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, 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.
Figures & tables
Figure 1: (a) Policy variance in the LIBERO environment. Repeated sampling reveals higher action variance in contact-rich regions, where task failures are most likely to occur. (b) Values of the proposed trigger. The trigger value aligns with task difficulty and the policy’s internal uncertainty. These observations motivate triggering repeated sampling and verification at critical decision steps.
Figure 2: Conceptual comparison between different paradigms. (a1) Sampling-and-selection with VLM: Directly predict the delta progress with the current observation and the action chunk produced by VLAs. (a2) Sampling-and-selection with generative models: Extract the delta progress after generating the estimated future. (b) Model Predictive Control (MPC) with latent world models: Execute action(s) that minimize the latent distance with the goal image. (c) Latent world model verifier (ours): Estimate the future latent and the delta progress in a unified latent space.
Figure 3: Overall pipeline. (a) Two-Axis Trigger. The VLA policy generates two initial action chunk samples, based on which the two-axis trigger decides whether to activate repeated sampling. (b) Latent World Model Verifier. If triggered, the verifier evaluates K candidate chunks and selects the one with the highest estimated delta progress for execution. The JEPA-based latent world model autoregressively takes states, actions, and observations as input and predicts delta progress in the latent space. (c) Attention Mask. The delta progress key is designed to be invisible from other queries to remove possible task-specific shortcuts for world modeling.
Figure 4: Real-world experimental setup. We conduct the real-world experiments on an AgileX PiperX dual-arm robot, and design three fine-grained tasks, namely, pen insertion, pipetting and apple placement.
Method
Validation Metrics
Completion Score (%) ↑
Rank Loss ↓
Huber Loss ↓
Pen Insertion
Pipetting
Apple Placement
Average
π0.5 [ 19 ]
–
–
45.0
52.5
55.0
50.8
+ RoboMonkey [ 23 ]
1.19
0.37
49.0
55.5
63.5
56.0
+ GPC-RANK [ 39 ]
1.18
0.30
46.0
54.5
58.0
52.8
+ Ours
1.09
0.25
60.0
63.5
73.0
65.5
Table 1: Comparison of different repeated-sampling-and-selection methods across tasks.
Figure 5: Latency analysis. (a) Additional verification latency per triggered decision for 10 samples under successive optimizations. The three optimization stages cumulatively reduce latency. KV caching provides the largest latency reduction. (b) Trigger rate (left axis) and the resulting average added latency per chunk (right axis) on the three real-world tasks.
Method
Mug
Switch
Seal
Mouse
Scan
Avg.
RoboTwin 2.0 Clean
Hy-VLA
38.0
44.0
74.0
88.0
88.0
66.4
+ RoboMonkey
34.0
48.0
70.0
88.0
86.0
65.2
+ GPC-RANK
42.0
50.0
78.0
88.0
86.0
68.8
+ Ours
48.0
50.0
82.0
96.0
92.0
73.6
RoboTwin 2.0 Clean2Random
Table 2: Results on RoboTwin 2.0 Clean and Clean2Random with always-on verification. Success rates are reported in percent.
Figure 6: Results on RoboTwin 2.0 Clean with Varying Trigger Thresholds. For trigger=r , we adjust a general trigger threshold for all the tasks, so that approximately an r fraction of decision steps trigger verification.
Method
Spatial
Object
Goal
10
Avg.
π0.5
98.8
98.2
98.0
92.4
96.9
+ RoboMonkey
98.6
99.0
97.6
94.0
97.3
+ GPC-RANK
97.6
96.8
97.8
89.0
95.3
+ Ours
98.8
98.8
98.6
94.0
97.6
StarVLA-GR00T
98.0
99.4
96.4
91.0
96.2
+ RoboMonkey
98.2
99.4
96.0
91.6
96.3
Table 3: LIBERO results. Our method boosts the base policy, especially on LIBERO-10, the task suite that has not been saturated.
Method
Rank ↓
Huber ↓
Pen ↑
Pipetting ↑
Apple ↑
Avg. ↑
VeriVLA
1.09
0.25
60.0
63.5
73.0
65.5
w/o WM
1.21
0.26
52.0
55.5
58.0
55.2
w/o pretrain
1.17
0.25
54.0
61.0
63.5
59.5
Table 4: Ablation on world modeling and predictor pretraining. Completion scores (%) are reported overall and per task. The visual encoder remains pretrained in all variants.
Method
Status AUC ↑
Variance Correlation ↑
k / add. latency (ms) ↓
random
0.487
–
–
DINO
0.702
0.630
0 / 0
ACE
0.588
0.716
2 / 14
Ttemp
0.704
0.789
1 / 0
Tcross
0.694
0.877
2 / 14
Ours
0.727
0.887
2 / 14
Table 5: Trigger performance analysis. k denotes the number of action samples; latency is additional per decision.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Trigger
World Modeling
Verification Pipeline
Prediction Target
RoboMonkey
always-on
×
one-stage
action preference
FOREWARN
always-on
✓
two-stage
latent, VLM narration
GPC-RANK
always-on
✓
two-stage
pixel, reward
ADV
always-on
×
one-stage
proximity to discrete action tokens
MG-Select
always-on
×
one-stage
masked-condition divergence
DreamAvoid
vision-based
✓
two-stage
pixel, delta progress
Appendix
Table 6: Comparison of test-time verification methods.
Hyperparameter
Value
Learning rate
5×10−4
Weight decay
0.04
Batch size
32
Optimizer
AdamW
JEPA loss weight
1
Rank loss weight
1
Appendix
Table 7: Training hyperparameters and resource configuration. The table specifically shows the real-world adaptation experiments. The actual training epochs and duration can largely depend on the scale of the dataset.
Figure 7: Critical Index and Critical Moments in LIBERO.
Figure 8: Critical Index and Critical Moments in RoboTwin.
Task
Stage
Completion criterion
Score
Pen Insertion
1
Grasp the pen
0.5
2
Insert the pen into the yogurt bottle
0.5
Pipetting
1
Move the right hand to the correct position at the right bottle
0.25
2
Successfully aspirate water with the left hand
0.25
3
Move the right hand to the correct position at the left bottle
0.25
4
Successfully dispense water with the left hand
0.25
Appendix
Table 8: Completion score breakdown for the three real-world tasks.
Vision-language-action (VLA) policies commonly execute long-horizon mobile manipulation through open-loop action chunks, issuing multiple actions without receiving new high-level visual input. A committed chunk therefore implies how observations should evolve, but accidental deviations can violate this expectation while the remaining actions continue to propagate the error: commit-time policy confidence cannot react to a deviation that occurs after dispatch, and observation-only anomaly scores lack an action-conditioned reference for separating expected effects from unexplained changes. We propose CheckVLA, which verifies execution with a separately trained, frozen action-conditioned world model. A conformally calibrated risk threshold bounds the episode-level probability of an unnecessary first intervention and determines when to intervene, its exceedance controls how strongly the rewritten suffix retains the superseded chunk, latency-aware hard prefixing restricts replacement to actions that remain deployable, and an event-driven keyframe bank preserves evidence of prior progress across repairs. On RoboCasa365, under a common training recipe and a matched invocation budget, CheckVLA attains a 36.1% average success rate against 27.6% for periodic replanning (+8.5 points). At a matched 5% episode-level false-alarm target, action conditioning raises timely recall to 77.9%, against 48.6% for an observation-only control and 37.9% for an action-shuffled control. These simulation results support action-conditioned verification as a way to restore feedback during chunked execution while keeping the repair consistent with inference latency.
Yushan Liu, Peibo Sun, Xintao Chao +8
Tsinghua University · Shanghai Jiao Tong University · Peking University +2
Vision-Language-Action (VLA) models are increasingly used as generalist robot policies, yet their evaluation still relies largely on static benchmarks that randomly sample task scenes. In high-dimensional embodied spaces, failures are sparse and clustered, so static benchmarking can underestimate robustness risks. We reframe VLA evaluation as an active failure-discovery problem and propose a failure-aware test-generation approach that combines diversity-driven exploration with surrogate models learned from observed executions. The method steers testing toward high-risk yet diverse scene regions. Across four state-of-the-art VLA models, it uncovers substantially more failures (up to +29.7 % over selected baselines) while revealing more diverse failure modes. This mean that, for instance, in the case of GR00T-N1.6, success rate dropped from 64.4% to 34.7%. More broadly, our findings call for a shift in VLA evaluation: from passive measurement on fixed task suites to adaptive, failure-seeking test generation that exposes the structure of model weaknesses before deployment.
Arusa Kanwal, Pablo Valle, Shaukat Ali +1
Mondragon University Mondragon, Spain · Simula Research Laboratory Oslo, Norway
Vision-Language-Action (VLA) models trained on large-scale data have made remarkable progress, but they remain vulnerable to distribution shifts at deployment time. Recent VLA models suggest that prompts can serve as an efficient interface for steering policy behavior, but existing prompt-based steering typically relies on external guidance. This raises a natural question: can test-time training (TTT) for VLA be achieved by optimizing a prompt, so that the steering interface itself can be learned and adapted from interaction? We address this question with TTT-VLA, a test-time training framework based on Latent Prompt Optimization (LPO). During training, the latent prompt is learned with an additional proxy task, providing an extra learned conditioning signal for policy learning. At test time, TTT is performed by collecting interaction data from the current environment and optimizing only the latent prompt on those data using the proxy task's self-supervised signal, without modifying the policy itself. Experiments on SimplerEnv demonstrate that the proposed method consistently improves task success rates in both single- and multi-embodiment settings. Further analysis shows that the gains arise primarily from correcting a small number of critical decisions rather than globally altering policy behavior. These results suggest that LPO provides an effective and practical pathway for deployment-time improvement of foundation manipulation policies.
Wenbo Zhang, Jianxiong Li, Shuai Yang +4
ByteDance Seed · The University of Adelaide · Tsinghua University +3