A Safe Action Is Not Enough: Feasible-Future Decoding for Vision-Language-Action Policies
VICS-G, a training-free reranker, lowers frozen VLA safety cost without retraining or online rollouts.
The paper argues that a locally safe action from a frozen vision-language-action policy can still leave no viable path to safe task completion. It derives a feasible-future target and approximates it with VICS-G, an alarm-triggered, training-free reranker that needs no online rollouts. Across six Safety-CHORES settings, VICS-G cut mean cumulative safety cost by 1.9% to 57.5% while staying within 2.5 percentage points of policy-sampling success and 0.82 steps of mean episode length.
- VICS-G reranks frozen VLA actions using feasible-future mass, without retraining.
- Mean cumulative safety cost fell 1.9% to 57.5% on six Safety-CHORES settings.
- Success stayed within 2.5 points and episodes within 0.82 steps of sampling.
From the original article
1 min read at huggingface.co
A safe action is not necessarily a viable one. A frozen vision-language-action (VLA) policy can favor a locally admissible move that leaves no policy-supported route to safe task completion. We call this the feasibility-likelihood gap: likelihood ranks the next move, while feasibility depends on the futures it leaves open.
By Tu Nguyen, Matthieu Zimmer, Vu Anh Vu · 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.