Learning Feasibility and Cost to Guide TAMP

Recent work in Task and Motion Planning (TAMP) has enabled a new class of algorithms that can better take advantage of off-the-shelf samplers and solvers to find solutions to sub-problems in a task plan, such as motion between configurations, or inverse kinematics solutions. However, not all sub-problems are equally valuable. Existing planners typically rely on heuristics to determine which sub-problem to attempt to solve next, unable to reason about the expected cost of doing so in the broader context of the full plan. In this work, we present a novel approach for TAMP, utilizing learned models to inform when to attempt to solve potentially expensive sub-problems. We test our approach in two simulated domains, as well as on a real Panda robot, showing improvement in planning and execution time compared to a heuristic driven baseline.