Skip Policy: Efficient Robot Manipulation via Action Relabeling
A new imitation learning approach called Skip Policy (SkiP) dynamically skips redundant control steps and refines actions only at key moments like contacts and grasps. The method uses action relabeling to leap over free-space segments in a single decision, without requiring a separate skip planner. This reduces computational waste in manipulation tasks.
Key facts
- Previous policies predict actions at every control step.
- Most steps in manipulation trajectories traverse free space with little task-relevant information.
- Key steps around contacts, grasps, and alignment require dense prediction.
- SkiP uses action relabeling to replace behavior cloning targets with actions at the next key segment entrance.
- The policy dynamically leaps over skip segments and refines actions in key segments.
- SkiP operates within a single unified network without a learned skip planner.
- The method automatically partitions demonstrations into key and skip segments.
- The paper is available on arXiv with ID 2605.15536.
Entities
Institutions
- arXiv