Failure Recovery

A robot noticing it made a mistake or is about to fail, and adjusting on its own to still finish the task.

Failure recovery means a robot detects something has gone wrong during execution — a missed grasp, a dropped object, getting stuck — and continues the task anyway by retrying, changing approach, or backing off to a safe state. An imitation-learning policy trained only on successful demonstrations tends to drift into states it has never seen the moment it deviates even slightly from the demonstrated trajectory, and small errors compound (compounding error), so recovery ability often determines whether a long-horizon task can be completed at all. There are two common approaches. One detects and explains the failure, then replans: REFLECT (CoRL 2023) uses a large language model to summarize the robot's experience and explain what went wrong, and AHA (2024) trains a vision-language model specifically to judge failures. The other teaches recovery actions to the policy directly: RaC (2025) has a human step in right before a failure, guide the robot back to a familiar state, then demonstrate the correction, and trains the policy on that kind of data.

ExampleWhen RaC's bimanual robot is hanging a shirt or packing a box, a human takes over right before a failure, guides the robot back to a familiar state, and demonstrates the corrective action; the policy learns to recover on its own from data like this.

Also called

Error Recovery

Related

Recovery and Correction Data · Human Intervention Data · Compounding Error · Human-in-the-Loop · RaC · Long-horizon Task

Sources

REFLECT: Summarizing Robot Experiences for Failure Explanation and Correction (arXiv 2306.15724)
AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation (arXiv 2410.00371)
RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction (arXiv 2509.07953)

As of

2025-09