acceptodds
Withdrawn from ICLR 2027

Latent Interface Purification for Robust World Action Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

World-Action Models and vision-language-action policies transform visual observations into robot actions through learned visual latents. This latent interface is a vulnerable point: small visual perturbations can shift the representation consumed by the action expert, causing errors to compound during closed-loop control. Existing defenses commonly require retraining the full policy or transform inputs in pixel space, which can be costly or disrupt task-relevant visual information. We introduce Latent Interface Purification, a plug-and-play defense inserted between a frozen visual encoder and a frozen action expert. LIP learns bounded, input-conditioned residual corrections in the native latent space while leaving the underlying policy unchanged. Its scale-normalized multi-scale architecture is initialized as an exact identity map and is trained to recover clean representations while preserving clean inputs. Across diverse world-action-model backbones and two closed-loop manipulation benchmarks, LIP substantially improves task success under universal visual attacks, with limited loss of nominal performance. It also transfers zero-shot across distinct environments and robotic embodiments, while adding minimal isolated preprocessing overhead of 1.26 ms. These results establish latent-interface purification as an effective, practical direction for improving the adversarial robustness of generative robot policies.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Latent Space, World Action Models, Robustness, Purification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_7VXWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Latent Interface Purification (LIP), a plug-and-play defense for World Action Models that operates between a frozen visual encoder and a frozen action expert. Instead of modifying input images or retraining the underlying policy, LIP learns bounded, input-conditioned residual corrections to recover clean visual representations from adversarially perturbed latents. The proposed method combines RMS-based latent normalization, multi-scale dilated convolutions, signed modulation, and identity initialization. It is trained using paired clean and attacked latent representations, together with an explicit clean-consistency objective to preserve nominal policy behavior. Experiments across 4 WAM backbones on LIBERO and RoboTwin demonstrate substantial recovery in closed-loop task success under universal adversarial perturbations.

Critical strengths
  1. The paper investigates a meaningful problem of improving adversarial robustness without retraining large robotic policies. Inserting a lightweight purifier between the visual encoder and action expert provides a practical alternative to full-policy adversarial training.

  2. The experimental results cover four WAM architectures and both single-arm and bimanual manipulation benchmarks (Table 1). For example, LIP improves FastWAM's attack success from 35.38% to 92.62% on RoboTwin and from 31.00% to 98.50% on LIBERO-Spatial, while incurring only a small clean-performance reduction.

  3. The paper includes comparisons against pixel-space and latent-space defenses, component ablations, cross-benchmark transfer, adaptive SPSA attacks, and perturbation-budget analyses. The reported 1.262 ms isolated inference latency in Line409 also supports the computational practicality of the proposed method.

Critical weaknesses
  1. LIP primarily combines established techniques, including RMS normalization, dilated residual convolutions, bounded latent corrections, and zero-initialized output projections. The signed modulation mechanism in Eq. (6) multiplies two learnable correction maps, but its advantage over directly predicting a bounded residual remains unclear. Although removing signed modulation reduces defended success from 93.92% to 88.50% in Table 4, this ablation also changes the network architecture and capacity.

  2. Although the threat model assumes attacker access to the visual encoder and action expert, and LIP is differentiable, the adaptive evaluation only uses SPSA with 1,800 queries (Table 5). Without gradient-based attacks optimized through the complete defended pipeline, it is unclear whether the reported robustness reflects LIP's effectiveness or insufficient attack strength. Multi-restart white-box PGD would provide a more convincing evaluation.

  3. The main comparison uses only a Latent MLP as the learned latent-space baseline, achieving 81.25% defended success versus LIP's 93.92% in Table 3. However, LIP has substantially more parameters (1.299M vs. 0.091M) and incorporates spatial convolutions as shown in Table 6. It remains unclear whether the improvement comes from LIP's specific architectural design or greater model capacity. Comparisons with parameter-matched convolutional latent denoisers would strengthen the claims.

Questions for authors
  1. What specific advantage does the signed modulation mechanism provide over directly predicting a bounded residual correction? Could the authors clarify its methodological novelty and evaluate it against a parameter-matched residual network without signed modulation?

  2. Given that LIP is differentiable, why is the adaptive evaluation limited to SPSA? How does LIP perform against stronger gradient-based attacks, such as multi-restart PGD optimized through the complete defended pipeline?

  3. Would LIP maintain its performance advantage over a parameter-matched convolutional latent denoiser trained using the same data and objective? This comparison would help distinguish the benefits of LIP's architecture from those of increased model capacity.

  4. Since LIP is trained primarily on universal adversarial perturbations, how well does it generalize to other attack types, such as per-observation adversarial perturbations or localized adversarial patches?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 7VXW (1/3)

We sincerely thank the reviewer for the professional and thorough feedback. Below, we respond to your four questions in sequence.

Question 1 (Reviewer 7VXW Q1, W1): Specific Advantages of Signed Modulation Mechanism Over Direct Bounded Residuals; Comparison with Parameter-Matched Residual Network

Response:

Methodologically, the key distinction lies in the control objective at the frozen WAM interface. Rather than inverting perturbations coordinate-by-coordinate for perceptual or generic feature reconstruction, LIP uses input-conditioned gating and signed modulation to determine which latent coordinates to adjust and in which direction—selectively rectifying features that govern downstream action generation.

Theoretically, Section 4.3 (Eq. 9) and Proposition 2 (Appendix A.1) prove that our joint recovery and clean-consistency objective upper-bounds both attacked-action error and clean-action distortion under local Lipschitz continuity (interpreted in Section 4.4, with multi-scale and gating components ablated in Table 4). Please refer to our Official Comment at the top of the replies for the full mechanism comparison table and 876-frame closed-loop diagnostic results.

Signed modulation decouples candidate correction from input-conditioned coordinate-wise selection. In z^=x−gψ(x)⊙Δψ(x)\hat z = x - g_\psi(x) \odot \Delta_\psi(x), gψ(x)∈[−1,1]g_\psi(x) \in [-1, 1] attenuates, preserves, or reverses candidate adjustments, while ∣Δψ(x)∣≤ρr(x)|\Delta_\psi(x)| \le \rho r(x) enforces the magnitude envelope. The network thus selects correction direction and magnitude according to local representation state. Zero-initialization of the candidate projection layer ensures strict identity behavior upon insertion, and clean consistency anchors nominal action execution.

While both direct bounded residuals and multiplicative modulation impose magnitude bounds, the mechanistic distinction lies in the input-conditioned multiplicative decomposition. The table below compares the complete architecture with an ablation removing signed modulation (parameter-matched to standard convolution):

ConfigurationParameters ↓\downarrowClean+Method ↑\uparrowDefended ↑\uparrowΔdef↑\Delta_{\text{def}} \uparrowedef↓e_{\text{def}} \downarrowdclean↓d_{\text{clean}} \downarrow
Undefended Baseline092.4048.88———
w/o Signed Modulation1.294M87.5088.50+39.620.0450.009
LIP (Ours)1.299M91.4593.92+45.040.0380.002

Full LIP achieves 93.92% defended success compared to 88.50% without signed modulation (a 5.42 pp gain); clean action distortion is 0.002 vs. 0.009. Thus, signed modulation contributes significantly to both task recovery and nominal action fidelity.


Question 2 (Q3, W3): Maintaining Performance Advantage Over Parameter-Matched Convolutional Latent Denoiser Under Identical Data and Objectives

Response:

Below, we isolate the comparison between the parameter-matched convolutional baseline and LIP. These comparisons confirm that the performance gains originate from structured spatial multi-scale correction and gating rather than model capacity.

MethodInterfaceClean+Method ↑\uparrowDefended ↑\uparrowΔdef↑\Delta_{\text{def}} \uparrowedef↓e_{\text{def}} \downarrowdclean↓d_{\text{clean}} \downarrow
Undefended BaselineNone92.4048.88———
RGB U-Net (Original, 0.222M)Pixel84.7573.50+24.620.0200.003
Scaled RGB U-Net (Matched, 1.368M)Pixel84.7574.25+25.370.0210.003
Latent MLP (Original, 0.091M)Latent88.5081.25+32.370.0420.008
Scaled Latent MLP (Matched, 1.364M)Latent88.0082.50+33.620.0310.022
Standard Conv (Matched, 1.294M)Latent89.7586.25+37.370.0460.011
LIP (Ours, 1.299M)Latent91.4593.92+45.040.0380.002
AuthorsOpenReview ↗

Response to Reviewer 7VXW (2/3)

Question 3 (Reviewer 7VXW Q2, W2): Why Adaptive Evaluation is Limited to SPSA Given Differentiability; Performance of Multi-Restart PGD Against the Complete Defended Pipeline

Response:

Implementation Mechanism, Physical Feasibility, and Evaluation of White-Box End-to-End Action-Loss PGD:

  1. Implementation Mechanism and Engineering Adaptation:
    Official open-source WAM repositories were released strictly for derivative-free inference deployment; backward computational graphs are severed via @torch.no_grad(), action .detach(), or truncation into NumPy arrays. To rigorously address the reviewer's concern regarding white-box adaptive attacks, we restructured the underlying execution pipelines, bypassed official inference decorators and truncation constraints, implemented model-specific low-level adaptations, and successfully established a closed-loop evaluation framework for end-to-end action-loss PGD backpropagated through multi-step diffusion denoising.

  2. Time Cost and Physical Real-World Feasibility Analysis:

  • Extremely High Time Cost: In online closed-loop rollouts, every replanning step requires end-to-end backpropagation through a 14B-parameter multi-step diffusion Transformer. A single rollout takes over 10× longer than nominal unattacked rollouts (a nominal task rollout takes ~40–60 s, whereas an online PGD rollout takes 550–600 s).
  • Minimal Physical Feasibility in Real Deployment:
    1. System Privileges and Intrusiveness: The attacker must obtain full white-box access to the robot's real-time inference system, including direct manipulation of GPU memory.
    2. Excessive Latency Violates Real-Time Control: The multi-second latency incurred by online backpropagation cannot satisfy the strict 10 Hz–30 Hz real-time control constraints of physical robots.
    3. Infeasibility of Passive Physical Deployment: Unlike physical UAPs (optical filters) or adversarial patches (printed stickers), online PGD cannot be passively deployed in physical environments via light reflection.
    4. Conclusion: Such end-to-end action PGD serves as a "theoretical extreme stress test" of policy and defense robustness rather than a practical deployment threat, and should not be used to dismiss the efficacy of LIP.
  1. Multi-Suite Benchmark Evaluation:
    Due to the substantial computational cost (10× testing duration), we systematically conducted benchmark experiments on our primary backbones, FastWAM and Light-WAM: covering all 4 suites of LIBERO, randomly selecting 2 tasks per suite, and evaluating 10 rollouts per task across 4 perturbation budgets (1/255,2/255,4/255,8/2551/255, 2/255, 4/255, 8/255). The results are summarized below:
ModelAttackClean1/2552/2554/2558/255
FastWAMPGD79/805/800/800/801/80
FastWAMPGD+LIP—75/8042/8016/800/80
Light-WAMPGD78/806/800/800/800/80
Light-WAMPGD+LIP—76/8038/8013/800/80
  1. Results Analysis and Objective Defense Boundaries:
    When subjected to this theoretical extreme white-box attack:
  • At small perturbation budgets (1/2551/255 and 2/2552/255): The undefended policy immediately collapses even under faint perturbations, while LIP demonstrates remarkable recovery capabilities.
  • At larger perturbation budgets (4/2554/255 and 8/2558/255): Because end-to-end action gradients nonlinearly exploit high-frequency pathways through the multi-step denoising network, residual errors exceed gripper tolerances and full recovery is challenging. This clearly and objectively delineates the defense boundaries of LIP under extreme white-box stress testing.
AuthorsOpenReview ↗

Response to Reviewer 7VXW (3/3)

Question 4 (Q4): Generalization to Other Attack Types (e.g., Per-Observation Adversarial Perturbations or Localized Adversarial Patches)

Response:

As shown in Table 12 below, when incorporating multi-corruption environmental augmented training, LIP's robustness under compound threats (adversarial attacks overlaid with environmental degradations) improves substantially. This lightweight data adaptation and incremental retraining (with merely 750 additional multi-environment frames) convincingly validates that LIP possesses strong architectural adaptability and plasticity, and can readily generalize to other attack modes and diverse perturbation patterns through minimal retraining.

Table 12: Closed-loop success and action errors under composite corruptions on FastWAM with multi-corruption augmented training.

CorruptionSeverityUndefended SR (%)LIP SR (%)Δ\Delta (pp)eundefe_{\text{undef}}eLIPe_{\text{LIP}}Δe\Delta eReR_e (%)
Gaussian noise + UAP132.5081.25+48.750.1470600.047106+0.099954+67.97
21.2576.25+75.000.2121700.056151+0.156018+73.53
30.0065.00+65.000.5182100.084876+0.433333+83.62
40.0045.00+45.000.7820100.134024+0.647986+82.86
50.0028.75+28.750.7771450.170649+0.606496+78.04
Motion blur + UAP135.0078.75+43.750.1989900.053755+0.145235+72.99
218.7581.25+62.500.2539170.073962+0.179955+70.87
30.0083.75+83.750.4764520.101922+0.374530+78.61
40.0086.25+86.250.5046790.123502+0.381177+75.53
50.0082.50+82.500.5164730.130119+0.386354+74.81
Salt-and-pepper + UAP10.0068.75+68.750.4473830.165685+0.281698+62.97
20.0040.00+40.001.1942000.437870+0.756330+63.33
30.0016.25+16.251.2922020.504537+0.787665+60.96
40.000.00+0.001.4415880.583327+0.858261+59.54
50.000.00+0.001.4414680.722571+0.718898+49.87
Brightness shift + UAP140.0083.75+43.750.1386450.065849+0.072796+52.51
237.5080.00+42.500.1505280.080502+0.070026+46.52
327.5076.25+48.750.1667890.102854+0.063935+38.33
48.7581.25+72.500.2040170.128996+0.075021+36.77
51.2566.25+65.000.2317980.153982+0.077816+33.57
Contrast drop + UAP10.0071.25+71.250.4425720.083551+0.359021+81.12
20.0068.75+68.750.4992620.102758+0.396504+79.42
30.0051.25+51.250.7433380.125332+0.618005+83.14
40.0035.00+35.000.7882070.190426+0.597780+75.84
50.000.00+0.001.2964490.575610+0.720839+55.60

Thank you for the detailed rebuttal and extra experiments. I appreciate the authors' efforts in addressing my concerns.

Howevery, I still have the remaining concerns:

  1. Why are the values ​​in Figure 3 and Figure 8 identical across all aspects—including the experimental setup—yet the confidence levels differ? Could the authors please provide some clarification? I am also not certain whether using confidence intervals to present performance results is standard practice. I didn't see any similar illustration in the relevant papers before.

  2. Even LIP would suffers severe performance degradation within bigger budgets like 4,8/255 under white-box PGD. Authors should make this failure boundary explicit in main text. Could you also clarify what practical threat model other than LIP is thought to defend against, and what kind of attack areas are expicitly out-of-scope?

Reviewer_1n8uOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LIP, a visual adversarial defense method for frozen WAMs. LIP introduces a lightweight residual network between the visual encoder and action expert. It uses scale normalization, multi-scale feature extraction, and signed modulation to correct attacked latents toward their clean representations. The model is trained on paired clean and attacked latents, with a clean-consistency loss to preserve the original policy behavior. Experiments on LIBERO and RoboTwin with four Wan-family WAM backbones report substantial improvements under universal adversarial perturbations, limited degradation on clean inputs, and strong cross-benchmark transfer performance.

Critical strengths
  1. The authors attempt to improve the adversarial robustness of WAMs by correcting latent representations between the frozen visual encoder and action expert, without retraining the original policy. The method is simple, easy to deploy, and adds little inference overhead.

  2. The experiments cover four WAM backbones and two simulation benchmarks, including cross-benchmark transfer, component ablations, and adaptive attack evaluations. LIP substantially improves task success under universal visual adversarial perturbations. The simulation results provide reasonable evidence for its effectiveness.

Critical weaknesses
  1. Using high-level representation consistency to guide adversarial denoising is already a well-established approach. LIP mainly applies this idea to the interface between the visual encoder and action expert in WAMs. The design is reasonable from an engineering perspective, but does not sufficiently address the unique aspects of action generation and closed-loop control in WAMs. The novelty is therefore relatively limited.

  2. The paper does not include real-world robot experiments. All closed-loop robot evaluations are conducted in simulation. Real-world environments involve more complex and interacting visual changes, and the current evidence is insufficient to support the claims about practical deployment and robustness across environments.

  3. The theoretical analysis mainly establishes bounded latent corrections and conditional bounds relating latent reconstruction error to relative action error of generated action chunks. It does not explain why the learned corrections generalize to unseen attacks or improve long-horizon closed-loop behavior. In particular, FastWAM achieves a higher defended success rate than its clean baseline on RoboTwin. This suggests that clean-latent recovery alone may not fully explain the performance gains and deserves further investigation.

  4. Some important evidence relies too heavily on the appendix, making the argument in the main paper less coherent. For example, the main text highlights positive results under natural noise corruptions but does not discuss the performance degradation under brightness and contrast changes reported in the appendix.

  5. Tables 1 and 2 report clearly inconsistent attacked baselines on LIBERO.

Questions for authors
  1. Could the authors provide real-world closed-loop robot experiments? This would be important to my assessment and could affect my final rating. The appendix shows that LIP performs inconsistently under brightness and contrast changes. Does the method remain effective when these visual changes occur together with adversarial perturbations? Without such validation, the claims about practical deployment and generalization should be more limited.

  2. In Table 1, FastWAM achieves a higher defended success rate than the clean baseline on RoboTwin. Is this difference statistically significant? Why does the defended policy outperform the clean policy? Could the authors further explain the source of this improvement?

  3. The appendix shows that LIP can increase action error under Brightness Shift and Contrast Drop. Could the authors present the complete results in the main text and clarify the range of perturbations for which the method is effective, rather than highlighting only the positive results?

  4. The attacked baselines on LIBERO differ substantially between Tables 1 and 2. Are these differences caused by different attacks or evaluation settings? Please clarify.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 1n8u (1/3)

We sincerely thank the reviewer for the professional and thorough feedback. Below, we respond to your critical weakness regarding novelty and your four questions in sequence.

Weakness 1 (W1): High-Level Representation Consistency is Well-Established; Unique Aspects of WAM Action Generation and Closed-Loop Control

Response:

The key distinction lies in the control objective at the frozen WAM interface. Unlike representation-consistency denoising that pulls features uniformly toward clean targets, LIP is co-designed around closed-loop WAM action generation in three ways: (1) it operates directly on latents consumed by the frozen action expert to minimize downstream action-chunk deviation rather than inverting perturbations coordinate-by-coordinate; (2) clean-path anchoring and exact identity initialization explicitly preserve nominal closed-loop behavior; and (3) Section 4.3 (Eq. 9) and Proposition 2 (Appendix A.1) prove that our objective upper-bounds both attacked-action error and clean-action distortion, while Section 4.4 characterizes how bounded corrections propagate across closed-loop steps.

Please refer to our Official Comment at the top of the replies for the full mechanism comparison table and closed-loop diagnostic validation.


Question 1 (Q1, Q3; W2, W4): Real-World Robot Closed-Loop Experiments; Robustness Under Concurrent Environmental Changes and Adversarial Perturbations; Complete Results for Photometric Degradations and Effective Boundaries

Response:

1. Current Status of Physical Robot Closed-Loop Evaluation:
Our lab’s robotic arm model does not match the robot hardware platform used in the training data for the four WAMs, making it difficult to reliably reproduce real-robot experiments within the short rebuttal period. If the paper is accepted, we will make every effort to add real-robot experiments to the final version; until then, we will restrict deployment-related conclusions to the simulated scenarios we have evaluated.

2. Complete Results and Effective Boundaries in the Main Paper (Table 9):
Regarding the slight decline under brightness shift, contrast drop, and motion blur: in the original manuscript, the action error percentages for motion blur, brightness shift, and contrast drop had a one-decimal offset during table compilation. This has been corrected and fully displayed in Table 9:

Table 9: Closed-loop success and action errors under natural corruptions on FastWAM.

CorruptionSeverityUndefended SR (%)LIP SR (%)Δ\Delta (pp)eundefe_{\text{undef}}eLIPe_{\text{LIP}}Δe\Delta eReR_e (%)
Gaussian noise167.5085.00+17.500.0558590.034851+0.021008+37.61
226.2582.50+56.250.1304100.051361+0.079049+60.62
30.0063.75+63.750.4991290.083480+0.415649+83.27
40.0047.50+47.500.7633590.123554+0.639805+83.81
50.0016.25+16.250.7924660.167477+0.624989+78.87
...
Contrast drop192.5092.50+0.000.0071760.007978-0.000802-11.18
288.7592.50+3.750.0126740.013142-0.000468-3.69
390.0086.25-3.750.0175340.017851-0.000317-1.81
478.7581.25+2.500.0227790.023018-0.000239-1.05
545.0047.50+2.500.0284120.028602-0.000190-0.67

Under brightness shift, contrast drop, and motion blur, LIP results in minor fluctuations or small decreases in closed-loop SR. Because undefended raw action errors under brightness shift and contrast drop are already very small, even when relative action error reductions show seemingly pronounced negative percentages like −14.79%, the corresponding absolute action deviations remain minute, leaving closed-loop task success largely unaffected. We attribute this phenomenon to the training distribution constraining generalization to these specific photometric shifts, though the actual effect on closed-loop SR is minimal. Detailed analysis is provided in revised Section 5.5 (Natural-Corruption Generalization paragraph) and Appendix B.5.

3. Composite Corruption Evaluation (Table 10):
To address whether LIP remains effective when visual environmental changes and adversarial perturbations co-occur, we conducted a comprehensive composite-corruption stress test: concurrently injecting L∞=16/255L_\infty = 16/255 universal adversarial perturbations (UAP) alongside 5 natural corruption families across 5 severity levels during every observation, covering 25 composite conditions and 4,000 closed-loop rollouts. The results are shown in Table 10:

AuthorsOpenReview ↗

Response to Reviewer 1n8u (2/3)

Table 10: Closed-loop success and action errors under composite corruptions on FastWAM.

CorruptionSeverityUndefended SR (%)LIP SR (%)Δ\Delta (pp)eundefe_{\text{undef}}eLIPe_{\text{LIP}}Δe\Delta eReR_e (%)
Gaussian noise + UAP132.5077.50+45.000.1470600.048195+0.098865+67.23
21.2572.50+71.250.2121700.062908+0.149262+70.35
30.0063.75+63.750.5182100.089496+0.428713+82.73
40.0042.50+42.500.7820100.134000+0.648010+82.86
50.0016.25+16.250.7771450.169107+0.608037+78.24
...
Contrast drop + UAP10.0070.00+70.000.4425720.085016+0.357555+80.79
20.0062.50+62.500.4992620.102953+0.396310+79.38
30.0048.75+48.750.7433380.148703+0.594635+80.00
40.0012.50+12.500.7882070.352422+0.435785+55.29
50.000.00+0.001.2964490.840067+0.456382+35.20

As shown in Table 10, when adversarial attacks and natural/photometric degradations co-occur, undefended policies suffer severe collapse, whereas LIP reduces action errors across all 25 composite conditions and restores substantial closed-loop task success in every corruption family.

4. Multi-Corruption Augmented Training Validation (Table 11 & Table 12):
To verify whether the minor fluctuations under pure brightness shift, contrast drop, and motion blur originated from training distribution boundaries, we retrained LIP with a compact augmentation of merely 750 additional multi-environment frames (150 frames per corruption family).

  1. Under pure natural corruptions (Table 11): Previously negative relative action-error reductions ReR_e almost entirely turn positive.
  2. Under composite corruptions (Table 12).

Table 11: Closed-loop success and action errors under natural corruptions on FastWAM with multi-corruption augmented training.

CorruptionSeverityUndefended SR (%)LIP SR (%)Δ\Delta (pp)eundefe_{\text{undef}}eLIPe_{\text{LIP}}Δe\Delta eReR_e (%)
Gaussian noise167.5090.00+22.500.0558590.021880+0.033979+60.83
226.2580.00+53.750.1304100.046296+0.084114+64.50
30.0065.00+65.000.4991290.080160+0.418969+83.94
40.0047.50+47.500.7633590.130000+0.633359+82.97
50.0028.75+28.750.7924660.169112+0.623354+78.66
...
Contrast drop192.5090.00-2.500.0071760.007525-0.000349-4.86
288.7592.50+3.750.0126740.012612+0.000062+0.49
390.0087.50-2.500.0175340.015689+0.001845+10.52
478.7583.75+5.000.0227790.017592+0.005187+22.77
545.0077.50+32.500.0284120.017380+0.011032+38.83

Table 12: Closed-loop success and action errors under composite corruptions on FastWAM with multi-corruption augmented training.

CorruptionSeverityUndefended SR (%)LIP SR (%)Δ\Delta (pp)eundefe_{\text{undef}}eLIPe_{\text{LIP}}Δe\Delta eReR_e (%)
Gaussian noise + UAP132.5081.25+48.750.1470600.047106+0.099954+67.97
21.2576.25+75.000.2121700.056151+0.156018+73.53
30.0065.00+65.000.5182100.084876+0.433333+83.62
40.0045.00+45.000.7820100.134024+0.647986+82.86
50.0028.75+28.750.7771450.170649+0.606496+78.04
...
Contrast drop + UAP10.0071.25+71.250.4425720.083551+0.359021+81.12
20.0068.75+68.750.4992620.102758+0.396504+79.42
30.0051.25+51.250.7433380.125332+0.618005+83.14
40.0035.00+35.000.7882070.190426+0.597780+75.84
50.000.00+0.001.2964490.575610+0.720839+55.60

These results substantiate the conclusion that earlier minor fluctuations arose from training data distribution boundaries, and demonstrate that with minimal data additions, LIP's robustness against both natural and composite corruptions can be significantly broadened.

AuthorsOpenReview ↗

Response to Reviewer 1n8u (3/3)

Question 2 (Q2, W3): Statistical Significance and Underlying Reason for FastWAM Defended SR Exceeding Clean Baseline on RoboTwin

Response:

This outcome is consistent with LIP's bounded, control-oriented corrections. Minor action chunk variations can alter thresholded grasp or contact outcomes, thereby modifying subsequent trajectory states and observations during rollout. The action error bound in Proposition 2 governs paired-input action fidelity, whereas 92.62% evaluates complete task rollout. Thus, the clean policy success rate serves as a nominal baseline reference rather than a hard ceiling for task success under bounded corrections.

Regarding generalization to unseen attacks: all evaluated UAPs were excluded from purifier training; Appendix B.3 demonstrates that when re-optimizing independent UAPs at every perturbation budget, defended action errors remain strictly below undefended curves. Defense-aware SPSA attacks and Appendix B.5 composite corruptions provide extensive evaluations on unseen perturbations, while revised Section 4.4 elucidates how bounded step-by-step corrections near grasp and contact thresholds govern long-horizon closed-loop behavior.


Question 3 (Q4, W5): Substantial Difference in Attacked Baselines on LIBERO Between Table 1 and Table 2

Response:

The substantial numerical difference arises from fundamental distinctions in the definition of the evaluated threat models:

  1. Table 1: Suite-Specific Attacks (Suite-Specific UAP)
    Perturbations are independently optimized on the offline frame sets of each individual LIBERO sub-suite (libero_spatial, libero_object, libero_goal, libero_10), targeting sub-suite-specific visual distributions. For models with aggressive temporal windowing and localized quantization like FasterWAM, sub-suite perturbation disruption is localized (82.38% attacked SR); for broad-attention models like StarWAM, disruption is severe (13.25% attacked SR); Light-WAM averages 60.00%, and FastWAM averages 52.25%.
  2. Table 2: Benchmark-Wide Transfer Attacks (Benchmark-Wide Transfer UAP)
    Perturbations are optimized across the full 40 tasks of the entire LIBERO benchmark to evaluate zero-shot transfer across tasks, model families, and robot embodiments.

We have explicitly clarified the distinct threat model definitions for Tables 1 and 2 in revised Section 5.3.

I thank the authors for their detailed responses and additional experiments. The new composite-corruption evaluations and manuscript revisions address some of my concerns and strengthen the experimental evidence.

However, my main concerns remain insufficiently resolved. For a study emphasizing embodied manipulation and practical deployment, the absence of real-world closed-loop robot experiments remains a significant limitation. The simulation results alone cannot fully establish reliability in real-world conditions. In addition, although the authors further emphasize the selective correction of action-critical features, the new theoretical analysis and experimental evidence do not directly establish this mechanism or sufficiently clarify the fundamental novelty over existing feature purification methods. The statistical significance of the defended performance exceeding the clean baseline also remains unaddressed.

I appreciate the authors' efforts and improvements, but given these remaining concerns, I will maintain my original rating.

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

This paper proposes Latent Interface Purification (LIP) to improve the adversarial robustness of World Action Models (WAMs). LIP adds a lightweight purification module between the frozen visual encoder and action expert. It does not require retraining the original policy. The method uses multi-scale convolutions, scale normalization, and signed modulation to correct perturbed latent features. It is trained using paired clean and attacked features, together with a clean-consistency loss.

Critical strengths
  1. LIP can be added to an existing WAM without retraining the encoder or action expert. This makes it easy to integrate into existing robotic systems.

  2. FastWAM's success rate on RoboTwin increases from 35.38% to 92.62% under attack. LIP also maintains performance on clean inputs, with only a small drop in success rate.

  3. The authors evaluate four WAM backbones on LIBERO and RoboTwin, covering both single-arm and bimanual manipulation. The cross-benchmark experiments and ablation studies are also useful. The additional computational cost is quite small.

Critical weaknesses
  1. Most components, including multi-scale convolutions, residual correction, normalization, and gating, are already widely used. Although their combination works well, the main technical novelty is not entirely clear. Comparisons with stronger latent-space denoising methods would be helpful.

  2. The adaptive attack evaluation uses SPSA with only 1,800 queries across three restarts. Since the proposed defense is differentiable, stronger white-box attacks, such as PGD, should also be considered. Without these experiments, it is difficult to assess the robustness of LIP against an attacker who knows the defense.

  3. Although four WAM backbones are evaluated, they all belong to the Wan-based model family. Each backbone also requires a separately trained purifier. It remains unclear whether LIP can generalize to other VLA architectures.

  4. FastWAM achieves 92.62% defended success on RoboTwin, compared with 85.00% clean success. The reason for this large improvement over the clean baseline is unclear. In addition, some results in Figure 13 appear to differ from those in Table 1.

Questions for authors
  1. Would a standard latent-space U-Net or residual CNN with a similar number of parameters achieve comparable results? What are the main technical differences between LIP and existing feature denoising methods?

  2. Have the authors tested stronger white-box adaptive attacks? For example, it would be useful to evaluate PGD against the full defended policy instead of relying only on query-based SPSA.

  3. How well does LIP perform on WAMs or VLA models with different architectures? Testing a different visual encoder or action architecture would help support the generalization claim.

  4. Why does the defended success rate sometimes exceed the clean baseline by such a large margin? Also, the aggregate success rates in Figure 13 appear to differ from those in Table 1. Could the authors check these numbers and explain the differences?

2 comments
AuthorsOpenReview ↗

Response to Reviewer JGpa (1/2)

We sincerely thank the reviewer for the constructive, professional, and thorough feedback. Below, we address your four questions in sequence, which also cover the critical weaknesses.

Question 1 (Q1, W1): Would a Parameter-Matched Standard Latent U-Net or Residual CNN Achieve Comparable Results? Technical Differences from Existing Feature Denoising Methods

Response:

The key distinction lies in the control objective at the frozen WAM interface. Unlike existing denoisers that prioritize perceptual fidelity or generic representation reconstruction, LIP targets downstream action fidelity by selectively rectifying action-critical latent features rather than inverting perturbations coordinate-by-coordinate.

Theoretically, Section 4.3 (Eq. 9) and Proposition 2 (Appendix A.1) prove that our joint recovery and clean-consistency objective upper-bounds both attacked-action error and clean-action distortion under a local Lipschitz condition (interpreted in Section 4.4 and ablated in Table 4). As further detailed in our Official Comment at the top of the replies and Table 13, while RGB U-Net and Latent MLP yield higher image and latent similarity, LIP achieves the lowest online action error (0.0523) and highest closed-loop success rate (92.00%).

Below, we compare the parameter-matched convolutional model against LIP in terms of overall design and computational overhead:

MethodInterfaceClean+Method ↑\uparrowDefended ↑\uparrowΔdef↑\Delta_{\text{def}} \uparrowedef↓e_{\text{def}} \downarrowdclean↓d_{\text{clean}} \downarrow
Standard Conv (Matched, 1.294M)Latent89.7586.25+37.370.0460.011
LIP (Ours, 1.299M)Latent91.4593.92+45.040.0380.002

LIP's defended success rate exceeds that of the parameter-matched standard convolution by 7.67 percentage points, while peak forward allocation is only 1.47 MB compared to 5.19 MB for standard convolution.


Question 2 (Q4, W4): Why Defended Success Rate Sometime Substantially Exceeds Clean Baseline; Discrepancy Between Figure 13 and Table 1

Response:

This observation is entirely consistent with LIP's bounded, control-oriented corrections. Minor deviations in action chunks can flip thresholded grasp or contact outcomes, subsequently altering downstream states and observations throughout the rollout. The action error bound in Proposition 2 governs paired-input action fidelity, whereas the 92.62% metric evaluates full closed-loop task execution. Consequently, the clean policy success rate represents a nominal baseline reference rather than a theoretical ceiling for defended task success under bounded corrections.

Figure 13 preserved early 300-frame experimental data, where the curve labeled Clean was actually Clean+LIP (92.88%), and the legend was not synchronized during compilation. Table 1 follows the final 1,000-frame protocol. We have replaced Figure 13 with the corrected plot in the revision.

AuthorsOpenReview ↗

Response to Reviewer JGpa (2/2)

Question 3 (Q2, W2): Evaluation of Stronger White-Box Adaptive Attacks (e.g., PGD Against the Full Defended Policy) Beyond SPSA

Response:

Implementation Mechanism, Physical Feasibility, and Evaluation of White-Box End-to-End Action-Loss PGD:

  1. Implementation Mechanism and Engineering Adaptation:
    Official open-source WAM repositories were released strictly for derivative-free inference deployment; backward computational graphs are severed via @torch.no_grad(), action .detach(), or truncation into NumPy arrays. To rigorously address the reviewer's concern regarding white-box adaptive attacks, we restructured the underlying execution pipelines, bypassed official inference decorators and truncation constraints, implemented model-specific low-level adaptations, and successfully established a closed-loop evaluation framework for end-to-end action-loss PGD backpropagated through multi-step diffusion denoising.

  2. Time Cost and Physical Real-World Feasibility Analysis:

  • Extremely High Time Cost: In online closed-loop rollouts, every replanning step requires end-to-end backpropagation through a 14B-parameter multi-step diffusion Transformer. A single rollout takes over 10× longer than nominal unattacked rollouts (a nominal task rollout takes ~40–60 s, whereas an online PGD rollout takes 550–600 s).
  • Minimal Physical Feasibility in Real Deployment:
    1. System Privileges and Intrusiveness: The attacker must obtain full white-box access to the robot's real-time inference system, including direct manipulation of GPU memory.
    2. Excessive Latency Violates Real-Time Control: The multi-second latency incurred by online backpropagation cannot satisfy the strict 10 Hz–30 Hz real-time control constraints of physical robots.
    3. Infeasibility of Passive Physical Deployment: Unlike physical UAPs (optical filters) or adversarial patches (printed stickers), online PGD cannot be passively deployed in physical environments via light reflection.
    4. Conclusion: Such end-to-end action PGD serves as a "theoretical extreme stress test" of policy and defense robustness rather than a practical deployment threat, and should not be used to dismiss the efficacy of LIP.
  1. Multi-Suite Benchmark Evaluation:
    Due to the substantial computational cost (10× testing duration), we systematically conducted benchmark experiments on our primary backbones, FastWAM and Light-WAM: covering all 4 suites of LIBERO, randomly selecting 2 tasks per suite, and evaluating 10 rollouts per task across 4 perturbation budgets (1/255,2/255,4/255,8/2551/255, 2/255, 4/255, 8/255). The results are summarized below:
ModelAttackClean1/2552/2554/2558/255
FastWAMPGD79/805/800/800/801/80
FastWAMPGD+LIP—75/8042/8016/800/80
Light-WAMPGD78/806/800/800/800/80
Light-WAMPGD+LIP—76/8038/8013/800/80
  1. Results Analysis and Objective Defense Boundaries:
    When subjected to this theoretical extreme white-box attack:
  • At small perturbation budgets (1/2551/255 and 2/2552/255): The undefended policy immediately collapses even under faint perturbations, while LIP demonstrates remarkable recovery capabilities.
  • At larger perturbation budgets (4/2554/255 and 8/2558/255): Because end-to-end action gradients nonlinearly exploit high-frequency pathways through the multi-step denoising network, residual errors exceed gripper tolerances and full recovery is challenging. This clearly and objectively delineates the defense boundaries of LIP under extreme white-box stress testing.

Question 4 (Q3, W3): Performance Across Different WAM or VLA Architectures to Support Generalization Claims

Response:

The four evaluated models cover diverse scales, latent interfaces, and inference configurations across the Wan-family WAMs, with each backbone utilizing an independently trained purifier. The identical design yields recovery gains across all four implementations, and cross-benchmark experiments further examine zero-shot embodied transfer within each backbone. Per-backbone training is highly lightweight: each purifier requires only 1,000 anchor frames, with average training times of 341.3 s on LIBERO and 365.5 s on RoboTwin (Table 7). LIP's integration prerequisite is access to the interface from visual latents to the action expert; extending to other models requires matching their tensor topology, scale, and action-conditioning pathway.

Reviewer_RpUzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the Latent interface purification to conduct the adversarial purification in the feature space. The proposed method learns a bounded residual correction in the latent space to rectify the adversarial effect in the feature space. Experiments are conducted on various world-action models and two manipulation benchmarks. Adversarial purification is well discussed in many previous works. While the proposed latent space purification achieves promising results for defending against universal adversarial noises, this work seems not to properly compare it with existing purification-based defense methods. Meanwhile, the threat models seem to be slightly narrow, only considers universal scenarios.

Critical strengths
  1. The discussed problem is interesting and important for the deployment of WAMs.

  2. The paper evaluates closed-loop task success across two manipulation benchmarks and four backbones.

  3. The reported recovery gains are substantial under the evaluated attacks with low inference latency.

Critical weaknesses
  1. My major concerns are regarding the scope of adversarial robustness discussed in this work. Section 3.2 restricts attacks to frame-invariant universal perturbations and grants access to the visual encoder and action expert, but does not clearly specify access to the purifier in the primary evaluation. The purification-based defenses are already proven to be pretty effective against $ l_\infty$-bounded pixel-space attacks. As LIP is differentiable, the author should discuss how LIP performs under various threatening scenarios, such as the attacker is aware of the purification module. Meanwhile, some feature-space-based attacks and patch-based attacks are also encouraged to be discussed in this paper.

  2. Second is the compared baselines are somehow outdate and insufficient. Table 3 compares LIP against classical image transformations, an RGB U-Net, and an in-house Latent MLP, none of which are sufficiently advanced. Moreover, these comparisons establish advantages over the tested implementations, but do not establish competitiveness with representative diffusion-based purification or stronger feature-space defenses. As this work utilizes the purification method, properly discussing and comparing with SOTA purification-based adversarial defense methods are required.

  3. The U-NET defense shows a critical performance drop on clean input. Could the author clearly clarify the reason and the training details for the U-Net model? Table 6 reports 1.299M parameters for LIP, compared with 0.222M for the U-Net and 0.091M for the MLP. This further challenges the fairness of the comparison.

  4. Feature-space purification, residual correction, identity initialization, multi-scale processing, and clean-consistency regularization all have substantial precedent. The application to a frozen WAM interface is valuable, but the current experiments do not sufficiently establish which proposed elements provide a distinctive advantage over simpler alternatives.

Questions for authors

Please see the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer RpUz (1/2)

We sincerely thank the reviewer for the constructive, professional, and thorough feedback. Below, we respond to your Critical Weaknesses point by point.

Weakness 1 (W1): Scope of Adversarial Robustness Threat Model: Purifier Access, Defense-Aware Adaptive Attacks, Feature-Space Attacks, and Patch-Based Attacks

Response:

Implementation Mechanism, Physical Feasibility, and Evaluation of White-Box End-to-End Action-Loss PGD:

  1. Implementation Mechanism and Engineering Adaptation:
    Official open-source WAM repositories were released strictly for derivative-free inference deployment; backward computational graphs are severed via @torch.no_grad(), action .detach(), or truncation into NumPy arrays. To rigorously address the reviewer's concern regarding white-box adaptive attacks, we restructured the underlying execution pipelines, bypassed official inference decorators and truncation constraints, implemented model-specific low-level adaptations, and successfully established a closed-loop evaluation framework for end-to-end action-loss PGD backpropagated through multi-step diffusion denoising.

  2. Time Cost and Physical Real-World Feasibility Analysis:

  • Extremely High Time Cost: In online closed-loop rollouts, every replanning step requires end-to-end backpropagation through a 14B-parameter multi-step diffusion Transformer. A single rollout takes over 10× longer than nominal unattacked rollouts (a nominal task rollout takes ~40–60 s, whereas an online PGD rollout takes 550–600 s).
  • Minimal Physical Feasibility in Real Deployment:
    1. System Privileges and Intrusiveness: The attacker must obtain full white-box access to the robot's real-time inference system, including direct manipulation of GPU memory.
    2. Excessive Latency Violates Real-Time Control: The multi-second latency incurred by online backpropagation cannot satisfy the strict 10 Hz–30 Hz real-time control constraints of physical robots.
    3. Infeasibility of Passive Physical Deployment: Unlike physical UAPs (optical filters) or adversarial patches (printed stickers), online PGD cannot be passively deployed in physical environments via light reflection.
    4. Conclusion: Such end-to-end action PGD serves as a "theoretical extreme stress test" of policy and defense robustness rather than a practical deployment threat, and should not be used to dismiss the efficacy of LIP.
  1. Multi-Suite Benchmark Evaluation:
    Due to the substantial computational cost (10× testing duration), we systematically conducted benchmark experiments on our primary backbones, FastWAM and Light-WAM: covering all 4 suites of LIBERO, randomly selecting 2 tasks per suite, and evaluating 10 rollouts per task across 4 perturbation budgets (1/255,2/255,4/255,8/2551/255, 2/255, 4/255, 8/255). The results are summarized below:
ModelAttackClean1/2552/2554/2558/255
FastWAMPGD79/805/800/800/801/80
FastWAMPGD+LIP—75/8042/8016/800/80
Light-WAMPGD78/806/800/800/800/80
Light-WAMPGD+LIP—76/8038/8013/800/80
  1. Results Analysis and Objective Defense Boundaries:
    When subjected to this theoretical extreme white-box attack:
  • At small perturbation budgets (1/2551/255 and 2/2552/255): The undefended policy immediately collapses even under faint perturbations, while LIP demonstrates remarkable recovery capabilities.
  • At larger perturbation budgets (4/2554/255 and 8/2558/255): Because end-to-end action gradients nonlinearly exploit high-frequency pathways through the multi-step denoising network, residual errors exceed gripper tolerances and full recovery is challenging. This clearly and objectively delineates the defense boundaries of LIP under extreme white-box stress testing.

Feature-space and patch attacks also inevitably involve the above optimization problem. To the best of our knowledge, there are currently no related attack methods targeting these WAMs. If you are aware of any, we would appreciate it if you could provide the relevant papers so that we can examine these extended threats under explicit objectives and access privileges.

AuthorsOpenReview ↗

Response to Reviewer RpUz (2/2)

Weakness 2 (W2, W3): Baseline Comparisons are Outdated and Insufficient; Lacks Comparison with Diffusion-Based Purification or Stronger Feature-Space Defenses; RGB U-Net Performance Drop on Clean Inputs, Training Details, and Parameter Fairness

Response:

We scaled RGB U-Net and Latent MLP to 1.368M and 1.364M parameters respectively, and added a parameter-matched standard convolutional baseline (1.294M) alongside the 1.299M LIP.

MethodInterfaceClean+Method ↑\uparrowDefended ↑\uparrowΔdef↑\Delta_{\text{def}} \uparrowedef↓e_{\text{def}} \downarrowdclean↓d_{\text{clean}} \downarrow
Undefended BaselineNone92.4048.88———
RGB U-Net (Original, 0.222M)Pixel84.7573.50+24.620.0200.003
Scaled RGB U-Net (Matched, 1.368M)Pixel84.7574.25+25.370.0210.003
Latent MLP (Original, 0.091M)Latent88.5081.25+32.370.0420.008
Scaled Latent MLP (Matched, 1.364M)Latent88.0082.50+33.620.0310.022
Standard Conv (Matched, 1.294M)Latent89.7586.25+37.370.0460.011
LIP (Ours, 1.299M)Latent91.4593.92+45.040.0380.002

When applying diffusion-based purification to WAMs, checkpoint choice, camera interfaces, sampling steps, and preservation of action-relevant information determine comparability. LIP is positioned as an efficient, real-time action restorer rather than a generic image denoiser: it executes a single forward pass at the native latent interface with a module latency of 1.26 ms and incremental forward allocation of 1.47 MB, compared with 14.89 ms and 560.65 MB for the scaled RGB U-Net. A fair comparison must jointly report closed-loop success, clean-action preservation, module and end-to-end latency, and the actual replanning budget. We welcome recommendations of specific diffusion or feature defense papers, public implementations, and configurations for focused adaptation and comparison.

  • RGB U-Net was trained on offline demonstration frames with pixel L1, Huber, and clean-consistency losses. Pixel-level reconstruction inherently distorts high-frequency details received by the downstream encoder, which explains the sharp decline in clean success from 92.40% to 84.75%.

Weakness 3 (W4): Precedents for Proposed Elements; Distinctive Advantages Over Simpler Alternatives

Response:

The key distinction lies in the control objective at the frozen WAM interface rather than generic representation reconstruction. Unlike purification methods that aim to invert perturbations across all pixels or features, LIP acts directly on the representation consumed by the frozen action expert to minimize downstream action deviation.

Theoretically, Section 4.3 (Eq. 9) and Proposition 2 (Appendix A.1) prove that our joint latent-recovery and clean-consistency objective upper-bounds both attacked-action error and clean-action distortion under a local Lipschitz condition, making latent recovery a control-grounded surrogate (interpreted in Section 4.4 and ablated in Table 4). As detailed in our Official Comment at the top of the replies, LIP achieves lower image/latent reconstruction fidelity than pixel denoisers yet reduces online action error far more effectively (0.0523 vs. 0.0987), raising closed-loop success from 84.00% to 92.00%.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Core Novelty of LIP

We thank all reviewers for their valuable feedback. To clarify how LIP fundamentally differs from prior image or feature denoising methods, we summarize the core distinction below:

  • Conceptual Analogy — Same Intervention Target ≠\neq Same Method: Adversarial attacks optimize image pixels through a model to alter task predictions, while image data hiding & watermarking also optimize image pixels through a model, but to embed secret information. Nobody would equate adversarial attacks with watermarking merely because both modify pixels. Likewise, although LIP operates by modifying the latent representation ztz_t, its optimization objective, recovery logic, and empirical behavior are fundamentally different from generic feature denoising.
  • Core Essence: LIP does not laboriously reconstruct the clean latent ztcz_t^c by retracing the perturbation path in reverse. It selectively corrects the action-critical latent features to reduce the downstream action deviation between clean and perturbed observations and thereby restore closed-loop success.
Comparison DimensionGeneric Image / Feature DenoisingLIP
Operational FootholdRGB image pixels ItI_t or generic intermediate visual featuresFrozen encoder–action-expert control interface ztz_t
Optimization ObjectiveVerbatim Reconstruction: Pull all coordinates equally toward clean targets (ItI_t or ztcz_t^c), detached from controlAction-Deviation Minimization: Suppress downstream action chunk drift ∥at:t+K−1d−at:t+K−1c∥2\|a^d_{t:t+K-1}-a^c_{t:t+K-1}\|_2
Recovery PathSymmetric Inversion: Reverse perturbations across the entire representation, blindly fitting dominant backgroundAsymmetric & Action-Targeted: Rectify action-governing latent coordinates via input-conditioned gating rather than verbatim inversion
Optimization LogicControl-Agnostic Optimization: Unweighted distance minimization, detached from policy control sensitivityAction-Error-Bounded Surrogate: Scale-bounded signed modulation gψ⊙Δψg_\psi \odot \Delta_\psi + clean anchor Lclean\mathcal{L}_{\mathrm{clean}}, backed by Propositions 1 & 2
Nominal PreservationLacks Identity Guarantee: Distorts fine contact details, dropping RGB U-Net clean SR to 84.75%Exact Identity Initialization + Clean Anchor: Preserves nominal clean SR at 91.45% with action distortion of only 0.002
Empirical ResultsHigh Fidelity but Poor Control: Low pixel/latent error fails to restore control, reaching only 74.25% defended SRHigh Action Fidelity & Recovery: Eliminates up to 80.27% of action error and attains 93.92% defended SR at 1.26 ms latency
  • Empirical Validation across 876 Closed-Loop Frames: To verify this distinction, we randomly selected 10 LIBERO tasks and ran 5 rollouts each, yielding 50 episodes and 876 decision frames. An active WAM executes the closed loop, while a frozen parallel WAM processes clean inputs so that environments stay strictly aligned. Along the pipeline, we measure decoded Image PSNR (dB) and SSIM, Latent Cosine Similarity and Relative L2 Error, Online Action Relative L2 Error and Cosine Similarity, and the final 50-episode Success Rate.
Defense MethodImage PSNR (dB) ↑Image SSIM ↑Latent CosSim ↑Latent Rel-L2 ↓Online Action Rel-L2 ↓Action CosSim ↑Closed-Loop SR ↑
Undefended Baseline34.800.83540.76780.70430.18430.945334.00%
RGB U-Net Denoiser42.670.97760.93920.34020.09870.971384.00%
Latent MLP Baseline36.270.95610.90870.41760.08910.978286.00%
LIP (Ours)33.480.94660.89670.44450.05230.989992.00%

Mechanism Interpretation: Experimental results show that LIP restores actions through a mechanism distinct from conventional denoising methods. RGB U-Net is guided by mean squared error and recovers representations point by point along the inverse noise path; its high fidelity mainly comes from fitting the static background that occupies most of the image. LIP operates on the same latent variable interface and, with multi-scale spatial context and input-conditioned sign modulation, directly rectifies the key components in the latent variable that determine the action.


Notice of Manuscript Revisions (All Changes Marked in Blue)

Per reviewer feedback, all revisions are highlighted in blue in the revised PDF. Key updates include: clarified threat model and control bounds, parameter-matched baselines, multi-level representation–action diagnostics (Table 13), and natural/composite corruption evaluations (Tables 9–12).

AuthorsOpenReview ↗

Requesting the reviewers to kindly see our responses and let us know any further query.

It would be great if the reviewers let us know whether there are any further comments? We are looking forward to respond as quickly as possible.

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

Related papers

Discussion (0)

Sign in to comment.