PaperlayerView exact revision on arXiv

Source-linked AI summary

VeriVLA: Adaptive Test-Time Verification for Flow-Based VLAs with Latent World Models

Xueli Sun, Yuxiang Lu, Yantong Liu, Xianzhe Fan, Xi Chen, Zhe Liu, Hengshuang Zhao

arXiv:2610.21985v1cs.RO

TL;DR

Stochastic flow-based VLA policies pose a deployment challenge, and evaluating action chunks requires predicting task progress. VeriVLA triggers extra sampling at critical states and uses a JEPA-based latent world model to rank candidates; across three real-world tasks, average completion score rises from 50.8% to 65.5%.

  • Problem

    Stochasticity in flow-based VLA policies creates a deployment challenge, while evaluating candidate action chunks requires estimating task progress.

    Source: §1 · pp. 1–2

  • Method

    A training-free two-axis trigger detects critical states, then a JEPA-based latent world model ranks candidate chunks by predicting future latents and delta task progress.

    Source: §1 · p. 2, §6 · p. 8

  • Results

    50.8% to 65.5% average completion score across three real-world tasks; experiments also report gains in RoboTwin 2.0 and LIBERO.

    Source: §4.2 · p. 7, §6 · p. 8

  • Takeaways & Limitations

    Reported gains span real-world tasks, simulation environments, and base policies.

    Source: §6 · p. 8

    Effectiveness depends on proposal diversity and quality, which remain limited; the frozen visual encoder may overlook subtle manipulation changes.

    Source: §5 · p. 8

Abstract

from arXiv · show

Flow-based Vision-Language-Action (VLA) policies are inherently stochastic: different noise seeds can produce action chunks of varying quality from the same observation. At test time, however, the policy typically commits to a single sample without an explicit estimate of its quality. In fine-grained manipulation, executing a suboptimal action can irreversibly disrupt task progress and lead to failure. To exploit policy stochasticity, we introduce a training-free two-axis trigger that uses two initial samples to measure within-chunk temporal variation and cross-sample variance, detecting the critical states. Once triggered, a JEPA-based latent world model ranks candidate action chunks by predicting action-conditioned future latent states and the corresponding changes in task progress. By operating in latent space, the verifier avoids pixel reconstruction and enables efficient and more accurate ranking. Across three real-world tasks, $\textit{i.e.,}$ pen insertion, pipetting, and apple placement), our method raises the average completion score from 50.8% to 65.5%, while adding 326 ms for sampling and verification per triggered decision on a single RTX 4090. Consistent gains are also observed in simulation, where the average success rate rises from 66.4% to 73.6% in the clean setting of RoboTwin 2.0, from 62.8% to 69.6% in its randomized setting, and from 96.9% to 97.6% on LIBERO.

1. Introduction

Flow-based VLAs can produce action chunks of varying quality from the same observation, yet cannot identify which sample is most likely to succeed before the robot commits. VeriVLA uses a training-free trigger to identify critical decisions and a latent world model to rank sampled actions.

1. Introduction

  • Different noise seeds produce action chunks of varying quality, but the flow-based VLA cannot determine which is most likely to succeed before committing to an action that may irreversibly alter the physical world.Poor actions can directly cause failure or accumulate errors that drift the trajectory out of distribution.Source: §1 · p. 1
  • A training-free trigger combines temporal action variation and cross-sample uncertainty to activate repeated sampling and verification at critical decision steps.It uses two sampled action chunks and tracks contact information and policy uncertainty.Source: §1 · p. 2
  • The verifier jointly predicts latent future states and task progress, avoiding pixel-level generation that adds latency and makes progress estimates depend on generation fidelity.Latent features are compact and dynamic-rich, while progress estimation is a scalar judgment.Source: §1 · p. 2
  • The paper treats action randomness as a resource for selecting among samples rather than a hazard to suppress.Its contributions combine latent-space verification with a trigger intended to reduce inference costs while preserving test-time selection benefits.Source: §1 · p. 2
  • The method improves real-world completion scores by 14.7 percentage points over the base policy, with 326 ms added per triggered round of sampling and verification.It uses a latent world model as verifier and a training-free, action-based two-axis trigger.Source: §1 · p. 2

2. Related Work

Prior test-time VLA enhancement includes online policy adaptation and frozen-policy action selection or augmentation. VeriVLA follows repeated sampling and selection, using latent-space prediction with an action-based trigger rather than VLM or pixel-level verification.

2. Related Work

  • Test-time training methods EVOLVE-VLA and TT-VLA optimize the policy online using dense rewards, while T2VLA derives intrinsic rewards from policy confidence for self-supervised adaptation.Source: §2 · p. 3
  • Frozen-policy methods include gradient-free repeated sampling with model-based verification and gradient-based action refinement using verifier-score gradients.Examples include RoboMonkey, GPC-RANK, DreamAvoid, and Guided Action Flow.Source: §2 · p. 3
  • VeriVLA follows repeated sampling and selection, jointly predicting dynamics and task-specific progress in a latent world model with an action-based trigger to allocate computation adaptively.Prior verifier-based methods instead use VLMs or pixel-level world models followed by downstream signal extraction.Source: §2 · p. 3
  • Latent world models support zero-shot planning through latent dynamics prediction and goal-image search, while VLA-integrated approaches incorporate latent prediction objectives or latent subgoals into VLA training.The cited approaches include DINO-WM, V-JEPA 2-AC, H-JEPA, VLA-JEPA, Being-H0.7, Copper-Policy, and LaWAM.Source: §2 · p. 3

3. Method

VeriVLA uses action-sample variation to decide when extra computation is warranted, then uses a latent world model to rank candidate chunks by estimated task-progress change. Its method defines the progress target, specifies a two-axis trigger, and adapts a JEPA-style verifier for candidate evaluation.

3.1. Problem Formulation

  • Action chunks are stochastic samples generated with distinct Gaussian-noise seeds, and their quality is defined as delta progress: the change in task progress caused by execution.Delta progress is presented as a more direct target for action evaluation than absolute progress.Source: §3.1 · p. 3

3.2. Overview

  • Two initial samples inform the trigger; without activation, the initial chunk is executed, while activation expands the candidate pool for verifier ranking.The verifier selects the chunk with the highest estimated delta progress.Source: §3.2 · p. 3, Figure 3 · p. 4

3.3. Two-axis Trigger

  • The trigger combines within-chunk temporal action variation with cross-sample action variance, using normalized chunks and two samples to measure divergence.The action chunks are normalized by average action magnitude to avoid bias toward fast-moving stages.Source: §3.3 · p. 4
  • The trigger calibrates reference distributions from rollout episodes and activates when either signal’s percentile exceeds the threshold τ0.τ0 is the trigger’s only hyperparameter.Source: §3.3 · p. 4

3.4. Latent World Model Verifier

  • The verifier adapts a pretrained JEPA world model with embodiment-specific state and action encoders and multi-step latent rollouts over uniformly divided action chunks.Sub-chunk rollouts provide denser supervision than predicting only the chunk endpoint while reducing inference overhead and mitigating error accumulation compared with per-action prediction.Source: §3.4 · pp. 4–5
  • A task-specific query conditions progress prediction, while an attention mask prevents that query from influencing task-agnostic latent dynamics prediction.The predictor imagines the next latent state before estimating the corresponding progress increment.Source: §3.4 · p. 5
  • Training jointly optimizes latent prediction and progress estimation, with Huber regression and pairwise ranking losses for progress estimates.The verifier also uses teacher forcing and self-forcing to balance learning efficiency and inference distribution alignment.Source: §3.4 · p. 5
  • When triggered, the verifier autoregressively estimates progress across sub-chunks and executes the candidate with the largest cumulative predicted progress.It ranks candidates by predicted delta progress without requiring a goal image.Source: §3.4 · pp. 4–5

4. Experiments

The experiments evaluate the trigger-verifier combination in real-world manipulation, RoboTwin 2.0, and LIBERO, with ablations and trigger analysis probing the design. They assess completion or success, generalization, and verification efficiency.

4.1. Real-world Experiments

  • The evaluation spans three real-world tasks and uses 50 trials per task across real-world, RoboTwin 2.0, and LIBERO experiments.The real-world base policy is fine-tuned with 50 demonstrations per task; test-time deployment uses one RTX 4090 GPU.Source: §4.1 · p. 6
  • A 14.7-percentage-point real-world completion-score gain over all baselines accompanies 326 ms of added latency per triggered decision.The latency excludes trigger latency and follows three cumulative optimizations.Source: §4.1 · p. 6, Table 1 · p. 6, Figure 5 · p. 7

4.2. Simulation Results

  • In RoboTwin 2.0, always-on verification improves success on all five clean tasks, while the trigger retains much of the gain on the first four; randomized-setting performance rises 6.8 percentage points.On LIBERO, the method improves LIBERO-10 performance while maintaining strong performance on other suites.Source: §4.2 · p. 7, Figure 6 · p. 7, Table 2 · p. 7

4.3.1. Effect of World Modeling.

  • Removing future-latent prediction lowers real-world completion score from 65.5% to 55.2%, indicating weaker performance without the dense world-modeling objective.The variant predicts delta progress only and becomes an autoregressive transformer without world modeling.Source: §4.3.1 · p. 7, Table 4 · p. 8

4.3.2. Effect of Predictor Pretraining

  • Without predictor pretraining, real-world completion scores fall, especially on tasks with stronger generalization requirements such as apple placement; degradation is relatively minor on pipetting.The pretrained predictor uses 23k DROID trajectories.Source: §4.3.2 · p. 8, Table 4 · p. 8

4.4. Trigger Analysis

  • The combined trigger reaches Status AUC 0.727 and Variance Correlation 0.887, exceeding either component, with 14 ms of added latency per decision.The temporal component has higher Status AUC, while the cross-sample component has higher Variance Correlation, indicating complementary signals.Source: §4.4 · p. 8, Table 5 · p. 8

5. Limitations and Future Work

The framework’s effectiveness is bounded by the diversity and quality of action proposals from the pretrained VLA, and by whether its frozen visual encoder captures manipulation progress.

5. Limitations and Future Work

  • Repeated sampling offers limited proposal diversity, and the verifier’s ranking depends on the quality of chunks produced by the pretrained policy.The verifier cannot steer that policy.Source: §5 · p. 8
  • A frozen visual encoder reduces training cost and preserves pretrained representations, but is not optimized for estimating manipulation progress.It may miss subtle object configurations, contact conditions, or task completion cues needed to distinguish actions.Source: §5 · p. 8
  • Future work could train an encoder–predictor pair within JEPA to optimize progress estimation.The paper proposes treating progress estimation as an optimization goal.Source: §5 · p. 8

6. Conclusion

VeriVLA adaptively verifies sampled action chunks with a latent world model and reports gains across real-world and simulated tasks. The related-work review distinguishes this approach from verifiers and other test-time enhancement methods.

6. Conclusion

  • 50.8% to 65.5%: average completion score across three real-world tasks, with 326 ms added per triggered decision for sampling and verification.The conclusion also reports 14 ms of trigger overhead and consistent gains across tasks, environments, and base policies.Source: §6 · p. 8

A.1. Test-time Verification

  • VeriVLA reads delta progress directly from predicted latents, unlike prior world-model verifiers that synthesize future pixels or align rolled-out latents with a fine-tuned VLM.It combines this verifier with an adaptive two-axis action-based trigger.Source: §A.1 · p. 11

A.2. Other Test-time Enhancement

  • Test-time training and RL methods update policy parameters using reward signals, while VeriVLA’s described approach ranks sampled chunks without that policy-update procedure.The cited methods require reward access, gradient computation, and tolerance for transient regression during adaptation.Source: §6 · p. 8, §A.2 · p. 11
  • Adaptive chunking methods reshape execution or decoding schedules, whereas VeriVLA’s trigger gates a verifier that can reject and replace a chunk.The comparison distinguishes progress evaluation from smoothing or resizing a trajectory.Source: §A.2 · p. 11
  • Runtime monitors predict failure, including methods using action consistency, VLA features, or action-conditioned latents; Foresight reports feature generalization across policies.Foresight predicts per-timestep failure scores from V-JEPA 2-AC latents.Source: §A.2 · pp. 11–12

B. Training Details

The appendix points to a table of training hyperparameters and resource configuration for real-world adaptation experiments, while noting that training duration depends on dataset scale.

B. Training Details

  • Table 7 lists training hyperparameters and resource configuration for the real-world adaptation experiments.The supplied passage does not give the table’s individual values.Source: Table 7 · p. 12, §B.1 · p. 12
  • Actual training epochs and duration can vary substantially with dataset scale.Source: Table 7 · p. 12

C.1. Evaluation Metrics

The appendix evaluates trigger alignment with critical task moments and policy variability, describes how simulator-derived critical labels are built, and lists baselines for trigger comparison.

C.1. Evaluation Metrics

  • Status AUC compares trigger scores with binary labels based on temporal proximity to simulator-defined critical moments, while Variance Correlation measures Spearman’s ρ against a K-sample reference variance.The deployment trigger uses k = 2 samples; the reference variance uses K samples.Source: §C.1 · p. 12

C.2. Critical-label Construction

  • Critical moments are contact-state flips involving task-relevant objects; Gaussian kernels around these events are summed and clipped to form a continuous critical index.The binary label thresholds the maximum index within each chunk window at 0.5; the appendix visualizes LIBERO and RoboTwin examples.Source: §C.2 · p. 12, Figure 7 · p. 13, Figure 8 · p. 13

C.3. Baselines

  • The trigger comparison includes a DINOv2-feature MLP, action chunk entropy, and the temporal and cross-sample variance trigger components.The MLP is trained with simulator signals and evaluated on LIBERO-Plus perturbations; quantitative results are reported in Section 4.4 and Table 5.Source: §C.3 · p. 13

D.1. Tasks

The real-world evaluation covers pen insertion, pipetting, and apple placement, with each task assessed using a completion score.

D.1. Tasks

  • The three real-world tasks are pen insertion, pipetting, and apple placement, and Table 8 defines their completion scores.Source: §D.1 · p. 13

D.2. Computational Cost

The computational-cost evaluation reports inference latency and GPU usage, with their breakdown in Table 9.

D.2. Computational Cost

Related papers

Loading 2610.21985v1…