism attaching a passive spring to the knee joint to achieve the smooth manipulation of the humanoid robots. 7 On the other hand, Morita et al. proposed an alternative mechanism 4 as shown in Fig.1 , which can compensate the gravitational forces perfectly and they apply it to the underactuated robotS and wheelchair. 8 Though these mechanisms are effective to compensate the gravitational effect, they may encumber the human motion and be unsafe if it is applied to the wearable equipment since the springs are barely stretched from the one link to another. In contrast with them, Herder et al. proposed a wheelchair with the gravity compensation system,9-11 in which the safety of the patients is secured by bearing off the springs from the arm link using the parallel linkage. Although its usefulness is proven in the clinical practice, this mechanism is not applicable to the wearable equipment.
1021
Thus, this paper proposes a new gravity compensation system to establish a safe wearable rehabilitation system. In the new mechanism, the drive members are embedded completely inside the link itself, so that it never disrupts the motion of limbs and is safe to use as a wearable rehabilitation systems.
2. Proposed gravity compensation system In this section, we construct a new gravity compensation system. At first, the gravity compensation system for the inverted pendulum is considered. When the pendulum is inclined at the moment mgf! sin is exerted around
e,
e
, fi:l.cdtonoor
Fig. 2.
Proposed gravity compensation system for 1 link system
the joint by gravity. Therefore the gravity compensation system must generate the counter moment -mgf! sin e around the joint by synthesizing the restoring forces of springs. In the proposed new gravity compensation system illustrated in Fig.2, the gear attached to the joint shaft is connected to the pulley 1 in the gear ratio of 1:2. Thus, when the joint rotates by [rad], the pulley 1 rotates by -e /2 [rad]. Since the pulley 1 and pulley 2 are connected with each other by the crossing wire, these pulleys rotate always in the counter directions by same angle. Hence, the spring hooked between two pulleys are stretched by 2h sin ｾ＠ and the restoring forces ±2Kh sin ｾ＠ are generated at the pins fixed on the pulleys. Caused by these restoring forces, the moment 2Kh2 sin ｾ＠ cos ｾ＠ is generated around the shaft of pulley 1. Since the same moment is exerted from the pulley 2 to the pulley 1 through the crossing wire, and pulley 1 is connected to the joint shaft in the gear ratio of 2:1, the moment generated around the joint is given by
e
T
= - 2K h 2 sin ｾ＠ cos ｾ＠ = - K h 2 sin e. 2
2
(1)
1022
Therefore, if the spring constant satisfies the relationship mgf!
(2)
K=T'
e
the counter moment -mgf! sin is driven on the joint, and the proposed system becomes to work as a gravity compensation system, Here, it should be noticed that the wire systems can be contained completely inside the links in the proposed structure. Therefore, it can be said that the proposed system is safer and preferable to the wearable rehabilitation systems than the previous ones. This system can be extended for the multi-link system in the same way as the previous works. 4 For the multi-link case, the gravity compensation system can be achieved by the system illustrated in Fig.3. In this system,
Fig. 3.
Proposed gravity compensation system for multi-link system
the second joint is connected to the 1st joint through the timing belt. Due to this structure, the pulleys in the second link rotate by ± ii, as shown in fig.3. Thus, the gravitational moments
1li2
(3) can be compensated by the proposed system, if the spring constants Kl and K 2 are set to K
K _ mgLl 1 -
respectively.
h2
'
_ mgL2 2 -
h2
'
(4)
1023
3. The experimental verification of the function of the proposed mechanism as a gravity compensation system
To verify the feasibility of the proposed mechanism, we construct an actual equipment as shown in Fig.4 and Fig.5. In these devices, the timing belt and another gear are used instead of the crossing wire in Fig.2 and Fig.3 to make assembling simple and make the performance precisely. The length of the links, mass properties and spring constants of the equipment are listed in Table1. Table 1. ment
link1 link2
Parameter of equip-
Mass (kg)
spring constants (N/mm)
0.46 0.54
7.95 15.9
Fig. 4.
The gravity compensation device for the 1Link system
Fig. 5.
The gravity compensation device for the 2Link system
To verify that the proposed device works as a gravity compensation system as expected, we examine whether the balance of the system is kept
1024
Fig. 6.
Gravity compensation by the proposed device
under the change of postures. In this experiment, the weight(500[g]) was put on the tip as an alternative to the upper body. As shown in Fig.6, it is verified that the proposed mechanism can surely compensate the gravitational moment and maintain the balance at any given posture.
4. Simulations In the previous section, it is confirmed that the system can compensate the gravitational moment successively. However, it is still unclear whether it can work effectively in the rehabilitation therapy to reduce the loads excerting on the patients joints. To show the effectiveness of the gravity compensation system, in this section, we examine the effectiveness of the proposed system in the bending and stretch exercise. In this simulation, we take the parameters as shown in Table.2. The trajectories of each joints are
1025
Table 2.
h 12 13
Link parameter for simulations.
Link mass (kg)
Link length (m)
Mass center (m)
3.51 7.41 40.0
0.395 0.401 0.750
0.228 0.210 0.300
Note: Mass center measured from the joint axis of the root.
E
C
eo ;:
-lC
ｾ＠
-20
6
0.0
O.S
1.0
Normal condition Fig. 7.
1.5
ｔｬｭ｣ＨｾＮｽ＠ ｔｬｭ｣ＨｾｃＮＩ＠
gravity supported
Joint moment profiles in the bending and strech exercise
taken to the simple sinusoidal form as
ｾｮ＠
81 = 82 =
ｾｮ＠
(cos
ＨｾｮｴＩ＠
(-cos
ＨｾｮｴＩ＠
+ 1)
+3)
(5) (6)
and the joint angle of the hip is chosen to maintain the center of gravity of over the ankle joint. The joint loads obtained in the normal exercise and those under the gravitational support are shown in Fig.7. From this figure, it is confirmed that the exerting loads are reduced dramatically by introducing the gravity compensation system. Maximum load exerting on the knee joint under the garvity support is about one-sixth of that in the normal case. From these simulation results, it is conceivable that the proposed system works effectively even in the rehabilitation therapy.
1026
5. Discussions In this paper, we proposed a new gravity compensation system. In the proposed system, all components are embedded in the links. Thus, it is safe and well-suited to the wearable power assists systems. From the experiments using actual equipment, it was shown that the proposed system can compensate the gravity completely for the arbitrary posture as expected from the static mechanical analysis. The effectiveness of the proposed device in the rehabilitation exercise was examined through some computer simulations. From these results, it is conceivable that the proposed device can work effectively in the rehabilitation therapy.
References 1. A. M. Dollar and H. Herr, IEEE Transactions on Robotics 24, 1 (2008). 2. A. H. Stienen, E. E. Hekman, F. C. V. der Helm, G. B. Prange, M. J. Jannink, A. M. Aalsma and H. V. der Kooij, Freebal: dedicated gravity compensation for the upper extremities, in Pmc. IEEE Int. Conf. on Rehabilitation Robotics, (Noordwijk, Netherlands, 2007). 3. G. F. F. S. A. Sisto and P. D. Faghri, IEEE Engineering in Medicine and Biology Magagine 27, 56 (2008). 4. T. Morita, F. Kuribara, Y. Shinozawa and S. Sugano, A novel mechanism design for gravity compensation in three dimensional space, in Proc. IEEE/ASME Int. Conf. on Advanced Intelligent Mechatronics, (Kobe, Japan, 2003). 5. Y. Ono and T. Morita, Journal of Robotics and Mechatronics 17, 553 (2005). 6. N. Ulrich and V. Kumar, Passive mechanical gravity compensation for robot manipulators, in Pmc. IEEE Int. Conf. on Robotics and Autometion, (Sacramento, USA, 1991). 7. S. Shirata, A. Konno and M. Uchiyama, Design and evaluation of a gravity compensation mechanism for a humanoid, in Proc. IEEE/RSJ Int. Conf. on Intelligent Robotics and Systems, (San Diego, USA, 2007). 8. Y. Morita, A. Akiba and T. Morita, An optimum link length design method based on shoulder movability criterion for torque compensated upper arm orthosis, in Pmc. JSME Conf. on Robotics and Mechatronics, Japan, (Tokyo, Japan, 2006). 9. R. B. B. M. Wisse, W. D. van Dorsser and J. L. Herder, Energy-free adjustment of gravity equilibrators using the virtual spring concept, in Pmc. IEEE Int. Conf. on Rehabilitation Robotics, (Noordwijk, Netherlands, 2007). 10. M. v. d. B. B. Mastenbroek, E. de Haan and J. L. Herder, Development of a mobile arm support(armon): Design evolution and preliminary user experience, in Proc. IEEE Int. Conf. on Rehabilitation Robotics, (Noordwijk, Netherlands, 2007). 11. J. L. Herder, Development of a statically balanced arm support: Armon, in Proc. IEEE Int. Conf. on Rehabilitation Robotics, (Chicago, USA, 2005).
A ROBOTIC WALKER WITH STANDING, WALKING AND SEATING ASSISTANCE D. CHUGO*, T. ASAWA *, T. KITAMURA * and K. TAKASE*
Graduate School of Information Systems, The University of Electro-Communications, Chofu, Tokyo 182-8585, Japan * E-mail: [email protected] http://www.taka.is.uec.ac.jp/chugo/ This paper proposes a robotic walker system with standing, walking and seating assistance function. Our system focuses on domestic use for aged person who needs nursing in their daily life. Our key ideas are two topics. The first topic is combination of standing assistance function and walking assistance function. In previous works, many assistance devices are specialized in only "standing-up operation" or "walking operation". However, in their daily life, elderly person needs standing, walking and seating assistance continuously by a same device. Therefore, our developing assistance system can support both operations by a small sized mechanism which is easy to use in the horne. The second topic is a seating position adjustment assistance. From questionnaires of nursing specialists, a seating position adjustment requires the elderly to walk backward and it is difficult operation for them. Furthermore, in many cases, a failure of this operation causes a fracture which has high risk to fall into bedridden life. Thus, our developing system can assist the elderly to adjust the seating position safety. The performance of our proposed system is verified by experiments using our prototype.
Keywords: Force assistance; Active walker; Standing up motion; Seating motion
1. Introduction
In Japan, the population ratio of senior citizen who is 65 years old or more exceeds 20[%] at January 2004 and rapid aging in Japanese society will advance in the future. 1 In aging society, many elderly people cannot perform normal daily household, work related and recreational activities because of decrease in force generating capacity of their body. Today, the 23.5[%] of elderly person who does not stay at the hospital cannot perform daily life without nursing by other people. 2 For their independent life, they need assistance system which enable them to perform daily life easily III their home even if their physical strength reduces.
1027
1028
Usually, their daily activities consist of standing, walking and seating operation continuously. Standing up motion is the most serious and important operation in daily life for elderly person who doesn't have enough physical strength. 3 ,4 Therefore, many researchers developed assistance devices for standing up motion. However, these devices are large scale and they are specialized in only "standing assistance". Therefore, the patient has to use other assistance device for their daily activities, for example when they want to walk, and these devices are not suitable for family use. 5 ,6 In this paper, we develop a robotic walker system with power assistance device for standing and seating motion. A robotic walker has two EC motors and it can recognize a target seat using laser range finders. Proposed power assistance device realizes the standing and seating motion using the support pad which is actuated by the manipulator with three degrees of freedom. Our system is enough small to use in the home and it can support the patient during standing, walking and seating operation continuously. Our key ideas are two topics. The first topic is combination of standing assistance function and walking assistance function. The second topic is a seating position adjustment assistance from walking to seating operation. Using our proposed system, the patient can use standing, walking and seating assistance continuously by a same device. We verify the performance of our proposed assistance system through experiments using our prototype.
2. System Configuration
2.1. Problem Specification We questions to nursing specialists about required assistance for aged people in their daily life. Their results are the followings. • Aged person requires standing, walking and seating assistance continuously by a same device. In typical required case, he stands up from the bed, he walks and he sits on the toilet by himself using the assistance system. • When he stands up, he requires power assistance for reducing the load and he also requires position assistance for maintaining his body balance. • When he sits on the target seat, he requires the position adjustment assistance. A failure of this motion causes a serious injury, therefore, this assistance is important. In our previous works, a reducing the load during standing is realized. 7
1029
Therefore, in this paper, we focus on (1) the assistance for stable posture during standing to walking motion and (2) the position adjustment assistance to target seating position.
2.2. Assistance Mechanism Fig.1(a) shows overview of our proposed assistance system. Our system consists of a support pad with three degrees of freedom and an active walker system. The support pad is actuated by proposed assistance manipulator mechanism with four parallel linkages. The patient leans on this pad during standing assistance. Our active walker is actuated by two EC motors on each front wheel and it has a laser range finder in its body. This LRF uses LED light with low cost and it can scan around 4[m]. Fig.l(b) shows our prototype. Our prototype can lift up the patient of 180[cm] height and 150[kg] weight maximum, and it can assist him during walking using actuated wheels.
(a) Overview of our system Fig. 1.
(b) Our prototype
Our assistance system
3. Standing Assistance for Stable Posture 3.1. Required Condition
In previous study, a lot of standing up motions for assistance are proposed. Kamiya8 proposed the standing up motion which uses remaining physical strength of the patients maximum based on her experience as nursing specialist. In our previous work, we analyze this standing up motion and find that Kamiya scheme is effective to enable standing up motion with smaller load. 9 For realizing the motion of Kamiya scheme, the conditions are discussed as follows.
1030
• When the elderly lifts off from the chair, it is required to reduce the knee load. • During standing up operation, it is required to maintain the standing up motion with stably posture. Our system consists of the assistance manipulator and the active walker. Considering with the characteristic of both devices, we design the assistance control system in the following policies. • The assistance manipulator assists the standing up motion of the patient for reducing the knee load. (We realized this condition in our previous work.?) • The active walker assists the body balance for stable posture. (This condition is our target in this paper. In our future work, same algor isms will be used for walking assistance.)
3.2. Force Control of Active Walker For realizing the required condition, the active walker maintains the COG of the patient as an index of body balance. We use PID controller as (1) and coordinate the COG. The coordination is shown in Fig.2(a).
e
ref
XCOG
=
[ref
XCOG
()
0 ,'"
S
Tout =
{
ref
'X COG
=
( ')
S ,'"
= tits
T TO
(ITI < TO) (ITI :::" TO)
XCOG -
XCOG
ref
(1)
ref
()] T
(2)
'X COG
1
(3)
(4)
where Tis force reference of the active walker and XCOG is the actual position of COG. is the position reference of COG (Fig.2(b), we measure the motion of Kamiya scheme. 8 ) and it is function of the movement pattern S as (2). The movement pattern is (3). ts is required time to the standing up operation and t is present time. kp,ki,kd are proportional, integral and derivative gain of PID controller, respectively. Toutis actual output force of the active walker as (4). For safety reason, output force is limited to TO.
x?!6G
1031
-Q,Z5 Movement Pattern (%)
(a) Coordination Fig, 2,
(b) References
(c) Relationship to force Sensor
The Position of Center of Gravity (COG)
3.3. Estimation of COG Our proposed force control of the active walker requires the information of the position of COG. Usually, COG is measured using a force plate, however, this device is not suitable in practical use. Thus, it is required to estimate the position of COG with the sensors equipped on this walker. In this study, we discuss the relationship between the COG and applied force to the pad in preliminary experiment using testers. Fig.2(c) shows the relationship between the position of COG (which is measured by a force plate) and COG on force sensors as (5) when the angle of support pad is 70[deg].
(5) where l is distance between two sensors as Fig.3(a). Testers are adult male with the special wearing equipment for the experience of the elderly.l0 From these results, both values seem to be proportional as Fig.3( c). (We show an approximation line of Tester A.) From this experiment, we can estimate the COG comparing with the measuring values of force sensors and these data which are derived in this preliminary experiment.
4. Seating position adjustment assistance We questions to nursing specialists about required assistance for aged people when they sit down. Their results are the followings.
1032
• Usually, the aged person can walk with a walker. In this case, he should walk without assistance for rehabilitation. • However, it is difficult for him to walk backward with exact position enough for seating. Therefore, system should assist this operation part. • System should assist him with his operation speed for safety reason. Too strong assistance causes his falling from the walker. Therefore, we design seating position adjustment assistance as follows. • System assists only for adjusting the seating position. The user walks by himself near the chair which they want to sit on and starts the assistance system. • System scans around the walker and find the chair using LRF. (The system has the shape data of the target chair previously. [xl) • If the system finds the target chair, it assists him in front of the seating position. • If the system cannot find the target chair, it announces the user to go around the target chair. • To assist him softly, the system uses damping control l l as (6). • For safety reason, output force is limited to TO as (4).
Ti
= ｣ｶｾ･ｦ＠
- B (F - Fa)
(6)
where ｶｾ･ｦ＠ is velocity control reference which is derived using the position information by LRF between the walker and the target chair. F is the applied load to the walker and Fa is the threshold which selects damping or velocity control. c and B are coefficients (B = 0 (if F < Fa)). i is wheel number. (i = 1,2)
5. Experiment Here, we verify the performance of our prototype system by the experiment. This experiment assumes typical action of elderly (a movement from chair to bed in same room) in their daily life. As the result of the experiment, our system can assist the patient as shown in Fig.3. The first, the tester stands up from the left chair with standing assistance of our proposed system (Fig.3(a)-(c)). The second, he walks to near the target bed himself and our system does not assist him (Fig.3( d)). Usually, it is difficult for the aged person to reach the suitable
1033
seating position. Therefore, in this experiment, the tester only walks around the bed. The third, our system adjusts the seating position and assists the sit down motion (Fig.3(e)-(h)). The height of the patient is 170[cmJ and he uses the special wearing equipment for the experience of the elderly.l0 Fig.4(a) shows the movement of our active walker utilizing proposed control. After 25[%J movement pattern, system starts to coordinate the body balance of the patient, because in this period, the assistant manipulator uses force control and the posture of the patient tends to change easily. Fig.4(b) shows the COG of the patient during standing up motion. For verifying effectiveness of our proposed scheme, we show the result without active walker function in Fig.4(b). From these results, we can verify that our system coordinate the body balance according to the reference of the nursing speeialists.
(f)
(e)
Fig. 3.
(h)
(g) Experimental Results
0.2
]
ｾ＠
0 , - -........
"g -0.04
15 -0.08
] Ｎｾ＠
ｾ＠
0
ｾ＠
-0.1
2 ｾ＠
-Reference
-0.2
"""""""Witout Proposed
Q
-0.12 Movement Pattern [%]
(a) Movement of the walker Fig. 4.
0.1
-0.3
Movement Pattern [%]
(b) COG during the standing up motion
Stability of the tester's body during Standing Assistance
1034
6. Conclusion In this paper, we develop a novel assistance walker system for family use. Our system is required to assist both standing up operation and walking operation. Furthermore, our system is required to adjust the seating position for safety sitting. In order to fulfill these conditions, we integrate the walking assistance function and the standing assistance function. Furthermore, we develop the seating position guidance function using LRF. Using our proposed system, the patient can use standing, walking and seating assistance continuously by a same device. This work is supported in part by Electro-Mechanic Technology Advancing Foundation. References 1. Japan current population estimates as of october 1, 2004 (2004), http://
www.stat.go.jp/english/data/jinsui/2004np/index.htm. 2. Japan annual reports on health and welfare 2001 social security and national life (2001), http://ww.mhlw.go.jp/toukei/saikin/hw/k-tyosa/ k-tyosa01/4-3.html. 3. N. B. Alexander, A. B. Schultz and D. N. Warwick, J. of Geometry: MEDICAL SCIENCES 46, M91 (1991). 4. M. A. Hughes and M. L. Schenkman, J. of Rehabilitation Research and Development 133, 409 (1996). 5. K. Nagai, 1. Nakanishi and H. Hanabusa, Assistance of self-transfer of patients using a power-assisting device, in Pmc. IEEE Int. Conf. on Robotics and Automation (ICRA '03), (Taipei, Taiwan, 2003). 6. A. Funakubo, H. Tanishiro and Y. Fukui, J. of the Society of Instrument and Control Engineers 40, 391 (2001). 7. D. Chugo, W. Matsuoka, S. Jia and K. Takase, Rehabilitation walker system for standing-up motion, in Proc. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS'07), (San Diego, US, 2007). 8. K. Kamiya, Development and evaluation of life support technology in nursing, in Proc. 7th RACE Symp., Research into Intelligent Artifacts for the Center of Engineering, (Chiba, Japan, 2005). 9. D. Chugo, E. Okada, K. Kawabata, H. Kaetsu, H. Asama, N. Miyake and K. Kosuge, Force assistance control for standing-up motion, in Proc. IEEE/RAS-EMBS Int. Conf. on Biomedical Robotics and Biomechatomnics (BioRob '06), (Pisa, Italy, 2006). 10. K. Takeya, Y. Kanemitsu and Y. Futoyu, Kawasaki Medical Welfare J. 11, 64 (2001). 11. T. Sugihara, K. Kawabata, H. Kaetsu, H. Asama, K. Kosuge and T. Mishima, Development of a reasonable force sensor for a standing-up and sitting motion support system, in Proc. of Robotics and Mechatronics Conf. 2004, (Nagoya, Japan, 2004).
STEP CLIMBING OF A FOUR-WHEEL-DRIVE OMNIDIRECTIONAL WHEELCHAIR MASA YOSHI WADA Department of Human-Robotics, Saitama Institute of Technology, 1690 Fusaiji, Fukaya, Saitama 369-0293, Japan This paper presents a development of a four-wheel drive (4WD) omnidirectional wheelchair with an advanced step climbing capability. For enhancing the mobility of standard wheelchairs, a new type of omnidirectional mobile platform with a 4WD mechanism is proposed. The 4WD system provides enhanced step climb capability however, chair inclination becomes large when either front or rear wheels are on a step because a wheelbase of the developed wheelchair is almost same as a standard one. This large inclination not only gives a high risk for falling but also brakes a load distribution condition between front and rear wheels which is required for climbing the step. To solve these problems relating to the chair inclination, we develop a chair tilting system for the 4WD omnidirectional wheelchair including a linear drive and a variable center of rotation mechanism and an inclination sensor.
1. Introduction The aging of society and the declining birth rate have become serious social issues world wide, especially some countries in Asia and Europe. In Japan, it reported that the population of over 65 years old would grow over 30% of total population in 2025[1]. In the aging society, it is suggested that demands for human support devices would increase dramatically for supporting moving capabilities of human and reducing labor of caregivers. Among the human support devices, we are developing an electric wheelchair with applying mobile robot technologies. A developed wheelchair is expected to be capable of using in multiple environments including indoors and outdoors. For the purpose, we have introduced a four-wheel-drive (4WD) mechanism for a mobile base and applied an omnidirectional control technique for the 4WD system. In this paper, analyses and a design of a 4WD omnidirectional wheelchair with a chair tilting mechanism will be presented together with fundamental tests using a wheelchair prototype. It is verified through experiments that the proposed wheelchair system can perform independent 3 DOF omnidirectional motions and high step climbing capability (90mrn) with a chair tilting system.
1035
1036
2. 4WD Omnidirectional Mobile System Figure 1 (a) and (b) show a schematic of a 4WD omnidirectional wheelchair and a prototype overview respectively. The wheelchair has two omni-wheels in front and two standard pneumatic tires in rear. A pair of an omni-wheel and a pneumatic tire mounted on the same side of the wheelchair are interconnccted by belt transmissions for rotating unison and driven by a common motor. Thus the 4WD system enables all four wheels to provide traction forces in rotating direction which configuration is invented in US in 1989[2] and recently applied to a wheelchair product[3]. In our proposed design, an additional third motor is installed on the 4WD platform for rotating a chair about the vertical axis which is also seen in Figure 1. Those three motors including the two wheel motors and the chair rotation motor enable a chair to realize independent 3DOF omnidirectional motions by a coordinated motion control. To realize such a mobile capability of a wheelchair, i.e. a holonomic and omnidirectional mobility, an onmidirectional control method, called "poweredcaster control"[4] is applied to the omnidirectional 4WD mechanism. Unlike conventional an-wheel-steering vehicles such as [6], it can move in an arbitrary direction from an arbitrary configuration of the mechanism.
(a)
(b)
Figure 1. A 4WD omnidirectional wheelchair and Prototype overview
To verify the omnidirectional mobility of the proposed system, a lateral motion was presented by the prototype wheelchair in which a chair orientation was maintained to be constant at all times. In experiments a small ball was attached to the chair to identify the center of a chair rotation axis. A stcreo camera system (Quick Mag IV from OKK Inc, Japan) detects a ball location in 3D coordinates. The camera system provides real-time video images with overwriting rectangle window(s) and the path of the target(s).
1037
Figure 2(a)-(f) shows a series of screen shots of video of an experiment. The prototype was controlled by a remote PC via a LAN connection. The prototype wheelchair moved from the right side of the picture frame to the left. A straight path is shown in the final capture, Figure 2(f), which was drawn by the wheelchair movement. The center of the chair was controlled by two wheel motors to locate on a reference straight line directing to the lateral direction at all times. It is also recognized that the wheelchair moved in lateral direction with maintaining the chair orientation constant hence the drive unit orientation varies during the motion. The 4WD mechanism moved backward (Figure 2(a) and (b)), changed its orientation (Figure 2(c)) and faced to the direction of motion (Figure 2(d)-(f)). This series of movement is called as a "caster flip" which is unique for the proposed onmidirectional control system.
(a)
(d)
(b)
(c)
(c)
(I)
Figurc 2. An example of omnidirectional motion: the latcral motion of the whcelchair prototype; it moves in sideways from the right side to the left of the picture frames while maintaining the chair orientation to be stable
3. Step climb capability of 4WD system 3.1. Static Analysis of Wheel-Step System
Compared with conventional omnidirectional wheelchairs [7], [8] and omnidirectional mechanisms [9], [10], the proposed 4WD onmidirectional system has an advanced step climb capability since the 4WD mechanism faces to the moving direction spontaneously and negotiates a step with a large curvature of a wheel.
1038
When the front wheels of a 4WD wheelchair touch a step edge as shown in Figure 3(a), the front wheels are restricted to rotate until either a slip or a step climbing motion occurs (we call this situation as "wheel-lock" in this paper). The front wheels can not rotate during the wheel-lock period however, the wheels maintain to provide drive torque to a step edge. Therefore a reaction of the drive torque is applied to a vehicle body in CCW direction about the center of the front wheels as shown in the figure. This reaction torque cause the increase of a load on the rear wheel axis, f, while a load on the front wheel axis is decreased by f since the total vehicle weight does not change, namely a front load plus a rear load must equal to a vehicle weight. This load change f enables the front wheels to climb the step easier and the rear wheels to provide larger propelling force for pushing the front wheels forward. Here the load change f is represented as, (1)
where Tr represents a driving torque on the front wheel shafts and L IS a wheelbase of the 4WD system. From Figure 4, a condition required for surmounting a step is derived as following FI + F, cos B :2: (WI -
f )sin B
(2)
where Ff and Fr are propelling forces generated by front and rear wheels respectively. Wfis a vertical load acting on front wheels and B is an angle formed by contact points A, B and a center of front wheel 0, as shown in Figure 4. From Eqs.(l) and (2), a minimum force for the front wheel climbing is given as following under the assumption that Ff =Fr =FI2, where F is a force provided by a wheel motor.
F> _
2sin B
Wj
(3)
1 + cos B + ｾ＠ sin B L
On the other hand, when the rear wheels reach to the step edge as shown in Figure 3(b), a drive torque acts to the vehicle body in CCW direction as well. However in this case, the load change f suppresses the rear wheels to climb a step since the vertical load to rear wheels is increased hence the propelling force produced by the front wheels is limited. Then a minimum force for the rear wheel climbing is given as,
1039
2sinB ｆｾ＠
r
1 + cos B - - sin B
(4)
WI.
L
Figure 5 shows required motor torques vs. a step height derived by results of the static analyses, Eqs.(3) and (4). When a influence of the reaction torque is not taken into account ( L ｾ＠ 00 ), step climb capabilities of the front and the rear wheels become identical which plot is indicated by a dashed curve in the middle in Figure 5. However, the reaction of drive torque extends the front wheel capability and suppresses the rear wheel climbing which are plotted by a solid curve and a dashed-dotted curve in Figure 5 respectively. Reaction Torque of T(
Reaction Torque of Tr
f
Wheel locked
(a) Front wheels contacting to a step edge
Wheel locked
T(
ＨｾＬ＠
(b) Rear wheels contacting to a step edge
Figure 3. Static analyses of wheel-step systems
Figure 4. Analysis of front wheel of 4WD system with a step
3.2. A Chair Tilting Mechanism A chair is inclined when either front or rear wheels are on the step as shown in Figure 6(a) and (b). For recovering the load distribution between front and rear wheels of the wheelchair, we developed a chair tilting mechanism. Figure 7(a) and (b) show the tilting mechanism in motion.
1040 Required motor torque for 4WD ｯｭｮｩｾｷｨ･ｬ｣｡ｲ＠ D=275mm, M=140kg, 1'=0.7, G=25 Wf:Wr = 0.45:0.55
4.5 c 4
£
3.5
ｾ＠
2.;
-
j
ＱＮｾ＠
2
0.5
o o
10 20 30 40 50 60 70 80 90 100110120130140150
Step height mm Figure 5, Required motor torques for 4WD wheelchair. (W!"W,=0,45:0,55)
A linear drive unit is installed on the back side of a chair for a tilting system which consists of a motor, a linear guide, a screw and a nut as shown in Figure 8. The nut can provide rated 950N force with O.2m1s velocity. It also has a self-lock function (non-back-drivable) for avoiding the falling of a chair even if an electric power supply is interrupted during operation. When the tilting mechanism pushes the chair and human up, a center of gravity shifts to the front side of the wheelchair which results in recovering the load distribution between front and rear wheels in almost same ratio as it stands on the horizontal floor. For instance, when only the front wheels are on a 90mm step with the tilting system is disabled, a load distribution ratio reaches 35:65 (front: rear, Figure 6(a)) while the ratio is alleviated to 45:55 when the tilting system is enabled(Figure 7(a)).
(a)
(b)
Figure 6. A chair inclination
(a) center
(b)
center
Figure 7, The tilting mechanism in motion
4. Step Climbing Experiments Figures 9 and 10 show experimental results in which the prototype wheelchair negotiated a 90mm step. In the experiment shown in Figure 9, the tilting system was disabled therefore a chair tilt angle was constant relative to the 4WD
1041
mechanism at all times. In this case, the back wheels of the wheelchair were not able to surmount the step and all four wheels slipped simultaneously. On the other hand when the tilting system was enabled, it is recognized in Figure 9 that the tilt angle of the chair was varied by the tilting mechanism and both front and back wheels successfully got over the top of the step.
Figure 8. A linear drive unit mounted on the back sidc of a chair for a chair tilting system
(a)
(b)
(c)
(d)
Figure 9. Climbing a 9cm step with tilting system disabled. Both front and rear wheels slip when rear wheels contacting the step edge and rear wheels fail to climb up the stcp
(a)
(b)
(c)
(d)
Figure 10. Climbing a 9cm step with tilting system enabled. Both front and rear wheels successfully climb up the step
5. Conclusions A 4WD omnidirectional wheelchair with tilting system is presented in this paper. First, an omnidirectional control, called a "powered-caster control" for realizing independent 3DOF motions was briefly described and performances of a prototype 4WD wheelchair were presented. Next, step climb capabilities of the 4WD system were discussed. The conventional 4WD system has a potential to make a wheelchair climb up a
1042 9cm-IOcm single step. However, due to the wrong weight distribution and reaction torque acting to the vehicle body, rear wheel of the wheelchair are difficult to surmount the high step. To clarify the difference between front wheels and rear wheels in step climb capability, each statics was analyzed. To change the load distribution, a chair tilting system was proposed. A prototype wheelchair was designed and built to verify the availability of the proposed system. The prototype design includes a 4WD omnidirectional platform with a chair tilting mechanism which can change the load distribution between front and rear. The chair tilting system on the prototype is effective for two objectives, one is to avoid falling of the wheelchair and the other is to recover a load distribution for surmounting a high step. The wheelchair with the developed chair tilting system successfully surmounted a 90mm step by enabling the tilting system, while it was not able to climb the step with disabled tilting control. By these results, it can be said that the validity of the proposed drive system is verified trough the series of experiments. Acknowledgement This project was supported by New Energy and Industrial Technology Development Organization (NEDO) Japan, research ID 05A06715a. References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.
11. 12.
National Institute of Population and Social Security Research, Japan. J. Farnam, US patent 4,823,900 (1989). Kanto Auto Works Ltd., http://www.kanto-aw.co.jp/jp/products/wheelchair/ M.Wada, and S.Mori, Proceedings of the 1996 IEEE International Conference on Robotics and Automation, 3671 (1996). M.Wada, A.Takagi and S.Mori, Proceedings of the 2000 IEEE International Conference on Robotics and Automation, 1531 (2000). E.Nakano and N.Koyachi, Proceedings of the International Conference on Advanced Robotics, 277 (1983). Fujian Fortune Jet Mechanical & Electrical Technology Co., Ltd. M.Wada and H. H. Asada, IEEE Trans on Robotics and Automation, 15-6, 978 (1999). S.Hirose and S.Amano Proc.ISRR, 253 (1993) K.Tadakuma, R.Tadakuma, J.Berengueres, Proceedings of the 2007 IEEEIRSJ International Conference on Intelligent Robots and Systems, 33 (2007). M.Wada, Proceedings of the 2007 IEEEIRSJ International Conference on Intelligent Robots and Systems 1196 (2007). M.Wada, Climbing and Walking Robots, 313, Advanced Robotic Systems International (2007).
STEERING CONTROL OF WHEELCHAIR ON TWO WHEELS SALMIAH AHMAD, M 0 TOKHI and KHALED M K GOHER Department ofAutomatic Control & Systems Engineering University of Sheffield, United Kingdom This paper discusses the steering control for a wheelchair on two wheels. A novel fuzzy logic control (FLC) system is designed and implemented for this highly nonlinear system. The wheelchair system is modeled in Visual Nastran software environment as a plant and controlled with the developed FLC in Matlab/Simulink environment. The steering motion takes place after lifting and stabilizing has been achieved. There are noticeable challenges in the modeling of the wheelchair where limited actuators are used for different functions, thus suitable controllers need to be developed. Results show that the FLC strategy works very well and gives good system performance.
1. Introduction Wheelchair on two wheels is a very interesting and challenging problem involving highly nonlinear system characteristics. In fact, as it is designed to mimic double inverted pendulum, the problem thus becomes more challenging for which an appropriate control strategy is required. The two-wheeled state of the wheelchair is realized by lifting the front wheels (casters) of the wheelchair so that it achieves an upright position with lifting the chair to a higher level. The current research is aimed to help disabled people who are using the wheelchair as the main means of transport for mobility, and cannot stand on their own due to permanent injuries in their lower extremities. Therefore, this will then help them reach certain levels of height in confined spaces, e.g. to pick and place things on shelves while sitting on the wheelchair. Furthermore, they will be able to participate in conversations at eye-to-eye level comfortably as normal people do. Researchers have accordingly shown a great deal of interest in advancing the current technology of wheelchair to let disabled people perform most of their daily life tasks independently. They have proved that using wheelchair as a means of mobility is more efficient than walking. Thus, it is important for wheelchair users to be independent, to perform tasks similar to normal people. Research work focusing on modeling and control of wheelchairs on two wheels is limited in the literature. Noticeable reports on wheelchairs are reported by Japanese researchers, mainly for their ageing society. Many researches, on the other hand, have focused on various system types involving inverted pendulum, such as inverted pendulum on cart [1-4], rotary inverted pendulum [5-6], and inverted pendulum on two wheels [7-8]. These research works differ
1043
1044
from one another mainly in terms of the control strategy adopted. Most of these have used conventional control methods such as PID, linear quadratic regulator (LQR), optimal control and nonlinear control. Only a few have proposed the use of fuzzy logic, neural networks and genetic algorithms. Some limited research work on wheelchairs on two wheels has been reported in the literature [9-14]. Among them, there are a number of research reports in the literature on wheelchairs with a mechanism whereby the user can climb over a step up to about 10 cm without a helper while maintaining control as an inverted pendulum. The technique uses conventional proportional-integral (PI) control with which the wheelchair can be maintained in the inverted pendulum position for about few seconds before climbing a step on the roadside. Throughout these studies, there is no evidence of use of intelligent control approaches such as fuzzy logic control for controlling the wheelchair on two wheels. Moreover, the models reported involve only single inverted pendulum scenarios whereas in this paper the current research considers a double inverted pendulum model, which is more complex and yet crucial.
2. Wheelchair Model The wheelchair with humanoid model was developed using MSC. Visual Nastran 4D (VN Software). The dimensions of the model are closely approximated from the actual wheelchair system whereby the humanoid model was designed and approximated by referring to [15]. The wheelchair was modeled in a basic form comprising two wheels, two casters, frames and axes connected to the seat. The main focus would be on the motor constraints connecting the two wheels to the horizontal axis, and a further motor connecting the horizontal frames to the seat. Therefore, there will be three independent input torques to be controlled in the system. The schematic diagram of the whole system is shown in Fig. 1. The system mimics a double inverted pendulum scenario and accordingly such a situation is the focus of this work. Torque 1 comprising of right torque (TorqueR) and left torque (TorqueL) represents the input torque to the wheels during lifting and stabilizing stage. In this paper, instead of having one input torque from the controller, the wheel has two independent input torques from the controller. Torque2 represents the torque between Linki and Link2, and will be used to cater for the whole weight of the human body. The weight here represents the human body, and in this case an average 70kg humanoid is used. On the other hand, steering torques given by controller is represented by Torque3 consisting of both right and left torque. The sensors are attached at the respective reference bodies for control and measurement. The inputs coming into the wheelchair system in Visual Nastran
1045
from the controller comprises of TorqueR (Nm), TorqueL (Nm) and Torque3 (Nm) with respect to certain condition. The output from the wheelchair system is fed back, and the feedback loop consists of the angular position of Link1 (Degree), angular position of Link2 (Degree) and steering position (Degree).
3. Steering Motion Controller Fuzzy logic is known to be very ideal for control of nonlinear systems, and specifically where the system is complicated to model, thus this research embarks on the development of fuzzy logic control (FLC) for steering motion of the wheelchair on two wheels. This steering controller is executed after the transformation of a wheelchair from four-wheeled state to two-wheeled state by lift the front wheels (casters) and the chair, and maintaining this inverted pendulum position. The details on lifting and stabilizing control for wheelchair on two wheels have been discussed in [16]. The same FLC used in lifting and stabilizing was used for forward or backward linear motion control, as presented in [17]. In this paper, steering motion for wheelchair is introduced, and two inputs two outputs of fuzzy logic controller are designed. Previously, with the system without steering motion control, the actuators at the two wheels were set to have the same output signal from the FLC but differing in direction. This was made to simplify the system modelling and control during the initial stage. However, for the steering motion control to be realized, the wheels need to have independent actuators, right wheel torque (TorqueR) and left wheel torque (TorqueL) thus increasing the number of degrees of freedom for the system. This requires the FLC to have two outputs instead of one with two inputs, increasing the complexity of the controller and the whole system. The FLC structure used for steering is shown in Fig. 2, where error is the orientation error of the wheelchair in degrees, as the wheelchair is steered relative to a reference point. There are 5 levels of membership function for the inputs and outputs, and those are NB, NS, Z, PS, PB resulting 25 rules. The output membership function was also set to have five levels, as can be seen in Fig. 3. Typical 5 rules that are fired are as follows, where (1) means 'AND' connective. The rest of the rules follow the same pattern. 1. If (Error is NB) and (Change_of-Error is NB) then (TorqueR is PB)(TorqueL is PB) (1) 2. fr (Error is NB) and (Change_of-Error is NS) then (TorqueR is PB)(TorqueL is PB) (1) 3. If (Error is NB) and (Change_of-Error is Z) then (TorqueR is PB)(TorqueL is PB) (1)
1046
4. If (Error is NB) and (Change_of-Error is PS) then (TorqueR is PS)(TorqueL is PS) (1) 5. If (Error is NB) and (Change_of-Error is PB) then (TorqueR is Z)(TorqueL is Z) (1)
Fig. 1. Schematic diagram of wheelchair system
&ror (5)
flc9b
(marrdani) 75 rules
TorqueL (5)
Systemflc9b: 2 inputs, 2 outputs, 75 rules
Fig. 2. Structure of fuzzy logic control for steering motion
1047 0.
0.
:E
:E
ｾ＠
ｾ＠
CD .0
CD .0
E CD E
E CD E
(5
(5
CD
CD
ｾ＠
Ol
ｾ＠
0
0
CD
Ol
CD
-1
-0.5
0 Error
0.5
0.
-1
-0.5 0 0.5 ChangeofE"ror
-1
-0.5
0.
:E
:E
＠ｾ CD .0
ｾ＠ CD
..Q
E CD E
E CD E
(5
(5
CD
CD
Ol
OJ CD
ｾ＠
ｾ＠
CD
0
0
-1
-0.5
0 TorqueR
0.5
0 TorqueL
0.5
Fig. 3. Membership levels for inputs and outputs of steering control
4. Results & Discussion The final position of the wheelchair system can be seen in Fig. 4. Figures 5 and 6 show typical results after the system is set to steer as much as 30" to the left from its initial position. It is noted in Fig. 5 that Errorl and Error2, both conesponding to Link! and Link2 error settled within 10 seconds while Enor3, which conesponds to the steering error, settled faster, in about 7 seconds with very small offset, which is quite acceptable. As can be seen from Fig. 6, the first graph represents wheel torque for both the right and left wheels resulting from the FLC for lifting and stabilizing Linkl. It can be seen that both have the same magnitude but differ in direction. The second graph on the right represents the torque for stabilizing Link2 where the actuator is placed between Link] and Link2. Both settled after about 10 seconds. The third graph cones ponds to the output torque from the FLC of steering controller, where the torque for the right wheel has the same values as the left wheel torque. Since wheel torque realizes several functions including lifting, stabilizing, linear motion and steering, the last graph (below right) represents the resultant controller output for lifting and stabilizing controller as well as steering controller. It can be concluded that FLC
1048 works very well on steering motion control of wheelchair on two wheels. Future work will concentrate on further improvement of system model and parameter optimization of the fuzzy logic controllers.
Fig. 4. Final wheelchair system position
-50
o
10
10
20
Time (s)
20
Time (s)
40
5
10
15
20
Time (s)
Fig. 5. Angular position error of Link 1 (Errorl), Link2 (Error2) as well as steering error (Error3) as a function of time
1049 ＲＰｾＭ
E ｾ＠
o
100
C\I
CD
:::J
eo
o
f-
-200
o
10
-100
20
o
lime (5)
E ｾ＠
50
ｾ＠
E
20
200
CD
:::J
0 C')
ｾ＠
CD
e-o
10
lime (5)
-50
c
0
.s
f-
-100
e-
o
:s -200 IｾＭ
10
lime (5)
20 ｾ＠
U)
0
10
20
lime (5)
Fig. 6. From Upper Left: Wheel torque from lifting & stabilizing control for Linkl, torque from lifting and stabilization control for Link2, wheel torque from steering controller and resultant wheel torque of Graph 1 and Graph 3 as a function of time
5. Conclusion An FLC strategy for steering motion of two-wheeled wheelchair has been developed. Two outputs of fuzzy control have been introduced and tested on the system. The proposed FLC has been successfully incorporated into the control scheme with predetermined membership type and related parameters. The results presented demonstrate that the FLC approach works very well on highly nonlinear systems such as the wheelchair system on two wheels and gives good system performance. Future work will concentrate on further improvement of system model, performance and parameter optimization of the fuzzy logic controllers.
References 1. Z. Lin, A. Saberi, M. Gutmann, and Y. A. Shamash, Linear controller for an inverted pendulum having restricted travel: A high-and-Iow gain approach, Automatica, vol. 32, pp. 933-937, 1996. 2. M. Bugeja, "Non-linear swing-up and stabilizing control of an inverted pendulum system, IEEE Region 8 EUROCON on Computer as a Tool, 2003.
1050
3. J. Yi, N. Yubazaki, and K. Hirota, Upswing and stabilization control of inverted pendulum system based on the SIRMs dynamically connected fuzzy inference model, Fuzzy Sets and Systems, vol. 122, pp. 139-152, 200l. 4. N. Muskinja and B. Tovornik, Swinging up and stabilization of a real inverted pendulum," IEEE Transactions on Industrial Electronics, vol. 53, pp. 631-639, 2006. 5. K. Furuta, M. Yamakita, and S. Kobayashi, Swing up control of inverted pendulum, International Conference on Industrial Electronics, Control and Instrumentation (IECON '91), 1991. 6. M. Yamakita, M. Iwashiro, Y. Sugahara, and K. Furuta, Robust swing up control of double pendulum, American Control Conference, 1995. 7. F. Grasser, A. D'Arrigo, S. Colombi, and A. C. Rufer, JOE: a mobile, inverted pendulum, IEEE Transactions on Industrial Electronics, vol. 49, pp. 107-114,2002. 8. K. Pathak, J. Franch, and S. K. Agrawal, Velocity and position control of a wheeled inverted pendulum by partial feedback linearization, IEEE Transactions on Robotics, vol. 21, pp. 505-513, 2005. 9. Y. Takahashi, S. Ogawa, and S. Machida, Front wheel raising and inverse pendulum control of power assist wheel chair robot, The 25th Annual Conference of the IEEE Industrial Electronics Society, 1999. 10. Y. Takahashi, S. Ogawa, and S. Machida, Step climbing using power assist wheel chair robot with inverse pendulum control, IEEE International Conference on Robotics and Automation (ICRA '00), 2000. 11. Y. Takahashi, T. Takagaki, J. Kishi, and Y. Ishii, Back and forward moving scheme of front wheel raising for inverse pendulum control wheel chair robot, IEEE International Conference on Robotics and Automation, 2001. 12. Y. Takahashi, N. Ishikawa, and T. Hagiwara, Soft raising and lowering of front wheels for inverse pendulum control wheel chair robot, IEEEIRSJ International Conference on Intelligent Robots and Systems (IROS 2003), 2003. 13. Y. Takahashi, N. Ishikawa, and T. Hagiwara, Inverse pendulum controlled two wheel drive system, 40 th SICE Annual Conference, 200l. 14. Y. Takahashi and O. Tsubouchi, Modern control approach for robotic wheelchair with inverse pendulum control, 5th International Conference on Intelligent Systems Design and Applications (lSDA '05),2005. 15. D. A. Winter, Biomechanics and Motor Control of Human Movement. New York: Wiley-Interscience, 1990 16. S. Ahmad and M. O. Tokhi, Fuzzy Logic Control of Wheelchair on Two Wheels, lASTED International Conference on Modeling, Identification and Control (MIC 08),2008. 17. S. Ahmad and M. O. Tokhi, Forward and Backward Motion Control of Wheelchair on Two Wheels, 3 rd IEEE International Conference on Industrial Electronics and Applications, (ICIEA 08), 2008.
SECTION-17 PLANETARY EXPLORATION AND LOCALIZATION
This page intentionally left blank
DEVELOPMENT OF AN UNDERGROUND EXPLORER ROBOT BASED ON PERISTALTIC CRAWLING OF EARTHWORMS HAYATO OMORI, TARO NAKAMURA and TAKA YUKI YADA
Faculty of Science and Engineering, Department of Precision Mechanics Chua University 1-13-27 Kasuga, Bunkyou-ku, Tokyo 112-8551 Japan Although shield tunnel construction and tunnel boring machines have developed greatly, these machines are still large in size and consume large amounts of energy. A robot small enough to be able to explore beneath the ground on its own would extend the range of underground investigation both under the Earth's surface as well as on the moon in the future. This study focuses on peristaltic crawling of earthworms as a locomotion mechanism for an underground explorer robot. In peristaltic crawling, extension and contraction waves are propagated in the anteroposterior direction by varying the thickness and length of the earthworm's segments, and a large surface area is brought into contact during motion. Furthermore, it requires no more space than that of an excavation part on the anterior of the robot. The proposed robot consists of several parallel link mechanisms. The robot could move on a plane surface and in vertical perforated dirt, and could climb a tube. The robot showed good performance in the experiments.
1. Introduction A variety of biomimetic robots have been developed including those resembling earthworms. An earthworm moves by peristaltic crawling. Figure 1 shows the locomotion pattern of an earthworm during peristaltic crawling. First, the earthworm contracts its anterior segments. This increases the friction between the segments and the surface, because the thicker segments are in contact with the surface during locomotion. The contraction propagates continuously towards the rear. This movement pulls the rear segments in the direction of movement. After the contraction has completed, the anterior segments of the earthworm are extended in the axial direction. Thus, this type of locomotion mechanism has the following advantages: 1.
It requires less space than other mechanisms, e.g. bipedal, wheel-based and
2.
It is expected to provide stability on the inside of a narrow pipe and in dirt.
meandering locomotion. Therefore, this mechanism is suitable for an underground explorer robot.
1053
1054
Some robotic and biological engineering studies have investigated peristaltic crawling robots based on earthworms [1]-(4]. Various actuation methods have been studied for achieving peristaltic crawling locomotion [5]-[8]. Researchers discussed the influences of various conditions on peristaltic crawling locomotion patterns, and examined the provision of stable adjustment to environmental variations. However, conventional robots do not have 3-DOF. Such robots could be used for various roles if they could be controlled in three dimensions. In this study, we developed a peristaltic crawling robot with parallel link mechanisms. The robot was covered with film to conduct experiments in the dirt. We conducted experiments on its forward movement and turning on a plane surface, together with observing it climbing up and down a tube. Thus, we confirmed that the robot can move in vertical perforated dirt. - where < r j , gj ,!J.i > is the sample color data for person and n j denotes the number of times the jth color tuple has occured in the sample data. Hence, the likelihood of the pixel to a person was computed efficiently as
P(pIDC)
= ｾｌ＠
j n j K(J",. (r - rj)Ku g (g - gj)KUb (b - b ) j
where p =
< r,g,b >
1158
If this likelihood is more than a particular threshold Pth, the pixel is classified as belonging to a person. This is done for every pixel in the patch and using a majority rule. The centriod of a person is computed from the patches classified as belonging to a person. Color model is then periodically updated by the new intensity values obtained from the cluster after identification is done. The robot velocities are controlled by the disparity and the angle of the centroid of person in image plane. The translational veolcity Vtx of robot is proportional to the disparity of the centroid and the rotational velocity Vrl of robot is proportional to the angle of centroid in the image plane. ｣ｬ､ｾ＠
Vtx
Vrl
ｃＲＨＩｾ＠
=
where ｾ＠ is disparity of the centroid of person and ＨＩｾ＠ centroid makes in the image plane at time t.
Oct
=t
an
-1
Ｈｊ［ｾＬ＠
is the angle the
-.f cx)
is the center of the image in x direction. Proximity of the vector ＨｸｾＬ＠ ｹｾＬ＠ ､ｾＩ＠ to its previous position is used as a consitency check, where (x;" ｹｾＩ＠ is the position of the centriod of a person in the image at time t. Figure 1 shows the robot following a person.
Cx
Fig. 1.
Robot following a person
4. Results The proposed method was implemented in C++ on a linux platform (FC7) with AMD Athlon 64-bit processor. The image resolution was kept at 320x240. The algorithm was extensively tested on our lab robot, SPAWN in indoor environments under different conditions such as similar background
1159
color, varying lightening conditions and presence of many static objects. Figure 2 and Figure 3 show that the robot is able to keep track of the person. The media files (format: .avi) showing the robot following a person can be obtained from the following website: http: / / students. iii t . ac. in/-ankurhanda/robot.html.
Fig. 2. Left images showing the motion segmentation after spatial relative velocity based filtering while right images showing the results of segmentation of person after fusing depth and color information
Fig. 3. Tracking of a person under different testing environment,left: same background color and right: poor lightening conditions
References 1. M. Piaggio, P. Fornaro, A. Piombo, L. Sanna and R. Zaccaria. An optical flow based person following behaviour. In Proceedings of the IEEE
ISIC/CIRNISAS Joint
ｃｯｲｾｦ･ｔＢｮ｣Ｌ＠
1998.
2. C. Schlegel, .1. Illmann, H . .1aberg, M. Schuster and R. Worz. Vision based person tracking with a mobile robot. In The British Machine Vision Con-
ference, 1998.
1160 3. Z.Zivkovic. Improved adaptive Gausian mixture model for background subtraction. International Conference Pattern Recognition, Vo1.2, pages: 28-31, 2004. 4. Wren C., A. Azarbayejani, T. Darrell, and A. Pentland. Pfinder: Real-Time Tracking of the Human Body. IEEE Transactions on Pattern Analysis and Machine Intelligence, Vo1.19, pages:780-785, 1997. 5. Y. Raja, S. McKenna, S. Gong. Object Tracking Using Adaptive Colour Mixture Models,Pmc. ACCV 98, Vol. I, pp 615-62266. 6. B.K. Horn and B.G. Schunck. Determining optical flow. Artificial Intelligence,VoI.17,pages: 185-203,1981. 7. A. Behrad, A. Shahrokni, S. A. Motamedi and K. Madani.A Robust Visionbased Moving Target Detection and Tracking System. In Proceedings of Image and Vision Computing conference (I VCNZ2001) , University of Otago, Dunedin, New Zealand, November 26-28, 2001 8. B. Jung and Gaurav S. Sukhatme. Detecting Moving Objects using a Single Camera on a Mobile Robot in an Outdoor Environment. In International Conference on Intelligent Autonomous Systems, pp. 980-987, Amsterdam, The Netherlands, Mar 2004. 9. H. Kwon, Y. Yoon, J. B. Park and A. C. Kak. Person tracking with a mobile robot using two uncalibrated independently moving cameras. In Proceedings of IEEE International Conference on Robotics and Automation (ICRA), 2005. 10. C. McCarthy and N. Barnes, "Performance of optical flow techniques for indoor navigation with a mobile robot", Proceedings of IEEE ICRA, pp 5093-5098, 2004. 11. Z. Chen and Stanley T.Birchfield. People Following with a robot using Binocular feature based tracking. In IEEE International conference on Intelligent Robots and Systems (IROS) , 2007. 12. J.L. Barron, D.J. Fleet, and S.S. Beauchemin, Performance of Optical Flow Techniques, Int 1. Computer Vision, 1994. 13. M.J. Black and P. Anandan, A Framework for the Robust Estimation of Optical Flow, Pmc. Int. Conf. Computer Vision, pp. 231-236, May 1993. 14. M.J. Black and P. Anandan, "The Robust Estimation of Multiple Motions: Parametric and Piecewise Smooth Flow Fields," Computer Vision and Image Understanding, Vol. 63, pp. 75-104, 1996.
DEVELOPMENT OF A SIMULATION ENVIRONMENT OF AN ENTERTAINMENT HUMANOID ROBOT DOING A HANDSTAND ON A HIGH BAR PEDRO TEODORO MIGUEL AYALA BOTTO CARLOS CARDEIRA JOSE sA DA COSTA JORGE MARTINS IDMECI/ST, TULisbon A v. Rovisco Pais 1049-001 Lisboa, Portugal email: {pedroteodoro@, ayalabotto@, carlos.cardeira@, martins@dem. sadacosta@dem.} ist.utl.pt.
LIMOR SCHWEITZER RoboSavvy Ltd, 37 Broadhust Gardens, London, United Kingdom, [email protected]. This paper presents an LQR controller approach for the simulation and controls of an affordable commercial humanoid robot doing a handstand on a high bar, by considering it as an underactuated 3-link inverted pendulum with off-centered masses. The developments presented include: i) the software development for interfacing the Matlab® Real Time Workshop Toolbox® with the humanoid robot servos; ii) the identification of the internal and external dynamic parameter of the humanoid servos and structure, respectively; iii) the dynamics modeling and simulation of the humanoid robot; iv) the deduction of the equations of motion for an underactuated n-link inverted pendulum. Simulation results proved the adequacy of LQR controller. Keywords: Humanoid robot; LQR; modeling; underactuated inverted pendulum; nonminimum phase system.
1. Introduction
Linear quadratic regulator (LQR) controller has proven its capability in the stabilization of 2-link underactuated pendulum, named pendubot in the vertical unstable position [1-3]. However, studies for a real humanoid robot while doing a handstand is still lacking. The main issues resembling this are the following: i) difficulties in obtaining the real physical properties of the robot (masses, inertia
1161
1162
tensors, centers of mass); ii) higher inertia tensors of the main parts of the robot as opposite to thin rods; iii) off-centered masses; iv) servo limitations. Therefore, this paper pretends to fill up this gap by undertaking a full analysis over the highly nonlinear system to be studied, in this case, thc Bioloid humanoid from Robotis.com and controlling it at his unstable vertical position by implementing the LQR optimal linear controller [4]. The system is considered to be composed of three main parts, arms, torso and legs, where the first link is passive and the hands rotate freely around a bar. The paper is structured as follows: Section 2 presents the humanoid. In section 3, the hardware and software architecture for communicating between PC using Matlab and servos is presented. Section 4 addresses the strategy undertaken in the identification of the physical properties of the humanoid structure and the internal behavior of the servos. In section 5, the equations of motion for a generalized n-link underactuated inverted pendulum and its analogy to the humanoid robot are described. The approach taken for the off-centered masses of the main parts of the robot and its simulation results are shown as well in section 5. Finally, the conclusions are presented in section 6.
2. Humanoid The humanoid robot considered is the Bioloid (Figure 1) from Korean manufacturer Robotis.com, due to its well designed servo controllers that provide current, voltage, position and temperature sensing. It has a well documented open controller board and the servo control protocol is well documented. The humanoid Robot has 18 degrees of freedom (DOF) powered by DC servos. Main Blocks
Schematic of joint and link
o Lower Arm o Arm III III Torso III Groin L£J Hip
III Leg III Leg III Ankle III Foot
YYaw
h
ZPitcn
X Roll
Figure 1. Bioloid humanoid
1163
3. Hardware and Software architecture The humanoid robot set up and the control architecture adopted is shown in Figure 2. The humanoid controller named CM-5 is connected to the controllers of thc servos through a RS485 bus and to the PC by RS232. The computer running Simulink / RTW / Real Time Windows Target has the device drivers to send/receive data to/from the servos using the appropriate protocol. To this end, a C-MEX S-function written in C to communicate with the CM-5 throughout UART (universal asynchronous receiver / transmitter) was created. Finally, a C program for Atmega128 for completing the serial communication bridge was created.
Figure 2. Hardware and software architecture
4. Humanoid identification 4.1. Mechanical properties identification An accurate static model of the Humanoid Robot can be obtained based on the physical properties of their components. Typically, by knowing the mass, center of mass and the inertia tensor of each element of the humanoid robot it is possible to get a quite reliable model that can be further used in simulation and control. For quantifying the masses of each element, a precision scale with a resolution of 0.05 grams was used. The centroid of each mass was then found by using the SolidWorks® software package, after the detailed elements of all the pieces involved were drawn in this 3D CAD software. It was assumed here that, except for the servos, all the pieces are of isotropic nature. A simple experiment has shown that the maximum error obtained for the geometric position of the centroid is of 0.5 mm on each Cartesian direction. Finally, the inertia tensor of each element was determined through the SolidWorks® software.
1164
4.2. Servo identification A set of output signals can be retrieved from the humanoid robot servos. These signals provide information regarding the actual servos angular position, angular velocity, DC current, temperature and voltage. The angular position, temperature and voltage signals are sampled at 100 Hz, while the angular velocity and load are sampled at a rate 10 times slower. For the identification of the dynamic behavior of the servos it was considered classical prediction error method [5] and the relation between the reference input velocity and the correspondent estimated velocity. Detailed results about the behavior, identification and validation of the servo can be found in [6].
5. Control and Simulation Results
5.1. Equations of motion In order to have the humanoid robot doing the handstand on a high bar, a special configuration of the humanoid robot was studied. In this configuration the humanoid is seen as being compound of three main blocks (arms, torso and legs) with two active joints (shoulder and hips), resembling an underactuated 3-link inverted pendulum (Figure 3). The angular displacement and velocity of the link i (li) is represented respectively by qi and iIi' Table 1 shows the humanoid robot mechanical parameters.
x Figure 3. Representation of the humanoid robot as an underactuatcd 3-link inverted pendulum
1165
Table 1. Humanoid robot mechanical parameters
Link 1 (Arms) Link 2 (Torso) Link 3 (Legs)
I (mm)
Ie (mm)
143.6 115.8 184.0
68.7 57.5 116.3
m (g) 367.6 98l.5 576.4
I (gem)
7890.7 32898.6 11328.0
The equations of motion for a generic n-link underactuated inverted pendulum, Equation (1), were deduced from the Euler-Lagrange equations [6,7]. Resolving for n=3 [2], the equations of motion for the humanoid robot are obtained.
mil [ｾｮｬ＠
m21
ml2 22 m
min 2n m
m n2
ｾｮ＠
j[ijlij2 j+ [9192 j+[hih2 j=[III2 j ｾｮ＠
ｾｮ＠ ｾＳ＠
(1)
;n
With:
i:5, j
n
hi =C i +
Lmij' j=i
Ib-CI sin [ ｱ､ＫｭｩｮＨｨＮｃＩＧ＠ｾ
and I. = I
Ie { Ii
i = k
I
i =F k
)
ｩｦｭｮＨ｢Ｌ･ＩｾＱ＠
(2)
1166
where mij are the inertial terms, ¢i are the gravitational terms, hi are the Corriolis and the centrifugal terms, are the input torques.
"i
5.2. State space model The above equations of motion of the system are highly non-linear. Therefore, and in order to use linear control algorithms, the system dynamics is linearized by using a first order Taylor's expansion at the vertical unstable equilibrium, q =[nl2,O,O]T and q =[O,O,O]T. Letting the state space vector x=[ q]- rrl2, q2' qo,' q] , q2 ' q3]T, yields:
x=Ax+Bu
(3)
where A is 6x6 matrix and B a 6x2 matrix.
5.3. Linear quadratic regulator The system presents unstable zeros and poles, being a non-minimum phase one. Therefore an optimal or nonlinear control technique is desirable. Since, LQR [2, 5] has proved its capability in similar problems [1-4] and because it is a popular stabilization technique which provides a linear state feedback control law for the system, it has been chosen. This control law has the following form:
(4) where the design gains are: kT =[-114.7206 92.6369
-63.3382 -41.0629
-23.2078 -13.3245
-25.2863 -15.8651
11.8534 -5.0644J (5) 8.1298 -2.2096
5.4. Results The plots shown in Figure 4 (a,c,e) show the behavior of the LQR controlled system when the state vector is defined as x =[ q] - nl2, q2' q3' q] , q2 ' Q3]T. As it can be seen, the system could not be stabilized. The reason is simple, since the effect of the real localization of the centers of gravity of the arms; torso and legs were not taken into account when building the model. Nevertheless, this problem can be solved, by finding and compensating the system with the angle of the resultant center of gravity (plots of Figure 12 (b,d,f), leading to a stabilized system. As it can be verified, the input torque is under 2 Nm, which is the limitation torque of two servos. Moreover, best results were achieved by tuning the angle compensations for each body. In this case, the input torque do not exceeded 0.3 Nm for both shoulder and hip joints.
1167
Angular Displacements
Ang u lar Displacements
lOOr-__________------
ｾ＠
00 ｾ＠
E
ｾ＠
_00
0...
.r..2
o ｾ＠
.lll
-100
"01-100
J(
o---=C-C---O:-:.2
0.3
-:DJ O
0;1
05
Tim e l.serond)
(.t\} \ViLhollt
｡ｮｰ［ｬｾ＠
Time (seoond)
ｃｏｴｮＮｰｎｉｈｬｩｾｪ＠
Angular Velocity
A ng u lar Velocity 0.4
ｯｾＭＬ＠ Ｎｾ＠
ｾ＠
,
'8"
-400
>
ＭｦＺｊｾ＠
t-
'"8
o
ＺＭｾ＠
o -0.2
.,-:;":.'''' -
,.,," ,. ,. ,,
.,.
¢
veb::nv of q, ___ veb::ity of q,.
-WJ
j ｾ＠
-:DJ
(ii
0.2
> -0.4
I
,
,
..
0.1
4
0.2 0.3 Time (serond)
Time (second)
Applied Torq ue
Applied Torque
Time (second)
Ti me (second)
1000
ｏｆＭｾ＠
tfl \'/ith angle
(,-Orn.r.. ｾｮＮｳ｡ｴｩｯ＠
Figure 4. Handstand simulation using LQR controller without and with angle compensation
1168
6. Conclusions In this paper it has been shown the possibility to control a real affordable commercial humanoid robot doing a handstand on a high bar by implementing an optimal LQR controller in simulation. The humanoid was treated as a three body serial chain in an inverted pendulum configuration. The system is underactuated, being the motion of the legs and torso prescribed in order to stabilize the full body of the humanoid above the high bar. Despite of the higher inertias of the main parts of the robot, and the off-centered centroids, the LQR controller stabilized the system, being the input control actions bellow than the maximum allowed torque of the servo (-INm). In future works, the servo dynamics will be considered in the system modulation.
Acknowledgments This work was suported by Robosavvy.com and partially funded by FCTFunda