AI

A Safe Action Is Not Enough: Feasible-Future Decoding for Vision-Language-Action Policies

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 13:18:24 UTC.

Briefings

A new VICS-G algorithm improves safety in vision-language-action policies by using a selective finite-candidate approximation to address the feasibility-likelihood gap. It reduces mean cumulative safety cost by 1.9%-57.5% across six Safety-CHORES settings while maintaining policy sampling success within 2.5 percentage points and reducing mean episode length by 0.82 steps. The method requires no policy retraining or online rollouts, offering a practical path to safer task completion.

Key points

• The feasibility-likelihood gap exists when a frozen VLA policy favors locally admissible moves that leave no policy-supported route to safe task completion.
• VICS-G lowers mean cumulative safety cost by 1.9%-57.5% across six Safety-CHORES settings while remaining within 2.5 percentage points of policy sampling in success and 0.82 steps in mean episode length.

Context

Vision-language-action policies face a feasibility-likelihood gap where locally admissible moves may not lead to safe task completion. This research addresses that gap using a selective finite-candidate approximation to recover viable candidates.

Limitations

The study uses a selective finite-candidate approximation for online evaluation, which may not capture all feasible futures. The exact next-block marginal derivation is restricted to safe task completion, limiting generalizability to unsafe scenarios. Results are based on six Safety-CHORES settings and do not account for all possible task variations.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 28/10/2026, 13:01:48 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeifowh7okkbz4ohteh3pmnaakudv3analvykbxo3vuzvrg7lwngnxy