Backtrack lets frozen vision-language-action policies retry failed motions
Original title: If at First You Don't Succeed, Backtrack: Lightweight Recovery for Frozen VLAs
On LIBERO-Long, failure rates fall 19–48% and success rises by up to 7.4 points; the head adds 0.48–0.88% parameters.
Backtrack adds a small failure-prediction head trained only on outcome-labeled rollouts of a frozen policy. It reads existing internal representations and, above a threshold, pauses, retraces recent arm motion, and returns control so the unchanged policy can retry. Tests cover SmolVLA and π0.5 with flow-matching experts and autoregressive π0-FAST on LIBERO. Gains are largest on long-horizon LIBERO-Long; shorter suites already succeed 85–99%. All twelve MetaWorld controller settings with SmolVLA beat the unassisted policy.
- Recovery needs no extra demonstrations, retraining, or a second model at inference.
- The head is trained only on outcome-labeled rollouts of the frozen policy.
- On three shorter LIBERO suites, where success is already 85–99%, the effect is smaller.
From the original article
1 min read at arxiv.orgVision-Language-Action (VLA) models perform well on robotic manipulation, yet remain brittle once their own actions carry them into states their training data barely covers. Recovering from such failures usually requires recovery demonstrations, policy retraining, or a separate model at inference time.
By Thomas Landeg · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.