Event
Backtrack: A Lightweight Failure Recovery Mechanism for Frozen VLAs
1 reports · 1 sources
Thomas Landeg and Thierry Tambe propose Backtrack, a lightweight failure detection and recovery mechanism for frozen vision-language-action models (VLAs), which they claim requires no recovery demonstrations, policy retraining, or additional inference models. The mechanism uses a failure prediction head that adds only 0.48%–0.88% parameters, reads the VLA's internal representations, and predicts whether the current episode will fail; when the failure probability exceeds a threshold, the controller pauses the policy, backtracks the robot arm's recently completed motion, and then returns control to the unmodified original policy so that it retries from an earlier configuration. The prediction head is trained only on outcome-labeled rollouts of the frozen policy. The authors evaluate Backtrack on LIBERO with three VLAs, covering two types of action generation: SmolVLA and π0.5, which use flow-matching action experts, and the autoregressive π0-FAST. The above results are all as stated by the authors.
Report timeline
- Backtrack: Lightweight Failure Detection and Backtracking Recovery for Frozen VLAs
The authors propose Backtrack, a lightweight failure detection and recovery mechanism for frozen VLAs that requires no recovery demonstrations, policy retraining, or additional inference models. It uses a failure prediction head that adds only 0.48-0.88% to the parameter count to read the VLA's internal representations; when the predicted failure probability exceeds a threshold, it pauses the policy, backtracks the robot arm's recent motion, and hands control back to the original policy for a retry.
arXiv cs.RO ·