Zoom Link

Abstract:
Heuristic search-based motion planning can be computationally costly in large state and action spaces. In this work we explore the use of generative models to learn contextual actions for successor generation in heuristic search. We focus on cases where the robot operates in similar environments, i.e. environments drawn from some underlying distribution. Intuitively, in such cases the robot is bound to observe similar looking local regions of the environment over the course of its operation. We use a conditional variational autoencoder (CVAE) to learn a distribution over contextual actions given this local map and a goal location. These contextual actions are used to help the search make faster progress towards the goal, and avoid or get out of local minima along the way. We show simulation results for kinematic planning problems in a variety of 2D environments for motion planning for a point-robot and a planar arm with up to 5 degrees-of-freedom. Our approach outperforms traditional search-based planning algorithms in terms of computational cost (number of expansions) while maintaining bounds on suboptimality.

Committee:
Maxim Likhachev (advisor)
Sebastian Scherer
Oliver Kroemer
Devin Schwab