A Safe Action Is Not Enough: Feasible-Future Decoding for Vision-Language-Action Policies
The paper targets actions that look safe now but make safe completion impossible later.
Nguyen and coauthors define this as a feasibility-likelihood gap in frozen vision-language-action policies. Their method, VICS-G, reranks candidate actions by estimating how much safe-completion future remains, without retraining the policy or doing online rollouts. Across six Safety-CHORES settings, it lowered mean cumulative safety cost by 1.9% to 57.5% while staying close to policy sampling on success and episode length. HF Daily Papers' note
Nguyen and coauthors define this as a feasibility-likelihood gap in frozen vision-language-action policies. Their method, VICS-G, reranks candidate actions by estimating how much safe-completion future remains, without retraining the policy or doing online rollouts. Across six Safety-CHORES settings, it lowered mean cumulative safety cost by 1.9% to 57.5% while staying close to policy sampling on success and episode length. HF Daily Papers' note
score 5