AutoIntervene: A New Framework for Robot Policy Switching

A novel framework named AutoIntervene has been launched to enhance the dependability of action-chunking visuomotor policies in robotics. These policies, which derive insights from demonstrations and forecast short sequences of actions instead of individual commands, may encounter perception inaccuracies and execution drift. This can lead to robots operating outside the demonstration range while still generating smooth yet inconsistent action segments. AutoIntervene mitigates this issue by enabling selective control transfer between the policy and an operator during deployment. It assesses proposed action chunks against a visual-action support memory created from successful task completions, merging visual similarity with consistency between suggested and reference actions. The framework employs phase-local support to manage policy-to-operator transfers within the current task phase and global support for reinstating policy control after operator recovery. Distinct switching thresholds are established for these two support types. The paper can be found on arXiv with the identifier 2608.07065, categorized as a cross announcement. This framework signifies progress toward more resilient autonomous systems by incorporating human oversight in a well-calibrated way.