REAP: Learning General Rearrangement Heuristics from Motion-Planner Feedback

Robotic rearrangement tasks require planning sequences of actions that fulfill high-level specifications while also being geometrically feasible. Symbolic abstractions make the combinatorial structure of rearrangement tasks tractable, but they usually do not fully capture the geometric constraints that determine which actions a robot can actually realize. These constraints are hard to express in the symbolic model. We propose REAP (Rearrangement with Elicited Action Preconditions), a framework in which a relational graph neural network heuristic is learned over a simple symbolic abstraction. A motion planner serves as a supervisor that elicits the hidden geometric preconditions of the problem during training. Deployed with greedy best-first search, the learned heuristic generalizes out-of-distribution along object count, plan length, and object arrangement, and can be trained jointly on several domains. Compared to classical heuristic search and learning-based baselines, REAP solves more problems, requires roughly an order of magnitude fewer motion-planner queries than comparable methods, and the action it ranks first is geometrically feasible in 94% of sampled states, against 22% for a random choice.