Adaptive%20Intelligent%20Mobile%20Robotics%20Leslie%20Pack%20Kaelbling,%20PI%20MIT%20Artificial%20Intelligence%20Laboratory

1
Adaptive Intelligent Mobile RoboticsLeslie Pack
Kaelbling, PIMIT Artificial Intelligence
Laboratory
Hierarchical Domain Decomposition for
Probabilistic Planning

• High-level goal determines reward at exit states
• Combine pre-computed value functions to
  determine near-optimal action

• Construct decomposition off line
• Solve for macro operators
• Plan for new goals in time logarithmic in
  plan length
• Trade optimality for efficiency

Vision-Based Navigation
Practical Reinforcement Learning

• Human guidance generates efficient exploration
• Locally weighted regression provides fast
  function approximation
• Uncertainty modeling and experience replay cause
  fast value propagation

Optical flow gives estimated distance to objects
Comparison of potential-field method to
empirically discovered human control laws for
local navigation
Corridor-following Task
Phase 1
Phase 2
Average training
Best possible
Current work acquiring topological maps based
on these primitives