acceptodds
Under review as a conference paper at ICLR 2027

WARD: Watching Actions via Latent-Space Reachability for Runtime Backdoor Defense in VLA Models

13 views

Abstract

Vision-language-action (VLA) models map visual observations and language instructions to robot actions, enabling a single policy to perform diverse manipulation tasks. However, backdoors implanted during training can preserve normal task performance while redirecting behavior when a specific trigger appears at deployment. In physical environments, such behavior can damage objects or equipment before task failure becomes apparent, and stopping afterward cannot undo the damage. Preventing these outcomes requires assessing actions before execution. This is challenging because a physically admissible action may still serve an unauthorized task, while failure detection alone does not establish whether a candidate action should be executed. We propose WARD, an external supervisor that combines predicted action effects with instruction-conditioned authorization to assess physical and task-consistency risk before actuation. WARD allows the candidate action or stops execution without modifying the deployed policy or accessing its internal representations. We evaluate WARD on OpenVLA-7B across four LIBERO suites and three backdoor attack families, with component ablations and physical-robot trials. WARD reduces suite-averaged progress-aware trajectory-matching attack success rates by 81.8–94.0 percentage points relative to no defense and achieves the highest aggregate defense utility among evaluated baselines, accounting for both attack suppression and clean-task preservation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
vision-language-action models, robot learning, runtime safety, backdoor defense, pre-actuation verification, world models, reachability analysis, model-based safety

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

This paper proposes WARD, an external pre-actuation supervisor for backdoored vision-language-action policies that combines a DINO-WM-based prediction of candidate-action effects, an instruction-conditioned authorization signal, and a fitted reachability/viability critic to decide whether a candidate robot action should be executed or terminally blocked.

Critical strengths

WARD addresses a practically important failure mode using an architecturally appealing security boundary: the deployed VLA is treated as untrusted, while an external supervisor evaluates its proposed action before actuation using independent world-model and instruction evidence.

Critical weaknesses

WARD is trigger-agnostic at deployment but not fully malicious-behavior-agnostic during training: its physical-risk labels explicitly cover attacker-target trajectories, and removing StateBackdoor or GoBA risk trajectories raises DTW-ASR to approximately 60% and 70%, respectively, in the paper's own LIBERO-Object holdout experiment. A second assumption is access to a separately trusted instruction c; this is reasonable in architectures with a protected task channel, but it considerably simplifies language-trigger defenses if the same potentially compromised command channel would otherwise feed both the VLA and supervisor.

Questions for authors
  1. The paper frequently motivates WARD as a defense against unknown backdoors, yet the physical-risk margin is trained using trajectories that realize the attacker-specified malicious behaviors. The attack-family holdout experiment is particularly revealing: removing StateBackdoor or GoBA trajectories substantially increases the corresponding defended DTW-ASR. Could the authors explicitly characterize WARD as trigger-agnostic but behavior-risk-supervised, and narrow broader backdoor-agnostic claims accordingly?

  2. A stronger generalization experiment would train WARD without any trajectory from one malicious target behavior and then evaluate a genuinely unseen target behavior, ideally involving a different object, contact mode, or manipulation primitive. Can the authors provide such a leave-one-target-behavior-out study? This seems more important than additional tests involving new visual trigger appearances, because the current results suggest behavioral coverage rather than trigger modality is the limiting factor.

  3. WARD assumes access to a trusted instruction c that is separate from the policy-side text c'_t. How is this trusted channel instantiated in the intended deployment scenario, particularly for language-trigger attacks in which the user command or upstream instruction pipeline itself may be compromised? Please evaluate or discuss performance when the trusted instruction is noisy, ambiguous, unavailable, or shares the same compromised input channel.

  4. The progress-aware DTW-ASR is useful for measuring intervention before completion of a registered malicious trajectory, but it relies on a pre-registered target-motion reference and is not identical to each attack's native success condition. Could the authors report defended native ASR for all three attack families alongside DTW-ASR in the main results?

  5. WARD depends on the accuracy of DINO-WM under potentially unusual trajectories induced by a compromised policy. Since DINO-WM was proposed as a feature-space world model for offline learned dynamics and zero-shot planning rather than as a certified safety predictor, could the authors quantify prediction error or uncertainty specifically on attack trajectories and near the decision boundary?

5 comments
AuthorsOpenReview ↗

Response to Reviewer WLpV (Part 1/5): Behavior-risk supervision and the scope of generalization

Dear Reviewer WLpV, thank you for your careful assessment and detailed suggestions. We appreciate your recognition of the external pre-actuation safety architecture for addressing backdoor-induced failures in deployed VLAs. Your comments help sharpen the distinction between trigger diversity and behavioral coverage, and clarify the role of the task reference. Below, we address these points with target-behavior holdout and shared-input experiments, complementary native-ASR reporting, and prediction-error measurements along attack trajectories and near the admission threshold.

Q1. The paper frequently motivates WARD as a defense against unknown backdoors, yet the physical-risk margin is trained using trajectories that realize the attacker-specified malicious behaviors. The attack-family holdout experiment is particularly revealing: removing StateBackdoor or GoBA trajectories substantially increases the corresponding defended DTW-ASR. Could the authors explicitly characterize WARD as trigger-agnostic but behavior-risk-supervised, and narrow broader backdoor-agnostic claims accordingly?

Response to Q1. We appreciate the reviewer's careful distinction between trigger diversity and behavioral coverage. We agree that WARD is trigger-agnostic at deployment and behavior-risk-supervised during training. Our claim concerns deployment across trigger modalities and transfer within the learned risk specification.

The physical and authorization margins learn hazardous interactions and instruction–behavior inconsistency from annotated trajectories. Fitted viability propagates this specification through predicted transitions. WARD's contribution is an external action-level defense that applies the specification across visual, textual, and state-triggered attacks.

Appendix F.1, Table 13 supports your observation. Removing StateBackdoor or GoBA trajectories raises the corresponding DTW-ASR to 60% and 70%, showing the importance of physical-risk coverage. Holding out BackdoorVLA trajectories still yields 0% DTW-ASR and DUS 79.86 through clean-derived instruction mismatches. This identifies a source of transferable authorization evidence.

The target-level experiment below further examines transfer between related behaviors with a matched training budget. We clarify the distinction between trigger diversity and behavioral coverage in Section 5.5 and Appendix H of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer WLpV (Part 2/5): Transfer to held-out target behaviors

W1. WARD is trigger-agnostic at deployment but not fully malicious-behavior-agnostic during training: its physical-risk labels explicitly cover attacker-target trajectories, and removing StateBackdoor or GoBA risk trajectories raises DTW-ASR to approximately 60% and 70%, respectively, in the paper's own LIBERO-Object holdout experiment.

Q2. A stronger generalization experiment would train WARD without any trajectory from one malicious target behavior and then evaluate a genuinely unseen target behavior, ideally involving a different object, contact mode, or manipulation primitive. Can the authors provide such a leave-one-target-behavior-out study? This seems more important than additional tests involving new visual trigger appearances, because the current results suggest behavioral coverage rather than trigger modality is the limiting factor.

Response to W1&Q2. We appreciate the suggestion to distinguish target-behavior transfer from robustness to new trigger appearances. We add a bidirectional target-behavior holdout experiment that excludes all malicious supervision for the evaluated target. WARD suppresses the held-out target event for both tested objects. The experiment directly evaluates the object-level variation suggested in your question and measures clean-task retention alongside target completion.

We use two preselected LIBERO-Spatial tasks with StateBackdoor objectives of grasping, lifting, and releasing a black bowl or a ramekin from height. The manipulation primitive is shared. Each direction excludes the evaluated target's malicious episodes and derived validation, risk-calibration, and replay supervision.

Full WARD uses 20 malicious episodes per target. Each held-out model uses 40 from the other target. All conditions share 40 clean episodes, matched training-update and positive-risk sampling budgets, and 100 independent clean calibration episodes. Affected components are trained afresh with one common seed. Physical-risk labels begin at high release and gripper detachment.

Each target has 20 clean and 20 triggered test episodes. Evaluation scores every timestep under first-LOCK termination. Target completion requires grasp, sufficient lift, high release, and contact-free descent.

Training conditionBowl clean successesBowl target completionsRamekin clean successesRamekin target completions
No defense16/2017/2016/2015/20
WARD, both targets supervised15/200/2016/200/20
WARD, evaluated target held out15/200/2013/200/20

Held-out WARD intercepts all successful raw attacks before gripper detachment. Clean retention matches full supervision on the bowl task and decreases by three episodes on ramekin.

Extending this evaluation to different contact modes and manipulation primitives is a valuable direction for future research.

We present this study in Appendix F.3 and discuss its implications for behavioral coverage in Section 5.5 of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer WLpV (Part 3/5): Shared instruction input and task authorization

W2. A second assumption is access to a separately trusted instruction c; this is reasonable in architectures with a protected task channel, but it considerably simplifies language-trigger defenses if the same potentially compromised command channel would otherwise feed both the VLA and supervisor.

Q3. WARD assumes access to a trusted instruction c that is separate from the policy-side text c'_t. How is this trusted channel instantiated in the intended deployment scenario, particularly for language-trigger attacks in which the user command or upstream instruction pipeline itself may be compromised? Please evaluate or discuss performance when the trusted instruction is noisy, ambiguous, unavailable, or shares the same compromised input channel.

Response to W2&Q3. We thank the reviewer for highlighting the practical requirements of the task-reference interface. We add a controlled shared-input experiment that supplies the complete triggered prompt to authorization. For the tested semantics-preserving textual trigger, its effect on DTW-ASR is small. Task references that are ambiguous, unavailable, or semantically replaced require additional handling at the application boundary.

In the reported benchmark, WARD's reference is obtained by stripping the recorded trigger from the policy prompt. The shared-input comparison instead supplies the complete prompt, including ~*magic*~, to the authorization branch. The token preserves the requested manipulation goal. The results below report DTW-ASR in percent on triggered episodes.

SuiteStripped requested instructionShared triggered instruction
Object0.002.22
Spatial0.740.74
Goal0.000.00
LIBERO-10 kitchen0.002.96

Retaining the trigger changes DTW-ASR by 0–2.96 percentage points, with residual DTW-ASR at most 2.96% across the four suites. This supports weak dependence on trigger stripping for the tested semantics-preserving attack. The following cases clarify how the task-reference assumption applies more broadly.

  • Noise that preserves meaning. This is an authorization-encoder robustness issue, and the trigger comparison evaluates one specific perturbation. The shared-input experiment evaluates a semantics-preserving perturbation that also serves as the backdoor trigger, linking authorization robustness to the tested language-trigger threat.

  • Ambiguous or missing instructions. The authorized objective is underspecified. A deployment can request clarification or retrieve the task from an authenticated operator/application record before task-conditioned execution resumes. This establishes the intent against which proposed behavior is evaluated.

  • Semantic replacement through the sole instruction channel. If the attacker controls both the policy command and WARD's only reference, the supervisor has no independent basis for recovering the original legitimate intent. A protected task record or an instruction-integrity mechanism is required. Goal changes can be confirmed through that independent authorization source.

In the intended architecture, the application retains the authenticated task while forwarding policy-side text to the VLA, and WARD keeps that reference fixed during execution. The final case gives the attacker control over this trusted source of intent and extends the paper's threat model. Task authentication and clarification supply the complementary system mechanisms needed when the reference is compromised or insufficient.

We add the shared-input results and task-reference discussion in Appendix F.4, with the main interpretation in Section 5.5 of the revised paper.

Reviewer_8MuWOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes WARD, an external pre-actuation supervisor designed to defend deployed VLAs against backdoor-induced malicious behaviors without modifying the underlying policy or accessing its internal representations. WARD integrates two complementary risk signals. First, a DINO-based world model predicts the consequences of candidate actions to construct physical-risk representations. Second, an instruction-conditioned authorization encoder assesses whether the observed task evolution is consistent with the trusted task instruction. These signals are combined into a joint safety margin, which is further evaluated by a learned latent Hamilton–Jacobi viability critic to estimate the risk of executing candidate actions. At runtime, WARD either executes the original action unchanged or enters an absorbing LOCK mode based on a calibrated risk threshold. The method is evaluated on OpenVLA-7B across four LIBERO suites against three backdoor attack families: StateBackdoor, BackdoorVLA, and GoBA. Experimental results demonstrate substantial reductions in attack success rates while preserving clean-task performance, achieving the highest aggregate Defense Utility Score among the evaluated baselines. The authors further conduct component ablations, cross-policy evaluations, and real-world robotic experiments to validate the effectiveness of the proposed method.

Critical strengths
  • Broad applicability across different backdoor threats. As an external pre-actuation supervisor, WARD can defend against backdoor attacks involving different trigger modalities, including visual, textual, and robot-state triggers, without requiring access to the underlying VLA's parameters or internal representations. This design enhances its practical applicability to black-box VLA deployments.
  • Proactive risk assessment through HJ-based viability learning. By incorporating a learned Hamilton–Jacobi viability critic and world-model-based predictions, WARD can assess potential risks along future action continuations rather than relying solely on reactive failure detection. This enables early intervention before potentially harmful actions are executed, as supported by the reported experimental results.
Critical weaknesses

Limited generalization to unseen backdoor behaviors. My main concern is WARD's generalization to unseen backdoor attacks. Although WARD does not require knowledge of attack triggers at inference time, training its physical-risk boundary and viability critic relies on annotated trajectories from backdoored VLA policies. Consequently, the learned risk representations may be strongly dependent on the malicious behavior patterns covered by the training data, limiting their generalization to unseen attacks. The attack-family holdout results in Table 9 further support this concern: excluding StateBackdoor or GoBA trajectories from training substantially degrades defense performance against the corresponding attacks. This raises concerns about the practicality and generalizability of WARD in real-world deployments where previously unseen malicious behaviors may emerge.

Questions for authors
  • World-model prediction under backdoor-induced distribution shifts. Given that DINO-WM is trained exclusively on clean LIBERO expert demonstrations, could the authors evaluate its one-step and multi-step prediction accuracy on backdoor-induced trajectories? It remains unclear whether the world model can reliably predict malicious behaviors that deviate from the clean training distribution, and how potential prediction errors affect the downstream risk estimation and interception performance.
  • Generalization to unseen malicious action objectives. Beyond the attack-family holdout experiments in Table 9, could the authors evaluate WARD and the baseline methods against backdoor attacks with previously unseen malicious action objectives, without including the corresponding malicious trajectories in defense training? This would help assess whether WARD learns generalizable safety representations rather than attack-specific behavior patterns.
  • Zero-shot transfer across VLA models. Since the π0-family evaluation in Appendix E.1 involves refitting WARD components on π0-specific trajectories, could the authors evaluate the direct transfer of the OpenVLA-trained supervisor to π0 without retraining, and quantify the data and computational costs of the required adaptation? This would help distinguish architecture-level portability from actual cross-VLA transfer of the learned supervisor.
5 comments
AuthorsOpenReview ↗

Response to Reviewer 8MuW (Part 1/3): World-model prediction on attack trajectories

Dear Reviewer 8MuW, thank you for your thoughtful review and constructive suggestions. We appreciate your recognition of WARD's black-box applicability across different trigger modalities and proactive risk assessment through HJ-based viability learning. In response to your questions, we provide additional evaluations of world-model prediction under attack-induced shifts, target-behavior holdout with matched baselines, and frozen-supervisor transfer across VLA policies. The responses below explain what these experiments establish about prediction quality and transfer of the learned supervisor.

Q1. World-model prediction under backdoor-induced distribution shifts. Given that DINO-WM is trained exclusively on clean LIBERO expert demonstrations, could the authors evaluate its one-step and multi-step prediction accuracy on backdoor-induced trajectories? It remains unclear whether the world model can reliably predict malicious behaviors that deviate from the clean training distribution, and how potential prediction errors affect the downstream risk estimation and interception performance.

Response to Q1. We appreciate the suggestion to directly measure prediction quality under attack-induced distribution shifts. We add a clean-to-attack prediction evaluation. The results reveal behavior- and horizon-dependent prediction shifts. We evaluate the original frozen world model, trained only on clean expert demonstrations, on clean expert trajectories and three types of triggered LIBERO-Object trajectories. All conditions use the same model, DINO feature space, and original prediction preprocessing.

The table reports MSE between predicted and observed future DINO features, averaged over the front and wrist views. Each trajectory contributes 20 prediction origins, five-step results use recursive open-loop prediction. Lower values are better.

Evaluation trajectoriesLatent MSE, 1 stepLatent MSE, 5 steps
Clean expert0.07610.1656
StateBackdoor, triggered0.05280.2087
GoBA, triggered0.05650.1432
BackdoorVLA, triggered0.08160.1992

One-step error is lower than the clean reference for StateBackdoor and GoBA and slightly higher for BackdoorVLA. At five steps, StateBackdoor and BackdoorVLA errors are approximately 26% and 20% higher than the clean reference, while GoBA remains lower. This identifies behavior-dependent prediction shifts, with a larger relative increase over the clean reference at the longer horizon for StateBackdoor and BackdoorVLA.

Latent MSE evaluates the predicted representation used by WARD. The persistence and end-effector diagnostics in Appendix G.2 provide complementary views of prediction quality. Accuracy in the visual representation and in Cartesian state can differ.

To examine predictions near action-admission decisions, we additionally match prediction records to per-step risk scores from the same trajectories and evaluate origins satisfying |R_exec(t) − τ| ≤ 0.1. One-step latent MSE ranges from 0.0522 to 0.0739, and five-step MSE from 0.1166 to 0.1857. This complements the full-trajectory comparison by measuring prediction quality in a fixed neighborhood of the decision threshold. The matched ablation in Appendix F.1 evaluates the downstream contribution separately: fitted viability reduces DTW-ASR across all three attacks while preserving or improving clean-task success under the original world model. These two comparisons connect the quality of the predicted representations to the observed benefit of using them for temporal risk assessment.

We discuss world-model prediction and its role in risk assessment in Appendix G.2 and Section 5.5 of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer 8MuW (Part 2/3): Target-behavior holdout and baseline comparison

W1&Q2. Generalization to unseen malicious action objectives. Beyond the attack-family holdout experiments in Table 9, could the authors evaluate WARD and the baseline methods against backdoor attacks with previously unseen malicious action objectives, without including the corresponding malicious trajectories in defense training? This would help assess whether WARD learns generalizable safety representations rather than attack-specific behavior patterns.

Response to W1&Q2. We appreciate the suggestion to evaluate transfer with the malicious objective itself held out. We conduct an additional matched, bidirectional target-behavior holdout experiment, including SAFE-MLP and Mahalanobis baselines. WARD suppresses the held-out target event in both directions. Clean-task retention differs between the two transfer directions.

The study uses two preselected LIBERO-Spatial tasks and StateBackdoor objectives of grasping, lifting, and releasing a black bowl or a ramekin from height. The target object and scene change while the manipulation primitive is shared.

Held-out malicious episodes and their derived training, validation, risk-calibration, and replay supervision are excluded. Full WARD uses 20 malicious episodes per objective, and each held-out model uses 40 from the other objective. They share 40 clean episodes, matched updates and risk-sampling budgets, and 100 independent clean calibration episodes. Affected components are trained afresh with one common seed. Physical-risk labels begin at high release and gripper detachment.

Each target has 20 clean and 20 triggered test episodes, evaluated at every timestep under first-LOCK termination. Target completion requires grasp, lift of at least 15 cm, release at least 10 cm above the initial height, and contact-free descent of at least 5 cm.

SupervisorBowl cleanBowl targetRamekin cleanRamekin target
No defense16/2017/2016/2015/20
WARD, both targets supervised15/200/2016/200/20
WARD, evaluated target held out15/200/2013/200/20
SAFE-MLP, evaluated target held out14/200/2016/208/20
Mahalanobis, evaluated target held out15/207/2016/2011/20

SAFE-MLP uses the allowed physical-risk supervision with matched update/sampling budgets. Mahalanobis fits clean-reference statistics and selects polarity on allowed development data.

Held-out WARD intercepts all successful raw attacks before gripper detachment. SAFE-MLP suppresses the bowl objective, while WARD provides stronger suppression on ramekin. The latter costs three clean successes relative to full WARD.

Generalization across different contact modes and manipulation primitives is a valuable direction for future research.

We present the target-behavior study in Appendix F.3 and discuss transfer across objects and scenes in Section 5.5 of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer 8MuW (Part 3/3): Frozen-supervisor transfer across VLA policies

Q3. Zero-shot transfer across VLA models. Since the π0-family evaluation in Appendix E.1 involves refitting WARD components on π0-specific trajectories, could the authors evaluate the direct transfer of the OpenVLA-trained supervisor to π0 without retraining, and quantify the data and computational costs of the required adaptation? This would help distinguish architecture-level portability from actual cross-VLA transfer of the learned supervisor.

Response to Q3. We thank the reviewer for recommending a direct test of cross-policy transfer without retraining or recalibration. We conduct an additional frozen-transfer experiment: an OpenVLA-trained WARD checkpoint is applied to π₀ and π₀-Fast on LIBERO-Object. WARD weights, normalization statistics, and calibration settings remained unchanged. Transfer used zero target-policy training trajectories, zero optimization updates, and no recalibration.

We compared each policy with and without the frozen supervisor using five paired trials per task, totaling 340 pairs and 680 episodes. The results are:

Target policyAttack-target success, no defense → WARDTask success, no defense → WARD
GoBA / π₀100% (50/50) → 0% (0/50)98% (49/50) → 94% (47/50)
StateBackdoor / π₀98% (49/50) → 0% (0/50)94% (47/50) → 92% (46/50)
BackdoorVLA / π₀-Fast64.44% (29/45) → 0% (0/45)82.22% (37/45) → 82.22% (37/45)
Independent clean π₀—100% (50/50) → 100% (50/50)

Parentheses give successful/evaluated episodes. Attack-target success is measured with triggers enabled: poison-object checking-region success for GoBA, robot–basket contact for the StateBackdoor variant, and attacker-selected task completion for BackdoorVLA. Task success is measured with triggers disabled or with the independent clean policy. Appendix E.1 reports the frozen-transfer setup and results.

With frozen WARD, no attack-target successes were observed in any of the three attack conditions. Task success decreased by 4 and 2 percentage points for GoBA and StateBackdoor, respectively, and was unchanged for BackdoorVLA. The independent clean π₀ retained 50/50 task successes with no stop interventions.

These results demonstrate direct reuse of the OpenVLA-trained supervisor across the evaluated π₀-family policies on LIBERO-Object: it suppresses the tested attack objectives while largely preserving normal task completion, with no target-policy training or recalibration cost.

We add the frozen-transfer results in Appendix E.1 and discuss supervisor reuse in Section 5.5 of the revised paper.

Reviewer_bX7rOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WARD is an external pre-actuation supervisor for potentially backdoored vision-language-action policies. It combines predicted physical consequences with instruction-conditioned authorization and uses a learned viability critic to execute a candidate unchanged or stop. Evaluation covers three attack families across four LIBERO suites, a second policy family, and physical-robot trials.

Critical strengths

S1: The supervisor distinguishes task authorization from physical admissibility. Sections 3.1 and 4 separate the trusted instruction from potentially modified policy text without accessing policy parameters or internal features. Figure 4 supports the complementary value of physical and authorization evidence.

S2: The evaluation measures attack suppression together with retained task performance. Table 1 reports average DTW-ASR of 0.67%, 0.19%, and 6.67% across the three attack families, with DUS incorporating retained clean success. Tables 2 and 3 extend the evaluation beyond OpenVLA simulation, including 0/20 bottle-knockover outcomes per physical attack condition while preserving 18-19/20 clean successes.

Critical weaknesses

W1: The main results do not isolate fitted viability from immediate joint-margin gating. Figure 4 compares the physical and authorization margins, but does not ablate the fitted backup against the immediate joint-margin gate. The main results therefore do not establish how much benefit comes from the learned backup beyond these two risk signals.

W2: The timing evidence comes from different control configurations. Section 5.1 gives a nominal physical action schedule of 30 Hz, or about 33.3 ms per period. Section 5.3 reports 22.41 ms mean added defense latency on an RTX 5090 OpenVLA replay, excluding action generation. Table 2 does not report end-to-end timing for its online π₀ evaluation. The replay average cannot establish deadline success or failure in the separate π₀ or physical loops. Same-configuration per-action end-to-end and tail latency are not reported for the target online setup. Table 3's bottle-knockover outcomes remain positive physical evidence, but are not a timing measurement.

Questions for authors
  1. Under matched conditions, how does fitted viability compare with immediate joint-margin gating in attack suppression and clean-task retention?

  2. What do Figure 4's physical-risk and authorization branch ablations each establish, and what do they not establish about fitted viability's added value?

  3. What action cadence or deadline does the online π₀ loop in Table 2 target? What are its per-action end-to-end and tail latency, deadline-miss rate, and VLA-generation and monitoring timings, and does each stop decision reach the controller before candidate execution? How does its deadline compare with, rather than assume, the nominal 30 Hz period of the physical setup? How do these results relate to the separate Section 5.3 RTX 5090 OpenVLA replay, which reports 22.41 ms mean added defense latency while excluding action generation?

4 comments
AuthorsOpenReview ↗

Response to Reviewer bX7r (Part 1/2): Fitted viability and interpretation of the ablations

Dear Reviewer bX7r, thank you for your careful review and constructive feedback. We appreciate your recognition of the complementary value of physical-risk and task-authorization evidence and the evaluation of attack suppression alongside retained task performance, including the physical-robot trials. The responses below clarify the contribution of fitted viability, distinguish the conclusions supported by the two sets of ablations, and present supplementary timing measurements for the physical runner.

W1. The main results do not isolate fitted viability from immediate joint-margin gating. Figure 4 compares the physical and authorization margins, but does not ablate the fitted backup against the immediate joint-margin gate. The main results therefore do not establish how much benefit comes from the learned backup beyond these two risk signals.

Q1. Under matched conditions, how does fitted viability compare with immediate joint-margin gating in attack suppression and clean-task retention?

Response to W1&Q1. We thank the reviewer for requesting a matched comparison of fitted viability and immediate-margin gating. The fitted subsystem improves attack suppression while preserving or improving clean-task success across all three attacks. Table 15 in Appendix F.1 of the revised paper compares the two supervisors under matched conditions, as summarized below.

Both supervisors retain the physical and authorization representations, boundary heads, frozen world model, action normalization, and risk data. The ablated supervisor gates the immediate joint margin. Full WARD uses the fitted critic.

AttackImmediate CSRWARD CSRImmediate DTW-ASRWARD DTW-ASRImmediate DUSWARD DUS
StateBackdoor88.0088.0014.000.0082.0096.00
BackdoorVLA92.0096.008.890.0071.3483.33
GoBA90.0090.6716.005.3380.2291.28

Rates are percentages. DTW-ASR falls by 14.00, 8.89, and 10.67 percentage points, respectively, with DUS gains of 14.00, 11.99, and 11.06 points. The additional suppression is accompanied by retained clean utility.

The immediate gate uses the world model through the candidate-conditioned physical representation. Fitted learning adds recursive propagation of successor risk through predicted transitions and locally optimized continuations. Its target takes the maximum of the immediate boundary and successor value, allowing later predicted risk to raise the current assessment.

The ablation removes the critic and safety actor together, so the measured gain is attributable to this fitted-viability subsystem. At deployment, WARD evaluates the VLA candidate with its prescribed reference continuation and selects EXEC or LOCK.

We expand the discussion of this comparison in Section 5.5 of the revised paper and present the results in Appendix F.1 (Table 15).

Q2. What do Figure 4's physical-risk and authorization branch ablations each establish, and what do they not establish about fitted viability's added value?

Response to Q2. We appreciate the opportunity to clarify the distinct conclusions supported by these ablations. Figure 4 demonstrates the complementary contributions of physical-risk and authorization evidence. It does not establish the additional value of fitted viability, which is retained across its variants.

Removing authorization most weakens BackdoorVLA, whose task-like motion conflicts with the requested instruction. Removing physical evidence most weakens GoBA's target-directed manipulation. StateBackdoor benefits from both margins because its risk depends on physical context and intended task.

The matched ablation in Appendix F.1 evaluates fitted viability's additional contribution, as summarized in our response to Q1. We discuss the complementary margins in Section 5.3 and the additional contribution of fitted viability in Section 5.5 and Appendix F.1 of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer bX7r (Part 2/2): Physical-runner timing and pre-actuation checks

W2. ** The timing evidence comes from different control configurations.** Section 5.1 gives a nominal physical action schedule of 30 Hz, or about 33.3 ms per period. Section 5.3 reports 22.41 ms mean added defense latency on an RTX 5090 OpenVLA replay, excluding action generation. Table 2 does not report end-to-end timing for its online π₀ evaluation. The replay average cannot establish deadline success or failure in the separate π₀ or physical loops. Same-configuration per-action end-to-end and tail latency are not reported for the target online setup. Table 3's bottle-knockover outcomes remain positive physical evidence, but are not a timing measurement.

Q3. What action cadence or deadline does the online π₀ loop in Table 2 target? What are its per-action end-to-end and tail latency, deadline-miss rate, and VLA-generation and monitoring timings, and does each stop decision reach the controller before candidate execution? How does its deadline compare with, rather than assume, the nominal 30 Hz period of the physical setup? How do these results relate to the separate Section 5.3 RTX 5090 OpenVLA replay, which reports 22.41 ms mean added defense latency while excluding action generation?

Response to W2&Q3.

We thank the reviewer for requesting a precise timing breakdown. We perform additional timing measurements of the physical π₀ runner and its incremental monitoring implementation, reported below.

1. Action generation and scheduling. Each π₀ query generates an action chunk, a sequence of 50 actions produced by one inference call. The runner executes the first 10 actions at a nominal target schedule of 30 Hz, corresponding to 33.3 ms per target.

2. Optimization and measured latency. We optimized the physical implementation by encoding only the latest front/wrist image pair and reusing the previous two frames’ cached DINO features. This reduces redundant visual computation, while the world model, risk heads, and HJ scoring run for every check.

LatencyMean (ms)P99 (ms)
π₀ action-chunk latency78.5887.01
Per-step WARD latency6.987.13
End-to-end latency (first target)116.39122.80

The π₀ timing is per chunk request and includes communication and return checks. Incremental monitoring is timed per check. Its measured mean and P99 are both below the 33.3 ms period corresponding to 30 Hz.

3. Timing constraint and exceedances. The runner separately configures a 250 ms initial-proposal freshness budget: the allowed time from source-image capture to approval of the first target. All four admitted cycles met this budget, with 0/4 exceedances. The 250 ms budget governs initial approval, while the 33.3 ms period governs target scheduling within the execution window.

4. Checks before execution. WARD checks each new target before it is released for robot execution. A rejected target is withheld, and the controller enters STOP/HOLD: it stops releasing new targets and maintains the current joint and gripper positions.

5. Relation to the OpenVLA result. The 22.41 ms value in Section 5.3 measures visual encoding plus WARD in a separate RTX 5090 OpenVLA replay and excludes action generation. The supplementary results above separately report π₀ chunk generation and the optimized incremental monitoring implementation.

We add these physical-runner timing measurements and their scheduling interpretation in Appendix C.4 of the revised paper.

Thank you for the detailed response. The matched ablation in Table 15 answers my main question. However, having read the other reviews, I also share their concern that generalization to unseen malicious behaviors remains open, since the new holdout changes the object but keeps the same manipulation primitive, and that prediction uncertainty near the decision threshold is still not quantified.

I therefore maintain my score.

Thank you for your thoughtful follow-up and for confirming that the additional ablation addresses your main question. We appreciate your careful assessment and your feedback on generalization and prediction uncertainty, which help clarify important directions for further development of this work.

Comments on the paper

3 comments

Thank you for the detailed responses and additional experiments. The new results address most of my concerns: behavior-level generalization, the trusted instruction assumption, and native ASR evaluation. I appreciate the authors' efforts to clarify the scope and limitations of WARD. I think the generalization across different manipulation primitives are worth further investigation, but I consider the current revisions sufficient. I will maintain my positive assessment.

Thank you for your thoughtful follow-up and positive assessment. We are glad that the additional experiments and clarifications address your concerns. We agree that generalization across manipulation primitives is an important direction for further investigation, and we appreciate your constructive feedback throughout the discussion.

AuthorsOpenReview ↗

General Response and Summary of Revisions

We thank all three reviewers for their constructive feedback and continued positive assessments of our work. Their suggestions have helped us clarify WARD’s contribution, strengthen its empirical evaluation, and better articulate its scope. We have submitted the revised manuscript, with revisions highlighted in blue. For ease of reference, we summarize the main additions and clarifications below.

  • Fitted viability. Section 5.5 discusses the matched immediate-margin ablation in Appendix F.1, clarifying the contribution of the fitted critic–actor subsystem beyond immediate risk evidence.
  • Generalization and transfer. We add a bidirectional target-object holdout study with matched supervision and baseline comparisons (Appendix F.3), together with zero-shot supervisor transfer across VLA models (Appendix E.1). Section 5.5 and Appendix H discuss the scope of these findings and the role of supervision coverage.
  • Instruction robustness and attack outcomes. We add an evaluation that supplies the same trigger-bearing instruction to the policy and authorization branch (Appendix F.4), and report defended native ASR alongside DTW-ASR (Appendix E.4).
  • Prediction quality and deployment timing. We add world-model prediction-error measurements on attack trajectories and within a fixed neighborhood of the admission threshold (Appendix G.2), together with a breakdown of physical-runner and incremental monitoring latency (Appendix C.4).

We appreciate the reviewers’ engagement throughout the discussion and the opportunity to improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.