Auditing Update Admission for Continual Embodied Agents

Summary

The paper examines how independent evaluation should decide whether a continual embodied agent may adopt a new policy update. It argues that an admission rule must measure both harmful-update control and the learning opportunities lost by rejecting useful updates, with the interaction budget stated explicitly. The authors identify a concrete failure in which a range-based confidence gate cannot certify unchanged performance on old tasks even when a substantial budget is available. They propose using a standard paired-binomial construction when outcome disagreements are rare, along with certified promotion of historical references and a round-level metric for missed learning opportunities. In a constructed one-step pushing diagnostic using 32 seeds, fresh paired checks admitted 31.6% of a common update stream after 2,000 episodes per stage, while the range-based gate admitted none. Despite that result, unconditional replay achieved better learning in closed-loop runs. A separate learned-dynamics stress test is designed to distinguish model bias from errors caused by feedback selection. The authors present the work as an admission-audit protocol supported by analytical and synthetic evidence, while noting that validation on physical robots and vision-language-action systems remains open.