Self-Reconfigurable Robot - A Platform of Evolutionary Robotics

2,376 views

Published on

Published in: Technology
0 Comments
4 Likes
Statistics
Notes
  • Be the first to comment

No Downloads
Views
Total views
2,376
On SlideShare
0
From Embeds
0
Number of Embeds
2
Actions
Shares
0
Downloads
248
Comments
0
Likes
4
Embeds 0
No embeds

No notes for slide

Self-Reconfigurable Robot - A Platform of Evolutionary Robotics

  1. 1. Keynote Speech Alife9 Sept. 14, 2004 Boston Self-Reconfigurable Robot - A Platform of Evolutionary Robotics Satoshi Murata Tokyo Institute of Technology / AIST murata@dis.titech.ac.jp
  2. 2. Outline Introduction Self-reconfigurable systems Modular transformer (M-TRAN) Demonstration of M-TRAN
  3. 3. Introduction
  4. 4. Hierarchy in biological system Homo/heterogeneous layers alternately appear in biological system (Masami Ito) Species hetero Individual homo Organ hetero Cell homo Organelle hetero Molecule homo
  5. 5. Heterogeneous systems Made of heterogeneous components Centralized Sequential Global interaction Design principle --- Reductionism
  6. 6. Homogeneous systems Made of homogeneous components Distributed Parallel Local Interaction Design principle --- Self-organization
  7. 7. Advantages of homogeneity Scalability Enlarge / reduce system size in operation Redundancy Fault tolerance Self-repair Flexibility Self-assembly Self-reconfiguration
  8. 8. Self-assembly in different scales Molecular self-assembly Small, simple, Proteins, DNA tiles, etc. a large number of elements, difficult to control Mesoscopic self-assembly Particles, bubbles, E-coli, etc. Robotic self-assembly Large, complicated, Modular robots a small number of elements, Mobile agents programmable
  9. 9. Self-reconfigurable systems
  10. 10. Self-reconfigurable systems Artifacts based on homogenous modular architecture Change their shape and function according to the environment (Self-reconfiguration) Able to assemble itself, and repair itself without external help (Self-Assembly, Self-Repair)
  11. 11. Homogeneous modular architecture The system made of many (mechanical) modules Each module is identical in hardware and software Each module has computational and communication capability Each module can change local connectivity
  12. 12. Self-assembly and self-repair Random shape Assemble target shape Detect failure Cutting off Reassemble
  13. 13. 2-D Regular Tessellations
  14. 14. 2-D Self-reconfigurable hardware Metamorphic robot (G.Chirikjian, JHU,93) Micro-module (MEL, 98) 2-D Crystaline (M.Vona, D.Rus, Dartmouth Col./MIT)
  15. 15. Fracta (Murata, 93) Solid state module based on hexagonal lattice
  16. 16. Basic operations of fracta
  17. 17. Self-assembly problem How to change connectivity among modules to achieve target configuration ? Random You must consider • Modules are homogeneous • Parallel and distributed • Only local communication • Physical constraints Given
  18. 18. Example: Self-assembly of fracta Parallel algorithm based on connection types and local communication Connection types Target shape
  19. 19. Program code o(K,K) K(o,K,K,s) s(K,K,K,K,K,K) Local configurations Exchange connection type with neighbors
  20. 20. Parallel distributed algorithm for self-assembly 1. Each module evaluates distance to the nearest target configuration in the program code 2. Modules compare the evaluation through simulated diffusion 3. Module which wins among the neighbors moves to random Type transition diagram direction defines metric among connection types
  21. 21. Difficulties in 3-D hardware More mobility in limited space Spatial symmetry requires more degrees of freedom More power/weight Mechanical stiffness
  22. 22. Space filling polyhedra Rhombic Truncated Regular cube dodecahedron octahedron
  23. 23. Lattice based designs Design based on cube Design based on rhombic dodecahedron 3-D Crystaline (M. Vona, D.Rus,Dartmouth, MIT) Proteo (M.Yim, PARC, 2000)
  24. 24. Lattice based designs Design based on cube Molecule (Kotay, Rus, Dartmouth/MIT) 3-D Universal Structure (MEL, 98)
  25. 25. Chain based designs PolyBot: M.Yim ,Xerox PARC CONRO: W-M.Shen, P.Will, USC
  26. 26. Lattice or chain ? Lattice based designs Reconfiguration is easy Motion generation is hard Requires many connectors & actuators Chain based designs Reconfiguration is hard Motion generation is easy Insufficient stiffness
  27. 27. M-TRAN (Modular Transformer)
  28. 28. M-TRAN(Modular Transformer) Hybrid of lattice and chain based designs Easy self-reconfiguration and robotic motion Two actuators Communication Stackable Battery driven
  29. 29. M-TRAN II
  30. 30. M-TRAN Module
  31. 31. Li-Ion battery Non-linear spring Light bulb PIC Connecting plate Main CPU Power supply circuit Permanent magnet PIC Neuron chip SMA coil Acceleration sensor M-TRAN II
  32. 32. M-TRAN I
  33. 33. Magnetic connection mechanism Magnet Light bulb Distance SMAcoil Non -linear spring SMA Actuator Force (a) Attraction by magnets Force Repulsion by springs Detach (b) (c) 0 10 20 30 40 50 60 70 80 90 100 a-b Temperature (ºC) Distance(mm)
  34. 34. New prototype M-TRAN III Hook connection mechanism • Quick • Reliable
  35. 35. Coping with complexity Because of physical constraints such as Maintain connectivity Avoid collision Limited torque Non-isotropic geometry of M-TRAN module makes self-reconfiguration very difficult Complexity can be relaxed by Automatic acquisition of rule set Heuristics (structured rule set) Periodical pattern in structure
  36. 36. Wall climbing 600 rules (no internal state) 18 rules (with internal state) Generated by software Hand-coded
  37. 37. Creeping carpet
  38. 38. Robot maker (structured rule set)
  39. 39. Rhythmic motion generation Central Pattern Generator (CPG) Connected neural oscillators Oscillators entrain phases mutually Feedback of physical interaction Mechanical interaction Motor control Angle feedback CPG Neural connection (CPG network)
  40. 40. CPG Antagonistically connected pair of nonlinear oscillators Output from other CPGs CPG y1i Extensor Neuron ue τ τ’ Σ u1i β v1i Joint angle feedback f1i Extensor m1 y1i = max(0, u1i ) – Input to Output to motor i w0 + m2 y2 i = max(0, u2 i ) Other CPGs f2i Flexor Joint angle feedback Σ u2i β v2i y2i τ τ’ ue Flexor Neuron Output from other CPGs (Taga 95, Kimura 99)
  41. 41. CPG network Generate stable walk pattern (limit cycle) y Inhibitory Excitatory z connection connection x CPG
  42. 42. CPG network tuned by GA Simulation space GA optimizes Given topology of robot Connection matrix of CPG Initial set of individuals Joint angles in initial posture Converge? Dynamics Simulation by evaluating Yes Generation +1 Energy consumption Mutation, crossover per traveled distance Selection Download to modules
  43. 43. Dynamics Simulation Before GA After GA Vortex simulator (CML)
  44. 44. -3 -2 -1 0 1 2 3 1 21 41 61 81 101 121 141 161 181 201 221 241 261 281 301 321 341 361 381 401 421 441 461 481 501 521 541 561 581 601 Forward 621 641 Obtained CPG network for 4-leg walker 661 681 Symmetric connection is obtained -1 +1
  45. 45. Real-time morphology control Adapt morphology suitable to the environment Rapidly-Exploring Random Trees (RRTs)
  46. 46. Self-reconfigurable robots ~ A new kind of artifacts Amoeba Reconnection to cluster Locomotive flow of periodic cluster Individual Producing individual agents Morphing Swarm
  47. 47. Conclusion Self-reconfigurable systems give a platform upon which we can investigate both individual adaptation and morphological evolution concurrently in a single framework. In this sense, self-reconfigurable systems open the new possibility of artifacts beyond natural evolution.

×