acceptodds
Under review as a conference paper at ICLR 2027

Uncovering Vulnerability of Vision-Language-Action Models under Joint-Level Physical Faults

Abstract

Robustness studies of Vision-Language-Action (VLA) models have largely focused on perceptual, task, and command-space perturbations, while changes in how the robot body physically realizes actions remain comparatively underexplored. We investigate this embodiment-side robustness gap through joint-level physical faults, including joint locks, range restrictions, and increased friction. These faults alter the physical conditions of action execution without directly perturbing policy outputs, introducing command–execution mismatch into the closed loop. Our evaluation of OpenVLA-OFT and \(\pi_0.5\) on LIBERO reveals substantial degradation that varies strongly with the affected joint and fault condition. Joint-dependent vulnerability is related to nominal joint usage and end-effector sensitivity, while increased friction reduces task success even without narrowing joint-position limits. We examine complementary mitigation strategies with the VLA backbones frozen: language-based fault feedback and fault-aware local kinematic correction provide no consistent task-level gains in the evaluated settings, whereas embodiment-side calibration partially mitigates degradation across both backbones while largely preserving fault-free performance. Together, our findings highlight changes in physical execution as an important dimension of VLA robustness that requires explicit evaluation alongside perceptual and task variations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Safety, Vision Language Action model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_D3LHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the robustness of VLA models to joint-level physical faults, including joint locks, range restrictions, and increased friction, which introduce mismatches between commanded and executed actions. The authors systematically evaluate two VLA backbones on LIBERO and analyze the impact of different fault conditions. They further examine language-based feedback and kinematic correction, finding limited benefits, and propose an embodiment-side action calibration method that partially recovers task performance while largely preserving fault-free behavior. Real-world experiments further validate the potential of the proposed calibration approach.

Critical strengths
  1. The paper investigates an underexplored dimension of VLA robustness: physical execution faults at the robot-joint level, complementing existing studies on visual, task-level, and action-space perturbations. Overall, it is a novel and practically relevant problem formulation.
  2. The authors evaluate two VLA backbones across four LIBERO suites, considering multiple fault types, severities, and affected joints. The analysis reveals substantial joint-dependent performance degradation and demonstrates that increased friction can impair task success even without restricting joint-position limits.
  3. The authors also explore language-based fault feedback, constrained inverse kinematics, and embodiment-side action calibration. These experiments provide useful insights into the limitations of directly communicating execution mismatches or applying local kinematic corrections, motivating alternative approaches to embodiment-aware adaptation.
  4. The proposed calibration method partially recovers task performance while keeping the VLA backbone frozen and largely preserving fault-free performance. The authors also validate the approach on a real-world robot, providing preliminary evidence of its practical applicability.
Critical weaknesses
  1. I am not entirely clear about how the physical perturbations are implemented in the experimental setup. Since a large portion of the experiments is conducted in simulation, could the authors clarify how the proposed joint-level physical perturbations are modeled and deployed in the simulated environment? In particular, how is increased friction implemented in the simulator, and what does its effect look like in practice? I encourage the authors to provide additional visualization examples.
  2. I am interested in understanding how these embodiment-site faults affect task execution over longer-horizon tasks. Do tasks fail from the very beginning due to the perturbations, or do errors gradually accumulate over time, eventually leading to task failure?
  3. The evaluation primarily focuses on individual joint-level faults, leaving the robustness of VLAs under simultaneous faults in multiple joints unclear. In real-world scenarios, mechanical degradation or actuator failures may affect multiple joints at the same time, and the resulting execution errors may interact. It would be valuable to evaluate the effectiveness of the proposed calibration method under multi-joint faults.
Questions for authors

See the weaknesses above.

5 comments
AuthorsOpenReview ↗

Response to D3LH: Implementation Detail about Physical Perturbation and Visualization on Friction Fault

Thank you for requesting clarification.

We evaluate a 7-DoF Franka Panda arm in LIBERO, which uses robosuite and the MuJoCo physics engine. VLA-predicted end-effector commands are executed through an operational-space pose controller (OSC_POSE) at 20 Hz, with replanning every eight steps.

Each fault is applied to one arm joint after the environment is reset to the selected benchmark initial state and remains active throughout the episode. We evaluate each arm joint separately. The fault is implemented by changing the selected joint’s simulator parameters.

Joint locks and range restrictions narrow the joint’s position limits around its initial position. Let qi0q_i^0 denote the initial position of joint ii, and let qi,min⁡q_{i,\min} and qi,max⁡q_{i,\max} denote its original limits. The imposed interval is

[max⁡(qi,min⁡,qi0−hi),min⁡(qi,max⁡,qi0+hi)].\left[\max(q_{i,\min}, q_i^0 - h_i), \min(q_{i,\max}, q_i^0 + h_i)\right].

For a joint lock, we set hi=10−4 radh_i = 10^{-4}\,\mathrm{rad}. This small nonzero interval effectively immobilizes the joint while avoiding a zero-width constraint.

For a range restriction, we set

\qquad r \in \{0.05, 0.03\}.$$ The ratio $r$ specifies the retained fraction of the original joint span before clipping to the original limits. **We model increased friction by scaling the selected joint’s MuJoCo `frictionloss` parameter.** $$f_i^{\mathrm{fault}} = \gamma f_i^{\mathrm{nominal}}, \qquad \gamma \in \{150, 200, 300\}.$$ The nominal value is 0.1, giving perturbed values of 15, 20, and 30, respectively. This parameter governs dry friction at the joint. Joint damping, actuator force limits, joint position limits, and contact friction between objects retain their nominal settings. Increased resistance can reduce joint displacement, cause trajectory deviations, or stall motion despite continued end-effector commands. The resulting changes in the robot’s motion affect subsequent observations and VLA predictions. **The supplementary visualization shows how increased friction affects different motions depending on the joint.** The examples exhibit patterns similar to those observed under joint locks. In the illustrated task, friction at $j_2$ and $j_3$ impedes lateral and vertical motion, respectively, while friction at $j_4$ and $j_5$ interferes with gripper alignment and grasping. Figure 6 in the appendix presents these results, and we added a detailed explanation of the experiment to Appendix E.
AuthorsOpenReview ↗

Response to D3LH: Fault Accumulation Test

We test whether the VLA can recover after the fault is removed. We vary how long the fault remains active, then remove it and let the VLA continue with the robot in a fault-free condition.

Task success generally declines with longer fault exposure, even after the fault is removed. Success remains high after brief exposure and tends to fall as the robot spends more time operating under the fault.

Limited recovery from accumulated execution errors motivates reference-guided calibration. At fault removal, the VLA must continue from the state the robot reached during faulty execution. The decline in success suggests that the VLA struggles to recover from those accumulated deviations on its own. Reference-guided calibration learns to correct the deviations as they arise.

Section 4.2 of the main paper presents the plot and describes the experiment in detail.

AuthorsOpenReview ↗

Response to D3LH: Multi-Joint Fault Evaluation

Thank you for suggestion of multi-joint fault evaluation. We extended the π0.5\pi_{0.5} evaluation to simultaneous faults in multiple joints using the existing embodiment-side calibrators. We tested both multiple joint locks and a joint lock combined with increased friction in another joint.

Simultaneous joint locks. Embodiment-side calibration achieves higher success rates in all 12 conditions, with mean gains of 7.48, 9.20, and 5.82 percentage points for the three configurations.

Each entry reports Base → embodiment-side calibration success rate (%).

Locked jointsSpatialObjectGoalLIBERO-10Mean
j2j_2,j4j_436.7 → 58.221.3 → 22.466.7 → 69.02.0 → 7.031.67 → 39.15
j2j_2,j5j_538.7 → 52.222.0 → 31.448.7 → 52.04.7 → 15.228.50 → 37.70
j4j_4,j5j_5,j6j_675.3 → 84.478.0 → 83.460.0 → 60.424.0 → 32.459.33 → 65.15

Simultaneous faults of different types. Embodiment-side calibration improves success in seven of the eight mixed-fault conditions and matches Base in the remaining condition. In each configuration, we locked one joint and increased friction in a different joint by a factor of 150.

Fault combinationSpatialObjectGoalLIBERO-10Mean
Locked j0j_0 + increased friction in j4j_414.4 → 39.076.0 → 81.077.2 → 77.245.0 → 50.053.15 → 61.80
Locked j2j_2 + increased friction in j5j_530.8 → 50.418.6 → 28.248.8 → 54.45.4 → 13.425.90 → 36.60

Recovery varies across configurations, and several LIBERO-10 conditions remain challenging. We added these evaluations to Appendix F and report the results in Tables 5 and 6.

Reviewer_WD7GOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the robustness of VLA models under joint-level physical faults, including joint locks, range restrictions, and increased friction. The authors find substantial and joint-dependent performance degradation, while the evaluated action-space robustness training, language feedback, and constrained inverse kinematics methods fail to provide consistent improvements. The paper further proposes J-PARC, which predicts action residuals using execution history and joint-regime representations. Experiments in simulation and on a real robot show partial recovery of task performance.

Critical strengths
  1. Joint faults are an important problem in robot control but have received relatively limited attention in current VLA robustness research. Investigating the closed-loop behavior of VLAs under physical execution faults is a valuable research direction.
Critical weaknesses
  1. The evaluation does not sufficiently distinguish physical task infeasibility from limitations in VLA robustness. Although the authors exclude joints that severely affect task feasibility in some analyses, the main evaluation results still include these conditions. The sensitivity analysis also provides limited explanation of performance differences across joint faults, making it difficult to establish how these faults affect the VLA policy itself.

  2. J-PARC uses reference-guided residual action correction, but the paper does not sufficiently explain how this design addresses the failure patterns observed earlier. The existing ablations also do not independently evaluate the contribution of reference-guided supervision.

Questions for authors
  1. Could the authors clarify how task feasibility is determined and further analyze VLA performance degradation and its causes under physically feasible fault conditions?

  2. How does the design of J-PARC relate to the observed failure patterns? Could the authors further validate the independent contribution of reference-guided supervision?

7 comments
AuthorsOpenReview ↗

Response to WD7G: Task Feasibility

We assess reachability for each task, initial state, and fault condition. Thank you for highlighting the distinction between physical feasibility and VLA robustness. We revised our categorical infeasibility statements and added a reference-guided reachability analysis under the imposed joint constraints (Appendix G), using reachability as a geometric proxy for feasibility. The main results characterize overall deployment performance, while Appendix G separately reports performance within the reference-reachable subsets. (weakness1, question1)

Substantial failures remain within the reachable subsets. Fault-only success is 62.83% under joint lock, 72.39% at joint-range ratio 0.03, and 77.73% at ratio 0.05. Therefore, restricted positional access alone does not explain the observed degradation. (question1)

The sensitivity analysis relates joint-level physical effects to differences in task success. We examine whether faults in joints that produce larger EEF displacements for comparable angular motion are associated with greater drops in success, accounting for how much each joint is used during nominal execution. This analysis characterizes the association between joint sensitivity and performance degradation. (weakness1)

Execution can drift substantially even when the reference positions are reachable. Within the reference-reachable joint-lock subset, we compare all available pairs of healthy and fault-only rollouts, regardless of task outcome. The distance between the end-effector positions in the paired rollouts increases from 3.51 mm at steps 0–5 to 85.48 mm at steps 62–102 while the joint remains locked. (weakness1, question1)

The evaluated VLAs show limited autonomous recovery after fault removal. The fault accumulation experiment in Section 4.2 tests whether the VLA can complete the task from the state reached during faulty execution. Final success generally declines with longer prior fault exposure, even after the fault is removed. (weakness1, question1)

AuthorsOpenReview ↗

Response to WD7G: Correspondence of Failure Pattern and Reference-Guided Residual Action Correction

The fault accumulation experiment motivates reference-guided supervision. Task success generally declines with longer fault exposure even after normal actuation is restored, suggesting that the VLA struggles to recover from deviations that have already accumulated. Removing the fault restores normal actuation, but it does not tell the VLA where to steer the robot to resume the task. The nominal trajectory provides that reference. We pair nominal and faulty rollouts from the same initial state and construct training targets that preserve the base action near the nominal trajectory and redirect translation toward the next nominal waypoint when execution deviates substantially. (weakness2, question2) We added the fault accumulation analysis to Section 4.2 of the main paper.

Differences in failure severity and persistence across joints motivate conditioning on the inferred fault regime. The fault accumulation experiment shows that success declines at different rates depending on the affected joint. The t-SNE visualization in appendix Figure 7 shows that the learned embeddings form joint-dependent clusters under both joint locks and increased friction, supporting their use as context for correction. (weakness2, question2)

Reference-guided calibration improves success even without regime or history inputs. In appendix Table 4, the base policy achieves 66.7% average success. Adding a reference-guided calibrator that uses only the current robot state and base action raises success to 71.1%. With reference-guided supervision held fixed, adding the regime representation raises success to 72.3%, and adding execution history alongside it raises success to 76.2%. (weakness2, question2)

Thank you for the response. Most of my concerns have been addressed. However, the existing ablations still do not independently validate the contribution of reference-guided supervision. In addition, the updated CIK results differ substantially from those in the original submission. I would appreciate clarification on whether the experimental setup has changed.

Reviewer_LHBXOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper evaluates vision-language-action models under persistent joint-level physical faults, including joint locks, restricted joint ranges, and increased friction. Unlike direct observation or action corruption, these interventions change how the robot physically executes the policy’s commands. The study evaluates OpenVLA-OFT and π0.5 across four LIBERO suites, characterizes performance differences across affected joints, and examines language-based execution feedback and fault-aware constrained inverse kinematics. It further trains a residual calibrator that modifies Cartesian translation commands between a frozen VLA and the underlying controller, using robot states and execution history. Supervision is derived from successful nominal trajectories and paired faulted rollouts. Simulation experiments and a real-robot bowl pick-and-place case study show that calibration can recover part of the performance lost under faults.

Critical strengths

1.A useful characterization of deployment-level reliability under physical faults.

The evaluation covers two VLA backbones, 40 LIBERO tasks, multiple fault conditions, and different affected joints. This provides concrete evidence about the sensitivity of deployed VLA systems to changes in physical execution capabilities. Such measurements are practically relevant because successful evaluation with an intact robot does not establish reliability after actuator degradation or joint constraints.

2.Evidence that some failures remain recoverable under the same physical fault.

The calibration results demonstrate that the observed degradation is not entirely explained by unavoidable mechanical incapacity. For example, under a locked joint (j_0), π0.5 improves from 57.4% to 72.2% success after calibration (Table 5). Evaluation on range restrictions and increased friction not encountered during calibration training, together with the limited real-robot case study, further supports the practical usefulness of execution-aware compensation.

Critical weaknesses

1.The substantive contribution associated with the VLA setting is insufficiently established relative to general execution compensation and residual policy adaptation.

The central problem—physical faults causing a mismatch between commanded and realized motion—also arises for other policies that issue Cartesian commands. The calibrator in Section 6 uses robot states, execution history, base actions, and joint-motion features, with supervision constructed from nominal trajectory waypoints. Its core mechanism does not directly use language, visual representations, or other properties of VLA models. Consequently, VLA primarily supplies the base commands, while the adaptation mechanism appears broadly applicable to policies sharing the same action interface.

This matters because the broader approach of retaining a base policy, inferring context from recent execution, and learning corrective actions is already established in work such as ReLCE [1]. ReLCE and this paper differ in their training objectives and information assumptions, so I am not claiming algorithmic equivalence. However, those differences do not by themselves establish a substantive methodological advance, and the main paper does not provide a sufficiently informative comparison under matched state information, supervision, and training data.

I recognize that the paper is positioned primarily as an empirical study. Nevertheless, it should establish what is learned about VLA deployment beyond the general consequences of altered robot dynamics and the benefits of external compensation. The current results do not sufficiently identify which aspects of VLA behavior explain the difficulty or the calibration benefit. This leaves both the VLA-related insight and the independent methodological increment insufficiently developed.

2.The evaluation does not adequately separate physical capability loss, low-level controller mismatch, and policy adaptation failure.

The authors appropriately acknowledge in Section 3 that some faults can make tasks physically infeasible, and they exclude (j_1) and (j_3) from parts of the analysis. However, the aggregate results in Table 1 still include these joints. Section 4 reports that locking (j_1) makes all LIBERO tasks infeasible and locking (j_3) makes 67.5% infeasible. The range-restriction experiments also retain only 3% or 5% of the original joint span. These are substantial changes to the robot’s available motion, and feasibility for the remaining fault–task combinations is not systematically established by the presented evidence.

The CIK experiment is a useful control, especially because it assumes correct identification of the faulty joint. However, a negative result for one local tracking implementation does not establish the achievable performance of fault-aware control or distinguish poor tracking of feasible commands from inappropriate policy-generated targets. Similarly, calibration gains establish partial recoverability without identifying the original failure mechanism.

As a result, the experiments characterize the reliability of the complete deployed system more clearly than they explain the VLA’s own adaptation limitations. A feasibility-controlled evaluation, accompanied by tracking measurements and a credible reference for fault-aware execution, is needed to determine the size and nature of the policy-level capability gap.

Questions for authors

1.What is the substantive research increment attributable to studying VLAs?

Which findings go beyond generic execution mismatch and residual compensation? A controlled comparison with a non-VLA policy using a comparable action interface, the same low-level controller, and comparable nominal performance would help determine whether the observed sensitivity and recovery patterns reveal a distinct limitation of the evaluated VLA systems.

2.How much degradation remains on fault–task pairs independently verified to be feasible?

How were the infeasibility claims for (j_1) and (j_3) established? Can you provide a reference trajectory, planner, or validated fault-aware controller for feasible cases, and report both end-effector tracking quality and task success? This would help distinguish unavailable physical motion, controller-level compensation failures, and ineffective policy action selection.

3.What benefit comes from the proposed calibration design beyond its additional information and supervision?

Please clarify the state and history signals actually consumed by each backbone and by the calibrator. Under the same fault rollouts, reference-waypoint supervision, and available state information, does the proposed design outperform a competitive generic residual adapter? Please identify any supplementary evidence that directly addresses this comparison.

3 comments
AuthorsOpenReview ↗

Response to LHBX: Central Contribution Associated with VLA

Thank you for the detailed review. Our central contribution is to clarify the limits of current VLA capabilities under physical faults and identify an effective adaptation route through analysis. (weakness1, question1) Three findings guide the calibration design.

First, the evaluated adversarial training does not improve joint-fault performance. Tessler et al. (2019) and VERONICA (Zhao et al., 2022) showed that adversarial learning or regularization can improve robustness to physical model changes. However, RobustVLA, reproduced on π0.5\pi_{0.5}, does not improve average success in any of our six joint-fault conditions. Explicit language feedback about intended motion, actual motion, and correction direction also provides no consistent improvement. These results motivate execution-side learning beyond the tested robustness and language interfaces. (weakness1, question1)

Second, fault heterogeneity motivates the fault-regime encoder. Failure severity and persistence vary across affected joints. We therefore supervise the encoder to distinguish joint-fault conditions and provide context for correction. Its embeddings form joint-dependent clusters under both locking and increased friction. (question3)

Third, the fault accumulation experiment in Sec. 4.2 reveals limited autonomous recovery in the evaluated VLA. Even after normal actuation is restored, task success generally decreases with longer prior fault exposure. The VLA can still struggle to complete the task from states reached during prolonged fault exposure, even after the fault is removed. (weakness2, question2).

This motivates nominal-trajectory supervision as an explicit recovery objective for deviations that the VLA does not reliably compensate for on its own. Paired nominal and faulty rollouts provide targets that preserve the base action near the reference trajectory and redirect translation toward it when meaningful deviations occur. (question3)

ReLCE (Nakhaei et al., 2024) learns residual corrections through reward-driven online RL. It does not use explicit fault conditions or nominal-trajectory supervision. Its training objective is therefore not aligned with our analysis. (weakness1) Our goal is to test whether directly addressing the identified recovery limitations can mitigate the observed failures. We therefore train the calibrator with fault-regime supervision and nominal-reference correction targets. The performance gains support this strategy, while the regime/history ablations support combining fault context with execution history. (weakness1, question3)

We believe that a contribution to VLA research does not necessarily require directly fine-tuning the VLA or using vision-language representations in the adaptation module. Our findings establish limits of the evaluated VLA robustness methods and language-feedback interfaces, and show that embodiment-side calibration can mitigate the observed joint-fault failures. Useful recovery with the pretrained VLA unchanged is an empirical finding of this investigation, providing concrete knowledge about where additional learning can extend current VLA capabilities. (weakness1, question1)

Reference

Chen Tessler, Yonathan Efroni, and Shie Mannor. Action robust reinforcement learning and ap- plications in continuous control. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Pro- ceedings of Machine Learning Research, pp. 6215–6224. PMLR, 09–15 Jun 2019. URL https://proceedings.mlr.press/v97/tessler19a.html.

Zhigen Zhao, Simiao Zuo, Tuo Zhao, and Ye Zhao. Adversarially regularized policy learning guided by trajectory optimization. In Roya Firoozi, Negar Mehr, Esen Yel, Rika Antonova, Jeannette Bohg, Mac Schwager, and Mykel Kochenderfer (eds.), Proceedings of The 4th Annual Learning for Dynamics and Control Conference, volume 168 of Proceedings of Machine Learning Research, pp. 844–857. PMLR, 23–24 Jun 2022. URL https://proceedings.mlr.press/v168/zhao22b.html.

Mohammadreza Nakhaei, Aidan Scannell, and Joni Pajarinen. Residual learning and context encoding for adaptive offline-to-online reinforcement learning. In Alessandro Abate, Mark Cannon, Kostas Margellos, and Antonis Papachristodoulou (eds.), Proceedings of the 6th Annual Learning for Dynamics & Control Conference, volume 242 of Proceedings of Machine Learning Research, pp. 1107–1121. PMLR, 15–17 Jul 2024. URL https://proceedings.mlr.press/v242/nakhaeinezhadfard24a.html

AuthorsOpenReview ↗

Response to LHBX: Reliability of the Complete Deployed System

Your comments helped us identify an issue in the low-level controller used in the original CIK evaluation. Thank you for pointing out the need to examine the controller’s role more closely. It used the full joint model to compute tracking torques even when a joint was locked. In a follow-up experiment, we evaluated a fault-aware controller that uses the known identity of the locked joint and computes tracking torques over the remaining joints. We report the new results in the main text and retain the original controller as the fault-only baseline. (weakness2, question2)

In Appendix G, we check whether the robot can reach positions along the nominal reference trajectory under the imposed joint constraints. We use this positional reachability check as a geometric proxy for feasibility. Even within the reference-reachable subsets, fault-only task success is 62.83% under joint lock, 72.39% at a joint-range ratio of 0.03, and 77.73% at 0.05. To understand how failures develop when the reference positions are reachable, we examine command tracking over the rollout. (weakness2, question2)

For the reachable joint-lock subset, we compare all available pairs of healthy and fault-only rollouts, regardless of task outcome. The distance between the end-effector positions in the paired rollouts increases from 3.51 mm at steps 0–5 to 85.48 mm at steps 62–102, as the joint lock remains active. (weakness2, question2)

In the follow-up to the CIK evaluation, we measured the distance between the realized and nominal end-effector (EEF) positions at steps 62–102 across four LIBERO suites. With the fault-aware controller, this distance decreases from 108.24 to 48.24 mm for j2j_2 and increases from 101.30 to 137.86 mm for j6j_6. Even when the locked joint is known, the effect of the controller on deviation from the nominal trajectory depends on which joint is locked. (question2)

The fault accumulation experiment in Section 4.2 examines whether the VLA can recover once the fault is removed. Final task success generally decreases with longer fault exposure, even after the fault is removed. At that point, the VLA must resume the task from the state the robot reached during faulty execution. The growing end-effector position gap in the paired rollouts supports this interpretation. Execution deviations accumulate while the fault is active, and the frozen VLA does not reliably recover from the states it reaches as a result. (weakness2)

These experiments show that failures persist even when the reference positions are reachable and that the controller can substantially affect how execution deviates from the nominal trajectory. They also reveal limited autonomous recovery in the evaluated VLAs. Even after the fault is removed, the VLAs struggle to resume task-directed behavior from states reached during prolonged fault exposure. (weakness2)

AuthorsOpenReview ↗

Response to LHBX: Follow-up on Our Response

Thank you again for your detailed suggestions. In our response, we clarified how the observed limited recovery and heterogeneous fault effect motivate nominal-reference supervision and how this objective differs from ReLCE. To help distinguish task feasibility, controller-level mismatch, and policy adaptation failure, we assessed physical reachability, compared end-effector tracking with fault-aware control, and tested recovery after fault removal. We would appreciate your feedback on whether these clarifications and analyses address your main concerns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Overview and Response Summarization

We thank the reviewers for their detailed feedback. Reviewers WD7G and D3LH highlighted the evaluation of physical faults in deployed VLAs as a strength. Reviewer LHBX questioned what studying physical faults in the VLA setting adds beyond general execution compensation (Weakness 1, Question 1). We address this concern by clarifying our central contribution. Our analysis identifies limits of the evaluated VLA robustness methods and language-feedback interfaces under joint faults and shows how those findings motivate execution-side calibration. Recovery with the pretrained VLA unchanged is an empirical finding of this investigation. It shows where additional learning can extend current VLA capabilities without directly fine-tuning the VLA or using vision-language representations in the calibration module.

The main revisions in response to the three reviewers’ feedback are summarized below.

  1. Fault accumulation and recovery (Section 4.2). In response to D3LH’s question about how errors develop over a rollout, we added an experiment that varies how long a joint fault remains active before removing it. Task success generally declines with longer prior fault exposure even after normal actuation is restored. Recovery therefore depends on the states the VLA reaches during faulty execution. This finding also addresses WD7G’s question about the calibration design. Nominal trajectories teach the calibrator where to steer when execution has drifted.

  2. Additional evidence in the supplementary material. To address questions from LHBX and WD7G about physical feasibility, we added an episode-level positional reachability analysis using nominal reference waypoints as a geometric proxy. In response to D3LH’s request for qualitative examples, we added joint-by-joint visualizations of increased-friction failures. To address WD7G’s question about how the calibration design relates to the observed failure patterns, we added the t-SNE visualization in Supplementary Figure 7. The embeddings form joint-dependent clusters under both joint locks and increased friction, supporting their use as context for execution-side calibration.

  3. Correction to the local-control evaluation. In re-examining the CIK evaluation, we found that the low-level controller computed tracking commands without accounting for the locked joint. We corrected the controller, reran the evaluation, and updated the results and discussion. We thank LHBX for raising detailed questions about low-level tracking that helped us identify this issue. Task success improves substantially for j0j_0 and j2j_2 and declines for j4j_4, j5j_5, and j6j_6. We added a detailed analysis in Section 5.3 of how closely the robot realizes the current command and how far subsequent VLA commands diverge from nominal behavior. Since the local-control evaluation uses a different setup, we moved its results from Table 5 to the appendix, where we present the full per-joint breakdown alongside the corresponding protocol.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.