Robotic rearrangement tasks require planning action sequences that satisfy high-level specifications while remaining geometrically feasible. Symbolic abstractions make the combinatorial structure tractable, but they usually do not fully capture the geometric constraints that determine which actions a robot can actually realize. In the paper "REAP: Learning General Rearrangement Heuristics from Motion-Planner Feedback," Daniel Swoboda and Hector Geffner propose REAP (Rearrangement with Elicited Action Preconditions) to address this gap.

REAP trains a relational graph neural network heuristic over a simple symbolic abstraction. During training, a motion planner acts as a supervisor and elicits the hidden geometric preconditions of the task. Deployed with greedy best-first search, the learned heuristic generalizes out-of-distribution along object count, plan length, and object arrangement, and it can be trained jointly on several domains.

Compared with classical heuristic search and learning-based baselines, REAP solves more problems and requires roughly an order of magnitude fewer motion-planner queries. In sampled states, the action it ranks first is geometrically feasible 94% of the time, versus 22% for a random choice. This digest is based on the single arXiv paper, so no independent comparison across sources is included.