acceptodds
Under review as a conference paper at ICLR 2027

ReflectVLA: Event-Anchored Paired Trajectory Conditioning for Corrective Re-execution

Abstract

Vision-Language-Action (VLA) methods have achieved strong performance for robotic manipulation. However, models trained on success-only datasets often fail when driven into out-of-distribution states. Existing correction strategies use different intervention interfaces, while paired trajectory conditioning directly represents the deviation between failed and successful behaviors for controllable re-execution. In this paper, we propose ReflectVLA, a framework that learns corrective re-execution from paired successful and failed trajectories. The core of ReflectVLA is an event-anchored Contrastive Trajectory Encoder (CTE) that learns paired trajectory representations, capturing how failed trajectories diverge from successful references. The deviation representation is encoded as correction tokens and injected into the VLA inference process. A correction adapter enables the policy to interpret these correction tokens and generate re-execution actions while preserving the base VLA behavior during normal execution. During inference, an external detector identifies an incipient failure. When the trigger arrives before an irreversible environmental change, ReflectVLA interrupts execution, rolls back to a visited pre-failure keyframe, and retrieves the most relevant successful reference from a Success Bank. The paired success–failure trajectory is jointly encoded to guide correction. Experiments demonstrate that ReflectVLA enhances recovery performance under this early, reversible intervention setting.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Failure Recovery, Contrastive Trajectory Learning, Corrective Re-execution

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_LKDKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ReflectVLA uses paired failed and successful trajectories to guide corrective re-execution. A detector trigger rollback and retrieval of a successful reference. Evaluation focuses on early, reversible failures in LIBERO.

Critical strengths
  1. Under shared triggers and budgets, the method recovers 452 episodes, compared with 97 for Retry and 239 for action replay (Table 2). This supports a benefit beyond simply trying again.

  2. The paper uses closed-loop rollback rather than restoring simulator states, and includes three-seed results and component ablations. 3.The logic is clear and the method is reasonable.

Critical weaknesses
  1. The main evaluation updates the Success Bank and allows two corrections, while the audit uses a fixed bank and the first trigger. Why their gains match exactly is unclear.

  2. The detector relies on simulator signals, and training failures are synthetic. It is still unclear how well the method handles natural failures with realistic detection

Questions for authors
  1. Can you explain how Tables 1–3 relate and report end-to-end results with a frozen training-only Success Bank?

  2. How does recovery change with noisy detection and natural failures, counting rollback within the step budget?

5 comments
AuthorsOpenReview ↗

Response to Reviewer LKDKPart 1

Response to Weakness 1 & Question 1:

Thank you for highlighting the ambiguity in our evaluation protocol. We have clarified the relationship between the system-level evaluation and shared-trigger audit, and additionally evaluated ReflectVLA with a frozen training-only Success Bank. As the manuscript has undergone revisions, the table numbering has been updated accordingly. All table references below are based on the latest version of the manuscript.

(1) Evaluation protocol and Success Bank usage. Both the main evaluation (Table 1) and the shared-trigger audit (Table 3) allow up to two rollback-and-correction attempts per episode under the same environment-step budget, including physical rollback actions. In the audit, the first trigger serves only as a shared intervention point. Subsequent triggers may still initiate a second correction.

In the main evaluation, each seed/task-specific Success Bank is initialized from training successes and updated with verified successful evaluation trajectories. In the shared-trigger audit, the bank snapshot and retrieved reference are held fixed across comparison conditions at each trigger, rather than freezing the training-only bank throughout evaluation. Thus, the audit controls retrieval differences without changing the underlying online-update protocol.

(2) Relationship between Tables 1, 3, and 8. Table 1 reports end-to-end success rates over 1,500 episodes per LIBERO suite (6,000 in total). Appendix Table 8 provides the rule-based detector statistics, identifying 1,364 true-positive triggers for OpenVLA. Table 3 reports conditional recovery on these same triggers.

Importantly, the Full and Retry results in Tables 1 and 3 are derived from the same matched episode-level evaluation records, rather than independent evaluations. Specifically, Full ReflectVLA recovers 452/1,364 triggered failures, while Retry recovers 97/1,364. Their per-suite differences in recovered episodes are 96, 71, 113, and 75, corresponding to end-to-end success-rate gains of 6.4, 4.7, 7.5, and 5.0 percentage points, respectively, after normalization by 1,500 episodes per suite. Overall, the 355 additional recovered episodes correspond to 355/6,000 = 5.92 percentage points, matching the 5.9-point gain reported in Table 1 for OpenVLA with the rule-based detector.

Therefore, the exact correspondence arises from the same underlying recovery outcomes reported with different denominators, rather than independently obtained identical improvements. Table 3 uses true-positive triggers as the denominator for conditional recovery, whereas Table 1 uses all evaluation episodes for end-to-end success.

(3) End-to-end evaluation with a frozen training-only Success Bank. We additionally evaluated ReflectVLA with online bank updates disabled, retaining only successful training demonstrations throughout evaluation. As reported in the revised Table 4, the frozen-bank variant achieves the following end-to-end success rates under the same single-seed setting:

SettingSpatialObjectGoal10Avg.
Frozen Bank92.693.484.059.882.45
Full Model92.693.884.660.682.90

The frozen-bank variant achieves 82.45%, only 0.45 percentage points below the online-updated model under the same setting. This indicates that the observed recovery improvements do not substantially depend on online Success Bank updates.

We have clarified these protocols and their relationship in Sections 3.5, 4.2, and 4.3, and Appendices A.2, A.4, and A.7. The frozen-bank results are included in Table 4 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer LKDK Part 2

Response to Weakness 2 & Question 2:

Thank you for raising this important concern. We have extended our evaluation to examine recovery from naturally occurring policy failures under imperfect, non-privileged failure detection, while explicitly accounting for physical rollback within a unified execution budget.

(1) Natural failures versus synthetic training data. We clarify that synthetic failure trajectories are used only for training the correction mechanism. During evaluation, failures arise from the autonomous execution of the original base VLA policies, without injecting OU noise or other artificial action perturbations. Thus, recovery is evaluated on naturally occurring policy failures in simulation, rather than on the synthetic failures used for training. These include unsuccessful grasps, placement deviations, and failed manipulation interactions. We additionally evaluate on LIBERO-Plus, which introduces distribution shifts in camera viewpoints, robot configurations, language instructions, lighting, backgrounds, visual noise, and object layouts.

(2) Recovery under imperfect, non-privileged detection. To address the reliance on privileged simulator signals, we additionally introduce a frozen Qwen3.5-27B-based VLM detector. Unlike the rule-based detector, it uses only causally available RGB observations, robot proprioception, executed actions, and language instructions, without accessing ground-truth object poses, contact states, or simulator task-completion signals.

As reported in Appendix Table 8, the VLM detector achieves 96.84% accuracy, 87.85% recall, and a 2.05% false-positive rate, reflecting imperfect detection with both missed and false triggers. Importantly, ReflectVLA maintains consistent improvements under this detector. On LIBERO (Table 1), the end-to-end success rates with rule-based versus VLM-based detection are:

BackboneRule-basedVLM-basedVLM gain over Retry
OpenVLA82.8%82.3%+5.8 pp
OpenVLA-OFT98.2%97.9%+0.8 pp
π0.598.9%98.7%+2.2 pp

Across all three backbones and four LIBERO suites, the number of recovered base-policy failures changes from 742/1,940 (38.25%) with rule-based detection to 705/1,940 (36.34%) with VLM-based detection, a decrease of 1.91 percentage points. Since the two detectors can trigger on different failure events, this comparison measures end-to-end recovery rather than conditional recovery on identical triggers.

Furthermore, under VLM-based detection on LIBERO-Plus (Table 2), ReflectVLA achieves 82.0% with OpenVLA-OFT and 89.9% with π0.5, outperforming their corresponding Retry baselines by 9.6 and 9.4 percentage points, respectively. These results provide additional evidence of recovery robustness under imperfect visual observations and distribution shifts.

(3) Physical rollback and unified execution budget. We clarify that all executed simulator control steps count toward the environment-step budget, including normal policy actions, gripper preparation, physical rollback, and corrective re-execution. For each task, all internally evaluated methods and their corresponding baselines use the same maximum environment-step budget, with at most two rollback-and-correction attempts per episode. Rollback is executed through closed-loop tracking of previously visited end-effector poses, without resetting the simulator or restoring object states. Therefore, ReflectVLA receives neither free rollback steps nor additional environment interactions compared with the corresponding Retry baseline.

Overall, these evaluations demonstrate that ReflectVLA retains its recovery advantage on autonomous policy failures under imperfect, non-privileged detection, with physical rollback fully included in the execution budget. We acknowledge that the current evaluation remains simulation-based and focuses on early, recoverable failures. Handling irreversible scene changes and real-world deployment remains future work.

We have clarified these details in Sections 3.5 and 4.2, Tables 1–2, and Appendices A.3–A.4, A.5.2, and A.6.1 of the revised manuscript.

AuthorsOpenReview ↗

Additional Experiments on Maniskill Plate25 Benchmark

To further strengthen our evaluation, we have added new experiments on the ManiSkill3 Plate25 benchmark (PutOnPlateInScene25Main-v3), as reported in Appendix A.5.3 (Table 7) of the revised manuscript.

Beyond the standard in-distribution setting, we evaluate three out-of-distribution (OOD) categories: visual, semantic, and execution perturbations. We conduct experiments with both OpenVLA-OFT and π0.5 backbones, comparing the original policies, their Retry variants, and ReflectVLA. Both Retry and ReflectVLA use the same VLM-based failure detector and rollback interface to ensure a controlled comparison.

The new results demonstrate consistent improvements:

  • OpenVLA-OFT: ReflectVLA achieves an OOD average success rate of 69.78%, compared with 57.15% for Retry (+12.63 percentage points).
  • π0.5: ReflectVLA achieves an OOD average success rate of 71.53%, compared with 58.51% for Retry (+13.02 percentage points).
  • OOD Execution: The improvements are particularly pronounced under execution perturbations, reaching +18.52 and +20.39 percentage points over the corresponding Retry baselines.

These additional results provide further evidence that the benefits of paired success–failure trajectory conditioning extend beyond the LIBERO benchmark and remain effective under different types of distribution shifts. While real-world deployment remains an important direction for future work, the ManiSkill3 experiments provide an additional independent simulation benchmark supporting the generalization and robustness of ReflectVLA.

We have incorporated these results and the corresponding analysis into Appendix A.5.3, Table 7 of the latest revised manuscript.

Reviewer_Le5ZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ReflectVLA first constructs a paired success-failure dataset, and then it uses it to construct a contrastive trajectory encoder, and then it trains this CTE and the Correction Adapter with the original VLA parameters frozen. The major motivation for ReflectVLA is that current correction mechanisms do not utilize the differences between successful or failed cases and lack corresponding generalization capabilities to action failures. Experiments show performance gains.

Critical strengths
  1. The paper is quite well written.
  2. The structure of the paper is clear.
  3. Figure 2 and Figure 3 are drawn pretty nicely.
  4. The main idea is worth exploring.
Critical weaknesses

For the method sections

The main concern I have for the method part is: In Section 3.5, it states that it uses “a lightweight rule-based detector over task-relevant signals and the simulator target-object position” to detect the failure cases, so this rule-based detector feels quite disconnected from the main method. First, this rule-based detector requires prior knowledge and action states in the simulation environments, which may hinder its generalization capabilities in real-world rollouts (what if you don’t have such prior knowledge or a clear target-object position in a real detailed data structure). Second, the reason for using OU noise, as explained in Section 3.3, is its temporal correlation. The main purpose of adding OU noise is to increase the diversity of failure trajectories to improve the model’s robustness and generalization to unseen failure scenarios. So what about other noises? Uniform Noise, Impulse Noise, or Correlated Gaussian Noise? The purpose of noise is to enrich diversity, so it would be helpful to test more noise types, and thus possibly simulate more potential failure cases.

For the experimental sections

1: One of the main concerns for the experimental section is that there are no real-robot rollouts. Since this is a paper about corrective re-execution, pure simulation environments may not sufficiently capture the failure cases the robot may face during real-world execution. Since training on paired success-failure trajectories is required, what about the sim-to-real gap here? In my understanding, ReflectVLA can be applied to real-world scenarios to see how it performs, and in real-world environments, you can see many more “failed trajectories” and thus further teach the robot how to react to these failed cases and expand its OOD capabilities.

2: The models and datasets I see here (I’ve checked both the main paper and the supplementary material) are quite limited. The models are limited to the OpenVLA family (OpenVLA and OpenVLA-OFT), so what about the Pi model series? Since there are different action policies for different VLA model families, conducting experiments solely on OpenVLA cannot fully demonstrate the efficiency of the proposed method. Also, for the dataset, it is the same problem: LIBERO only may not be sufficient enough. Referring to Table 6 in the paper, OpenVLA-OFT itself can reach up to around a 0.95 success rate on four different suites in LIBERO, so the gain is quite marginal on this dataset. ReflectVLA can be applied to LIBERO-Pro to check its performance.

I understand that due to limited resources, it’s not quite reasonable to ask the authors to complete both full-range real-world rollouts and extensive simulation experiments, but experiments solely on the OpenVLA series and on the LIBERO dataset are not sufficient enough to support a strong paper.

Questions for authors

Please refer to the weaknesses section.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Le5Z Part 1

Response to Weakness 1 (Method)

We agree that simulator-specific failure signals may limit real-world applicability. We therefore add a non-privileged VLM detector and clarify the rationale for our failure synthesis.

(1) Failure detection without privileged simulator information. The original rule-based detector uses task-specific simulator signals, but detection is modular: the detector decides when and where to intervene, whereas the CTE and correction adapter learn how to re-execute. We additionally implement a frozen Qwen3.5-27B VLM detector using only causal RGB observations, robot proprioception, executed actions, and language instructions. It receives no ground-truth target-object positions, contact/articulated-object states, task-completion signals, or future observations. It also selects recovery keyframes from observed execution history rather than privileged simulator states (Section 3.5; Appendix A.3.2).

The VLM detector achieves 96.84% accuracy, 87.85% recall, and a 2.05% false-positive rate (Table 8). ReflectVLA retains consistent gains over VLM-triggered Retry (Table 1):

BackboneRetry (VLM)ReflectVLA (VLM)Gain
OpenVLA76.5%82.3%+5.8 pp
OpenVLA-OFT97.1%97.9%+0.8 pp
π0.596.5%98.7%+2.2 pp

This shows that the learned correction mechanism remains effective without privileged failure detection. Real-robot deployment, however, remains to be evaluated.

(2) Motivation for OU noise and event-conditioned synthesis. Our goal is to reproduce representative policy failure patterns, not merely to maximize perturbation diversity. Before dataset construction, we extensively inspected autonomous failures across OpenVLA and π-series policies. Recurring failures included accumulated end-effector drift, pose misalignment, missed grasps, inaccurate placements, and unsuccessful contact interactions.

These observations motivated temporally correlated OU noise to produce gradual, accumulated action deviations. In preliminary qualitative trials, frame-independent Gaussian noise and unstructured random perturbations often yielded less coherent motions that resembled observed failures less closely. Moreover, our synthesis is not limited to OU noise: we introduce event-specific, task-conditioned modifications for grasp and placement offsets, contact errors, altered pushing/pulling directions, and insertion/release failures, with multiple severity levels. Each candidate is re-executed under simulator dynamics, and only verified task failures are retained (Section 3.3; Appendix A.1).

We agree that uniform, impulse, and alternative temporally correlated noises could cover additional failure modes. We do not claim OU is universally optimal or that these preliminary observations replace a controlled noise-type ablation.

(3) Effectiveness beyond synthetic failures. Synthetic perturbations are used only for training; evaluation instead uses failures arising from autonomous base-policy execution, without injecting artificial action noise. Thus, we test transfer beyond the synthesized failure distribution. Across the evaluated backbone–suite settings, ReflectVLA recovers 742/1,940 (38.25%) base-policy failures with rule-based detection and 705/1,940 (36.34%) with non-privileged VLM detection (Appendix A.5.2). On LIBERO-Plus, it outperforms VLM-triggered Retry by 9.6 pp with OpenVLA-OFT and 9.4 pp with π0.5 (Table 2). Together, these results support the effectiveness of the current OU-plus-event synthesis on naturally occurring policy failures in simulation, including under distribution shifts.

AuthorsOpenReview ↗

Response to Reviewer Le5Z Part 2

Response to Weakness 2 (Experiments):

We thank the reviewer for this important comment. We agree that real-robot evaluation is essential for fully assessing corrective re-execution, as physical environments introduce sensing uncertainty, contact dynamics, and failure modes that may not be fully captured in simulation. While our current evaluation remains simulation-based, we have substantially extended the experiments and implementation to better assess the practical applicability of ReflectVLA and reduce its dependence on simulator-specific assumptions.

(1) Generalization from synthetic training failures to autonomous policy failures. We first clarify that synthetic success–failure pairs are used for training, whereas evaluation is conducted on naturally occurring failures from autonomous base-policy execution, without injecting artificial action perturbations. Therefore, the correction mechanism is not merely evaluated on the same synthesized failure distribution used for training.

As reported in Appendix A.5.2, ReflectVLA recovers 742/1,940 (38.25%) base-policy failures with rule-based detection and 705/1,940 (36.34%) with VLM-based detection across the evaluated backbone–suite settings. Furthermore, we extend our experiments to π0.5 and LIBERO-Plus, where ReflectVLA improves the corresponding Retry baselines by 9.6 pp with OpenVLA-OFT and 9.4 pp with π0.5 under VLM-based detection (Table 2). These results provide evidence that correction learned from synthetic trajectories transfers to autonomous policy failures under diverse distribution shifts in simulation.

(2) Reducing simulator-specific assumptions for practical deployment. To address the reliance on privileged simulator signals, we additionally introduce a frozen Qwen3.5-27B-based VLM detector (Appendix A.3.2). It uses only causally available RGB observations, robot proprioception, executed actions, and language instructions, without accessing ground-truth object poses, contact labels, or simulator task-completion signals. Importantly, the same CTE and correction adapter remain effective under this non-privileged detector.

Moreover, our rollback procedure does not reset the simulator or restore object states. Instead, it physically tracks previously visited end-effector poses using closed-loop control, while object dynamics continue evolving normally. All executed actions, including rollback, are counted within the same task-specific environment-step budget (Appendix A.4). Although evaluated in simulation, this design avoids relying on an oracle state-restoration operation that would be unavailable on physical robots.

(3) Incorporating real-world failures for further adaptation. We agree with the reviewer that naturally collected real-world failures offer a valuable opportunity to further improve corrective capabilities. ReflectVLA's paired-trajectory formulation provides a natural framework for such an extension: verified successful robot demonstrations could populate the Success Bank, while recorded real-world failure trajectories could be aligned with successful references by task and interaction event to construct additional training pairs. These pairs could then be used to further adapt the CTE and correction adapter, reducing reliance on purely synthetic failures.

We emphasize that this real-world failure-driven adaptation is a promising extension rather than a capability already validated in the current experiments. ReflectVLA presently targets early, recoverable execution deviations, and physical deployment would require additional validation of contact safety, rollback feasibility, and robustness to irreversible scene changes.

Overall, while we acknowledge the absence of real-robot rollouts, the revised experiments demonstrate transfer from synthetic training failures to autonomous policy failures, effectiveness under non-privileged visual detection, and robustness to diverse simulation distribution shifts. These results strengthen the evidence for practical applicability without overstating the current sim-to-real conclusions.

We have clarified the scope and implementation in Sections 3.5 and 4.2, Tables 1–2, and Appendices A.3–A.4, A.5.2, and A.6.1 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Le5Z Part 3

Response to Weakness 3 (Experiments):

We thank the reviewer for this valuable suggestion. We agree that evaluating only OpenVLA-family models on standard LIBERO is insufficient to establish the generality of ReflectVLA, particularly given the near-saturated performance of OpenVLA-OFT. We have therefore substantially extended our evaluation to include the π0.5 backbone and the more challenging LIBERO-Plus robustness benchmark.

(1) Generalization across different VLA architectures.

Following the reviewer's suggestion, we additionally integrate ReflectVLA with π0.5, a flow-based VLA model with a different action-generation architecture from the autoregressive OpenVLA and OpenVLA-OFT. Our framework now supports three backbones covering discrete autoregressive and continuous action-generation paradigms, while keeping the pretrained base policies frozen.

As shown in the revised Table 1, ReflectVLA consistently improves the corresponding Retry baselines across all three backbones and both detector implementations. Specifically, with π0.5, ReflectVLA improves the average success rate from 96.7% to 98.9% (+2.2 pp) under rule-based detection and from 96.5% to 98.7% (+2.2 pp) under VLM-based detection. These results demonstrate that the proposed paired-trajectory conditioning mechanism is not restricted to the OpenVLA family or a single action-generation paradigm.

(2) Addressing performance saturation: Why LIBERO-Plus instead of LIBERO-PRO?

We agree that standard LIBERO is approaching saturation for stronger backbones, making further improvements difficult to assess. We appreciate the suggestion to evaluate on LIBERO-PRO. We chose to prioritize LIBERO-Plus because its robustness evaluation is particularly aligned with the operating scope of ReflectVLA.

LIBERO-Plus introduces controlled distribution shifts across seven dimensions: camera viewpoints, robot initial states, language instructions, lighting, backgrounds, sensor noise, and object layouts. These changes directly stress-test whether the correction mechanism remains effective under variations in observations, robot configurations, and execution conditions, while preserving the underlying task objectives.

In contrast, LIBERO-PRO places additional emphasis on generalization beyond memorized task configurations, including changes to target objects and task definitions. Such settings are complementary and valuable, but some require task-level re-grounding or replanning rather than the event-local corrective re-execution under preserved task semantics studied in this work. Our choice of LIBERO-Plus therefore reflects the specific generalization question we aim to test, rather than a claim that it replaces or is more challenging than LIBERO-PRO.

(3) Substantial improvements under distribution shifts.

The new LIBERO-Plus results directly address the reviewer's concern about marginal improvements on saturated benchmarks.

As shown in Table 2, under non-privileged VLM-based failure detection:

BackboneBaseRetryReflectVLAGain over Retry
OpenVLA-OFT68.9%72.4%82.0%+9.6 pp
π0.574.5%80.5%89.9%+9.4 pp

Notably, for OpenVLA-OFT, the improvement over Retry increases from +0.8–1.0 pp on standard LIBERO to +9.6 pp on LIBERO-Plus. The gains are also consistent across all seven perturbation dimensions. For example, ReflectVLA (OFT) improves performance under robot initial-state perturbations from 43.8% to 66.0% (+22.2 pp), while ReflectVLA (π0.5) improves performance under object-layout perturbations from 75.5% to 90.7% (+15.2 pp).

These results provide substantially stronger evidence that ReflectVLA remains effective when the underlying policies encounter challenging distribution shifts, rather than merely yielding small improvements on near-saturated standard LIBERO tasks.

Overall, the revised experiments demonstrate consistent improvements across distinct VLA families and substantially larger recovery gains under challenging robustness conditions. We acknowledge that LIBERO-Plus and LIBERO-PRO evaluate complementary aspects of generalization, and do not claim to have covered the latter. Nevertheless, the new results directly address the concerns regarding backbone diversity, benchmark saturation, and the limited scope of our original evaluation.

We have incorporated these experiments into Section 4.1, Section 4.2, and Tables 1–2 of the revised manuscript.

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

This paper proposes ReflectVLA, it adds a corrective re-execution pathway to a frozen VLA. The authors collect the failed data are synthesized from LIBERO success demos, the task-specific action modifications is injected into the translational and rotational actions inside an event window [t_pre, t_end]. Only rollouts that fail are kept. A shared-weight Contrastive Trajectory Encoder (CTE) encodes the failed and successful segments: images and state at the three anchors, plus actions compressed into three tokens. For the training, the base model's action loss is used, with the successful demo's actions as targets. Two auxiliary losses are added: a ranking loss and a cosine-margin contrastive loss. And for the inference,

Critical strengths
  1. Retry shares the detector, rollback target and budgets. This is the right control for asking whether conditioning does more than give the policy a second attempt.
  2. The scope is actually honest, no privileged scene reset, failure cases are shown, and generation and rollback parameters are given in full.
Critical weaknesses
  1. It seems that the performance improvement is incremental, if we decompose the table 2, the Full - Retry gives a 5.9% improvement, while failure-prefix-only condition already accounts for around 3.2% improvement, means only 2.7% success-rate gain for pairing.
  2. As I understand Sections 3.3 and 3.4, the training reference is the failed branch's own source demonstration, whose actions are also the training target, so the action loss does not need the failure branch and the model never sees an imperfect reference in training, which may explain why random retrieval brings the success rate down to about the Retry level?
  3. Retry may be a fairly easy baseline, because restarting what is probably a deterministic policy from nearly the same state will often reproduce the failure (only 7.1% recovery), so comparisons with stochastic retry, best-of-N sampling from the rollback state, or methods such as CycleVLA (Ma et al., 2026), and FPC-VLA (Yang et al., 2025) would make the gains more convincing.
  4. The evidence is also somewhat narrow: the synthetic failures are perturbed demonstrations replayed open-loop, with no check that they resemble the policy's own failures, all experiments are in simulation on LIBERO with OpenVLA and its OFT variant, and the OFT gains over Retry are small.

[1] Ma, C., Yang, G., Lu, K., Xu, S., Byrne, B., Trigoni, N., & Markham, A. (2026). CycleVLA: Proactive self-correcting vision-language-action models via subtask backtracking and minimum Bayes risk decoding (arXiv:2601.02295). arXiv. https://doi.org/10.48550/arXiv.2601.02295 [2] Yang, Y., Duan, Z., Xie, T., Cao, F., Shen, P., Song, P., Jin, P., Sun, G., Xu, S., You, Y., & Liu, J. (2025). FPC-VLA: A vision-language-action framework with a supervisor for failure prediction and correction (arXiv:2509.04018). arXiv. https://doi.org/10.48550/arXiv.2509.04018

Questions for authors

Could the authors add more baselines & benchmarks to showcase the method is actually generalized well?

5 comments
AuthorsOpenReview ↗

Response to Reviewer ZHar Part 1

Response to Weakness 1:

We thank the reviewer for this insightful comment. We agree with the numerical decomposition, but would like to distinguish the marginal inference-time benefit of the successful reference from the overall benefit of learning with paired trajectories.

First, the 2.7 pp gain measures an inference-time intervention on an already pair-trained model. In Table 3, Failure-prefix-only removes the successful reference at inference while retaining the CTE and correction adapter trained with paired success–failure trajectories. Therefore, this comparison does not evaluate a model independently trained without successful references. On the same 1,364 true-positive triggers, Full ReflectVLA improves conditional recovery from 21.3% to 33.1%, corresponding to 162 additional recovered episodes. The 2.7 pp figure is obtained by normalizing these additional recoveries by all 6,000 evaluation episodes, rather than by the triggered failures.

Second, we additionally train independent single-trajectory alternatives (Table 4). The Success-only and Failure-only Encoder variants are separately trained using only their respective trajectory branches, with the same action supervision. Under the same single-seed evaluation setting, they achieve 79.15% and 78.40% end-to-end success rates, compared with 82.90% for Full ReflectVLA. Thus, paired conditioning outperforms independently trained failure-only conditioning by 4.50 pp and success-only conditioning by 3.75 pp. Furthermore, retaining both trajectory branches while removing the auxiliary ranking and contrastive losses still achieves 80.45%, indicating that dual-trajectory conditioning remains useful even without explicit auxiliary supervision.

These results address two complementary questions: Table 3 measures the contribution of the successful reference at inference, whereas Table 4 evaluates the benefit of training with both trajectory branches. We do not treat the two gains as additive or directly interchangeable. Together, they provide evidence that the value of paired trajectory conditioning extends beyond the 2.7 pp inference-time difference.

Response to Weakness 2:

We thank the reviewer for raising this important concern. We agree that the successful reference in our synthetic training pairs originates from the failed branch's source demonstration, whose actions also serve as training targets. Therefore, the action loss alone does not exclude a potential reference-action shortcut. Our auxiliary ranking and contrastive objectives encourage success–failure discrimination, but we do not regard them alone as proof against shortcut learning.

Importantly, our independently trained ablations provide direct evidence that the failure branch contributes beyond the successful reference. As shown in Table 4, the Success-only Encoder achieves 79.15%, compared with 82.90% for Full ReflectVLA under the same single-seed setting. Thus, using the failure trajectory provides a 3.75 pp improvement even when the successful reference is already available.

The shared-trigger analysis further argues against a pure action-replay explanation. Under identical true-positive triggers (Table 3), Full ReflectVLA achieves 33.1% conditional recovery, compared with 19.5% for Success-reference-only conditioning and 17.5% for direct retrieved-action replay. Moreover, removing successful-reference action tokens at inference still yields 27.7% recovery. These results indicate that corrective re-execution benefits from the observed failure history and trajectory context, rather than being reducible to direct replay of reference actions.

Regarding Random Retrieval, we agree that reference quality is important. ReflectVLA retrieves successful references using task identity, compatible interaction stages, and initial/pre-event state similarity (Appendix A.2). Random retrieval within the same task can disrupt this alignment, causing unrelated trajectory differences to enter the CTE. Accordingly, Random Retrieval decreases end-to-end success from 82.90% to 76.45% in the same single-seed evaluation (Table 4). This demonstrates sensitivity to reference alignment, but does not by itself distinguish alignment-dependent correction from potential shortcut learning.

Finally, at inference, the successful reference is retrieved from the Success Bank rather than being the exact source demonstration used to generate the current failure, providing evidence of transfer beyond the original synthetic training pairs. Nevertheless, we acknowledge that training currently uses source-matched references and does not explicitly model imperfect reference alignment. Therefore, our claims concern correction with task- and event-aligned retrieved references, rather than robustness to arbitrary reference mismatches.

AuthorsOpenReview ↗

Response to Reviewer ZHar Part 2

Response to Weakness 3:

We thank the reviewer for this important suggestion. We agree that deterministic Retry alone is insufficient to establish the effectiveness of paired-trajectory correction, since restarting a nearly deterministic policy from a similar state may reproduce the same failure. To address this concern, we have added stronger stochastic and multi-candidate retry baselines under a controlled shared-trigger evaluation, together with comparisons to CycleVLA and FPC-VLA.

(1) Controlled comparisons with stronger retry baselines. We introduce two additional baselines (Table 3 and Appendix A.7):

  • Stochastic Retry: Instead of deterministic re-execution, actions are independently sampled from the original OpenVLA policy with temperature T=1.0 and top-p=1.0.
  • Retry + MBR: At each decision step, we sample N=8 candidate actions and select the candidate minimizing the mean pairwise L2 distance in normalized action space. Since OpenVLA predicts one action at a time, the selection horizon is H=1. No privileged success labels or additional environment rollouts are used.

All methods are evaluated on the same 1,364 true-positive failure triggers, sharing the initial intervention point, rollback controller and target, remaining environment-step budget, and maximum of two correction attempts. Importantly, all physically executed actions, including rollback, count toward the same step budget. MBR incurs additional inference computation but no additional environment interactions.

The resulting pooled conditional recovery rates are:

MethodRecovery (%)
Deterministic Retry7.1
Stochastic Retry10.0
Retry + MBR (N=8)15.5
Full ReflectVLA33.1

Stochastic Retry and Retry+MBR improve recovery over deterministic Retry, confirming the reviewer's concern that deterministic re-execution is a relatively weak baseline. Nevertheless, ReflectVLA still outperforms Stochastic Retry and Retry+MBR by 23.1 and 17.6 percentage points, respectively. These matched comparisons demonstrate that the recovery improvement cannot be explained solely by introducing stochasticity or selecting among multiple sampled actions.

(2) Comparisons with CycleVLA and FPC-VLA. Following the reviewer's suggestion, we have additionally included the published results of CycleVLA and FPC-VLA in Table 1, and CycleVLA on LIBERO-Plus in Table 2.

For CycleVLA, although its implementation is publicly available, the official trained checkpoints were unavailable. We attempted to reproduce its reported performance through two training attempts but were unable to match the published results. Rather than reporting an inadequately reproduced baseline, we therefore include the original authors' reported results. For FPC-VLA, we could not obtain a publicly released implementation, so we likewise report its published results.

For context, on standard LIBERO, ReflectVLA with π0.5 achieves 98.9% using rule-based detection and 98.7% using VLM-based detection, compared with 98.5% reported by CycleVLA (π0.5). On LIBERO-Plus, ReflectVLA (π0.5) achieves 89.9%, compared with CycleVLA's reported 86.0%. We also include FPC-VLA's published LIBERO result (86.9%) for completeness. Since these studies differ in backbone configurations and evaluation protocols, we present these published results as contextual references rather than strictly controlled head-to-head comparisons.

We have incorporated these results and clarifications into Tables 1–3, Section 4.2, and Appendix A.7 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer ZHar Part 3

Response to Weakness 4:

We thank the reviewer for highlighting the limited scope of the original evaluation. We have substantially extended our experiments and would like to clarify the motivation behind our synthetic failure construction and, importantly, the distinction between synthetic training failures and naturally occurring evaluation failures.

(1) Synthetic failures are designed to approximate observed natural failure patterns. Our failure synthesis strategy was not chosen arbitrarily. Before constructing the dataset, we extensively inspected naturally occurring execution failures across multiple VLA models, including the OpenVLA family and π-series models. We observed recurring failure patterns such as accumulated end-effector drift, missed grasps, inaccurate placements, contact misalignment, and unsuccessful articulated-object interactions.

These observations motivated our use of temporally correlated Ornstein–Uhlenbeck (OU) noise, which produces gradual action deviations rather than independent random perturbations. We further introduce event-specific, task-conditioned action modifications to simulate interaction-dependent failures, including grasp offsets, placement errors, unsuccessful contact, and incorrect manipulation directions.

(2) Recovery is evaluated on naturally occurring policy failures, not synthetic perturbations. We emphasize that synthetic failures are used for training, whereas evaluation uses failures arising from the autonomous execution of the original base VLA policies, without injecting OU noise or other artificial action perturbations. The correction mechanism is therefore tested on failures generated by the policies themselves, rather than merely on the synthesized failure distribution.

This distinction is important: although our training data are synthesized from perturbed demonstrations, the evaluation directly examines whether the learned correction mechanism transfers to autonomous policy failures. As reported in Appendix A.5.2, ReflectVLA successfully recovers 742 of 1,940 (38.25%) base-policy failures with rule-based detection and 705 of 1,940 (36.34%) with non-privileged VLM-based detection across the evaluated backbone–suite settings. These results provide direct evidence of transfer from synthetic failure training to naturally occurring failures in simulation.

(3) Broader evaluation across backbones and distribution shifts. To address the limited experimental scope, we additionally evaluate ReflectVLA with π0.5, extending beyond the OpenVLA family, and introduce experiments on LIBERO-Plus under diverse distribution shifts.

On standard LIBERO (Table 1), ReflectVLA consistently outperforms the corresponding Retry baselines across OpenVLA, OpenVLA-OFT, and π0.5. In particular, ReflectVLA (π0.5) improves success from 96.7% to 98.9% with rule-based detection and from 96.5% to 98.7% with VLM-based detection.

We acknowledge that the gains with OpenVLA-OFT on standard LIBERO are relatively small (+1.0 pp with rule-based detection and +0.8 pp with VLM-based detection), partly reflecting the high baseline success rates and limited remaining headroom.

To examine performance beyond this near-saturated setting, we further evaluate on LIBERO-Plus (Table 2), which introduces camera, robot-state, language, lighting, background, visual-noise, and layout perturbations. Under VLM-based detection, ReflectVLA improves OpenVLA-OFT from 72.4% to 82.0% (+9.6 pp) and π0.5 from 80.5% to 89.9% (+9.4 pp) over their corresponding Retry baselines. These substantially larger gains demonstrate that the correction mechanism remains beneficial under more challenging distribution shifts, where the base policies have greater room for improvement.

We have clarified the failure construction procedure in Section 3.3 and Appendix A.1, and expanded the evaluation and analysis in Section 4.2, Tables 1–2, and Appendix A.5.2 of the revised manuscript.

Response to Questions:

We have substantially expanded both the baselines and benchmarks to provide stronger evidence of ReflectVLA's generalization.:(1) Stronger baselines. (2) Broader backbone evaluation. (3) Additional robustness benchmark.**

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.