Transcript Adaptive Intelligent Mobile Robotics Leslie Pack Kaelbling, PI

Adaptive Intelligent Mobile Robotics
Leslie Pack Kaelbling, PI
MIT Artificial Intelligence Laboratory
Hierarchical Domain Decomposition for Probabilistic Planning
• High-level goal determines reward at
exit states
• Construct decomposition off line
• Solve for macro operators
• Combine pre-computed value
functions to determine near-optimal
action
• Plan for new goals in time
logarithmic in plan length
Planning
• Trade optimality for efficiency
Skill Learning
Built-in Behaviors
Vision-Based Navigation
Optical flow gives estimated distance to objects
Practical Reinforcement Learning
• Human guidance generates efficient exploration
• Locally weighted regression provides fast function approximation
• Uncertainty modeling and experience replay cause fast value
propagation
Phase One
Phase Two
Environment
Environment
R O
A
Supplied
Control
Policy
Comparison of potential-field method to empirically discovered human
control laws for local navigation
R O
Supplied
Control
Policy
Learning
System
Learning
System
Corridor-following Task
Average steps to goal
Phase 1
Steps to goal
125
Phase 2
Average training
105
85
“Best” possible
65
Current work: acquiring topological maps based on these primitives
-25
-15
-5
5
Training runs
15
25
A