New hybrid framework lets robots undo actions with symbolic planning and RL

Inverting a robotic action—like pushing a cube back to its original position—is surprisingly hard. Simply reversing the motor trajectory or reversing symbolic state transitions often fails under continuous dynamics. Researchers Yigit Yildirim, Giuseppe Rauso, Riccardo Caccavale, and Alberto Finzi tackle this with a hybrid approach: first, they extract STRIPS-like operators from demonstrations using soft geometric predicates. For each extracted operator, they construct an inverse restoration objective that preserves preconditions, restores delete effects, and negates add effects.

A task planner attempts to satisfy the inverse objective using available action primitives. If any symbolic predicates remain unresolved, a residual operator learning problem is solved via Reinforcement Learning (a Soft Actor-Critic policy). Evaluated on the ManiSkill3 PushCube environment, the symbolic inverse generates a coarse pick-and-place plan, while the residual RL policy refines the cube's pose to satisfy the remaining predicates. The result: a physically grounded inverse skill that turns an approximate symbolic inverse into a precise restoration.