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

A Safe Action Is Not Enough. Feasible-Future Decoding for Vision-Language-Action Policies. By Tu Nguyen and six others, from the Huawei Heisenberg Research Center, Huawei Noah's Ark Lab, and two other institutions. Published on the sixth of October. No code repository is linked to it.

A safe action is not necessarily a viable one. Picture a frozen vision-language-action policy choosing its next move. It can prefer a move that is locally admissible, breaking no rule right now, yet leaves no route the policy supports toward completing the task safely. The authors name this the feasibility-likelihood gap. Likelihood ranks the next move. Feasibility depends on the futures that move leaves open, and likelihood does not look at those.

To bring the futures into the decision, the authors derive an exact quantity: the next-block marginal of the trajectory law jointly produced by the policy and the environment, conditioned on history and restricted to safe task completion. In plain terms, it is what the policy's next block of actions would look like if only the futures ending in safe completion were allowed to count.

The derivation exposes a term that depends on the candidate being considered, which the authors call the feasible-future mass. It carries two pieces of information. Its support records whether safe completion is still possible at all, under the frozen continuation process, after a given candidate. Its magnitude measures how much weighted safe-completion mass remains. One part says yes or no to a future. The other says how much of it there is.

Evaluating that target exactly is impractical online. So the authors develop a selective approximation over a finite set of candidates, and establish conditions under which it recovers the best viable candidate that was retained. The guarantee is relative in two ways: to the frozen policy, and to the candidates the approximation keeps.

In practice the theory becomes VICS-G, a reranker that is triggered by an alarm and needs no training. Across six Safety-CHORES settings it lowers mean cumulative safety cost by between one point nine percent and fifty-seven point five percent. Success stays within two point five percentage points of plain policy sampling, and mean episode length within zero point eight two steps.

The method asks nothing new of the policy. There is no retraining and no online rollout. What changes is the ranking rule: a move is judged by the safe futures it keeps available rather than by its likelihood alone, and that rule is anchored to an exact target defined relative to the policy itself.