Can robots undo what they just did?

Reversing a robot's actions isn't just rewinding—physics gets in the way. This work combines symbolic planning (like STRIPS operators extracted from demos) with residual reinforcement learning to undo manipulation tasks. A planner first sketches the inverse moves; when that fails to fully restore object state, RL refines the outcome. On pushing tasks, symbolic picks place the cube roughly; RL fine-tunes pose precision to satisfy remaining constraints.