Keywords: Task and Motion Planning, Learning
TL;DR: We train several small models to guide search in task and motion planning.
Abstract: Recent work in Task and Motion Planning (TAMP) has enabled a new class of algorithms to better take advantage of off-the-shelf black-box samplers and solvers to find plans. However, many approaches to solving TAMP problems must balance the need to explore new regions of the search space with the computational cost of querying solvers for sub-problems that may be unlikely to succeed. In this work, we present a novel approach for solving TAMP problems, utilizing learned models trained from experience to inform when to attempt to solve potentially expensive sub-problems. We take advantage of highly optimized classical planners by learning representations that can be integrated with existing abstractions to guide search in long-horizon TAMP domains.