acceptodds
Under review as a conference paper at ICLR 2027

Taming VLAs under Robot Execution Errors: Self-Compensation and Stress Testing

Abstract

Vision-language-action (VLA) policies often fail when a robot's executed motion deviates from their commanded action. Such execution errors arise from the robot's mechanics and operating conditions, such as wear and payload changes. We propose self-compensating VLA, a deployment-time adaptation method that enables a VLA policy to pre-compensate for the robot's execution errors when generating commands. Without task rewards or labels, it updates the policy online using the residual between the action commanded by a VLA and the motion executed by the robot. To stress-test VLA robustness across execution conditions that are impractical to cover with physical robots alone, we introduce RoboStress, a controlled simulation benchmark. It combines established joint-level models of friction, backlash, compliance, and gravity-compensation error into seven deployment scenarios whose execution errors depend on the robot's state and motion history. On RoboStress, self-compensating VLA achieves higher average task success than both the base policies and methods that build in robustness during training. On two physical robot arms with different usage histories, it raises the average task success rate by more than 30 percentage points on each arm, and the gains extend to objects not seen in the task demonstrations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, robot execution errors, deployment-time adaptation, robot manipulation, benchmark

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_zJx3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores an important and practical challenge in the deployment of VLAs, where the deviation between commanded actions and the actual execution outcomes of the robot. The idea of compensating for such execution errors is both attractive and well justified, as it directly addresses a real-world problem that has not yet received sufficient attention. In addition, it introduce RoboStress, a simulation benchmark that models four joint-level sources of execution error, thereby systematically evaluating the robustness of VLAs under different deployment conditions. Extensive experiments conducted in simulation and on real robots validate the effectiveness of the proposed compensation method. Although its technical approach is relatively simple and intuitive, this paper is clearly motivated and thoroughly evaluated. And most importantly, it highly focuses on practical deployment applications in robotics.

Critical strengths
  1. Practical and well-motivated problem. The paper addresses an important real-world challenge in VLA deployment. The proposed self-compensation method is simple yet appealing, requiring neither task rewards nor modifications to low-level controllers.
  2. Solid benchmark and extensive experiments. RoboStress provides a physically motivated benchmark with four joint-level error models. Extensive simulation and real-world experiments demonstrate the effectiveness of the method, making the work a solid and practically relevant contribution.
Critical weaknesses

The analysis of online adaptation efficiency is limited. Although the proposed method demonstrates promising improvements through online adaptation, it remains unclear how many episodes are needed to achieve meaningful performance gains. The current evaluation uses 20 episodes per task in simulation and 25 episodes per task on the real robot, and averages the success rates over these episodes. However, it does not show how performance evolves during this adaptation process. Providing success rate curves as a function of the number of adaptation iterations would help clarify the adaptation speed and better demonstrate the practical effectiveness of the method.

Questions for authors
  1. Limited real-world evaluation under varying execution conditions. The real-world generalization experiments focus on appearance variations of bowls and plates. However, since the paper primarily addresses execution errors caused by physical factors, evaluating objects with different weights instead would be more relevant and provide stronger evidence of the method's effectiveness under varying execution conditions.
  2. The proposed pseudo-targets assume that compensating for execution errors improves task performance, which may not always hold when the original actions are suboptimal or during contact-rich interactions. Although the ablations demonstrate empirical effectiveness, further analysis of the relationship between execution error reduction and task success would help clarify the reliability and limitations of this approach.
3 comments

We appreciate your insightful feedback and constructive suggestions, which help improve our paper substantially. We have addressed all the comments and included the additional experiments in the revised paper, where the changes are marked in blue. Please find our responses to the comments below. Unless otherwise stated, all additional experiments were conducted using π0.5\pi_{0.5}. Simulation results are averaged over the seven deployment scenarios of RoboStress, and real-robot results are obtained on Task 1 with Robot A.


Weakness 1. Analysis of online adaptation efficiency

Thank you for the constructive suggestion. We report how the success rate changes as the numbers of episodes and updates increase, both on the real robot and in simulation. The results are also shown in Figure R4.

On the real robot, the table below and Figure R4(a) report the success rate (%) over 25 consecutive episodes in five-episode windows. Our method achieves the same success rate as the base policy in the first five episodes and outperforms it in every window after the first.

Episodes1-56-1011-1516-2021-25
Base6040802040
Ours608010010080

In simulation, the table below and Figure R4(b) report the success rate (%) on the first task of LIBERO-Spatial in four-episode windows, averaged over the seven scenarios. Our method is comparable to the base policy in the first 12 episodes and outperforms it by 7.1 and 10.7 points in the last two windows.

Episodes1-45-89-1213-1617-20
Base75.078.689.371.475.0
Ours75.078.685.778.685.7

Figure R4(c) shows how the gain changes when the adaptation continues beyond the first task. It covers the ten tasks of each suite, across which the LoRA parameters are carried. The curve is the success rate of our method minus that of the base policy over all episodes so far. The gain increases overall as the number of updates increases.

In short, the gain of our method appears within a single task, after about 5 episodes on the real robot (about 60 updates) and after about 12 episodes in simulation (about 40 updates). It then continues to increase as the updates accumulate over the ten tasks. We have included these analyses in Section 5.3 of the revised paper, with Table 7 and Figure 6, and the details in Appendix D.7.


Question 1. Real-world evaluation with different object weights

Thank you for the valuable comment. As suggested, we conducted a real-robot experiment where weights of 500 g and 1,000 g are placed in the bowl, so that the load changes while the bowl and the grasp remain the same. The setup is shown in Figure R5, and the table below reports the success rate (%) over 25 episodes for each added weight.

Added weight0 g500 g1,000 g
Base484432
Ours848472

As shown in the table, our method outperforms the base policy by 36, 40, and 40 points with added weights of 0 g, 500 g, and 1,000 g, respectively. As the weight increases, the success rate of the base policy decreases from 48% to 32%, whereas our method achieves at least 72%. These results suggest that our method is effective under varying execution conditions. We have included this experiment in Section 5.3 and Figure 9 of the revised paper.

Question 2-1. Relationship between execution error reduction and task success

We agree that compensating for execution errors may not improve task performance when the action intended by the policy is itself suboptimal, since the task can fail even if the robot executes it accurately. We would like to note that our method aims to make the robot execute the action intended by the policy, not to correct the intended action itself. It is thus complementary to methods that improve action selection.

As suggested, we examined the relationship between execution error reduction and task success. Specifically, on RoboStress, we computed the reduction in execution error and the gain in success rate of our method over the base policy on each of the ten tasks, averaged over the suites. The table below reports the correlation between the two.

PearsonSpearman
Correlation0.830.68

The two are positively correlated, which indicates that the success rate improves more on the tasks where the execution error is reduced more. We have included this analysis and discussion in Appendix D.8 of the revised paper.


Question 2-2. Reliability under contact-rich interactions

Regarding contact-rich interactions, we evaluated our method under constrained contact by adding a pressing phase to four pick-and-place tasks. In the pressing phase, which lasts for 60 steps after grasping, we keep the policy active and add a downward offset to its commands so that the gripper is pushed against the table and its downward motion is blocked. Video R1 shows this setting, with the object interaction and pressing phases marked. To examine whether the residuals measured during contact mislead adaptation, we also evaluated two variants of our method that ignore them when updating the policy. The first ignores the residuals measured while the gripper holds the object, during both object interaction and pressing. The second ignores only those measured during pressing. Both use contact labels from the simulator to identify these residuals. Without pressing, the second variant is identical to our method. The table below reports the success rate (%).

MethodWithout pressingWith pressing
Base54.633.6
Ours55.739.8
Ours ignoring all contact residuals53.435.0
Ours ignoring pressing residuals–34.5

As shown in the table, our method outperforms the base policy under pressing, and ignoring the contact residuals does not lead to better results. These results suggest that our method remains effective under contact, and that the residuals during contact do not mislead adaptation in this setting. Our method also improves the success rate of the drawer-opening task on both arms in Table 6, where the gripper motion is constrained by the drawer. We have included this analysis in Appendix D.5 and Table D11 of the revised paper, and will add this contact setting to RoboStress.

AuthorsOpenReview ↗

A gentle reminder

Dear Reviewer zJx3,

We sincerely appreciate your effort and the time you have dedicated to reviewing our submission. We have addressed your concerns in our rebuttal, and it would be helpful if you could give us any further feedback before the author-reviewer discussion phase ends. Your opinion would be invaluable to us in improving our work, and we are more than willing to provide further clarifications or respond to any additional questions you might have. Thank you for your consideration.

Best regards,

Authors

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

The paper proposes Self-Compensating VLA, which adapts robot policies online by measuring the discrepancy between commanded actions and actual robot motion using proprioceptive feedback. The authors use this residuals to construct pseudo-targets, and Lora finetune the VLA to account for the discrepancy.

The paper also introduces RoboStress, a benchmark simulating realistic execution errors from friction, backlash, compliance, and gravity-compensation mismatch. Experiments are conducted on the LIBERO benchmark and two physical robots show improved task success over baseline VLAs and existing robustness methods.

Critical strengths
  1. I really like the fact that the authors used 2 piper arms -- one that was 1 year old, and a new one. This makes their experiments really exciting for me.

  2. The idea proposed is simple, and straightforward to implement.

  3. The paper is well written and easy to go through in a single sitting.

Critical weaknesses
  1. The compensation signal assumes that command execution discrepancies indicate errors that should be corrected by modifying future actions. However, during contact-rich manipulation, such discrepancies can also arise from environmental constraints, object interactions and contact. Currently the proposed method doesn't seem to distinguish these effects from actuator-induced execution errors. Could applying the same correction during constrained contact lead to overcompensation or unstable behavior? A contact-versus-free-space ablation, comparing residual magnitudes and adaptation outcomes, would help clarify this assumption.

  2. The direct residual correction baseline uses only the previous chunk's mean residual, despite the paper emphasizing state- and history-dependent execution errors. Could the authors compare against an online state-conditioned residual predictor (a lightweight MLP should suffice) that uses the same proprioceptive feedback and online buffer while keeping the VLA frozen? A history-conditioned variant would also be informative. This would isolate whether adapting the VLA itself provides benefits beyond learning a lightweight execution-error compensation model.

  3. For the real world evaluatation on the two robots A and B -- the authors motivate adaptation using robot wear and different usage histories, but the measured execution discrepancies of the new and old robots are very similar. Can the authors provide stronger evidence that aging meaningfully changes execution dynamics and that their method adapts to these differences? Providing per-joint / task-specific error breakdown would be helpful.

  4. Please report std. deviation for all the results - it's hard to make conclusions about any of the results without uncertainty estimates.


Right now, I'm leaning towards weak rejection of the paper - but, if authors can address my concerns above, I'm happy to move it to Weak Accept.

Questions for authors
  1. Since LoRA parameters are carried across episodes, can the authors report success rates as a function of episode number? How many episodes or gradient updates are required before the method meaningfully improves performance?
8 comments

We appreciate your insightful feedback and constructive suggestions, which help improve our paper substantially. We have addressed all the comments and included the additional experiments in the revised paper, where the changes are marked in blue. Please find our responses to the comments below. Unless otherwise stated, all additional experiments were conducted using π0.5\pi_{0.5}. Simulation results are averaged over the seven deployment scenarios of RoboStress, and real-robot results are obtained on Task 1 with Robot A.


Weakness 1. Effect of constrained contact on adaptation

We sincerely appreciate this insightful comment. As you point out, the residual itself does not distinguish actuator-induced errors from motion constrained by contact. However, since our method adapts the VLA itself, the compensation is conditioned on the image and the robot state, and can thus differ between free motion and contact.

As suggested, we compared residual magnitudes and adaptation outcomes between free motion and constrained contact. Specifically, we added a pressing phase to four pick-and-place tasks. In the pressing phase, which lasts for 60 steps after grasping, we keep the policy active and add a downward offset to its commands so that the gripper is pushed against the table and its downward motion is blocked. Video R1 shows this setting, with the object interaction and pressing phases marked. The table below reports the normalized residual (%).

MethodFree motionObject interactionPressing
Base20.836.030.0
Ours19.411.923.2

As shown in the table, the residual of our method is lower than that of the base policy in all three phases; the gap is particularly substantial in the object interaction and pressing phases. Figure R2 shows the residual over time in one episode, where both methods start from the same initial state. With our method, the residual decreases during pressing and remains low after the contact is released, whereas that of the base policy increases again. These results demonstrate that our method does not amplify, but rather reduces the residual under constrained contact.

To examine whether such residuals mislead adaptation, we also evaluated two variants of our method that ignore the residuals measured during contact when updating the policy. The first ignores the residuals measured while the gripper holds the object, during both object interaction and pressing. The second ignores only those measured during pressing. Without pressing, the second variant is identical to our method. The second table below reports the success rate (%).

MethodWithout pressingWith pressing
Base54.633.6
Ours55.739.8
Ours ignoring all contact residuals53.435.0
Ours ignoring pressing residuals–34.5

As shown in the second table, our method outperforms the base policy under pressing, and ignoring the contact residuals does not lead to better results. These results suggest that the residuals during contact do not mislead adaptation in this setting. Our method also improves the success rate of the drawer-opening task in Table 6, where the gripper motion is constrained by the drawer.

Regarding your concern about overcompensation or unstable behavior under constrained contact, we did not observe either. Overcompensation would make the policy press harder against the table as updates accumulate. We thus measured the contact force exerted on the table. During pressing, the mean contact force is comparable, 29.5 N with our method and 27.5 N with the base policy, as shown in Figure R3. With our method, the contact force shows no increasing trend as updates accumulate, with a Spearman correlation of 0.01. Unstable behavior would appear as oscillation after the contact is released. After pressing, the end-effector reverses its direction at 2.5% of the steps with our method and 2.4% with the base policy.

We have included this analysis in Section 5.3, Figure 8, and Appendix D.5 of the revised paper, and will add this contact setting to RoboStress.

Weakness 2. Comparison with state- and history-conditioned residual predictors

Thank you for the constructive suggestion. We compared our method with a state-conditioned and a history-conditioned residual predictor. Both are lightweight MLPs trained online with the same proprioceptive feedback and online buffer as our method, while the VLA is kept frozen.

MethodSuccess rate (%)
Base49.1
State-conditioned MLP50.9
History-conditioned MLP54.4
Ours56.6

As shown in the table, using history improves the predictor’s success rate by 3.5 points. Nevertheless, our method still outperforms the history-conditioned predictor by 2.2 points. We believe this is because the predictors estimate the correction only from the state, action, and history, whereas our method learns the compensation within the VLA and thus also exploits the image, language, and multimodal representation of the VLA. This result suggests that adapting the VLA itself provides benefits beyond a lightweight compensation model. We have included this comparison in Section 5.2 and Table 5 of the revised paper.


Weakness 3. Difference in execution errors between Robots A and B

We agree that the mean residual alone does not reveal the difference between the two arms. As suggested, we report the execution error (%) of each joint in Tasks 1 and 3, measured with the base policy and our method on each arm.

TaskJointRobot A, BaseRobot A, OursRobot B, BaseRobot B, Ours
Task 1J111.06.98.48.2
J26.65.77.25.7
J38.99.08.46.2
J48.74.714.814.1
J57.96.58.86.9
J610.76.723.820.3
Task 3J17.25.17.96.5
J26.76.47.95.7
J38.58.48.86.4
J49.010.816.413.3
J58.47.710.59.0
J68.86.957.256.9

As shown in the table, Robot B, which has been used for one year, shows larger execution errors than Robot A at J4 and J6 in both tasks under the base policy. The largest difference is 48.4 points at J6 in Task 3, while the differences at joints other than J4 and J6 are at most 2.6 points.

Our method reduces the execution error in 22 of the 24 cases, and at every joint of Robot B in both tasks. At J4 and J6 on Robot B, the reductions range from 0.3 to 3.5 points, although the error at J6 in Task 3 remains large. These results show joint-specific differences between the two arms and reductions in execution error on both, consistent with the improved success rates in Table 6. However, this comparison does not isolate aging as the cause of these differences. We have summarized these results in Section 5.3 and included the table in Appendix D.6 of the revised paper.


Weakness 4. Standard deviations of the results

Thank you for the comment. We report the mean and standard deviation of the main results in Table 2 with π0.5\pi_{0.5} over three seeds. As shown in the table below, our method achieves the best success rate in all seven scenarios. The mean ± standard deviation ranges of our method and the second-best method overlap under Thermal Drift-Backlash and Aged Transmission. Averaged across scenarios, however, our method achieves 56.6 ± 0.7%, compared with 51.0 ± 0.7% for DR, the second-best method overall.

ScenarioBaseDRRobustVLAOurs
Clean97.0 ± 0.597.1 ± 0.496.6 ± 0.495.8 ± 0.4
Heavy Payload44.1 ± 0.142.1 ± 0.438.9 ± 1.056.8 ± 1.5
Thermal Drift-Stribeck54.1 ± 0.454.6 ± 0.755.9 ± 0.360.5 ± 0.3
Thermal Drift-Backlash83.4 ± 0.784.5 ± 1.083.1 ± 0.384.8 ± 1.1
Aged Transmission33.2 ± 1.139.7 ± 1.540.2 ± 1.241.8 ± 1.6
Aged Joint-Uniform29.5 ± 0.936.0 ± 1.133.8 ± 2.138.6 ± 1.2
Aged Joint-Shoulder43.9 ± 2.446.4 ± 1.543.0 ± 1.051.2 ± 1.7
Aged Joint-Elbow55.6 ± 0.354.0 ± 0.853.9 ± 1.662.3 ± 1.2
Avg. (excluding Clean)49.1 ± 0.551.0 ± 0.749.8 ± 0.356.6 ± 0.7

We have included this table in Appendix D.4 of the revised paper, and will also report the standard deviations of the other simulation results.

Question 1. Success rate over episodes and updates needed for improvement

Thank you for the valuable comment. As suggested, we report how the success rate changes as the number of episodes increases, both on the real robot and in simulation. The results are also shown in Figure R4.

On the real robot, the table below and Figure R4(a) report the success rate (%) over 25 consecutive episodes in five-episode windows. Our method achieves the same success rate as the base policy in the first five episodes and outperforms it in every window after the first.

Episodes1-56-1011-1516-2021-25
Base6040802040
Ours608010010080

In simulation, the table below and Figure R4(b) report the success rate (%) on the first task of LIBERO-Spatial in four-episode windows, averaged over the seven scenarios. Our method is comparable to the base policy in the first 12 episodes and outperforms it by 7.1 and 10.7 points in the last two windows.

Episodes1-45-89-1213-1617-20
Base75.078.689.371.475.0
Ours75.078.685.778.685.7

Figure R4(c) shows how the gain changes when the adaptation continues beyond the first task. It covers the ten tasks of each suite, across which the LoRA parameters are carried. The curve is the success rate of our method minus that of the base policy over all episodes so far. The gain increases overall as the number of updates increases.

In short, the gain of our method appears within a single task, after about 5 episodes on the real robot, or about 60 updates, and after about 12 episodes in simulation, or about 40 updates. It then continues to increase as the updates accumulate over the ten tasks. We have included these analyses in Section 5.3, Table 7, and Figure 6 of the revised paper, with details in Appendix D.7.

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

This paper studies VLA robustness to robot execution error from mechanics and operating conditions. It measures the residual between commanded action and executed displacement via proprioception, constructs compensated pseudo-targets, and updates LoRA adapters on the action expert online under a flow-matching objective, base frozen, no reward or labels. The proposed method beats base policies and two training-time robustification methods.

Critical strengths
  1. Execution error is something every deployment hits and benchmarks ignore. This paper is cleanly separated it from synthetic action noises. The DTW comparison shows Gaussian perturbation can not fit real payload deviations. It worths study in practice.
  2. The sim-benchmark-plus-hardware loop is more complete than most sim-only or hardware-only papers.
Critical weaknesses
  1. No comparison against direct competitors in test-time adaptation.
  2. Simulation gains are small and uneven. Benefits concentrate in hardware and heavy-wear settings. Should be more clear about where the method does and does not apply.
Questions for authors
  1. Please add an EVOLVE-VLA or TT-VLA-style test-time adaptation baseline. If such methods also work here, what is your advantage over them?
  2. Did online LoRA updating ever degrade or oscillate on hardware machine? Waht were the anchor weight and learning rate used and their working range?
  3. How were the severity factors of the other six scenarios set? With a different set, would the ordering change?
  4. What is the mechanism behind direct correction (23.6%) falling far below base (41.7%) in Table 4?
4 comments

We appreciate your insightful feedback and constructive suggestions, which help improve our paper substantially. We have addressed all the comments and included the additional experiments in the revised paper, where the changes are marked in blue. Please find our responses to the comments below. Unless otherwise stated, all additional experiments were conducted using π0.5\pi_{0.5}. Simulation results are averaged over the seven deployment scenarios of RoboStress, and real-robot results are obtained on Task 1 with Robot A.


Weakness 1 / Question 1. Comparison with test-time adaptation methods

Thank you for the constructive suggestion. We compared our method with TT-VLA and EVOLVE-VLA both on RoboStress and on the real robot, following the protocols of Table 2 and Table 6 of the paper, respectively. Since the official code of EVOLVE-VLA has not been released, we re-implemented it following the paper.

MethodRoboStress (%)Real robot (%)
Base49.148
TT-VLA49.756
EVOLVE-VLA50.164
Ours56.684

As shown in the table, our method outperforms both methods by at least 6.5 points on RoboStress and 20 points on the real robot. We attribute this to the difference in the adaptation objective. Specifically, TT-VLA and EVOLVE-VLA adapt the policy to increase the estimated task progress, which reflects execution errors only indirectly. In contrast, our method adapts the policy to compensate for execution errors, which are directly measured through proprioception and used to supervise the action. Our method also requires no task reward or progress estimator. We have included this comparison and discussion in Sections 5.2 and 5.3 of the revised paper, with Tables 5 and 8.


Weakness 2. Uneven simulation gains and applicability of the method

Our method improves the success rate of π0.5\pi_{0.5} in all seven deployment scenarios by 7.5 points on average, although the gains vary across scenarios.

Across the 35 scenario-severity settings in Table D8, our gain is positively correlated with the base policy’s performance drop from Clean, with a Pearson correlation of 0.88. For example, the gain increases from 3.4 to 12.7 points as the severity of Heavy Payload increases. This result suggests that our method recovers part of the performance lost to execution errors.

On the other hand, the gain is limited where the errors have a smaller effect on the base policy. Under Thermal Drift-Backlash, the base policy retains a success rate of 83.4%, and our method improves it by only 1.4 points, compared with 6.4 to 12.7 points in the other six scenarios. Under Clean, where no execution error is injected, our method slightly lowers the success rate from 97.0% to 95.8%. We have clarified these conditions in Appendix D.2 of the revised paper, with a summary in Section 5.2.

Question 2-1. Stability of online LoRA updates on hardware

In our real-robot experiments, we observed neither a progressive decline in success rate nor oscillation as updates accumulated. The table below reports the success rate (%) over 25 consecutive episodes. Our method achieves at least 80% in every five-episode window after the first, and is at or above the base policy in all windows.

Episodes1-56-1011-1516-2021-25
Base6040802040
Ours608010010080

To examine oscillation, we recorded the joint angles at 200 Hz for 25 episodes per method, during which our method performed 304 LoRA updates. We flagged a potential oscillation when a joint made three or more direction reversals of at least 0.5° within 0.6 sec, then inspected each flagged event in the video. No oscillation was found in any episode of either method, including the chunks executed right after an update. We also measured small high-frequency motion in each episode as the root mean square (RMS) of the joint motion above 3 Hz. As shown in Figure R1, it remains comparable to that of the base policy and does not increase with the number of updates, with a Spearman correlation of −0.04. In the figure, the line shows the mean of the base policy and the band covers 95% of its episodes. We have included these analyses in Section 5.3, Table 7, and Figure 7 of the revised paper, with details in Appendix D.7.


Question 2-2. Anchor weight, learning rate, and their working range

We used a learning rate of 10−610^{-6} for π0\pi_0 and 10−510^{-5} for π0.5\pi_{0.5}, with an anchor weight of 10−410^{-4}, both in simulation and on the real robots. As requested, we examined their working range on the real robot. With π0.5\pi_{0.5}, we varied each of them while fixing the other at its default value. The tables below report the success rate (%) over 25 episodes for each setting.

Learning rate10−710^{-7}10−610^{-6}10−510^{-5} (default)10−410^{-4}10−310^{-3}
Success rate (%)6476847256
Anchor weight10−610^{-6}10−510^{-5}10−410^{-4} (default)10−310^{-3}10−210^{-2}
Success rate (%)7280847264

As shown in the tables, our method outperforms the base policy, whose success rate is 48%, in a wide range of learning rates and anchor weights. We have stated this setting in Section 5.1 and Appendix B.6, and included these results in Appendix D.3 of the revised paper.


Question 3-1. Rationale for the severity factors

The severity factors of each scenario are presented in Appendix A.5. We chose them as stress-test conditions that remain physically plausible according to the literature. Specifically, friction ×2 scales the friction torque while preserving its velocity dependence, as joint friction is known to increase with wear and to change with temperature [1]. Backlash ×15 results in a total gap of 1.5° to 2.2° depending on the joint, which is comparable to the backlash of about 2° reported for a geared robotic joint [2]. Stiffness ÷5 reduces the stiffness to 20% of its nominal value, which is comparable to the reduction to one-fourth of the nominal stiffness reported for a worn strain wave gear [3].

In the six scenarios, the two Thermal Drift scenarios gradually increase either friction or backlash within an episode to examine errors that change over time. Aged Transmission combines friction and backlash on all joints, and Aged Joint additionally increases compliance and applies the factors to different sets of joints. We have included this rationale in Appendix A.2 of the revised paper.

[1] Modeling and Experiment Design for Identification of Wear in a Robot Joint Under Load and Temperature Uncertainties Based on Friction Data, IEEE/ASME Transactions on Mechatronics 2014

[2] Backlash Elimination Control for Robotic Joints with Dual-Motor Drive, Actuators 2024

[3] A Preliminary Experimental Study on the Effects of Wear on the Torsional Stiffness of Strain Wave Gears, Actuators 2022

Question 3-2. Method ordering under different severity levels

To examine whether the ordering changes with a different set of severity factors, we extended the severity analysis in Appendix D.2 of the paper to DR and RobustVLA. In this analysis, each scenario is evaluated under five sets of severity factors, L1 to L5 in Table D7, where L5 is the setting of Table 2. The table below reports the success rate (%) of π0.5\pi_{0.5} averaged over the four LIBERO suites. Our method achieves the best success rate in 33 of the 35 settings. The two exceptions are L1 and L3 of Thermal Drift-Backlash, where it is within 1.4 points of the best method. In contrast, the ordering among the other three methods varies with the severity level.

ScenarioMethodL1L2L3L4L5
Heavy PayloadBase80.974.265.455.944.1
DR78.572.463.953.242.1
RobustVLA79.270.962.652.438.9
Ours84.379.372.865.056.8
Thermal Drift-StribeckBase60.558.856.755.554.1
DR60.259.257.955.054.6
RobustVLA62.160.057.557.855.9
Ours65.263.663.062.260.5
Thermal Drift-BacklashBase92.490.890.287.783.4
DR93.890.888.686.284.5
RobustVLA92.290.586.684.983.1
Ours92.490.989.187.884.8
Aged TransmissionBase56.252.045.939.033.2
DR58.450.644.639.239.7
RobustVLA56.651.846.042.540.2
Ours62.259.253.346.741.8
Aged Joint-UniformBase57.550.243.836.729.5
DR59.951.544.637.936.0
RobustVLA59.951.546.036.433.8
Ours62.859.852.147.338.6
Aged Joint-ShoulderBase68.463.556.451.343.9
DR69.566.158.253.546.4
RobustVLA67.964.658.250.543.0
Ours74.270.064.559.351.2
Aged Joint-ElbowBase66.364.261.059.055.6
DR65.462.457.657.554.0
RobustVLA66.264.960.656.653.9
Ours72.368.466.363.962.3

We have included these results in Appendix D.2 and Table D8 of the revised paper.


Question 4. Degradation of direct correction in Table 4

Direct correction keeps the policy frozen and subtracts the mean residual of the previous chunk from all commands of the next chunk. In other words, it assumes that the execution error stays the same across chunks, whereas the error actually varies with the pose, speed, and direction of motion. We attribute the degradation to this mismatch, which causes the correction to add error to the commands instead of removing it.

To address this limitation, we additionally evaluated a state-conditioned correction on π0\pi_0, following the setting of Table 4. In this variant, a lightweight MLP is trained online to predict the residual from the current state while the policy is kept frozen. Although this variant avoids the degradation of direct correction, it still underperforms our method, as shown in the table below.

MethodSuccess rate (%)
Base41.7
Direct correction23.6
State-conditioned MLP42.5
Ours44.2

We have included this analysis in Section 5.2 and Table 4 of the revised paper.

AuthorsOpenReview ↗

A gentle reminder

Dear Reviewer ZSmh,

We sincerely appreciate your effort and the time you have dedicated to reviewing our submission. We have addressed your concerns in our rebuttal, and it would be helpful if you could give us any further feedback before the author-reviewer discussion phase ends. Your opinion would be invaluable to us in improving our work, and we are more than willing to provide further clarifications or respond to any additional questions you might have. Thank you for your consideration.

Best regards,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their valuable comments and constructive feedback. We are grateful that the reviewers recognized the following strengths of our work: (1) the practical and well-motivated problem of execution errors in VLA deployment, (2) the simple method that requires no task rewards, and (3) the solid experiments on both RoboStress and real robots.

We revised the paper according to the feedback and marked every change in blue. The main updates are summarized below, and we respond to each reviewer's comments in the corresponding thread.

(1) New comparisons

  • Online adaptation methods: TT-VLA and EVOLVE-VLA on RoboStress and on the real robot (Sections 5.2 and 5.3, Tables 5 and 8).
  • Residual predictors: State- and history-conditioned MLPs (Section 5.2, Tables 4 and 5).
  • Severity levels: DR and RobustVLA across five levels (Appendix D.2).

(2) New real-robot experiments

  • Object weight: Added weights of 500 and 1,000 g (Section 5.3, Figure 9).
  • Hyperparameters: Working range of the learning rate and anchor weight (Appendix D.3).
  • Per-joint execution errors: Comparison between Robots A and B (Section 5.3, Appendix D.6).

(3) New analyses

  • Adaptation over episodes: Success rate and stability as updates accumulate (Section 5.3, Table 7, Figures 6 and 7, Appendix D.7).
  • Constrained contact: Residual, success rate, and contact force under pressing (Section 5.3, Figure 8, Appendix D.5).
  • Execution error and task success: Correlation between the two (Appendix D.8).
  • Applicability: Conditions under which the gain is large or limited (Section 5.2, Appendix D.2).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.