Learning from Runtime Feedback through Failure-Bank Self-Evolution for Vision-Language-Action Models
FailBank converts runtime safety feedback into persistent policy gains for vision-language-action robot models.
FailBank is a four-stage self-evolving framework that turns runtime feedback into persistent policy improvement for vision-language-action models. A fixed control-barrier-function safety module acts as an observe-only teacher, and outcome-aware admission converts useful proposals into corrective targets for guarded LoRA updates. On the VLA-Arena benchmark, FailBank improved task success by 8.5 and 6.9 percentage points across two backbones while reducing policy-induced cumulative cost by 35.6% and 23.8%. Compared with runtime shielding, success rose by 25.4 and 9.5 points at comparable cost.
- FailBank uses a CBF module as an observe-only teacher
- Guarded LoRA updates come from admitted corrections
- Success rises 8.5 and 6.9 points on two backbones
- Policy-induced cost falls 35.6% and 23.8%
From the original article
1 min read at huggingface.co
Vision-language-action (VLA) models generalize broadly across robotic manipulation tasks, but complex environments require balancing task success with unintended contact. Runtime shields can correct individual actions, but they leave the underlying policy unchanged, so repeated disagreements may create a persistent policy-shield mismatch that blocks task progress.
By Mingyue Cui, Zheyuan Liu, Yihan Zhu · huggingface.co
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.