PaperlayerView exact revision on arXiv

Source-linked AI summary

RoboDrop: Curating VLA Post-Training Data via Local Gradient Compatibility

Runze Xu, Yuanfan Xu, Cuijie Xu, Shuang Dai, Yining Li, Yu Wang, Jincheng Yu

arXiv:2609.10021v1cs.RO

TL;DR

RoboDrop tackles the difficulty of curating noisy, task-specific robot demonstrations for VLA post-training. It audits samples with local gradient compatibility against task- and visually matched validation data during training, then filters episodes; across simulated and real-robot settings, curated data improves policy outcomes, including real-robot success from 35.0% to 67.5%.

  • Problem

    Task-specific VLA post-training data can contain execution mistakes, sensor drift, and temporal misalignment, making reliable supervision difficult to identify manually or with corruption-specific cleaning methods.

  • Method

    RoboDrop scores candidate samples online during a one-epoch warm-up using gradient compatibility with task-semantic and visually matched validation samples, aggregates scores by episode, and filters the data.

  • Results

    Across controlled corruptions, simulated suboptimal demonstrations, and real-robot errors, RoboDrop more accurately distinguishes unreliable demonstrations and improves downstream policies, with real-robot rollout success rising from 35.0% to 67.5%.

  • Takeaways & Limitations

    Training-trajectory-aware, context-conditioned supervision auditing is an effective approach to robust VLA post-training data curation.

  • Takeaways & Limitations

    RoboDrop relies on a clean reference set whose collection or verification may still require expert effort.

Abstract

from arXiv · show

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 \textsc{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 $35.0\%$ to $67.5\%$. These results establish training-trajectory-aware, context-conditioned supervision auditing as an effective approach to robust VLA post-training.

1 INTRODUCTION

RoboDrop addresses unreliable supervision in task-specific VLA post-training by auditing observation–action pairs with context-conditioned, trajectory-aware gradient compatibility. Across controlled corruptions, simulated suboptimal demonstrations, and real-robot errors, it improves data discrimination and downstream policy success.

  • The framework targets localized observation–action errors that trajectory-level outcomes and retrieval similarity cannot reliably identify within otherwise relevant demonstrations.Its goal is a fine-grained, context-conditioned assessment of each pair’s contribution to post-training.
  • RoboDrop scores candidate samples online by comparing their gradients with task-semantic and visually matched validation samples during a one-epoch warm-up.Scores are aggregated at the episode level for filtering.
  • RoboDrop consistently improves episode-level discrimination and downstream policy performance across controlled corruptions, simulated suboptimal demonstrations, and real-robot collection errors.On real-robot data, post-training success rises from 35.0% to 67.5%.
  • RoboDrop evaluates three imperfect-supervision sources: controlled observation–action corruptions, naturally suboptimal simulated demonstrations, and non-expert real-robot data.

2 RELATED WORK

Prior robot-data curation methods operate at trajectory, retrieval, transition, or policy-utility levels, while gradient-based attribution offers a tractable alternative to expensive retraining-based valuation. RoboDrop is positioned against these approaches as a more localized and training-aware auditing method.

  • Prior methods select robot data using trajectory quality, retrieval similarity, task progress, redundancy, or estimated utility for a target policy.
  • QoQ applies gradient-based influence at a fixed checkpoint, while RoboDrop’s motivation is to account for local and evolving supervision quality.
  • Data Shapley quantifies expected marginal utility but typically requires repeated retraining, whereas influence and gradient estimators provide tractable approximations.

3 PRELIMINARIES

The paper formulates VLA post-training as flow matching over task-specific candidate episodes, where demonstrations may contain locally unreliable observation–action supervision. Its curation objective is to retain episodes that support downstream policy success, evaluated through clean target-task data.

  • The candidate collection consists of task-specific episodes whose samples contain multimodal observations, language instructions, and action chunks.
  • Flow matching interpolates Gaussian noise with demonstrated actions and trains a velocity field conditioned on the observation, instruction, and flow time.
  • The resulting parameters are obtained by minimizing the flow-matching objective from the pretrained initialization.
  • The ideal curation objective is selecting episodes that maximize downstream task success despite locally unreliable supervision from teleoperation or sensing errors.

4 ROBODROP

RoboDrop curates VLA post-training data by measuring each candidate sample’s gradient compatibility with task- and visually matched clean validation behavior along a warm-up trajectory. It aggregates these scores by episode and filters episodes with weak compatibility before fresh post-training.

  • 4 ROBODROP: RoboDrop’s overview combines online sample scoring, episode-level aggregation, and automatic filtering before final post-training on retained candidate episodes and clean reference data.The pipeline uses a one-epoch warm-up policy only for auditing; the final policy is trained from a fresh pretrained copy.
  • 4.2 FROM VALIDATION LOSS TO GRADIENT COMPATIBILITY: A differentiable gradient-alignment proxy estimates whether a candidate update improves clean validation behavior without requiring environment interaction or rollout evaluation.Positive alignment predicts reduced clean validation loss, while normalization makes the score directional rather than magnitude-based.
  • 4.3 TASK-SEMANTIC AND VISUAL-CONTEXT NEIGHBORHOOD REFERENCE: RoboDrop replaces a global validation reference with task-semantic and visually matched local gradients, avoiding incompatible mixtures across manipulation phases and tasks.The method retrieves semantically compatible validation samples, selects visually similar neighbors, and aggregates their gradients with similarity-based weights.
  • 4.4 ONLINE IN-RUN SCORING AND EPISODE-LEVEL AGGREGATION: RoboDrop scores each sample once at its encountered optimization step during a shuffled one-epoch warm-up, using cached local validation references that refresh periodically.Episode-level averages therefore capture compatibility at multiple points along the evolving training trajectory rather than at a fixed checkpoint.
  • 4.5 OPTIONAL AUTOMATIC FILTERING: Episode scores are converted into filtering decisions by identifying a lower-mean Gaussian component as unreliable when a two-component model is supported.The posterior-confidence threshold controls the retention–purity trade-off, after which the warm-up policy is discarded and a fresh pretrained policy is post-trained on retained episodes.

5 EXPERIMENTS

RoboDrop is evaluated on synthetic corruptions, naturally suboptimal demonstrations, multiple VLA backbones, and real-robot collection errors. Across these settings, it improves episode discrimination and downstream post-training performance over competing selection methods and unfiltered or randomly filtered data.

  • 5.2 CONTROLLED CORRUPTIONS ON LIBERO: RoboDrop raises LIBERO rollout success to 91.6% under temporal corruption and 95.1% under action corruption, exceeding unfiltered and random-removal baselines.Under temporal corruption, success rises from 84.0% unfiltered and 83.4% with random removal; under action corruption, from 91.3% and 92.0%.
  • 5.2 CONTROLLED CORRUPTIONS ON LIBERO: On LIBERO-10, RoboDrop reaches 96.1% AUROC under temporal corruption versus 81.8% for the strongest baseline and achieves the strongest average performance across both corruptions.DataMIL and QoQ perfectly separate controlled action corruption but deteriorate under temporal misalignment.
  • 5.3 NATURALLY VARYING DEMONSTRATIONS AND BACKBONE GENERALITY: Across Robomimic-MH tasks, RoboDrop achieves the highest mean AUROC and balanced accuracy, exceeding the strongest baseline by 4.25 and 4.02 points.Episode length remains competitive, but local gradient compatibility provides the strongest overall discrimination.
  • 5.3 NATURALLY VARYING DEMONSTRATIONS AND BACKBONE GENERALITY: RoboDrop maintains mean AUROC above 85% across π0.5-Base, GR00T, and X-VLA, while a weaker Florence-initialized X-VLA still reaches 80.12% AUROC.The pretrained X-VLA improves average AUROC and balanced accuracy by 5.23 and 3.14 points over the weaker initialization.
  • 5.5 ABLATION STUDIES: Ablations show that local online gradient references outperform full-validation, fixed-checkpoint, visual-similarity, and unnormalized-gradient alternatives.Fixed-checkpoint scoring reaches 83.50 AUROC, while DINO and DINO + Action reach 75.86 and 79.73 AUROC; unnormalized gradients reduce AUROC and balanced accuracy by 5.89 and 6.22 points.

6 CONCLUSION

RoboDrop audits VLA post-training data with context-conditioned local gradient compatibility along the warm-up trajectory and improves policy success across diverse imperfect-supervision settings. Its remaining limitation is reliance on a clean reference set that may require expert effort to collect or verify.

  • RoboDrop audits supervision through context-conditioned local gradient compatibility evaluated along the warm-up trajectory, improving data-quality discrimination and downstream policy success across controlled, simulated, and real-robot settings.The framework is presented as a practical tool for VLA data curation.
  • A clean reference set remains necessary, and collecting or verifying it may still require expert effort.Reducing this dependence is identified as an important direction for future work.
  • LIBERO contains four manipulation suites with ten tasks each, covering spatial, object, goal, and long-horizon generalization.The controlled benchmark includes separate task suites for distinct manipulation capabilities.

A.3 REAL-WORLD DATASET

The real-world dataset contains four ARX-arm manipulation tasks with proficient and non-expert teleoperation episodes, while Figure 4 depicts their representative execution sequences.

  • The real-world dataset has four tasks with 100 teleoperated episodes each, including 80 proficient and 20 non-expert demonstrations per task.Ten proficient episodes per task are reserved for clean validation, leaving 90 candidate episodes.
  • Figure 4 presents the principal execution stages for each of the four real-world manipulation tasks.The figure provides representative task sequences rather than quantitative results.

B.1 DETAILED LIBERO RESULTS

On LIBERO, RoboDrop maintains strong discrimination across all controlled-corruption settings and improves post-training success, with task- and backbone-specific variation among methods.

  • 93.38% average rollout success follows removal of the lowest-scoring 10% of episodes, versus 87.63% unfiltered and 87.69% random removal.The gain is +7.63 points under temporal corruption and +3.88 points under action corruption; the oracle reaches 95.69%.
  • 96.20% average AUROC and 93.13% best balanced accuracy show strong RoboDrop separability across all eight LIBERO suite–corruption settings.The result covers four suites and both action and temporal corruption types.
  • RoboDrop achieves the best AUROC on CAN and SQUARE, stays near the strongest method on LIFT and TRANSPORT, and has the best balanced accuracy on three tasks.Baseline leaders vary by task: Scizor leads LIFT and QoQ leads TRANSPORT.
  • Compatibility transfers across architectures despite task variation: π0.5-Base leads CAN and SQUARE, GR00T is competitive on LIFT, and X-VLA leads TRANSPORT.The Florence-initialized X-VLA remains above chance but is weaker on average.

B.3 DETAILED REAL-WORLD FILTERING RESULTS

RoboDrop achieves consistently strong real-world non-expert identification and adapts its automatic filtering to unknown contamination rates, while qualitative examples span diverse execution failures.

  • RoboDrop obtains the strongest AUROC and best balanced accuracy on every real-world task, whereas baseline strength varies by task.This indicates no single trajectory-level heuristic consistently captures non-expert behavior.
  • Across 10%–40% contamination, BIC-selected filtering reaches macro-average F1 scores of 85.6%–93.0%, with precision above 84.0% and recall above 88.2%.The same posterior confidence threshold q = 0.8 is used without tuning to the contamination rate.
  • Without knowing contamination rates, automatic filtering removes 17–21 episodes per task and achieves 93.78% precision, 90.00% recall, and 91.70% F1.The procedure maintains strong accuracy without a fixed removal budget.
  • Figure 5 compares episode-level cosine-score distributions for worse corrupted or non-proficient demonstrations against better clean or proficient demonstrations.The violin plots include score distributions and individual episode scores.
  • Figure 6 highlights failures including overturned cups, misaligned placement, unstable handovers, ineffective drawer contact, and prolonged low-progress motion.These examples illustrate diverse low-quality trajectories identified by RoboDrop.

C ABLATION STUDIES

The ablations examine neighborhood size, gradient parameters, CountSketch compression, refresh frequency, and visual retrieval for RoboDrop. Results support the default design choices and show that DINO-nearest frames provide more aligned gradient directions than random frames.

  • Visual neighborhood and gradient parameters: 88.33 average AUROC with action-expert gradients exceeds 87.02 with full-VLA gradients, while Kvis = 10 exceeds Kvis = 1 at 88.33 versus 86.54.The larger visual neighborhood is more stable across tasks, although Kvis = 1 is strongest on TRANSPORT.
  • CountSketch dimension: 88.33% mean AUROC and 81.83% balanced accuracy make 4096 CountSketch dimensions the strongest tested compression setting.Performance remains stable across compression levels, while 4096 dimensions provide substantial compression without sacrificing the episode-ranking signal.
  • Validation-gradient refresh interval: M = 200 achieves the best AUROC and balanced accuracy, whereas intervals beyond 400 steps generally reduce performance.Refreshing every 100 steps is comparable but does not improve over M = 200, supporting the default refresh interval.
  • Visual retrieval validation: DINO-nearest frames have substantially higher cross-episode gradient-direction similarity than randomly sampled frames.Alignment decreases as progressively less similar frames enter the neighborhood, while the random baseline remains low.

D TRAINING DETAILS

RoboDrop uses a one-epoch warm-up to score candidate samples, then discards that warm-up policy and post-trains a fresh copy on automatically retained episodes. The reported setup includes specified compute times across simulated and real-robot datasets.

  • Warm-up Training: Each candidate is scored exactly once during a one-epoch warm-up, using Kvis = 10 cached validation gradients refreshed every M = 200 steps.Candidates are shuffled before warm-up, and scoring occurs when each sample enters a minibatch.
  • Compute: A one-epoch warm-up takes approximately 2 hours on three LIBERO tasks, 4 hours on LIBERO-10, 5.4 hours across Robomimic, and 4.5 hours across real-robot tasks.These timings use NVIDIA A100 GPUs with a global batch size of 32.
  • Post-training: The post-processing rule retains all episodes under a single-Gaussian score model, or removes episodes assigned to the lower-mean component with posterior probability above q = 0.8.After filtering, the warm-up policy is discarded and a fresh copy of the pretrained policy is post-trained on retained episodes.
Loading 2609.10021v1…