acceptodds
Under review as a conference paper at ICLR 2027

Are Vision-Language-Action Models Robust to One-Step Observation Perturbations?

Abstract

Understanding the safety risks of vision-language-action (VLA) models is essential for their deployment in the physical world. Existing safety research has mainly considered persistent perturbations that are applied continuously to observations throughout an episode. However, momentary observation corruption, in which observations are severely perturbed only briefly within an episode, remains an underexplored safety threat. To address this gap, this work investigates robustness to one-step perturbations applied at a single time step per episode. Our experiments reveal that these perturbations substantially degrade VLA performance and that their impact depends on the action chunk execution length. Based on them, we propose CARE, which dynamically selects the execution length based on consistency with the previously predicted action chunk. CARE improves robustness with low computational overhead while preserving clean performance by selecting shorter execution lengths only under perturbations.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Vision-Language-Action Model, Robustness, Action Chunking

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_UeVoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the robustness of Vision-Language-Action (VLA) models against "one-step observation perturbations," where visual or proprioceptive inputs are severely corrupted at a single inference step within an episode. Across 8 perturbation types and 5 VLA models, the authors find that single-step corruptions at critical moments severely degrade task success rates, and longer action chunk execution lengths amplify this vulnerability. To mitigate this, they propose CARE, a training-free method that dynamically shortens the execution length when the newly predicted action chunk diverges from the unexecuted portion of the previous chunk. Experiments on LIBERO, Meta-World, and a bimanual robot setup validate the approach under synthetic single-step corruptions.

Critical strengths
  1. Clear mechanistic insight into action chunking under anomalies
  2. Good writing
  3. Comprehensive evaluation across backbones
Critical weaknesses
  1. Assuming severe observation corruption at strictly a single isolated inference step is detached from real-world physical control. In practice, hardware glitches, illumination shifts (e.g., auto-exposure recovery), motion blur, or communication drops rarely occur as an isolated single-step impulse in the middle of a trajectory and immediately vanish. As shown in Appendix D.4, when perturbations realistically persist for just 5–10 steps, CARE's effectiveness degrades substantially.

  2. Even under the idealized one-step threat model, CARE sets the minimum execution length to Ht=1H_t = 1, meaning the robot still executes the first corrupted action ata_t before replanning. In real-world control, a single erroneous action (e.g., premature gripper opening under blur in Section 6.4) is sufficient to cause irreversible task failure.

  3. Accepting the corrupted chunk AtA_t pollutes the reference buffer AprevA_{prev} for step t+1t+1, causing a cascading discrepancy even after clean observations are restored.

Questions for authors
  1. Given that real-world sensor corruptions (e.g., camera overexposure, packet loss bursts) typically span multiple consecutive frames rather than a single step, how can CARE be adapted to handle realistic, temporally correlated disturbances where replanning at t+1t+1 still encounters corrupted inputs?

  2. CARE causes a notable 11.6% drop in clean success rate on π0\pi_0-FAST (Table 2). Does discrete autoregressive token sampling inherently induce higher inter-chunk variance that falsely triggers short execution horizons under clean observations?

  3. When a severe discrepancy is detected at step tt (r0>τr_0 > \tau), why does CARE execute the corrupted action ata_t (Ht=1H_t=1) instead of falling back to the unexecuted clean actions A^\hat{A} from the previous chunk?

7 comments

We thank Reviewer UeVo for the thoughtful and constructive feedback. Your comments have helped us sharpen the scope of our robustness claims and clarify CARE’s limitations and the interpretation of our results.


W1-1: A single, isolated, severe observation corruption

Thank you for your suggestion. We agree that real-world observation corruptions can persist across multiple steps. However, their duration in steps depends on the disturbance and the sensing and control frequencies, making it difficult to specify a single representative duration. We therefore use one-step corruption, which allows us to focus on the timing of corruption without varying its duration, and to evaluate vulnerability at the temporal lower bound of attack duration.

This allows us to examine two important aspects of robustness. First, it enables temporal analysis, revealing how the impact of corruption depends on when it occurs, as shown in Figure 3. Second, it shows that even the shortest attack can cause serious harm. A single corrupted input can lead to task failure despite subsequent observations being clean, demonstrating that sustained attacks are not necessary to expose this vulnerability.

Our protocol is therefore designed to isolate key aspects of robustness, rather than to fully reproduce real-world corruptions. We will clarify this motivation in the revision.


W1-2 & W2; Q1 & Q3: Limitations of CARE

We thank the reviewer for these helpful comments on CARE’s limitations.

W1-2 & Q1: Robustness to multi-step corruption. We agree that CARE has limited effectiveness against multi-step corruption. Research on VLA robustness to one-step observation perturbations is still at an early stage, and our work provides a controlled analysis of this vulnerability and a simple mitigation method. Although CARE’s effectiveness decreases under 5- and 10-step perturbations, it still provides greater robustness than Vanilla in most evaluated settings (Appendix D.4).

W2 & Q3: Irreversible failures caused by the first action. We agree that CARE cannot always prevent irreversible failures caused by the first action. We view this limitation as a trade-off for CARE’s simple design, which requires no additional training or inference for execution length selection. By relying solely on discrepancies between existing action chunks, CARE can limit prolonged execution of potentially erroneous actions, but cannot determine which prediction is correct or whether the first action is safe. Setting Ht=1H_t=1 allows replanning at the next step, but does not guarantee that the action executed before replanning is safe.

The reviewer’s suggestion to fall back to the unexecuted portion A^\hat{A} of the previous chunk when r0>τr_0>\tau is a useful extension that could address some of these failures. However, a large discrepancy between chunks can also arise when an actual change in the environment requires an updated action plan. In such cases, falling back to the previous prediction could prevent a necessary update. Because discrepancy alone cannot distinguish these situations, always falling back to the previous prediction could improve robustness to corruption at the cost of clean task performance.

A possible extension is to combine CARE with an observation-reliability mechanism that determines when to retain a previous plan and when to accept a newly predicted chunk. CARE requires only minor changes to the execution procedures of action-chunking VLA models, making integration with such complementary mechanisms a promising direction for future work.


W3: Transient reference contamination and its empirical impact

Thank you for pointing this out. We agree that the corrupted chunk AtA_t becomes the reference for the next prediction, so the discrepancy can remain elevated even after clean observations are restored. We observed that CARE selects an execution length of 1 at the corrupted step and again at the following prediction, after which normal execution lengths resume. Although the reference contamination temporarily affects execution-length selection, it does not prevent CARE from preserving task performance in our evaluations as shown in Table 2.

Q2: Sensitivity of π0\pi_0-FAST to Execution Length

Thank you for raising this interesting question. We suspect that the clean-performance drop primarily reflects π0\pi_0-FAST’s sensitivity to execution length, rather than autoregressive sampling falsely triggering short horizons. Figure 4 shows that its performance drops substantially even with fixed execution lengths outside a favorable range. This suggests that varying the execution length is itself challenging for this model, which may explain the degradation under CARE. Autoregressive token sampling could contribute to this sensitivity, but our experiments do not isolate its role.


We thank Reviewer UeVo again. We also appreciate their recognition of the mechanistic insight into action chunking and the comprehensive evaluation. We hope our responses clarify the paper’s contributions to robustness evaluation. We believe the finding that even a single corrupted observation can cause lasting failure offers a useful perspective on VLA robustness.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer UeVo,

Thank you again for your valuable comments and feedback on our submission. With two days remaining in the discussion period, we would like to kindly follow up and ask whether our rebuttal has sufficiently addressed your concerns. Should any questions or concerns remain, we would be grateful for the opportunity to provide further clarification.

Thank you very much for your time and consideration.

Reviewer_1JB5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the robustness of existing VLA models to one-step observation perturbations. The analysis shows that perturbation timing significantly affects task success and that longer action execution lengths can amplify the impact of perturbations. Based on these findings, this paper proposes CARE, a training-free method that adaptively selects execution lengths using temporal consistency between consecutive action predictions.

Simulation and real-world experiments show that CARE improves robustness with low computational overhead while largely preserving clean task performance.

Critical strengths
  • Systematic evaluation and analysis. This paper systematically evaluates one-step observation perturbations across multiple VLA models, perturbation types, and intervention timings. It reveals the vulnerability of existing VLAs to such perturbations and analyzes the impact of action-chunk execution length on robustness. The motivation is clear.

  • Simple and practical method. CARE is training-free and requires no additional model inference. Its effectiveness is demonstrated across multiple models, simulation benchmarks, and real-world experiments.

Critical weaknesses
  • Insufficient validation of one-step corruption as a distinct robustness challenge. This paper emphasizes one-step corruption as a distinct robustness challenge from persistent corruption, but it remains unclear whether existing methods designed for persistent corruption can effectively handle this setting. Although RobustVLA is included, the baseline coverage remains limited. Evaluating one or two additional persistent-corruption robustness baselines under the proposed one-step protocol would better justify the motivation and necessity of the proposed approach.

  • Overstated selectivity of execution length adaptation. This paper repeatedly claims that CARE selects shorter execution lengths "only under perturbations," as stated in the abstract. However, Table 3 and Figure 5 show that CARE also frequently shortens execution lengths under clean observations, contradicting this claim. This suggests that the temporal-consistency mechanism can also be triggered by natural variations in action predictions, potentially leading to unnecessary replanning under clean observations.

  • Incomplete computational cost evaluation under perturbations. Section 6.3 reports episode policy time only under clean observations. However, the computational cost under perturbations remains unclear, and this setting is more important to CARE. The authors are encouraged to add more results to support the claim.

  • Insufficient positioning relative to prior work. At L126, the paper claims to be the first to use temporal consistency across action predictions for adaptive execution length selection. Prior methods such as MoH [1] and A^3 [2] also use prediction consistency to adapt execution length. The authors should clarify how CARE differs from these approaches.

  • Imprecise problem framing. The abstract and introduction frame the work as a safety problem, using terms such as "safety risks" and "safety threat." Yet the evaluation primarily focuses on task robustness without directly assessing safety-related outcomes. Since task failure does not necessarily imply a safety violation, the strong safety framing is not fully supported by the experimental evidence.

[1] Jing et al., Mixture of Horizons in Action Chunking, ICML 2026. [2] Chen et al., Dynamic Execution Commitment of Vision-Language-Action Models, arXiv preprint arXiv:2605.11567, 2026.

Questions for authors

In the released implementation, the LIBERO evaluation configuration uses all perturbation indices for Vanilla, but only the first 15 for RobustVLA, AAC, and CARE. Were these settings used to produce the results in Table 2? If so, why were different ranges of perturbation indices used to compute Bottom10%-SR?

4 comments

We thank Reviewer 1JB5 for the thoughtful and constructive feedback. Your comments have helped us strengthen the experimental evidence through additional baseline comparisons and computational cost evaluations. We would like to clarify our claims, evaluation protocol, and positioning relative to prior work.


W1: Comparison with an additional robustness baseline

Thank you for your suggestion. We agree that additional baselines can strengthen our findings on one-step perturbations. However, the limited availability of existing robustness methods applicable to this setting should also be considered. Based on our literature search, we could not find an additional method beyond RobustVLA that is applicable to our setting and has an official implementation.

Nevertheless, to address your concern, we additionally evaluated STRONG-VLA [1], a fine-tuning method for improving robustness to multimodal perturbations. Since we could not find an official implementation, we implemented it based on the original paper.

Table W1 reports results under one-step data loss across the four LIBERO suites. STRONG-VLA improves over Vanilla on four of the five models, but CARE outperforms STRONG-VLA on all five models. These results indicate that robustness fine-tuning does not necessarily transfer to severe one-step corruption. We plan to include evaluations under additional perturbations and a more detailed discussion in the revised paper.

Table W1: Robustness to one-step data loss across four LIBERO suites (Bottom10%-SR (%))

ModelVanillaSTRONG-VLACARE
π₀7.47.953.5
π₀-FAST49.557.664.2
π₀.₅47.251.492.4
SmolVLA15.415.559.8
X-VLA5.03.088.4

[1] Yuhan Xie et al., “STRONG-VLA: Decoupled Robustness Learning for Vision–Language–Action Models under Multimodal Perturbations,” ACM MM 2026.


W2: Overstated selectivity of execution length adaptation

Thank you for bringing this imprecise wording to our attention. We agree with your assessment and will revise the wording from “only under perturbations” to “when consecutive action predictions disagree” to more accurately describe CARE's behavior.

We would also like to note that replanning under clean observations is not inherently undesirable. Action predictions can vary even without perturbations, and prior work [2,3] demonstrates the benefits of adaptive execution lengths in clean settings. Our experiments also show preserved clean performance across most models and only a small increase in measured computational cost. These results suggest that shortening execution lengths under clean observations has only a minor impact in practice.

[2] Pan, Yi, et al. "VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon." arXiv preprint arXiv:2607.01804 (2026).

[3] Yuanchang Liang, Xiaobo Wang, et al. “Adaptive action chunking at inference-time for visionlanguage-action models.” CVPR2026.


W3: Additional computational cost evaluation

Thank you for this valuable suggestion. We agree that evaluating computational cost under perturbations is important and have conducted additional experiments under data loss perturbations, as shown in Table W3.

The results show that CARE maintains computational costs close to Vanilla while remaining substantially more efficient than baselines under perturbations. Although CARE's episode policy time increases slightly (from 2.345 s to 2.537 s), this is likely due to more frequent replanning caused by shorter execution lengths under perturbations. We will include these results in the revision.

Table W3: Computational Cost Evaluation Under Data Loss Perturbations

MethodInference Time (ms)Episode Policy Time (s)
CleanData LossCleanData Loss
Vanilla170.85166.002.2842.288
AAC719.48709.819.8539.825
VLA-Corrector229.69234.187.1687.663
CARE (Ours)168.14167.662.3452.537

W4: How CARE differs from prior work

Thank you for the opportunity to clarify the differences from these related methods. We agree that MoH and A³ also use prediction consistency for adaptive execution length selection. We clarify the differences below.

Comparison with MoH. Unlike MoH, CARE requires neither additional training nor modifications to the policy architecture. MoH generates action predictions across multiple horizons, fuses them through a learnable linear layer, and selects execution lengths based on cross-horizon consistency. In contrast, CARE can be directly applied to pretrained policies without additional training or architecture modifications.

Comparison with A³. Unlike A³, CARE requires no additional model inference for execution length selection and is applicable to a broader range of policy architectures. A³ generates multiple candidate action chunks, selects a representative chunk based on prediction consistency, and verifies its actions through conditional re-decoding to determine the longest executable prefix. This verification requires additional model inference and is implemented through flow-matching inpainting, which may limit its direct applicability to other architectures. In contrast, CARE selects execution lengths using existing action predictions without additional inference or architecture-specific mechanisms.

We will revise the Related Work section to clarify these distinctions.


W5: Discrepancy between evaluation metrics and safety risks

Thank you for this important observation. We agree that task success rate does not directly measure safety, and that task failure does not necessarily imply a safety violation. However, task success rate is a standard metric in existing VLA robustness studies, and establishing a comprehensive benchmark for safety-related outcomes is an important open question beyond the scope of this work. We acknowledge the importance of this distinction and will revise the abstract and introduction to frame our contribution as robustness to one-step observation corruption. Potential safety implications will be presented as motivation rather than as experimentally established outcomes.


Q1: Clarification of perturbation indices in implementation

Thank you for pointing this out. For methods that adapt the execution length, the total number of inference steps varies across rollouts and is not known in advance. We therefore cannot list all perturbation indices beforehand. We set the default to 15 because the average number of inference steps in clean Vanilla rollouts was approximately 14.6. For the evaluations in Table 2, we adjust this value according to the number of inference steps in the clean run. We will clarify this point in the implementation documentation.


We thank Reviewer 1JB5 again. We also appreciate your recognition of the clear motivation, systematic analysis, and practical value of CARE as a simple, training-free method. We hope that these additional results and clarifications address your concerns.

I appreciate the responses from the authors. Most of the my concerns have been solved. However, I agree with Reviewer 9zKp about the novelty of this paper. Although the authors have claimed again the main contribution of this paper, it seems that the novelty of this paper is not significant enough.

Thank you for your follow-up. We are pleased that our responses have addressed most of your concerns.

Regarding novelty, we have provided further clarification in our response to Reviewer 9zKp. If concerns remain after considering that response, we would appreciate it if you could specify which aspects you find insufficiently novel.

We would be happy to address those points directly. Thank you again for your time and consideration.

Reviewer_9zKpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the robustness of vision-language-action models to observation corruption applied at exactly one inference step per episode. It evaluates eight image, state, and combined perturbations across five VLA models, examines sensitivity to intervention timing, and investigates how action-chunk execution length affects the resulting performance degradation. The proposed CARE method uses discrepancies between temporally overlapping action predictions to select the execution length without additional training or additional policy predictions for the selection itself. Experiments on LIBERO, with additional Meta-World and real-robot evaluations, show substantial vulnerability at sensitive intervention times and demonstrate that CARE can improve robustness while largely preserving clean performance for most evaluated models.

Critical strengths

1.Useful systematic characterization of temporally localized failures.

Evaluating separate intervention times across multiple VLA models and perturbation modalities provides useful evidence that sensitivity is concentrated at particular stages of execution. The execution-length analysis and real-robot experiments strengthen the practical relevance of this observation. This is a meaningful contribution because aggregate robustness evaluations can obscure vulnerable moments within an otherwise successful trajectory.

2.A simple and practically relevant mitigation.

CARE reuses existing action predictions to adjust execution length, avoiding additional training and additional policy predictions for the selection procedure. Its improvements are substantial in several settings: for example, on (\pi_{0.5}), the data-loss Bottom10%-SR increases from 47.2% to 92.4%, while clean success remains approximately unchanged (Table 2). These results support the practical value of adapting execution length when predictions become inconsistent.

Critical weaknesses

1.The novelty of the proposed setting and its associated insight is not sufficiently established relative to the closest prior work.

The paper motivates its contribution primarily through a contrast with persistent perturbations. However, CronusVLA, already cited in the submission, explicitly evaluates constant, cyclic, and sparse observation corruptions, including full-image occlusion, in SimplerEnv-OR [1]. Thus, intermittent corruption and severe temporary loss of visual information are already part of the VLA robustness literature. The present protocol of exactly one intervention per episode, together with a scan over intervention times, is a useful refinement, but the paper does not sufficiently explain what distinct scientific conclusion this refinement establishes beyond prior temporal-corruption evaluations.

The qualitative explanation that erroneous predictions become more damaging when executed for longer without feedback also follows established limitations of open-loop action chunking, discussed, for example, in SGAC [2]. I acknowledge the inference-schedule distinction discussed in Appendix A.4 and CARE’s practical efficiency. Nevertheless, the current evidence supports a focused empirical extension and an efficient mitigation more clearly than it establishes a substantially new vulnerability or mechanism. This limits the significance of the central contribution in my assessment.

2.The experimental comparisons do not adequately isolate the temporal distinction on which the motivation relies.

Figure 1 contrasts persistent, small perturbations with a single, severe corruption, changing both temporal structure and severity. This ambiguity also affects the comparison with RobustVLA. Appendix C.3 describes visual augmentations followed by image PGD, and explicitly states that state observations are not perturbed during training. Consequently, its limited performance under severe image corruption or state failures does not establish that the difficulty arises specifically from the one-step temporal structure; severity and modality mismatch remain alternative explanations. Section 6.2 itself acknowledges the severity mismatch.

Appendix D.4 usefully extends motion blur to multiple steps, but the CARE-only duration curves for one corruption type do not resolve this issue. A matched comparison across durations, using the same corruption type, severity, and modality, is needed to substantiate the claimed distinction. Additionally, Bottom10%-SR appropriately measures particularly vulnerable intervention times, but without the corresponding average over intervention times, the typical impact of an isolated fault remains unclear. These limitations weaken the interpretation of the reported vulnerability and baseline failures, even though CARE’s improvements over Vanilla are useful.

[1] Li et al. CronusVLA: Towards Efficient and Robust Manipulation via Multi-Frame Vision-Language-Action Modeling. 2025. [2] So et al. Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking. 2025.

Questions for authors

1.Relationship to prior temporal-corruption evaluations.

What specific conclusion becomes identifiable through exactly one intervention per episode that was not established by cyclic or sparse corruption evaluations such as SimplerEnv-OR? Please clarify the scope of the novelty claim and explain whether the main advance is the intervention-timing characterization, the execution-length analysis, or a distinct failure mechanism.

2.Separating temporal structure from severity and modality.

Can you provide a comparison that holds corruption type, severity, and modality fixed while varying its duration, including at least Vanilla and CARE? How should the RobustVLA results be interpreted given that its implementation does not train on state corruption and is not matched to the severe failures used here?

3.Typical risk and recovery.

Using the existing intervention-time evaluations, can you report success averaged over intervention times, with uncertainty estimates, alongside Bottom10%-SR? What evidence distinguishes failures that remain recoverable with additional execution time from failures that are irreversible? This would help determine how broadly the critical-time results support the paper’s motivation.

4 comments

We thank the reviewer 9zKp for the thoughtful and constructive feedback. Your comments have helped us sharpen our contribution claims, strengthen the experimental evidence, and clarify the interpretation of our results.


W1 & Q1: Novelty of our setting and insights

Thank you for highlighting the need to clarify our contribution relative to the closest prior work. Our evaluation and analysis establish three findings:

  1. A single corrupted observation can substantially reduce task success.
  2. The impact of the same corruption type and severity strongly depends on when it occurs.
  3. Vulnerability to a single corrupted observation depends on the action-chunk execution length.

Below, we clarify which aspects extend beyond the evaluations and conclusions reported in prior work.

Comparison with CronusVLA. We agree that SimplerEnv-OR already covers intermittent corruption, including cyclic and sparse disturbances. These evaluations vary the ratio of corrupted to clean frames, with settings such as Constant (1:0), Cyclic (1:1), and Sparse (1:3, 1:5). However, because these evaluations do not independently control the exact number and timing of corrupted observations, they do not directly answer two questions: how many corrupted observations are needed to substantially reduce task success, and how its impact depends on when it occurs.

Our protocol addresses these questions by applying exactly one corruption event per episode and systematically varying its timing. Our results show that even a single corrupted observation can substantially reduce task success when all subsequent observations are clean (finding 1), and that the magnitude of this reduction depends strongly on the intervention time, even for the same corruption type and severity (finding 2).

In particular, a single corrupted observation is the smallest possible corruption intervention and more realistic than periodic corruption, since real-world disturbances often occur briefly rather than at regular intervals. We therefore believe that showing that even this minimal corruption can substantially reduce task success provides an important reference point for robustness evaluation.

Comparison with SGAC. We agree that the underlying limitations of open-loop action chunking have been investigated. SGAC demonstrates the benefits of closed-loop reactivity under environmental disturbances. However, achieving this reactivity requires policy inference at every step, and SGAC does not quantify how execution length affects vulnerability to a single corrupted observation.

Our analysis shows that longer execution generally amplifies the performance degradation caused by a single corrupted observation (finding 3). This finding points to a simple, practical mitigation: shortening execution while retaining multi-step execution between policy calls. It motivates CARE’s design, requiring no additional policy inference for length selection.

Our contribution is therefore a quantitative characterization of execution length as a practical means of mitigating single-corruption vulnerability, rather than the discovery of a fundamentally new failure mechanism. In the revision, we will clarify this distinction.


W2-1 & Q2-2: How should the RobustVLA results be interpreted?

Thank you for highlighting this confound. We agree that the original comparison did not isolate temporal structure from perturbation severity. To address this concern, we conducted additional experiments at five blackout levels while keeping the one-step intervention fixed. As shown in Table W2-2, RobustVLA performs substantially better under weaker corruption, even with the same one-step intervention. This suggests that perturbation severity is a major factor behind its limited performance under severe failures.

Table W2-2. Robustness (π0.5, LIBERO-goal, Bottom10%-SR %)

Method0.0 (full blackout)0.10.50.8 (RobustVLA setting)1.0 (clean)
Vanilla36.042.281.589.596.8
RobustVLA50.556.080.088.096.0

Regarding modality, our reproduction initially included state corruption during training, but we omitted it because it did not yield satisfactory performance. We therefore acknowledge that the state-corruption evaluations involve a modality mismatch. However, RobustVLA also shows limited robustness under severe image-only corruption, indicating that the observed difficulty is not confined to unseen state corruption.

These results do not change our original point that defenses designed for persistent perturbations do not adequately generalize to severe one-step perturbations. As explained in the Introduction, realistic faults such as sensor dropouts can cause a severe loss of observation information even when the intervention is momentary. Our setting therefore involves both intervention timing and perturbation magnitude.

W2-2 & Q2-1: Corruption duration

Can you provide a comparison that holds corruption type, severity, and modality fixed while varying its duration, including at least Vanilla and CARE?

Thank you for pointing out this ambiguity. In Appendix D.4, duration is measured in control steps, rather than policy inference steps. Vanilla replans every H=10H=10 or 1515 control steps, so corruption lasting d=1,5,d=1,5, or 1010 steps affects only the initial inference. Vanilla therefore gives the same results across these durations under matched rollout randomness, whereas CARE may receive corrupted observations again because it can replan earlier. We have clarified this point and added Vanilla’s results from Table 2 to Figure 13 for comparison under the same corruption settings.

W2-3 & Q3: Additional evaluations

Thank you for these helpful suggestions. We address average performance and failure recoverability separately below.

Average SR evaluation. We report the mean and standard error across 40 task-level success rates, each averaged over all evaluated intervention times. Table W2-3 shows large gaps between Average SR and Bottom10%-SR in many settings. For example, under data loss, π₀.₅ achieves an Average SR of 84.8 ± 2.0%, compared with a clean SR of 97.4% and a Bottom10%-SR of 47.2%. Consistent with Figure 3, these large gaps indicate that perturbations at certain intervention times cause particularly large performance drops. We will add Average SR with its standard error in the revision.

Table W2-3: Robustness evaluation (Bottom10%-SR / Average SR ± standard error (%))

ModelCleanOverexp.BlackoutBlurPacket lossZero fillSpikeRandomData loss
π₀70.953.9 / 71.8 ± 3.748.1 / 69.8 ± 3.751.4 / 70.7 ± 3.750.2 / 70.2 ± 3.85.9 / 47.8 ± 3.013.8 / 50.0 ± 3.019.6 / 50.8 ± 2.97.4 / 48.4 ± 3.1
π₀-FAST89.844.4 / 79.5 ± 2.245.0 / 80.1 ± 2.236.7 / 78.9 ± 2.375.6 / 88.6 ± 2.079.7 / 90.4 ± 1.872.9 / 88.6 ± 1.976.1 / 89.2 ± 1.949.5 / 81.3 ± 2.6
π₀.₅97.488.0 / 96.7 ± 0.745.9 / 84.8 ± 2.064.9 / 90.1 ± 1.591.1 / 97.2 ± 0.690.1 / 97.3 ± 0.791.6 / 97.4 ± 0.791.0 / 97.3 ± 0.647.2 / 84.8 ± 2.0
SmolVLA82.633.3 / 73.5 ± 3.134.8 / 74.4 ± 3.244.1 / 75.4 ± 3.269.5 / 81.8 ± 3.553.3 / 76.7 ± 3.352.3 / 76.9 ± 3.358.5 / 79.2 ± 3.415.4 / 63.4 ± 3.1
X-VLA96.490.5 / 96.7 ± 1.028.6 / 74.8 ± 2.638.5 / 79.4 ± 2.241.5 / 83.3 ± 2.015.0 / 63.8 ± 3.12.8 / 42.5 ± 3.010.4 / 47.0 ± 2.45.0 / 45.4 ± 3.3

Distinguishing recoverable and irreversible failures. We evaluated Vanilla under data loss with twice the episode length: 280→560 steps for LIBERO-Spatial and Object, 300→600 for Goal, and 520→1040 for LIBERO-10. For a matched comparison, we restricted both evaluations to their shared intervention timings. The results show that Bottom10%-SR changes little with additional execution time. It suggests that many failures are effectively irreversible within the tested execution budget.

Table Q3: Effect of doubling the episode length on Bottom10%-SR (%)

ModelVanillaVanilla with 2× episode length
π₀7.259.25
π₀-FAST42.5047.50
π₀.₅46.7551.75
SmolVLA13.2513.25
X-VLA5.006.50

We thank the reviewer 9zKp again. We also appreciate the positive assessment of our temporally localized evaluation and CARE’s practical value. We hope our responses clarify how examining when a single corrupted observation causes failures offers a useful perspective on VLA robustness and helps develop a simple, efficient mitigation.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer 9zKp,

Thank you again for your valuable comments and feedback on our submission. With two days remaining in the discussion period, we would like to kindly follow up and ask whether our rebuttal has sufficiently addressed your concerns. Should any questions or concerns remain, we would be grateful for the opportunity to provide further clarification.

Thank you very much for your time and consideration.

Dear Reviewer 9zKp,

Thank you again for your valuable comments and constructive feedback. We have provided additional experimental results and clarifications to address your concerns.

With fewer than four hours remaining in the discussion period, we would greatly appreciate it if you could review our rebuttal and share any further feedback.

We would be happy to provide any additional clarification before the discussion concludes.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.