acceptodds
Under review as a conference paper at ICLR 2027

Adapting Vision-Language-Action Models to Unknown Visual Disruptions During Execution

Abstract

Visual disruptions can arise while a robot is executing a task, leaving a vision-language-action (VLA) policy to respond without knowing the disruption type or timing. We introduce Self-supervised Adaptation from Leftover Trajectories (SALT), which uses the leftover trajectory, the unexecuted part of the previous action chunk, as self-supervision for test-time adaptation. Because consecutive chunks overlap in time, the leftover provides a temporally aligned target for the current prediction over the same future control interval. At the onset of a visual shift, the leftover can retain a plan formed before the input became corrupted, so updating the policy toward it anchors the adaptation across the shift (Transition Anchoring). SALT keeps the adapted policy and regenerates the current chunk, whose leftover becomes the target at the next replan, carrying the correction forward along the execution trajectory (Sequential Correction Propagation). Supervision comes entirely from the policy's own predictions, requiring no disruption annotations, expert actions, or target-domain demonstrations, and a lightweight adaptation gate calibrated only on nominal trajectories decides when updates begin. On LIBERO-10, SALT increases average success across five persistent visual corruptions from 43.9% to 53.2% with SmolVLA and from 58.7% to 66.0% with GR00T N1.7, while largely preserving nominal performance. On a real robot, it raises task progress averaged over digital and physical disruptions from 0.49 to 0.61.

Primary area:
applications to robotics, autonomy, planning
Keywords:
VLA, TTA, adaptation, Vision-Language-Action Model, Test-time Adaptation, SALT, Temporal Alignment, Robotics, Robot, Self-supervision

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_oeHJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SALT, a test-time adaptation method for VLA models facing unexpected visual disruptions. It uses the unexecuted actions from the previous action chunk as self-supervision to update the policy online. The method is evaluated on two VLA backbones across several LIBERO benchmarks, along with real-robot experiments.

Critical strengths
  1. The main idea is simple and interesting. Reusing leftover actions for online adaptation avoids the need for additional demonstrations or external feedback.

  2. The experiments show consistent improvements across SmolVLA and GR00T N1.7, with limited impact on nominal performance.

  3. The ablations are fairly comprehensive and help explain the importance of temporal alignment and sequential updates.

Critical weaknesses
  • The comparison with BID is useful, but it does not fully answer this question. A simpler baseline would be to directly combine the previous leftover actions with the current prediction, without updating the model. It would be interesting to see whether SALT still offers a clear advantage. Also, the Self-training baseline supervises all H actions, while SALT only uses the L overlapping actions. Using the same supervision mask would make the comparison fairer.

  • While task progress improves, the actual success rates remain quite low. Based on Table 6, SALT succeeds in only 47 out of 240 trials, including just 1 out of 80 on the Stack task. Another concern is the computational cost. SALT requires synchronous gradient updates, and the reported inference latency increases considerably. It would help to report the actual execution delays and task completion time on the real robot.

  • The adaptation gate can also trigger during normal execution. Although average clean performance remains similar, some episodes might become worse after unnecessary adaptation. A more detailed analysis of these cases, particularly around grasping or object contact, would be helpful.

Questions for authors
  1. How does SALT compare with simple action blending or temporal ensembling without parameter updates?

  2. Can the authors provide more details on real-world failure cases and actual execution latency?

  3. When the gate activates incorrectly, does adaptation ever turn a successful episode into a failure?

3 comments
AuthorsOpenReview ↗

Author Response [1/3]

Dear Reviewer oeHJ,

Thank you for your thoughtful comments and constructive suggestions! We appreciate your recognition of SALT’s simple idea of reusing leftover actions without additional demonstrations or external feedback, the consistent improvements across SmolVLA and GR00T N1.7 with limited impact on nominal performance, and the comprehensive ablations on temporal alignment and sequential updates. Your feedback has helped us strengthen our baseline comparisons, report real-robot execution latency and failure cases, and analyze unnecessary adaptation around grasping and object contact. We have also uploaded a revised manuscript, with all changes highlighted in blue. Below, we address each weakness (W1–W3) and question (Q1–Q3).


Weakness 1) The comparison with BID is useful, but it does not fully answer this question. A simpler baseline would be to directly combine the previous leftover actions with the current prediction, without updating the model. It would be interesting to see whether SALT still offers a clear advantage. Also, the Self-training baseline supervises all H actions, while SALT only uses the L overlapping actions. Using the same supervision mask would make the comparison fairer.

Thank you for these suggestions. With these baselines evaluated under the same gate, SALT still offers a clear advantage over both simpler baselines without parameter updates and self-training with the same supervision mask, showing that its gains come from learning from leftover targets. The table below summarizes the success rates.

MethodParameter updatesSmolVLA (%)GR00T (%)
Leftover execution✗47.262.0
Frozen temporal ensembling✗48.862.9
Self-training (same supervision mask)✓43.356.5
SALT✓53.266.0

For self-training, we have additionally matched the supervision mask to SALT’s by supervising only the L overlapping actions. We have added the comparisons with leftover execution and temporal ensembling, along with their implementation details, to Section 6.2 and Appendix G.3.


Weakness 2) While task progress improves, the actual success rates remain quite low. Based on Table 6, SALT succeeds in only 47 out of 240 trials, including just 1 out of 80 on the Stack task. Another concern is the computational cost. SALT requires synchronous gradient updates, and the reported inference latency increases considerably. It would help to report the actual execution delays and task completion time on the real robot.

We appreciate the reviewer’s concern about the low success rates, particularly on Stack. However, our results indicate that these rates partly reflect the difficulty of the long-horizon tasks, and that SALT still substantially improved binary success over the baselines. Even under clean conditions, Frozen achieved 80%, 50%, and 20% success on Two Cats, Drawer, and Stack, respectively (Appendix E). Under corruption, SALT succeeded in 47 out of 240 trials, more than doubling the 21 of Frozen and exceeding the 32 of BID (Table 7). To give a clearer picture, we have added the binary success rates alongside task progress (Section 5.3). The actual execution delays and task completion time on the real robot are detailed in our response to Question 2.

(Continued in the next comment)

AuthorsOpenReview ↗

Author Response [2/3]


Weakness 3) The adaptation gate can also trigger during normal execution. Although average clean performance remains similar, some episodes might become worse after unnecessary adaptation. A more detailed analysis of these cases, particularly around grasping or object contact, would be helpful.

We analyzed every clean episode in which the gate activated and compared it with the same episode under Frozen. Using contact information from MuJoCo [r1], we define contact as the gripper touching a task-relevant object and grasping as both finger pads touching the same object, following robosuite’s grasp check [r2].

BackboneGate activatedBoth succeedBoth failSuccess → failureFailure → success
SmolVLA53/200 (26.5%)40931
GR00T N1.725/200 (12.5%)18412
  • Changes in episode outcomes. Unnecessary activation changed the outcome in only 4 of 53 SmolVLA episodes and 3 of 25 GR00T episodes, and the changes went in both directions (3 worse and 1 better on SmolVLA, 1 worse and 2 better on GR00T).

  • How the degraded episodes failed. None of the degraded episodes failed at the moment of activation or on an object being held. On SmolVLA, SALT released the first object at the same replan as Frozen, and the episode diverged later, at the second object. On GR00T, the gate was activated at the first replan, before any contact. We observed no immediate grasp/contact failures at gate activation in these episodes.

  • Activation during grasping or contact. Activation rarely coincides with contact/grasp onset, about half as often as expected under random timing, and none of these episodes degraded. In the 18 SmolVLA episodes where the gate activated while an object was held, we observed no drops of that object.

    BackboneAt contact/grasp onsetRandom timingDegradedWhile holding an object (dropped)
    SmolVLA7 (13.2%)24.0%018 (0)
    GR00T N1.72 (8.0%)23.6%00

[r1] Emanuel Todorov, Tom Erez, and Yuval Tassa. MuJoCo: A physics engine for model-based control. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2012.

[r2] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martín-Martín, et al. robosuite: A modular simulation framework and benchmark for robot learning. arXiv:2009.12293, 2020.


Question 1) How does SALT compare with simple action blending or temporal ensembling without parameter updates?

Following the reviewer’s suggestion, we have evaluated temporal ensembling, which blends overlapping action predictions without parameter updates. We have also evaluated direct execution of leftover actions. Both use the same gate as SALT. SALT outperformed temporal ensembling and leftover execution on both backbones.

MethodSmolVLA (%)GR00T (%)
Temporal ensembling48.862.9
Leftover execution47.262.0
SALT53.266.0

These results support the benefit of learning from leftovers beyond ensembling or directly executing action predictions. We have added these comparisons to Section 6.2 and Appendix G.3.

(Continued in the next comment)

AuthorsOpenReview ↗

Author Response [3/3]


Question 2) Can the authors provide more details on real-world failure cases and actual execution latency?

  • Real-world failure cases. Common real-world failures involved precise placement, such as placing the block inside the drawer or stacking it on another block, corresponding to the stages illustrated in Figure 12. Under mesh camera occlusion, Frozen often failed to grasp the gray cat, whose visibility was reduced by the mesh (Figure 5), whereas SALT completed more trials (5/20 vs. 1/20, Table 7).

  • Actual execution latency. For execution latency, we have analyzed real-robot logs. Including execution and communication, the median replan interval was 1.70 s on SALT update replans, compared with 1.24 s for Frozen. Across each method’s successful trials, median completion times were 62.3 s versus 60.7 s on Two Cats and 31.8 s versus 27.6 s on Drawer (SALT versus Frozen). To reduce this overhead, we have optimized the implementation with identical computation results, reducing update-replan compute latency from 529 ms to 124 ms for SmolVLA and from 735 ms to 226 ms for GR00T in simulation benchmarks on an H200. We have added the latency measurements and implementation optimizations to Appendix F.1.


Question 3) When the gate activates incorrectly, does adaptation ever turn a successful episode into a failure?

Yes. Across 400 clean episodes, four Frozen successes became failures under SALT, while three failures became successes. We provide the detailed episode-level analysis, including activation around grasping and contact, in our response to Weakness 3.


In summary, SALT outperformed simpler baselines without parameter updates and self-training with the same supervision mask, supporting the benefit of learning from leftover targets. On the real robot, SALT more than doubled the binary success of Frozen, and we have reduced update-replan compute latency from 529–735 ms to 124–226 ms. Finally, unnecessary activation during normal execution rarely changed outcomes, with changes in both directions, and rarely coincided with contact/grasp onset, and none of those episodes degraded. We sincerely hope that our responses could properly address all your concerns. If so, we would deeply appreciate it if you could raise your score. If not, please let us know your further concerns, and we will continue actively responding to your comments and improving our submission.

Best regards,

Authors

Reviewer_T9UmOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SALT, a test-time adaptation method that helps vision-language-action (VLA) models handle unexpected visual disruptions during tasks. SALT uses the unexecuted part of the previous action chunk as a self-supervised target: it anchors the policy to pre-disruption behavior (Transition Anchoring) and carries corrections forward through later updates (Sequential Correction Propagation). A lightweight gate determines when adaptation starts, without disruption labels or target-domain demonstrations. On LIBERO-10, SALT improves average success across five visual corruptions from 43.9% to 53.2% with SmolVLA and from 58.7% to 66.0% with GR00T N1.7, while largely preserving clean-condition performance. Real-robot tests also show greater task progress under digital and physical disruptions.

Critical strengths
  • This is a well-motivated idea: overlapping action chunks guide online policy updates, extending prior work that uses overlap mainly for action selection or smoothing. It is intuitive, requires no extra supervision, and works with different action-chunked backbones.
  • The ablations support SALT’s effectiveness. Self-training on current predictions underperforms the frozen policy, while shuffled or misaligned targets lose the benefit. The leftover target nearly matches the oracle clean-input target (53.2% vs. 56.9%). Figure 9 also shows that transition anchoring and continued adaptation both matter.
  • The paper evaluates two VLA backbones, multiple corruption types and severities, and additional LIBERO suites. Paired episodes and statistical tests strengthen the simulation results. The real-robot experiments are especially valuable, testing physical disruptions (such as lights being turned off and cameras covered) not just synthetic image corruptions.
Critical weaknesses
  • The adaptation gate is unreliable. On GR00T N1.7, it fails to activate in 246 of 1,000 corrupted episodes and activates in just 31.5% of motion-blur episodes and 31.0% of fog episodes (Table 10). On SmolVLA, it activates in 26.5% of clean episodes. In Table 12’s gradual and temporary corruption experiments, the system adapts at every replan because the gate cannot reliably respond to these shifts. As a result, the full deployment pipeline is validated mainly for abrupt, persistent corruption. The need for separate nominal calibration data for each suite or robot also deserves discussion.
  • Real-world gains are measured mostly through partial progress. While encouraging, Table 6’s binary success counts offer a more limited picture: SALT completes few trials in several conditions, particularly Stack. The progress metric also rewards early sub-goals, such as touching an object. Confidence intervals or statistical tests would help establish how reliable the real-world gains are. SALT’s 2.9-point improvement over BID on GR00T is modest.
  • Deployment latency remains unclear. SALT requires synchronous gradient updates and action regeneration. On an H200, replanning takes 559 ms for SmolVLA and 724 ms for GR00T, while latency under corruption is 2.9–3.4 times that of the frozen policy. The paper does not clearly describe its compute setup or report end-to-end adaptation latency on a real robot, both important for online control.
  • The paper should clarify the method's scope and baselines. SALT relies on a useful pre-disruption plan, limiting its use for initial corruptions or changes that invalidate the plan. The simulations cover only LIBERO-Plus sensor-noise corruptions. BID also has substantially higher inference costs than SALT, and there is no latency-matched comparison or frozen-policy temporal-ensembling baseline to separate learning from leftover data from simply reusing overlapping actions.
Questions for authors
  • For GR00T N1.7, could the authors report success rates separately for episodes with and without gate activation, broken down by corruption type? This would show how much improvement comes from adaptation.
  • Could the authors test gated SALT on gradual and temporary corruptions, perhaps with a simple gate-release mechanism? Since the current experiments use always-on adaptation, they are hard to compare with the main results.
  • Could the authors report confidence intervals or statistical tests for real-world progress scores and give the binary success rates in Table 6 more prominence?
  • What hardware was used for adaptation in the real-robot experiments, and how long did updates take? Did synchronous updates cause noticeable pauses?
  • How sensitive is the adaptation gate to its calibration data and threshold? For instance, would a threshold calibrated on LIBERO-10 work on LIBERO-Goal, and how much would performance change with fewer nominal calibration trajectories?
4 comments
AuthorsOpenReview ↗

Author Response [1/4]

Dear Reviewer T9Um,

Thank you for your thoughtful comments and constructive suggestions. We appreciate your recognition of SALT’s motivation, ablation studies, and real-robot evaluation! Your feedback has helped us clarify the adaptation gate’s role and strengthen our empirical analysis. We have also uploaded a revised manuscript, with all changes highlighted in blue. Below, we address each weakness (W1-W4) and question (Q1-Q5).


Weakness 1) The adaptation gate is unreliable. On GR00T N1.7, it fails to activate in 246 of 1,000 corrupted episodes and activates in just 31.5% of motion-blur episodes and 31.0% of fog episodes (Table 10). On SmolVLA, it activates in 26.5% of clean episodes. In Table 12’s gradual and temporary corruption experiments, the system adapts at every replan because the gate cannot reliably respond to these shifts. As a result, the full deployment pipeline is validated mainly for abrupt, persistent corruption. The need for separate nominal calibration data for each suite or robot also deserves discussion.

We thank the reviewer for the careful analysis of the gate. We agree that the current gate is not a precise disruption detector (Section 7; Appendix F). It serves as a lightweight deployment trigger (adding 3–9 ms per replan; Table 10) for reducing unnecessary updates on nominal inputs, while SALT’s technical contribution is leftover-supervised adaptation, which can be driven by any activation rule (Appendix F).

We have additionally analyzed the impact of the gate. SALT with the gate improved success over Frozen across all tested corruption types among activated GR00T episodes (Question 1), under gradual and temporary corruptions (Question 2), and across the tested calibration sizes and cross-suite transfer settings (Question 5). Detailed results are provided in the corresponding responses below.


Weakness 2) Real-world gains are measured mostly through partial progress. While encouraging, Table 6’s binary success counts offer a more limited picture: SALT completes few trials in several conditions, particularly Stack. The progress metric also rewards early sub-goals, such as touching an object. Confidence intervals or statistical tests would help establish how reliable the real-world gains are. SALT’s 2.9-point improvement over BID on GR00T is modest

SALT’s real-world gains are not limited to partial progress. SALT improved average progress over Frozen from 0.49 to 0.61 (Figure 6), and also achieved the highest binary success counts. As reported in Table 7, SALT completed 47/240 trials (19.6%), compared with 21/240 (8.8%) for Frozen and 32/240 (13.3%) for BID, showing that its gains extend beyond early sub-goals. Stack remained challenging even without visual corruption, with Frozen succeeding in only 4/20 clean trials (Appendix E), so complete successes on this task were rare for every method. We have additionally computed confidence intervals for these gains, as detailed below.

  • Confidence intervals. We have computed 95% confidence intervals for real-world progress. SALT improved progress over Frozen by 0.125 (95% CI [0.086, 0.161]), and the interval excludes zero. We have added these results to Appendix E.
  • 2.9-point improvement over BID on GR00T. Since SALT is identical to Frozen when the gate is inactive (see our response to Question 1), we have additionally analyzed the episodes where the gate is active and SALT performs adaptation. In this subset, SALT improved over BID by 5.0 points (63.1% vs. 58.1%, p = 0.012), compared with 2.9 points overall. We have added these results to Appendix H.2.

Weakness 3) Deployment latency remains unclear. SALT requires synchronous gradient updates and action regeneration. On an H200, replanning takes 559 ms for SmolVLA and 724 ms for GR00T, while latency under corruption is 2.9– 3.4 times that of the frozen policy. The paper does not clearly describe its compute setup or report end-to-end adaptation latency on a real robot, both important for online control.

Please see our response to Question 4.

(Continued in the next comment)

AuthorsOpenReview ↗

Author Response [2/4]


Weakness 4) The paper should clarify the method's scope and baselines. SALT relies on a useful pre-disruption plan, limiting its use for initial corruptions or changes that invalidate the plan. The simulations cover only LIBERO-Plus sensor-noise corruptions. BID also has substantially higher inference costs than SALT, and there is no latency-matched comparison or frozen-policy temporal-ensembling baseline to separate learning from leftover data from simply reusing overlapping actions.

SALT targets disruptions that occur after nominal execution has begun. Below, we clarify SALT’s scope. We have also evaluated BID configured for latency closer to SALT’s, as well as frozen-policy baselines that reuse overlapping actions without parameter updates.

  • Method scope. As stated in Section 3.2, SALT targets episodes that begin under nominal visual conditions, where a useful pre-disruption plan is available. Within this scope, we respond to the reviewer’s three points as follows.

    • Initial corruptions. No pre-disruption plan exists in this setting, and always-on SALT left aggregate success nearly unchanged (32.5% → 32.9% on SmolVLA; Appendix I). This setting falls outside SALT’s focus on unexpected visual disruptions arising during execution. Prior work explores one-shot adaptation to shifted environments using target-domain demonstrations [r1, r2], which are not assumed available in our setting.
    • Changes that invalidate the plan. When the task state changes, the previous plan may no longer be appropriate. If the new state is observable, the natural response is to replan from the latest observation, and how quickly the policy can respond depends on the replanning frequency and execution horizon [r3, r4] (Section 3.1). SALT targets a different setting, where the task state is unchanged but corrupted observations lead to erroneous actions.
    • Disruption types. Beyond LIBERO-Plus sensor-noise corruptions, our real-robot evaluation also included physical lighting changes and camera occlusion (Section 5.3).
  • A latency-matched comparison with BID. To approximate a latency-matched comparison, we have reduced BID’s sampling budget to two candidates and evaluated both methods under motion and Gaussian blur.

    PolicyMethodSuccess (%)Latency (ms)
    SmolVLABID (2 candidates)46.0577
    SmolVLASALT55.8559
    GR00TBID (2 candidates)62.8650
    GR00TSALT67.0724

    At comparable latency, SALT achieved 9.8 and 4.2 points higher success than BID on SmolVLA and GR00T, respectively.

  • Frozen-policy temporal-ensembling baseline. We have additionally evaluated a frozen-policy temporal-ensembling baseline, which averages overlapping predictions from the most recent 16 chunks, and a leftover-execution baseline, which executes the cached leftover once and then resumes Frozen. Both use the same gate as SALT without parameter updates. All values are success rates (%) averaged over the five corruption types.

    PolicyFrozenLeftover executionTemporal ensemblingSALT
    SmolVLA43.947.248.853.2
    GR00T58.762.062.966.0

    SALT outperforms both baselines on both backbones. These results support the benefit of learning from leftovers beyond reusing or averaging overlapping action predictions. We have added these alternative uses of the leftover to Section 6.2 and Appendix G.3.

[r1] Kang, Taewook, et al. "Domain arithmetic: One-shot vla adaptation under environmental shifts." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.

[r2] Li, Weiqi, et al. "Vla models are more generalizable than you think: Revisiting physical and spatial modeling." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

[r3] Yuejiang Liu, Jubayer Hamid, Annie Xie, Yoonho Lee, Max Du, and Chelsea Finn. Bidirectional decoding: Improving action chunking via guided test-time sampling. In International Conference on Learning Representations, volume 2025, pp. 4594–4627, 2025.

[r4] Yuanchang Liang, Xiaobo Wang, Kai Wang, Shuo Wang, Xiaojiang Peng, Haoyu Chen, David Kim Huat Chua, and Prahlad Vadakkepat. Adaptive action chunking at inference-time for vision-language-action models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20802–20811, 2026.

(Continued in the next comment)

AuthorsOpenReview ↗

Author Response [3/4]


Question 1) For GR00T N1.7, could the authors report success rates separately for episodes with and without gate activation, broken down by corruption type? This would show how much improvement comes from adaptation.

Following the reviewer’s suggestion, we have reported GR00T success rates on the episodes of Table 1 separately for episodes with and without gate activation, broken down by corruption type. Unlike the gate-timing analysis, which counts post-corruption activations, this analysis also includes early activations.

Gate-activated episodesGate-inactive episodes
Corruption# of episodesFrozen → SALT# of episodesFrozen = SALT
Motion blur7974.7 → 79.7 (+5.0)12177.7
Gaussian blur20046.0 → 55.5 (+9.5)0–
Zoom blur19753.3 → 64.0 (+10.7)333.3
Glass blur20040.0 → 54.0 (+14.0)0–
Fog7885.9 → 87.2 (+1.3)12273.0
Overall75453.4 → 63.1 (+9.7)24674.8

Among activated episodes, SALT improved success over Frozen for every corruption type. Among non-activated episodes, SALT was identical to Frozen, since SALT does not update the policy when the gate is inactive. The improvement therefore comes from adaptation. Most non-activated episodes come from motion blur (121) and fog (122), the two corruptions under which the frozen policy loses the least success in Table 1 (from 81.5% on clean episodes to 76.5% and 78.0%). We designed the activation gate to reduce unnecessary computation (Appendix F). The lower activation rates for motion blur and fog suggest that the gate avoids many updates under these relatively mild corruption types, whereas it activates in nearly all episodes under Gaussian, zoom, and glass blur, which degrade the frozen policy most severely.

We have added these results to Appendix H.2.


Question 2) Could the authors test gated SALT on gradual and temporary corruptions, perhaps with a simple gate-release mechanism? Since the current experiments use always-on adaptation, they are hard to compare with the main results.

Following the reviewer’s suggestion, we have evaluated gated SALT with SmolVLA under gradual and temporary corruptions. To make these results directly comparable with the main results, we used the same gate, threshold, calibration, and adaptation settings as in Table 1. We have also evaluated a simple gate-release mechanism under both corruptions, which resumes the original model policy after a detected transition followed by two consecutive replans within the nominal score band.

SettingGradual success (%)Temporary success (%)
Frozen37.657.2
Gated SALT44.062.8
Gated SALT + release42.164.1

Gated SALT improved over Frozen under both corruptions, from 37.6% to 44.0% under gradual and from 57.2% to 62.8% under temporary corruption, so its gains are not limited to the abrupt, persistent corruption in the main results. With the gate-release mechanism, SALT further raised success under temporary corruption from 62.8% to 64.1%, as the release prevents unnecessary updates once the input returns to clean conditions. Under gradual corruption, however, the release lowered success from 44.0% to 42.1%, since it can reset the adaptation even though the corruption persists. The effect of release thus depended on the setting, while gated SALT outperformed Frozen under both corruptions. We have added these results to Appendix I.


Question 3) Could the authors report confidence intervals or statistical tests for real-world progress scores and give the binary success rates in Table 6 more prominence?

SALT’s real-world progress gains are statistically significant, and we have given the binary success rates in Table 7 more prominence in the revised manuscript.

  • Confidence intervals. Following the reviewer’s suggestion, we have computed 95% confidence intervals. SALT improved real-world progress over Frozen by 0.125 (95% CI [0.086, 0.161]), with the interval excluding zero. We have added these results to Appendix E.
  • Binary success rates. We have highlighted aggregate binary success rates in the main results of Section 5.3. SALT achieved 19.6% (47/240), compared with 8.8% (21/240) for Frozen and 13.3% (32/240) for BID. Per-condition counts remain in Table 7.

(Continued in the next comment)

AuthorsOpenReview ↗

Author Response [4/4]


Question 4) What hardware was used for adaptation in the real-robot experiments, and how long did updates take? Did synchronous updates cause noticeable pauses?

In the real-robot experiments, a robot-side client controls the SO-101 arm and sends observations to a policy server, which runs inference and adaptation on an NVIDIA H200. We have additionally analyzed pauses on the real robot with the execution logs. The five SmolVLA update steps in each replan took a median of 430 ms in total. On SALT replans that include an update, the median wait between action chunks was 860 ms, compared with 410 ms for Frozen. Including execution and communication, the median replan interval was 1.24 s for Frozen and 1.70 s for SALT on replans that included an update. To reduce this overhead for online control, we have additionally optimized the update implementation without changing the computation results (details in Appendix F.1), decreasing replanning latency with updates on an H200 from 529 ms to 124 ms for SmolVLA and from 735 ms to 226 ms for GR00T (measured in the same session, Table 11). We have added the compute setup details to Appendix E and real-robot latency measurements to Appendix F.1.


Question 5) How sensitive is the adaptation gate to its calibration data and threshold? For instance, would a threshold calibrated on LIBERO-10 work on LIBERO-Goal, and how much would performance change with fewer nominal calibration trajectories?

Following the reviewer’s suggestion, we have varied calibration size, tested transfer from LIBERO-10 to LIBERO-Goal, and varied the threshold multiplier.

  • Sensitivity to the calibration data. Each calibration set determines both score statistics and the threshold. We report success and the percentage of episodes with gate activation under corrupted and clean conditions. Corrupted activation counts episodes whose first gate activation occurs at or before the shift boundary, and clean activation counts episodes with any gate activation.

    Evaluation suiteCalibrationSALT success (%)Frozen success (%)Corrupted activation (%)Clean activation (%)
    LIBERO-1012 trajectories53.343.999.025.5
    LIBERO-1020 trajectories53.243.998.926.5
    LIBERO-1040 trajectories52.843.993.29.0
    LIBERO-GoalLIBERO-1076.473.196.38.0
    LIBERO-GoalLIBERO-Goal76.673.189.80.0

    These results suggest that fewer nominal trajectories and cross-suite calibration can still support effective adaptation in the tested settings.

  • Sensitivity to the threshold. We have varied the multiplier applied to the calibrated threshold (default 0.9) for SmolVLA on LIBERO-10, keeping the 20-trajectory calibration statistics fixed.

    MultiplierThresholdSALT success (%)Frozen success (%)Corrupted activation (%)Clean activation (%)
    0.855.4152.743.999.332.5
    0.90 (default)5.7353.243.998.926.5
    0.956.0552.943.997.617.5

    Raising the multiplier from 0.85 to 0.95 lowered clean activation from 32.5% to 17.5%, while corrupted success stayed within 52.7–53.2% and clean success within 70.0–70.5% (Frozen 71.5%). Within the tested range, the multiplier substantially changes clean activation rates, while task success varies little.

Calibration and threshold choices affect activation rates, but SALT outperforms Frozen under corrupted inputs across all tested settings. We have added these results to Appendix H.3.


In summary, SALT improved over Frozen for every corruption type among GR00T gate-activated episodes, under gradual and temporary corruptions, and across the tested calibration settings. On the real robot, SALT achieved the highest binary success, with a progress gain over Frozen whose confidence interval excludes zero. SALT outperformed both BID at comparable latency and frozen-policy temporal ensembling. Additionally, we optimized the update implementation, reducing update-replan compute latency by 3.3–4.3×. We sincerely hope that our responses could properly address all your concerns. If so, we would deeply appreciate it if you could raise your score. If not, please let us know your further concerns, and we will continue actively responding to your comments and improving our submission.

Best regards,

Authors

Reviewer_oaokOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes SALT, a self-supervised adaptive method for VLA testing in the presence of unknown visual disturbances. Its core idea is to utilize the leftover trajectory that has not been executed in the previous action block, to supervise the prediction of the current action block within the same future time interval, and to propagate the correction along the execution trajectory through Transition Anchoring and Sequential Correction Propagation. SALT does not require type annotation of disturbances, expert actions, or demonstrations from the target domain. It has improved task performance under visual disturbances in both LIBERO-10 and real robot experiments, while maintaining the performance in normal environments basically.

Critical strengths
  1. The paper found that there is temporal overlap between adjacent action blocks and transformed the unexecuted leftover trajectory into supervision during testing. This signal directly comes from the current rollout, avoiding reliance on expert actions, paired clean images, or pre-collected target domain data.

  2. Transition Anchoring uses the action plan generated before the interference occurs to cross the visual change boundary, while Sequential Correction Propagation continuously updates the target and retains the adapted state. Both solve the problems of initial cross-domain anchoring and subsequent trajectory correction, ensuring the integrity of the logic.

  3. The experimental results demonstrate the consistency of improvement. Under five types of continuous visual interference in LIBERO-10, the success rate has increased while largely preserving nominal performance. Real robot experiments also cover digital blurring, dim lighting, and occlusion of the camera, enhancing the practical credibility of the results.

Critical weaknesses
  1. The SALT model assumes that interference only affects visual input, while the task state, action space, and robot dynamics remain unchanged. Therefore, the previously generated action plan can still serve as a reliable supervision. If there are simultaneous occurrences of object movement, contact failure, action execution errors, or changes in task state, the old trajectory may have become invalid, and continuously fitting it might even reinforce incorrect actions. The paper lacks validation and failure detection mechanisms for such compound interference.

  2. It is impossible to distinguish between "visual interference" and "real state changes" internally. SALT only triggers adaptation based on changes in visual features, but changes in lighting, camera occlusion can also cause significant feature changes. The former may be suitable for maintaining the old plan, while the latter may require re-planning. The old plan can cause confusion in the model's learning of causality.

  3. Under interference conditions, the average replan delay of SALT is approximately 2.9-3.4 times that of the frozen strategy; when an update occurs, a single replan on H200 takes about 559-724 ms, and the robot needs to wait for optimization and action re-generation to complete. The paper has not fully evaluated this impact.

Questions for authors

Please refer to Weakness

4 comments
AuthorsOpenReview ↗

Author Response [1/2]

Dear Reviewer oaok,

Thank you for your thoughtful comments and constructive suggestions! We appreciate your recognition of SALT’s use of the leftover trajectory as supervision without expert actions or target-domain data, the design of Transition Anchoring and Sequential Correction Propagation, and the consistent improvements in both LIBERO-10 and real-robot experiments. Your feedback has helped us clarify the scope of SALT and the role of the adaptation gate, and further assess the impact of replan delay. We have also uploaded a revised manuscript, with all changes highlighted in blue. Below, we address each weakness (W1-W3).


Weakness 1) The SALT model assumes that interference only affects visual input, while the task state, action space, and robot dynamics remain unchanged. Therefore, the previously generated action plan can still serve as a reliable supervision. If there are simultaneous occurrences of object movement, contact failure, action execution errors, or changes in task state, the old trajectory may have become invalid, and continuously fitting it might even reinforce incorrect actions. The paper lacks validation and failure detection mechanisms for such compound interference.

We would like to clarify the distinction between “visual shifts” and “real state changes”, which require different responses.

  • Visual shifts. Changes in lighting or camera occlusion alter what the policy sees without changing the underlying task state. The previous action plan can therefore remain useful, but the policy may produce incorrect actions due to the corrupted observation. SALT addresses this problem by adapting the policy using the leftover plan as supervision.
  • Real state changes. When an object moves or a grasp fails, the previous plan may no longer be appropriate. If the new state is observable, the natural response is to replan using the latest observation to account for the change. As discussed in Section 3.1, how quickly the policy can respond depends on the replanning frequency and execution horizon [r1, r2].

SALT specifically addresses visual interference, using the leftover from the most recent plan as supervision to mitigate errors caused by corrupted observations. However, under compound interference where a real state change and a visual shift occur simultaneously, the most recent plan may no longer match the changed state, and the validity of leftover supervision cannot be guaranteed.

To examine this case, we conducted an additional experiment on SmolVLA, combining visual corruption with a 5 cm displacement of the object nearest to the gripper at the shift boundary. On the 605 episodes where the object was successfully moved, SALT's gain over Frozen decreased from 5.4 percentage points (44.1% vs. 38.7%) to 1.4 points (21.7% vs. 20.3%). This suggests that real state changes weaken the benefit of Transition Anchoring, but SALT maintained aggregate performance comparable to Frozen, potentially because each executed chunk is still regenerated from the current observation rather than copied from the leftover. Changes in robot dynamics or action space are also outside our scope.

A validation and failure detection mechanism that identifies invalid leftover plans and suspends adaptation would be a valuable extension of SALT, and we have noted it as a direction for future work in Section 7.

[r1] Yuejiang Liu, Jubayer Hamid, Annie Xie, Yoonho Lee, Max Du, and Chelsea Finn. Bidirectional decoding: Improving action chunking via guided test-time sampling. In International Conference on Learning Representations, volume 2025, pp. 4594–4627, 2025.

[r2] Yuanchang Liang, Xiaobo Wang, Kai Wang, Shuo Wang, Xiaojiang Peng, Haoyu Chen, David Kim Huat Chua, and Prahlad Vadakkepat. Adaptive action chunking at inference-time for vision- language-action models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20802–20811, 2026.

(Continued in the next comment)

AuthorsOpenReview ↗

Author Response [2/2]


Weakness 2) It is impossible to distinguish between "visual interference" and "real state changes" internally. SALT only triggers adaptation based on changes in visual features, but changes in lighting and camera occlusion can also cause significant feature changes. The former may be suitable for maintaining the old plan, while the latter may require re-planning. The old plan can cause confusion in the model's learning of causality.

If we understand correctly, the reviewer’s concern has three parts: (1) SALT cannot internally tell visual interference from real state changes, (2) lighting changes and camera occlusion also shift visual features and therefore trigger adaptation, and (3) when the state has actually changed, adapting toward the old plan may confuse what the model learns about causality. We address each point below.

  • Visual interference vs. real state changes. The gate’s role is to reduce unnecessary updates and their computation, not to identify the cause of a shift (Section 4.3, Appendix F). In clean SmolVLA episodes, where no visual interference is present, the gate was activated in 26.5% of LIBERO-10 episodes, but SALT’s success stayed within 1.0 points of Frozen (70.5% vs. 71.5%; Appendix H). Updates triggered without visual interference therefore had little effect on aggregate task success.
  • Lighting changes and camera occlusion. For visual disturbances that do not alter the underlying task state, the leftover plan can remain useful as a supervision signal. SALT therefore uses this plan to supervise adaptation once the gate activates. Our real-robot experiments include both physical lighting changes and camera occlusion (Section 5.3), and SALT improved success over Frozen in both, from 16.7% to 28.3% with the lights off and from 10.0% to 23.3% under camera occlusion.
  • Confusion caused by the old plan. We examined the practical impact of this concern in the compound-interference experiment described in our response to Weakness 1, where visual corruption coincides with an unexpected object displacement that can invalidate the previous plan. SALT maintained aggregate success comparable to Frozen (21.7% vs. 20.3%), with no observed aggregate performance degradation relative to Frozen in this setting.

If our reading differs from what the reviewer intended, we would appreciate further clarification.


Weakness 3) Under interference conditions, the average replan delay of SALT is approximately 2.9-3.4 times that of the frozen strategy; when an update occurs, a single replan on H200 takes about 559-724 ms, and the robot needs to wait for optimization and action re-generation to complete. The paper has not fully evaluated this impact.

We agree that synchronous updates introduce additional replan delay. To address this concern, we have analyzed its impact on the real robot and reduced the latency.

  • Impact of replan delay. In addition to the runtime measurements in Appendix F, we have analyzed the execution logs of the real-robot experiments in Figure 6, which use the original implementation. On replans with updates, the median waiting time between action chunks was 860 ms, compared with 410 ms for Frozen. The real-robot results in Figure 6 include this waiting time. Despite the replan delay, SALT improved average task progress over Frozen by 0.125. We have added these results to Appendix F.1.
  • Single replan latency on H200. We have optimized the update implementation with identical computation results (details in Appendix F.1). With this optimization, update-replan latency decreased from 529 ms to 124 ms for SmolVLA and from 735 ms to 226 ms for GR00T (measured in the same session, Table 11). Further accelerating adaptation, for example through asynchronous updates, is a promising direction for future work.

In summary, SALT improved success over Frozen under real-robot lighting changes and camera occlusion, where the adaptation gate is intended to activate. SALT’s real-robot gains were achieved with the replan delay included, and we have optimized the update implementation, reducing update-replan compute latency by 3.3–4.3×. Finally, we have added compound interference as a limitation and validation and failure detection mechanisms as a future direction in Section 7. We hope these responses have addressed your concerns.

Best regards,

Authors

Reviewer_oaokOpenReview ↗

Response to authors

I appreciate the authors' detailed response, which addresses most of my concerns. Overall, I am happy to maintain my positive rating.

Dear Reviewer oaok,

Thank you very much for your time and for reviewing our rebuttal. We sincerely appreciate your constructive feedback and continued support!

Sincerely,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.