Sequence-Based Plan Feasibility Prediction for Efficient Task and Motion Planning. (arXiv:2211.01576v1 [cs.RO])

Robots planning long-horizon behavior in complex environments must be able to
quickly reason about the impact of the environment's geometry on what plans are
feasible, i.e., whether there exist action parameter values that satisfy all
constraints on a candidate plan. In tasks involving articulated and movable
obstacles, typical Task and Motion Planning (TAMP) algorithms spend most of
their runtime attempting to solve unsolvable constraint satisfaction problems
imposed by infeasible plan skeletons. We developed a novel Transformer-based
architecture, PIGINet, that predicts …

arxiv motion planning planning prediction