acceptodds
Under review as a conference paper at ICLR 2027

RoboDrop: Curating VLA post-training data via local gradient compatibility

Abstract

Vision–language–action (VLA) models acquire broad generalization through large-scale pretraining, yet adapting them to a new task and robot embodiment still requires post-training on newly collected data. Unlike pretraining, post-training targets task- and embodiment-specific adaptation, making it particularly sensitive to data quality. In practice, collected robot datasets often contain heterogeneous errors, including execution mistakes, sensor drift, and timestamp misalignment, which can impair post-training and policy performance. Manual inspection is costly, while existing data-cleaning methods are typically tailored to particular corruption types. To address these challenges, we introduce RoboDrop, a data-curation framework that audits supervision using local gradient compatibility measured along the training trajectory as a proxy for its effect on post-training performance. During a one-epoch warm-up run, RoboDrop scores each candidate sample online by comparing its gradient with those of task-semantic and visually matched validation samples. The resulting sample scores are aggregated at the episode level, and a simple automatic post-processing rule converts them into filtering decisions. We evaluate RoboDrop on controlled observation–action corruptions, naturally suboptimal demonstrations in simulation, and real-robot datasets containing non-expert collection errors. Across these settings, RoboDrop more accurately distinguishes unreliable demonstrations than prior methods, while post-training on the curated data consistently yields stronger downstream policies, with average real-robot rollout success rising from to . These results establish training-trajectory-aware, context-conditioned supervision auditing as an effective approach to robust VLA post-training.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Data Curation, Robot Learning

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_3My3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes RoboDrop, a data-filtering framework for improving VLA model post-training by filtering unreliable robot demonstrations. The method estimates sample quality through local gradient compatibility with clean validation samples matched by task semantics and visual context. During a one-epoch warm-up, RoboDrop computes sample-level compatibility scores along the training trajectory, aggregates them at the episode level, and filters low-quality demonstrations using fixed-ratio or automatic selection. Experiments on LIBERO, Robomimic-MH, and real-robot tasks demonstrate improved discrimination of unreliable demonstrations and downstream policy performance.

Critical strengths
  1. The paper addresses data quality in VLA post-training, an important challenge for adapting pretrained models to new robotic tasks.
  2. The proposed local gradient compatibility provides an intuitive, model-aware approach to identifying unreliable demonstrations.
  3. The method is evaluated on synthetic corruptions, naturally suboptimal demonstrations, and real-robot tasks, with comparisons against multiple baselines.
Critical weaknesses
  1. In practical settings, how to define or sample reference episodes? The authors mention that they sample demonstrations produced by proficient operators. It's not clear what's the criteria of sampling or just random sampling. Besides, the assumption of proficient operators produces high-quality demonstrations while less-skilled operators produces flawed demonstrations is not rigorous. Less-skilled operators may produce valuable failure recovery trajectories.
  2. The method relies on matching candidate samples to clean validation demonstrations, which may limit the model to fit to the reference distribution in a stronger way. In particular, failure recovery often involves states rarely observed in successful demonstrations. Could RoboDrop incorrectly filter valuable recovery behaviors simply because they are out of distribution relative to the reference set?
  3. How does the method distinguish operation validity and style? Low gradient compatibility does not necessarily indicate incorrect supervision, as different operators may adopt distinct but equally valid action strategies. The current evaluation largely relies on operator proficiency as a proxy for data quality. It remains unclear whether RoboDrop can distinguish genuinely incorrect actions from valid demonstrations with different execution styles.
  4. Although the paper emphasizes fine-grained supervision auditing, sample-level scores are ultimately averaged for episode-level filtering. Moreover, the experiments primarily evaluate episode-level corruption or quality labels, without directly assessing the detection of localized errors within otherwise valid trajectories.
Questions for authors
  1. Has the BIC-based filtering strategy been evaluated on entirely clean datasets (0% contamination)? Can it reliably avoid removing valid demonstrations when no filtering is necessary?
  2. In Robomimic-MH, RoboDrop uses π0.5 for data scoring but X-VLA for downstream post-training. How does this compare with using the same backbone for both scoring and training?
Reviewer_ACPhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a new algorithm RoboDrop, which curates post training datasets for VLAs. Their method uses a small validation set of clean demos to construct local reference gradients based on retrieving the closest reference states over the first warmup epoch of training to identify and filter out bad demonstrations. The paper evaluates on synthetically corrupted sim datasets as well as mixed quality human collected data in both sim and real.

Critical strengths

I found this paper to be well presented, as well as their evaluation tasks and baselines are comprehensive. Their method was clear to follow and distinct from the related works.

Critical weaknesses

RoboDrop’s algorithm relies on a clean validation set of target post training episodes to define a proxy for the subset of the candidate demos that achieve the highest policy success. However, collecting a validation set that has this property seems to me to have almost the same difficulty as collecting and filtering the perfect demonstration set which’s challenges are discussed in the paper. I think additional experiments on the diversity and quality of this validation set and how it affects real success rates would be valuable.

I also did not find the final method clearly motivated, as while there is a first order motivation for using the gradient alignment to measure changes in validation loss, there was no clear justification about why the flow validation loss is a good estimate of the task success rate. I think this concern would be especially relevant when high quality demonstrations have diverse behaviors or when the validation set is not perfectly constructed.

Questions for authors

For the libero results, the AUROC for Scizor is much worse than BR but performed about the same, why is this?

Can you report error bars for the success rate numbers potentially over different validation sets to help evaluate the significance of the performance gains? While there are some ablations of certain design decisions, some sensitivity analysis based on the validation set composition and optimality would help show robustness.

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

This paper proposes RoboDrop, a data curation method for VLA post-training based on local gradient compatibility. It uses task-semantic and visual similarity to retrieve clean reference samples, compares their gradients with those of candidate samples during a one-epoch warm-up run, and aggregates sample scores to filter unreliable demonstrations. An optional BIC-based procedure automatically determines the filtering budget. Experiments on LIBERO, Robomimic-MH, and real-world robot tasks demonstrate improved data quality identification and downstream policy performance.

Critical strengths
  1. Interesting idea of local gradient compatibility. Constructing reference gradients from semantically and visually similar clean samples is a meaningful improvement over global gradient comparison. The ablation experiments support the effectiveness of this data curation design.

  2. Comprehensive empirical validation. The paper evaluates multiple corruption types, VLA backbones, and real-world robot tasks. The improvements in downstream policy success demonstrate practical value beyond data quality classification.

Critical weaknesses
  1. Potential limitations in reference matching. RoboDrop constructs local reference gradients based on task semantics and visual similarity. However, visually similar observations do not necessarily correspond to the same valid actions, especially in manipulation tasks involving different interaction stages or contact states. Moreover, the clean validation set may not cover all valid action modes. In such cases, even high-quality demonstrations could receive low compatibility scores and be incorrectly filtered out. The robustness of RoboDrop to these situations remains unclear.

  2. Trajectory-level averaging may hide critical local errors. Although RoboDrop aims to identify locally unreliable supervision, it averages sample-level compatibility scores to make trajectory-level filtering decisions. A small number of critical errors, such as incorrect grasping actions, may be diluted by many correct steps, allowing unreliable trajectories to remain in the training set.

Questions for authors

Q1. Visual-context ambiguity: Appendix C.1 shows that DINO similarity correlates with gradient alignment on average. How does RoboDrop handle visually similar states requiring different valid actions, such as grasping versus releasing?

Q2. Missing reference coverage: How robust is RoboDrop when the clean reference set lacks certain valid action modes? Could the authors evaluate whether high-quality but underrepresented demonstrations are incorrectly filtered?

Q3. Localized errors: How does mean aggregation perform when only a few critical steps in a trajectory are corrupted? Could the authors compare it against lower-percentile or worst-segment aggregation?

These questions are central to my assessment, particularly Q2 and Q3. Evidence that RoboDrop remains robust under these conditions would meaningfully strengthen the paper and could lead me to revise my rating upward.

1 comment

Thank you for the detailed responses and the additional experiments. Regarding Q1, the additional experiments provide useful evidence that visually similar contexts with different valid actions can still exhibit compatible gradients. The comparison between clean pairs, random pairs, and clean–corrupted pairs supports the authors' argument that gradient compatibility is more informative than action similarity alone. Although the degree of separation varies across tasks, I find the additional analysis helpful in clarifying the behavior of RoboDrop in such scenarios.

Regarding Q2, the Right-only reference experiment is particularly useful, as it demonstrates that RoboDrop does not simply discard all valid action modes absent from the reference set. The retention of 45/54 proficient demonstrations from the two unrepresented modes is encouraging. However, the retention rate for the Left mode (10/17) remains noticeably lower than that for the Front mode (35/37), suggesting that the method may still be sensitive to substantial reference bias. Nevertheless, the similar downstream rollout success across the three configurations provides additional evidence of robustness. I encourage the authors to acknowledge this remaining sensitivity as a limitation in the final version.

Regarding Q3, the additional experiments effectively address my concern. The localized-corruption results show that worst-window aggregation can improve sensitivity to localized errors, while the results on naturally collected demonstrations support full-trajectory mean as a more robust default when corruption patterns are unknown. I think the results are reasonable and consider this concern largely addressed.

Overall, I appreciate the additional analyses and clarifications. The authors have addressed most of my concerns, and I lean toward Weak Accept.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.