Slide 1 - CMU Gamedev Server

Document Sample
Slide 1 - CMU Gamedev Server Powered By Docstoc
					Footstep Planning
  for the Honda
ASIMO Humanoid

 Joel Chestnutt, Manfred Lau,
German Cheung, James Kuffner,
Jessica Hodgins, Takeo Kanade
• To achieve navigation
  autonomy, stable
  walking bipeds need:
  – Ability to walk through
    real environments
  – To take advantage of
    their biped capabilities
  – To be cautious of the
    limitations of their legs
     Two extremes of approach

                                     [M.Kallmann et al. 04]

Ignore details of legged
locomotion, treat as       General motion planning
wheeled mobile robot.      for all degrees of freedom
                    Other approaches
[O.Lorch et al. 00]

                      Modify path and step
                      length while walking to
                      reactively avoid obstacles.
                                                    [Gutmann et al. 05]

[Okada et al. 05]
       Our approach: Encode the
        capabilities of the biped
• Use a representation that           [Chestnutt et al. 03]
  captures the robot’s abilities,
  but without the details of how it
  will be accomplished
• Generate a plan within the
  limits of the biped’s
• Allow the walking controller to
  handle the details of executing
  the path.
Planning algorithm
        A* Search – returns the optimal
        path with an admissible heuristic
   The Honda ASIMO: a different
 representation of ability is needed
• Robust, reliable walking
• Only a high-level API
  – No direct control of foot
  – No control of swing leg
• Results of commands are
  dependent upon robot
• Limited state information
Influence of robot dynamics
  on the effects of actions

              right                 right
              foot                  foot

  Standing still      Walking forward
   Dealing with the effects of actions
• Need a new function
   T: S £ A £ E! S
• For ASIMO, the state of
  the robot can be
  completely determined
  by the last two
TASIMO(ai-2, ai-1, ai, e) = si+1
TASIMO determined experimentally
   using motion capture data
     Now able to find a series of
 footsteps that ASIMO can execute
• Open-loop plan
• No re-planning or
• Dynamic effects
  already accounted for
  in the plan
      Dynamic environments
• Known dynamic environments
  – Predictable motion
  – Plan with time as part of the state
  – Change environment representation to be
    able to predict future terrain
• Unknown environments
  – Include sensing of environments
  – Real-time replanning to account for new
  Planning for predicted motion
• Include time in states
  and actions.
• Plan into the future
  based on what the
  environment should
  evolve to.
  Planning for unknown motion
• Incorporate sensing as the
• Limit planning time, but
  continually replan during
  walking so that a new plan
  is generated each step
• Reuse old plans when
• ASIMO presented some unique challenges
  – Only high-level control available
  – State-dependent actions
  – Limited state information
• New representation deals with state-
  dependence, and allows us to plan for using
  ASIMO’s legged capabilities
• Can deal with known dynamic environments in
  advance, and unknown dynamic environments
  through real-time replanning
                Thanks to
•   Honda Motor Company
•   Philip Michel
•   Koichi Nishiwaki
•   Satoshi Kagami
•   James Kuffner
•   Manfred Lau
•   German Cheung
•   Jessica Hodgins
•   Takeo Kanade
 A mobile robot heuristic trades off
    some optimality for speed
• Not an admissible
  heuristic, but generally
  more informed
• Can severely
  overestimate or
  underestimate in certain
• Performs better the more
  the optimal path can be
  followed by a mobile