Explore an integrated strategy for planning, perception, state-estimation, and action in complex mobile manipulation domains through this lecture by MIT's Tomás Lozano-Pérez. Delve into planning in the belief space of probability distributions over states using hierarchical goal regression. Discover a vocabulary of logical expressions describing sets of belief states and learn how a small set of symbolic operators can generate task-oriented perception for manipulation goals. Witness the implementation of this method in simulation and on a real PR2 robot, demonstrating robust and flexible solutions to mobile manipulation problems involving multiple objects and substantial uncertainty. Gain insights into topics such as POMDPs, complex conditional effects, geometric heuristics, hierarchical planning, belief updates, and motion planning with uncertainty.
Overview
Syllabus
Intro
Real robot meets real world
Action selection is the driving problem
POMDP: Optimal solution is complex...
Steps toward a principled approximation
Complex conditional effects — Implicit predicates
Geometric - Extended "delete" heuristic
Postponing preconditions creates hierarchy
Hierarchical plan
Hierarchical — Goal Regression
Heuristic for regression planning
Outcome uncertainty — Replanning
Update belief based on perceptual information
Belief update and inference strategy
Uncertainty about geometry
Motion planning with uncertainty
Logical representations of uncertain geometry!
Finding the soda (blue box)
Trying to move the soup can out of the way
Picking up the soda box
Meta-cognitive learning: improving planning
Taught by
Paul G. Allen School
