Robot Learning

Failure Recovery

Failure recovery is a robot's ability to detect that a task execution has gone wrong and take corrective action to return to a state from which the task can be completed. Approaches range from hand-coded retry logic and finite state machines to learned recovery policies trained on failure data, and vision-language models that verify subgoal completion and replan. Recovery data is often deliberately collected, since expert demonstrations rarely contain mistakes.

Why it matters for physical AI

Long-horizon autonomy is limited less by peak success rate than by compounding errors; policies that recognize and correct their own failures are what separate lab demos from robots that run unattended for hours.