1 / 49

Multiagent Control of Modular Self-Reconfigurable Robots

Multiagent Control of Modular Self-Reconfigurable Robots. Tad Hogg HP Labs. Hristo Bojinov Jeremy Kubica Arancha Casal PARC’s modular robotics group. topics. modular robots multi-agent control results. modular robots. collections of modules each module is a robot self-reconfigurable

lucinda
Download Presentation

Multiagent Control of Modular Self-Reconfigurable Robots

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Multiagent Control of Modular Self-Reconfigurable Robots Tad Hogg HP Labs Hristo Bojinov Jeremy Kubica Arancha Casal PARC’s modular robotics group

  2. topics • modular robots • multi-agent control • results

  3. modular robots • collections of modules • each module is a robot • self-reconfigurable • modules can change connections • so overall robot changes shape • “modular self-reconfigurable” robots • MSR

  4. why change shape? • adjust shape to task • e.g., locomotion • wheel, spider, snake, … • e.g., manipulation • match “finger” size to object size

  5. topics • modular robots • Proteo • Prismatic • future possibilities • multi-agent control • results

  6. Proteo • rhombic dodecahedron • space filling

  7. Proteo • modules move over neighbors each edge of cube is a diagonal of RD face

  8. topics • modular robots • Proteo • Prismatic • future possibilities • multi-agent control • results

  9. prismatic MSR robots • modules connect via arms • extending arms moves neighbors • examples • Crystalline robot (Dartmouth) • moves in 2 dimensions • TeleCube (PARC) • moves in 3 dimensions

  10. TeleCube • cubes • 6 independent arms • 2:1 length ratio

  11. physical move virtual move contract expand neighbors cooperate to move

  12. actual modules topics • modular robots • Proteo • Prismatic • future possibilities • multi-agent control • results

  13. devices for “smart matter” • micro-electomechanical (MEMS) • bacteria • molecular • quantum sensor + computer + actuator

  14. made with photolithography e.g., programmable force fields (open loop) hard to assemble micromachines (MEMS)

  15. biotechnology: program bacteria e.g., T. Knight, R. Weiss at MIT AI Lab limited abilities biological machines

  16. programs for bacteria • gene regulatory networks • engineered changes give some program control over behavior

  17. ribosomes: make proteins in cells protein motors move material in cells ATP synthase rotor size: 10nm molecular machines protein DNA mRNA See Nature, 386, 299 (1997)

  18. carbon nanotubes and buckyballs strong, light, flexible, electronic devices easy to make hard to arrange molecular machines

  19. complex molecules for robot parts currently: only theory hard to make hard to assemble potential: cheap, fast, strong parts molecular machines example medical applications: R. Freitas, Jr., Nanomedicine, 1999 example designs: E. Drexler, R. Merkle, A. Globus

  20. potential: much faster algorithms e.g., factoring very difficult to build many worlds inside quantum computers quantum search heuristic amplitudes while solving a 10-variable 3-SAT instance with 3 solutions Java demo: www.hpl.hp.com/shl/projects/quantum/demo

  21. potential: detail control over materials e.g., interfere two ways to absorb light => transparent very difficult to build quantum machines See T. Hogg and G. Chase, Quantum smart matter, 1996 www.arxiv.org/abs/quant-ph/9611021 S. Lloyd and L. Viola, Control of open quantum systems dynamics, 2000 www.arxiv.org/abs/quant-ph/0008101

  22. example: coin weighing puzzle quantum sensor finds bad coin in single try quantum machines See B. Terhal, J. Smolin, Single quantum querying of a database, 1997 www.arxiv.org/abs/quant-ph/9705041

  23. devices: summary • smaller devices • harder to make • harder to connect, assemble • greater potential capability • but need many, cheap devices • statistical or systems view

  24. challenge: How to build? • physical/engineering constraints • unreliable parts • misconnected • limits early technology • economics • build cheaply

  25. challenge: How to use? • information/computational constraints • limited, changing info from environment • computational complexity • e.g., planning optimal device use • limits even mature technology control

  26. topics • modular robots • multi-agent control • results

  27. control challenge • coordinate many modules • sensor & actuator errors • decompose programming task to only need • local info (small scale) • high-level task description (large scale) • e.g., grasp object of unspecified shape cf., H. Simon: nearly decomposable systems

  28. control before hardware? • many, small modules don’t yet exist • hence, hardware details unknown • but can study general issues • control may simplify hardware design • e.g., manage in spite of defects • identify compute/communicate tradeoffs sensor + computer + actuator

  29. physics vs. size gravity friction Brownian motion thermal noise decoherence MEMS molecular quantum faster smaller harder to build

  30. multi-agent control • matches control to physics • different agents for each scale • matches control to available info • rapid response to local info • manager agents: overall coordination • without need for details

  31. motivation: biology social insects, multicellular organisms, ecology reliable behavior from unreliable parts examples termite mounds embryo growth cf. incentive issues noncooperative agents economics, common law, …

  32. motivation: teams • robot soccer • insect-like robot teams • e.g., foraging • MSR robots have • tighter physical constraints • direct access to neighbor locations • e.g., no need for vision to find neighbors

  33. topics • modular robots • multi-agent control • results • computational ecology • Proteo • Telecube

  34. computational ecology • dynamical behavior of simple agents • asynchronous, local decisions • delays, imperfect information • “mean-field” statistical theory • see B. Huberman, The Ecology of Computation, 1988 • apply to actual robot behaviors • see K. Lerman et al. in Artificial Life, 2001

  35. techniques • finite-state machine for each module • simple script, some randomness • local communication • create gradients through structure • “scents”

  36. topics • modular robots • multi-agent control • results • computational ecology • Proteo • Telecube

  37. example: growing a chain • modes: • SLEEP, SEARCH(red), SEED(yellow), FINAL(white) • initially: all in SLEEP, randomly pick one SEED • seed: • picks growth direction • emits scent • attracts modules

  38. growing a chain descend gradient + propagate scent emit scent=0 SEARCH SEED if neighbor is seed detect scent if neighbor became seed SLEEP FINAL propagate scent

  39. scent S+1 S S-1 • set S=min(neighbors)+1 • move around neighbor until lower value found • if seed found: become new seed SEED S = 0

  40. structures • recursive branching • multilevel arms • grow around object • using contact sensors See H. Bojinov et al., Multiagent Control of Self-reconfigurable Robots, 2000 www.arxiv.org/abs/cs.RO/0006030

  41. topics • modular robots • multi-agent control • results • computational ecology • Proteo • Telecube

  42. locomotion • make snake shape • move toward goal • barrier • follow wall • find gap • higher-level control: general direction • building on low-level agent behavior see: Kubica et al, Proc. ICRA 2001

  43. object manipulation • exert forces to move object • based on contact with object • “scent” recruits other modules • modules on surface form rigid shell

  44. summary • simple agents perform basic tasks • reconfiguration • locomotion • manipulate objects • apply to different hardware types • Proteo: surface motions • TeleCube: internal motions

  45. future directions • quantify capabilities • design more complex behaviors • implement on hardware

  46. quantify capabilities • examples of agent-based control • are only specific instances • quantify • how robust? • how accurate? • what cost? • e.g., power use

  47. agent design • combine with higher-level agents • e.g., switch among low-level behaviors • automate agent design • e.g., genetic algorithms (FXPAL)

  48. test on hardware • various existing robots • few, fairly large modules • large number of tiny modules • don’t yet exist • wait for hardware vs. simulate? • understand likely hardware capabilities • e.g., MEMS, …

  49. conclusions • agent-based control for MSR robots • gives robust low-level behaviors • simplifies higher-level task control • biological system models • suggest module rules • useful even if not biologically accurate www.hpl.hp.com/shl/people/tad

More Related