VLA action chunks get a validity check before execution: PAV predicts each prefix's outcome and lifts LIBERO average success to 94.65%
Synopsis
Vision-language-action (VLA) policies emit a whole action sequence per query, but how much of it to execute was previously judged only indirectly, through generation stability or prediction agreement. PAV instead predicts, inside the policy's own task-conditioned representation, what a candidate prefix will lead to, and a verifier trained on trajectory-level outcomes scores whether the task can still be completed. On the four LIBERO suites, PAV averages 94.65% success at 14.0 policy queries per episode, above direct verification without future prediction (84.55%, 22.3 queries) and prediction in DINOv2 features (87.00%, 22.2 queries).
Interpretation
When a robot executes an action chunk from a VLA policy, it can now judge whether the task remains completable after a candidate prefix before committing to a length. Earlier adaptive methods chose the execution length from generation stability or prediction agreement, and neither signal establishes that executing the prefix still permits successful task completion. Across the four LIBERO suites, PAV averages 94.65% success at 14.0 queries per episode, versus 84.55% and 22.3 queries for direct verification without future prediction and 87.00% and 22.2 queries for prediction in DINOv2 features.
The judgment rehearses the consequences of candidate actions in the policy's internal representation, and a verifier converts that rehearsal into a probability of successful continuation. The predictor targets features the policy extracts from observations and instructions, aligns them to observed future representations by cosine distance, and adds a same-task action-swap ranking constraint so predictions stay sensitive to action content. On held-out data, predictions conditioned on the executed actions reach a cosine similarity of 0.9401 with observed future representations, against 0.5304 for same-task swapped actions, 0.4816 for reversed actions and 0.4103 for zero actions.
The verifier sees only trajectory-level success or failure, so per-step local validity is treated as latent and trained by matching the product of step scores to the final outcome. A failed rollout can still contain valid prefixes, and lowering every step of a failed trajectory would penalize those valid actions. A localization prior concentrates the validity reduction on fewer steps in failed rollouts; removing it gives 83.65% average LIBERO success at 25.60 queries, against 94.65% and 14.0 queries for the full objective.
Perspective
Evaluation covers single-arm manipulation with two flow-based VLA policies on the four LIBERO suites, dual-arm manipulation on all 50 RoboTwin 2.0 tasks, and three real tasks on a 7-DoF Franka Panda: placing, stacking and button pressing, with 30 trials per real task. The verifier is trained separately per policy configuration and benchmark, and one verifier is trained per real task.
The authors note that the verifier receives only trajectory-level outcomes, so the validity of an individual extension stays latent; the localization prior only concentrates the score reduction within failed rollouts and assigns no per-step failure label. The overhead measurement covers verifier GPU computation only, 0.45%–1.24% of action-generation time, and excludes end-to-end control-loop latency. On RoboTwin, the fixed-horizon baseline comes from a separate evaluation, so episode-level initial conditions are not guaranteed to be paired with the other methods.