Atomic-Probe Governance for Skill Updates in Compositional Robot Policies

A recent study presents a paired-sampling cross-version swap protocol aimed at examining the impact of substituting specific skills in a robot's library on compositional results. In a dual-arm peg-in-hole challenge, researchers identified a dominant-skill effect: one ECM reaches an atomic success rate of 86.7%, while others fall at or below 26.7%. Incorporating this ECM can enhance composition success by as much as 50 percentage points. This boundary is further illustrated through a basic pick task, where all atomic policies achieve a saturation of 100%. Notably, off-policy behavioral distance metrics do not effectively predict these outcomes.