Learning Low-Latency Action Verification for Frozen Vision-Language-Action Models

Wen-Han Hsieh, Ken Jiankun Zheng, Chia-Chien Chen, David M. Chan, Trevor Darrell

November 2026

Real-world task: a Franka Panda puts the plushie on the blue plate.

TL;DR A compact verifier, trained only on whether a frozen pi0.5 policy's own past rollouts succeeded, scores two sampled action chunks and the robot executes the better one. On a real Franka this raises average success on three tasks from 44.7% to 67.3%; on LIBERO it adds 3.5 points for 5.4% more median query latency.

What they do

VLA policies such as pi0.5 can sample many different action chunks for the same observation, but deployments usually execute just one. HDH (Hindsight Decision Model with History) samples two candidates from a frozen VLA at each query and picks one with a small learned verifier. The verifier learns only from the success or failure of the frozen policy’s own autonomous rollouts. It needs no expert comparisons, synthetic preferences or predicted futures, and the VLA is never updated.

How it works

Data. Roll out the frozen policy (100 rollouts per task in simulation, 150 retained real-robot rollouts) and label every executed action prefix with its episode’s final outcome.
Verifier. Inputs are early SigLIP features of two camera views from the frozen policy, proprioception and the 8-step executed prefix. Action tokens self-attend, cross-attend to the observation tokens, and an MLP outputs a failure logit. One verifier per task, about 481K parameters.
History. HDH adds a GRU token summarizing the observations and executed actions of up to three previous queries, to expose task progress.
Selection. Two chunks are sampled with independent flow-matching noise; the robot executes the first 8 actions of the higher-scoring one, then queries again.

Results

LIBERO-10, four tasks: 82.3% average success, vs 78.8% for the frozen baseline and 77.5% for picking randomly between the same two candidates. Without history (HDM) it reaches 80.3%.
Median query time is 718 ms vs 681 ms for the baseline. GPT-5 and GPT-6 used as verifiers take 27,674 ms and 7,403 ms and do not reach a higher average success.
Real Franka Panda with pi0.5-DROID, 50 trials per task: plushie on plate goes from 64% to 94%, cube out of drawer from 38% to 52% (not significant, p = 0.23), strawberry into closet from 32% to 56%.

Limitations

Every task needs its own verifier and rollout data, and terminal outcomes are coarse supervision. Only two candidates per query were tested. Latency was measured in simulation, not broken down on the real robot, and the evaluation covers few tasks on one robot setup.

Videos

LIBERO-10 simulation: putting the alphabet soup and tomato sauce in the basket.

LIBERO-10 simulation: putting the black bowl in the bottom drawer and closing it.

Figures

From Figure 2: Overview of HDH. Given the instruction, camera images and proprioception, a frozen VLA samples candida... — From Figure 2: Overview of HDH. Given the instruction, camera images and proprioception, a frozen VLA samples candidate action chunks; a compact verifier trained on terminal rollout outcomes gives each a predicted task-success score, and the robot executes the first H actions of the highest-scoring candidate.

From Figure 3: The three real-world tasks: placing a plushie on a plate, placing a strawberry in a closet and closing... — From Figure 3: The three real-world tasks: placing a plushie on a plate, placing a strawberry in a closet and closing the door, and removing a cube from a drawer and closing the drawer.

From Figure 4: Accuracy-latency trade-off on LIBERO, with latency on a log scale. HDH has the highest average success... — From Figure 4: Accuracy-latency trade-off on LIBERO, with latency on a log scale. HDH has the highest average success, 3.5 points above the baseline for 5.4% more median query latency.

Abstract

Vision-language-action (VLA) models have made significant progress in general-purpose robotic manipulation. However, sampled action sequences can be ineffective, while evaluating alternatives with computationally expensive verifiers adds latency at each policy query. In this work, we propose HDH, a lightweight framework that uses hindsight from a frozen VLA's own deployments to improve action verification. The framework has three stages. First, we deploy the policy and collect its autonomous rollouts. We then train a compact action verifier to assign each action chunk a predicted task-success score. Finally, the frozen VLA generates action candidates, and the verifier selects the candidate with the highest score for execution. Across four LIBERO-10 tasks, HDH achieves an average success rate of 82.3\% and on three real-world manipulation tasks, HDH improves average success over the baseline policy by 22.6\%.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong