acceptodds
Under review as a conference paper at ICLR 2027

BadWAM: When World Action Models Dream Right but Act Wrong

33 views

Abstract

World-action models (WAMs) are emerging as a promising foundation for embodied control: rather than predicting actions alone, they learn representations that couple action generation with future world prediction. This coupling is often viewed as a source of robustness, interpretability, and safety, as a robot's action can in principle be checked against its imagined future. In this paper, we show that this assumption is fragile. We introduce **BadWAM**, a unified framework for modeling **World-Action Drift Attacks**: a new class of WAM-specific adversarial attacks that use small visual perturbations to break the alignment between what a WAM imagines and what it executes. BadWAM characterizes this attack surface along two natural criteria: attack strength and stealthiness. When the adversary prioritizes disruption, BadWAM instantiates an **action-only adversarial attack**, which directly drives the model toward task-failing actions. When the adversary additionally prioritizes stealth, BadWAM instantiates an **imagination-preserving adversarial attack**, which seeks to induce harmful action shifts while keeping the model's predicted future close to its clean imagination. Together, these two attacks capture a spectrum of WAM-specific failures: from overt action hijacking to stealthier cases where the model appears to imagine a plausible future but executes a desynchronized action. We evaluate BadWAM across multiple WAM variants and benchmarks. Results show that our attacks substantially reduce task success rates under different execution setups. For example, our action-only attack reduces the model performance from 96.5% to 43.1% success on LIBERO. The results of our imagination-preserving attack further exposes a WAM-specific vulnerability: moderate future-preserving regularization can maintain strong attack performance while reducing future imagination drift.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
World Action Model, World Action Drift Attack

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_3pWCOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces BadWAM, a black-box adversarial attack framework targeting World-Action Models.The authors investigate the potential misalignment between action generation and future imagination under adversarial visual perturbations.

Critical strengths
  1. This paper investigates adversarial vulnerabilities in the World-Action Model (WAM) paradigm, with a particular focus on the relationship between action generation and future imagination. This provides an interesting perspective on the security of embodied foundation models.

  2. The authors conduct experiments on LIBERO and RoboTwin across multiple WAM variants. The results demonstrate that adversarial perturbations can substantially degrade task performance. The additional analyses of attack transferability and perturbation budgets further strengthen the empirical evaluation.

Critical weaknesses

W1. Insufficient Justification and Validation of Imagination-Preserving Attacks.

The central motivation for imagination-preserving attacks remains insufficiently supported. The paper assumes that keeping adversarial future predictions close to clean predictions improves attack stealthiness, yet no imagination-based safety monitor or detection mechanism is evaluated to substantiate this claim. More fundamentally, preserving the predicted future does not establish its correctness. The experiments only measure the distance between clean and adversarial predictions, without verifying their accuracy against actual future states or examining their consistency with the outcomes of adversarial actions. Furthermore, the action-only attack consistently achieves stronger task disruption, while imagination preservation offers no demonstrated practical advantage beyond reducing prediction drift. Consequently, the paper does not convincingly establish that imagination preservation creates a meaningful or practically exploitable vulnerability, substantially weakening its central contribution.

W2. Insufficient Experimental Validation and Limited Methodological Novelty.

The proposed attacks primarily combine established action-deviation objectives, future-prediction regularization, and zeroth-order optimization. However, comparisons are largely restricted to random perturbations and a white-box reference, without stronger black-box attack baselines under matched query budgets. More importantly, the complete absence of physical experiments substantially limits the practical significance of the findings, particularly given the paper's emphasis on vulnerabilities in embodied robotic systems. Although the authors cite potential physical damage as justification, prior studies [1, 2] have demonstrated that physical adversarial attacks on VLA-controlled robots can be evaluated under controlled conditions. Such validation is particularly important for assessing whether the proposed perturbations remain effective under real-world perception, dynamics, and execution uncertainties. Without stronger baseline comparisons or any physical validation, the paper provides insufficient evidence of either a substantial methodological advance or a practically relevant robotic safety threat.

[1] Wang, T., et al. Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics. ICCV, 2025.

[2] Lu, H., et al. When Robots Obey the Patch: Universal Transferable Patch Attacks on Vision-Language-Action Models. CVPR, 2026.

Questions for authors

As weaknesses

7 comments
AuthorsOpenReview ↗

Response to reviewer 3pWC [1/2]

It’s our great honor to receive Reviewer 3pWC's thoughtful comments and kind words to our work. We would like to address the concerns as below.


[W1] Justification and validation of imagination-preserving attacks

A1: Thanks for the valuable comment. We added two analyses to distinguish future-output preservation from prediction correctness.

  • First, we conducted a matched evaluation over all 40 LIBERO tasks using the Joint WAM, ϵ=0.06\epsilon=0.06, B=16B=16, and identical task and trial initializations:
ObjectiveTask successAction distanceFuture distance
Action-only61.5%0.60614.014
Img-preserving (λ=0.05\lambda=0.05)65.9%0.57113.035

Imagination preservation reduces future distance by 0.979 and action distance by 0.035, while increasing task success by 4.4.

  • Second, we directly compare the predicted future with the subsequently realized observation and evaluate a future-output detector. Both detectors are calibrated at a 5% clean false-positive rate using disjoint calibration and evaluation episodes.
ConditionOutput detection ↓\downarrowConsistency detection ↓\downarrowFailed episodes escaping output detection ↑\uparrowPrediction-to-realization SSIM ↑\uparrow
Clean2.5%2.5%--0.865
Action-only27.5%72.5%62.5%0.812
Img-preserving7.5%70.0%86.7%0.831
Random search10.0%65.0%71.4%0.815

Compared with action-only, imagination preservation increases prediction-to-realization SSIM from 0.812 to 0.831, reduces output detection from 27.5% to 7.5%, and increases the fraction of failed episodes escaping output detection from 62.5% to 86.7%. Consistency detection remains comparable (72.5% versus 70.0%). These results establish the practical role of the preservation objective: it makes attack-induced failures substantially less visible from the predicted output while improving prediction-to-realization alignment. We will revise the paper to distinguish this measured property from the stronger claim that the predicted future is semantically correct.

AuthorsOpenReview ↗

Response to reviewer 3pWC [2/2]

[W2] Methodological contribution and stronger baselines. Physical experiments and deployment scope.

A2: Thanks for the valuable comment. We provide the following clarifications and additional experimental results:

  • We additionally evaluated existing VLA attacks (the attacks are from the two papers that reviewer 3pWC mentioned in the review) on LIBERO:
AttackTask Success Rate$\downarrow$
UADA [1]94.6%
TMA [1]95.1%
UPA [1]98.8%
UPA-RFAS [2]98.3%
BadWAM (ours, action-only)61.5%
BadWAM (ours, img-preserving)63.0%

As can be observed, the transferred VLA attacks retain 94.6–98.8% target task success, whereas the action-only and imagination-preserving BadWAM objectives reduce success to 61.5% and 63.0%, respectively. This comparison suggests that attacks developed for conventional VLA policies do not automatically exploit the WAM action–imagination interface.

  • Thanks for pointing us to the two relevant physical VLA studies [1,2]. They focus on persistent physical patches for VLA systems, and we fully agree that real-robot evaluation would broaden the practical contribution of our work. However, we currently do not have access to a physical robot arm or the associated camera required for real-robot experiments. We will therefore clarify the scope of our conclusions, and discuss the related physical studies. We'll add the following discussions in the revision:

    1. Our current experiments establish the vulnerability through controlled closed-loop simulation. Simulation-only robotic attacks have also appeared in recent top-venue studies, including DP-Attacker [3] and BadVLA [4].

    2. Given BadWAM’s original and novel formulation of the attack surface and its comprehensive evaluation across benchmarks and WAM variants, we sincerely hope that reviewer 3pWC could reconsider the overall assessment in light of these clarifications and additional results.


[1] Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics. ICCV 2025.

[2] When Robots Obey the Patch: Universal Transferable Patch Attacks on Vision-Language-Action Models. CVPR 2026.

[3] Diffusion Policy Attacker: Crafting Adversarial Attacks for Diffusion-based Policies. NeurIPS 2024.

[4] BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization. NeurIPS 2025.


Thanks again for the constructive feedback. If reviewer 3pWC has any remaining concerns, we are happy to clarify further.

Dear Reviewer 3pWC,

We are grateful for the care and thought you have put into reviewing our work. We hope our rebuttal has offered helpful clarification regarding your concerns. If there are any aspects we could explain more clearly, we would sincerely appreciate the opportunity to do so.

Your kind consideration and the time you have generously given to our work mean a great deal to us.

Best regards,

Authors of Submission 5001

Reviewer_7ZryOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BadWAM, a framework for black-box visual attacks on world-action models. It optimizes input perturbations using action-output queries. When predicted futures are also exposed, it penalizes changes to those predictions, aiming to cause execution failure while keeping imagination close to the clean output. The experiments compare three FastWAM variants across 40 LIBERO tasks and 50 RoboTwin tasks, with 20 trials per task. In Table 1, the action-only WAM's LIBERO success rate falls from 96.5% to 43.1%. Joint and IDM variants also remain vulnerable when the future-preservation term is added.

Critical strengths
  1. The research question is practically relevant. I appreciate the focus on the relationship between executed actions and predicted futures, beyond changes in actions alone. This tests the assumption that a plausible prediction implies trustworthy execution. The action-disruption and future-preservation objectives also make the two levels of attacker access clear (§3–4).
  2. The closed-loop evaluation is reasonably broad. The two benchmarks, three variants, repeated trials, transfer tests, and parameter sweeps provide stronger evidence than isolated failure videos (Tables 1–2, Appendix B.6). I also find the random perturbations and white-box reference useful for putting attack strength in context.
Critical weaknesses
  1. A smaller prediction distance does not establish semantic correctness or evasion of safety checks. D_img measures the difference between clean and attacked predictions. It does not test whether those predictions remain consistent with the task and physical outcome, and no actual safety monitor's miss rate is reported. I therefore cannot infer that the model “dreams right” simply because its prediction stays close to the clean one. In the RoboTwin Joint WAM experiment in Appendix B.6.5, covering 12 tasks with 10 trials each, the default preservation weight reduces future distance to 0.96 times that of the action-only attack, while reducing action distance to 0.86 times. I do not think this alone establishes a useful stealth advantage. I would like to see evidence from safety checks, or a claim explicitly limited to changes in output distance.
  2. The key stealth comparison lacks consistent, verifiable measurements. I could not reconcile the values in §5.1 with Figure 6: the text reports a reduction from 14.01 to 13.04, but the visible curves lie at approximately 13.7–14.6 and do not show 13.04. The task success rates under the two attacks are 61.5% and 65.9%. Appendix B.2 says this comparison uses a larger preservation weight but does not give its value. Appendix B.5 also leaves the choice and normalization of each model's future tensor unclear. Without paired uncertainty estimates and a baseline for variation across repeated clean predictions, I cannot judge whether the distance improvement is stable and meaningful.
Questions for authors
  1. Additional core evidence: With a future-based monitor calibrated to a fixed false-positive rate on clean inputs, does the imagination-preserving attack yield a higher rate of task failures that escape detection than the action-only attack? Please also check whether the clean predictions are correct, so that preserving a prediction is not equated with preserving a correct future.
  2. Clarification and statistical evidence: Which models, weights, and statistics correspond to Figure 6, the 14.01/13.04 values in the text, and Appendix B.2? Please specify the distance definitions and control of generation randomness, and report paired confidence intervals across tasks.
  3. Deployment clarification: The query search in §5.3 takes seconds to tens of seconds per replan. Is the simulation paused during optimization, and is the attacker assumed to be able to make the robot wait? This determines whether the conclusion applies to a digital setting that permits pausing for queries or to real-time deployment.
4 comments
AuthorsOpenReview ↗

Response to reviewer 7Zry [1/2]

It’s our great honor to receive Reviewer 7Zry’s thoughtful comments and kind support for our work. We would like to address the concerns as below.


[W1, Q1] Prediction distance, semantic correctness, and monitor detection

A1: Thanks for the valuable comment. We added an eight-task LIBERO diagnostic using disjoint clean calibration episodes and held-out evaluation episodes. The output detector identifies abnormal future-output changes, while the consistency detector compares the predicted future with the subsequently realized observation. Both operate at a 5% clean false-positive rate.

ConditionOutput detection ↓\downarrowConsistency detection ↓\downarrowFailed episodes escaping output detection ↑\uparrowPrediction-to-realization SSIM ↑\uparrow
Clean2.5%2.5%--0.865
Action-only27.5%72.5%62.5%0.812
Img-preserving7.5%70.0%86.7%0.831
Random search10.0%65.0%71.4%0.815

As can be observed, action-only and random search produce similar prediction-to-realization alignment (0.812 versus 0.815 SSIM). Imagination preservation increases SSIM to 0.831, reduces output detection from 27.5% to 7.5%, and increases the fraction of failed episodes escaping output detection from 62.5% to 86.7%. Consistency detection remains comparable (72.5% versus 70.0%), showing that preservation makes failures substantially less visible in the predicted output while improving alignment with the realized observation.


[W2, Q2] Figure 6, distance definitions, and statistical evidence

A2: Thanks for identifying this discrepancy. We clarify the aggregation, metric, and statistical protocol below.

1) Aggregation discrepancy. Here we make a mistake and we'll fix it: figure 6 and the text use the same matched runs and per-replan best-so-far distance, but apply different aggregation. Specificly, the original figure pools all replans, thereby assigning more weight to longer tasks, whereas the text first averages within each task and then weights all 40 tasks equally.

To be more specific, let dmterfinald_{mter}^{\mathrm{final}} denote the final best-so-far future distance for method mm, task tt, episode ee, and replan rr, and let RmteR_{mte} denote the number of replans. The original figure computes

Dmpool=∑t,e∑r=1Rmtedmterfinal∑t,eRmte,D_m^{pool} = \frac{\sum_{t,e}\sum_{r=1}^{R_{mte}} d_{mter}^{\mathrm{final}}}{\sum_{t,e}R_{mte}},

which gives 14.604/13.774 for the action-only/imagination-preserving objectives. The task-balanced statistic is

Dmtask=140∑t=140∑e∑r=1Rmtedmterfinal∑eRmte,D_m^{task} = \frac{1}{40}\sum_{t=1}^{40}\frac{\sum_e\sum_{r=1}^{R_{mte}}d_{mter}^{\mathrm{final}}}{\sum_eR_{mte}},

which gives 14.014/13.035. Both aggregations show lower future distance under imagination preservation; they differ only in task weighting. We will redraw Figure 6 using task-balanced aggregation so that its endpoints match the reported values.

2) Experimental setting and distance definition. Both objectives use the Joint WAM, ϵ=0.06\epsilon=0.06, B=16B=16, identical task and trial initializations, and all 40 LIBERO tasks with 20 trials per task. Action-only uses λ=0\lambda=0, while imagination-preserving uses λ=0.05\lambda=0.05. For decoded future frames VV, we compute

Dimg=1T∑t=1T∥Vtadv−Vtclean∥2.D_{img} = \frac{1}{T}\sum_{t=1}^{T}\|V_t^{adv}-V_t^{clean}\|_2.

The first conditioning frame is excluded. The remaining RGB frames are normalized to [0,1][0,1] and resized to 64×12864\times128. Clean and adversarial queries use the same inference seed. Under task-balanced aggregation, the matched results are:

ObjectiveTask successAction distanceFuture distance
Action-only61.5%0.60614.014
Img-preserving (λ=0.05\lambda=0.05)65.9%0.57113.035
AuthorsOpenReview ↗

Response to reviewer 7Zry [2/2]

3) Statistical evidence and relation to the main table. Paired task evaluation in the above table shows that imagination preservation reduces future distance by 0.979, and action distance by 0.035, while increasing task success by 4.4. These results quantify the trade-off between future preservation and attack strength.

Figure 6 uses λ=0.05\lambda=0.05, whereas the main table uses λ=0.015\lambda=0.015 at the same B=16B=16, yielding 65.9% and 63.0% task success, respectively. The higher success at λ=0.05\lambda=0.05 is consistent with stronger future preservation weakening action disruption. We will also clarify that eight paired finite-difference updates require 16 perturbation queries, plus one clean reference query.


[Q3] Query latency and deployment assumptions

A3: Thanks for the valuable comment. The simulator pauses during adversarial optimization at each replanning step and executes the selected action chunk only after the search completes. Our experiments are conducted in closed-loop simulation, following a protocol similar to recent studies such as DP-Attacker [1] and BadVLA [2]. We will add a discussion clarifying this protocol, its scope, and real-time extensions.

[1] Diffusion Policy Attacker: Crafting Adversarial Attacks for Diffusion-based Policies. NeurIPS 2024.

[2] BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization. NeurIPS 2025.


Thanks again for the constructive feedback and kind support. If reviewer 7Zry has any remaining concerns, we are happy to clarify further.

Dear Reviewer 7Zry,

Thank you again for taking the time to share your valuable comments and encouraging feedback. We have carefully considered the points you raised and hope that our rebuttal conveys our responses clearly. We would warmly welcome any further guidance you may wish to offer and would be happy to elaborate wherever helpful.

Your thoughtful attention to our work means a great deal to us.

Best regards,

Authors of Submission 5001

Area_Chair_FJFEOpenReview ↗

Please engage in the discussion with authors

Dear Reviewer 7Zry,

Please engage in the discussion with authors, thank you!

AC

Reviewer_Bwr7OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates adversarial attacks on WAMs, focusing on the potential inconsistency between the models' predicted future states and generated robot actions. The paper proposes BadWAM, a black-box query-based attack framework that disrupts robot task execution by applying bounded perturbations to visual inputs. The framework includes two attack methods, including the action-only attack reduces task success rates by maximizing action deviation, while the imagination-preserving attack further constrains changes in predicted future states to improve attack stealthiness. The paper conducts experiments on different WAM variants on LIBERO and RoboTwin, and the results show that the attacks reduce task success rates, while the imagination-preserving attack keeps predicted futures relatively stable.

Critical strengths
  1. The paper focuses on safety issues in emerging WAM-based robotic systems. The potential inconsistency between predicted future states and executed actions provides a meaningful perspective for studying the reliability and safety of WAMs.

  2. BadWAM is a black-box attack method, making it more applicable to practical deployment scenarios.

  3. Experiments show that adversarial perturbations can transfer across different WAM variants, with only a relatively small reduction in attack effectiveness.

Critical weaknesses
  1. My main concern is the insufficient justification for WAM-specific safety claims. The core motivation of this paper is that inconsistencies between actions and imagination in WAMs may undermine future-based safety verification. The imagination-preserving attack aims to alter actions while keeping predicted futures relatively unchanged, potentially evading safety checks. However, the current method and experiments do not fully support this motivation. To be specific:
  • The paper targets task failure and uses task success rate as its primary evaluation metric. However, task failure does not necessarily indicate a safety violation or demonstrate that a safety monitor has been bypassed. If the focus is on imagination-based safety checks, the paper should clarify what these checks detect and how reduced task success reflects security risks. If imagination is instead used to predict task completion, its relationship to the safety verification emphasized in the paper needs further explanation.

  • The necessity of imagination preservation remains insufficiently justified. The authors argue that preserving predicted futures improves attack stealthiness, but only measure the distance between clean and attacked imaginations, without demonstrating that smaller imagination drift reduces detection rates. Since imagination preservation can also weaken attack effectiveness, its practical benefits need further validation.

  • The paper does not directly validate its core claim of action–imagination desynchronization. The current metrics measure action and imagination deviations separately relative to clean predictions, but WAM actions and imagined futures are not guaranteed to be fully consistent. Therefore, larger action deviations and smaller imagination deviations do not necessarily imply inconsistency at the physical execution level. A more direct evaluation would compare the attacked model's predicted future with the actual outcomes of its actions.

  1. The experiments mainly compare against random perturbations and a white-box reference, lacking fair comparisons with existing VLA adversarial attacks and stronger black-box methods under comparable attack budgets.

  2. The paper lacks robustness evaluations against common input preprocessing and purification defenses, such as Gaussian filtering, JPEG compression, and DiffPure [1].

[1] Nie et al. Diffusion Models for Adversarial Purification.

Questions for authors
  1. What is the direct connection between task failure and the safety risks emphasized in the paper?

  2. How can action–imagination desynchronization be directly measured and validated?

  3. Does imagination preservation actually reduce the detection rate of future-based safety monitors?

  4. How does BadWAM compare with existing VLA adversarial attacks and other black-box baselines under the same attack budget?

  5. How robust is BadWAM against common input preprocessing and purification defenses? Do Gaussian filtering, JPEG compression, or DiffPure significantly affect its attack effectiveness?

8 comments
AuthorsOpenReview ↗

Response to Reviewer Bwr7 [1/2]

We thank reviewer Bwr7 for the thoughtful comments and kind words to our work. We would like to response to the questions as below.


[W1, Q1] Connection between task failure and security risks

A1: Thanks for the valuable comment. We respectfully make the following explanation:

  • Task success measures closed-loop execution integrity: whether bounded adversarial inputs turn a normally successful WAM into a failed execution. This captures manipulation, availability, and reliability risks.

  • For future-based checks, the security-relevant event is an attack-induced failure that remains unflagged. We therefore evaluate both future-output changes and prediction–execution consistency, separating three questions: whether the task fails, whether the predicted future changes detectably, and whether that prediction matches the subsequently realized observation. The corresponding diagnostic results are reported in A2 below. We will revise the paper to state this connection explicitly.


[W1, Q2–Q3] Necessity of imagination preservation and direct validation of desynchronization

A2: Thanks for the insightful comment. We conducted an eight-task LIBERO diagnostic using disjoint clean calibration episodes and held-out evaluation episodes. Both detectors were calibrated at a 5% false-positive rate. The output detector identifies abnormal future-output changes; the consistency detector identifies disagreement between the predicted future and the subsequently realized observation.

ConditionOutput detection ↓Consistency detection ↓Failed episodes escaping output detection ↑Prediction-to-realization SSIM ↑
Clean2.5%2.5%--0.865
Action-only27.5%72.5%62.5%0.812
Img-preserving7.5%70.0%86.7%0.831
Random search10.0%65.0%71.4%0.815

The action-only attack achieves a prediction-to-realization SSIM of 0.812, close to but slightly below random search at 0.815. Imagination preservation raises SSIM to 0.831, providing a clear improvement over both baselines, while remaining below the clean value of 0.865. It also reduces output detection from 27.5% to 7.5% and increases the fraction of failed episodes escaping output detection from 62.5% to 86.7%. Consistency detection remains high and comparable (72.5% versus 70.0%), indicating that imagination preservation makes the attack substantially less visible from the predicted output and improves prediction-to-realization alignment, while the remaining mismatch is still detectable after execution.

A separate matched evaluation over all 40 LIBERO tasks further quantifies the preservation trade-off:

ObjectiveTask successAction distanceFuture distance
Action-only61.5%0.60614.014
Img-preserving65.9%0.57113.035

Imagination preservation reduces future distance by 0.979, while action distance decreases by 0.035 and task success increases by 4.4. This quantifies the attack-strength/preservation trade-off: stronger preservation produces a statistically consistent reduction in future drift at the cost of some attack strength.

AuthorsOpenReview ↗

Response to reviewer Bwr7 [2/2]

[W2, Q4] Comparisons with existing attack baselines

A3: Thanks for the valuable suggestion. We evaluated four VLA attacks on LIBERO:

VLA attackProjected full-suite target task success
UADA [1]94.6%
TMA [1]95.1%
UPA [1]98.8%
UPA-RFAS [2]98.3%

The projected target success remains high, ranging from 94.6% to 98.8%, indicating weak transfer to the evaluated WAM. This motivates an objective that directly targets the WAM action–imagination interface.

[1] Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics.

[2] When Robots Obey the Patch: Universal Transferable Patch Attacks on Vision-Language-Action Models.


[W3, Q5] Input preprocessing and purification defenses

A4: Thanks for the important consideration. We evaluated different defenses on LIBERO using the Joint WAM, with 20 trials per task. Each clean result uses the same tasks and preprocessing as its attacked counterpart.

DefenseClean successTask success under attack
None98.5%63.0%
Gaussian blur98.9%90.9%
Resize-crop ensemble76.6%6.0%
JPEG-noise ensemble95.5%90.3%
JPEG-7095.8%91.1%

Gaussian blur provides the best overall clean-utility/robustness trade-off: it preserves 98.9% clean success while achieving 90.9% success under attack, corresponding to an 8.0 pp attack-induced drop. The JPEG-noise ensemble and JPEG-70 achieve 90.3% and 91.1% success under attack, corresponding to smaller drops of 5.2 pp and 4.7 pp from their respective clean baselines. These smaller clean-to-attack gaps come with lower clean success: 95.5% for the JPEG-noise ensemble and 95.8% for JPEG-70, compared with the 98.5% undefended baseline. Resize-crop is less suitable because it reduces clean success to 76.6% and attacked success to 6.0%.


Thanks again for the thoughtful and constructive feedback. We would definitely love to further interact with reviewer Bwr7 if there are any further questions.

Dear Reviewer Bwr7,

Thank you once again for your thoughtful comments and kind words about our work. We hope that our rebuttal has helped clarify the points you raised. Should anything remain unclear, we would be grateful for the opportunity to provide further explanation.

Your kind consideration truly means a great deal to us.

Best regards,

Authors of Submission 5001

Comments on the paper

2 comments
AuthorsOpenReview ↗

PDF Rebuttal Revision Update

Dear Reviewers,

Thanks for your constructive feedback. We have updated the manuscript PDF to incorporate the revisions discussed in our rebuttal (marked in blue). The main additions are:

  1. Future-output detection and prediction–execution consistency analyses, including prediction-to-realization SSIM.
  2. Comparisons with existing VLA attacks and full-benchmark preprocessing-defense results.
  3. An updated Figure 6 with consistent task-balanced aggregation, clarified experimental settings, and statistical analysis.
  4. A precise definition of the predicted-future distance and its computation protocol.
  5. Additional related work on recent VLA attacks and closed-loop adversarial evaluation.
  6. A discussion of our simulation protocol and future real-robot evaluation.

We hope these revisions address the reviewers’ concerns and make the paper clearer and more complete. We'd love to further interact with the reviewers if there are any remaining concerns.

Sincerely,
Authors of Submission 5001

Area_Chair_FJFEOpenReview ↗

Follow-Up on the Authors’ Rebuttal

Dear reviewers,

The authors have posted their rebuttal. Please take a moment to read their response and leave a brief follow-up comment on whether it addresses the concerns you raised in your review. If any points remain unclear or you have follow-up questions, please raise them during the rebuttal period so the authors have a chance to respond. Please also indicate whether the rebuttal changes your overall assessment of the paper.

Thank you !

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.