RoboDrop: Curating VLA Post-Training Data via Local Gradient Compatibility
【Paper Quick Read】Detecting 'Bad Demonstrations' in Robot Learning via Gradients: RoboDrop for VLA Data Curation
RoboDrop is a data curation framework for Vision-Language-Action (VLA) models that evaluates data quality by comparing gradient directions between candidate samples and clean validation data. It improves success rates on real robots from 35.0% to 67.5% by automatically excluding corrupted episodes. The method uses local gradient compatibility to identify and remove suboptimal demonstrations, achieving high performance in both simulated and real-world tasks.
RoboDrop introduces a gradient-based method for auditing VLA data during post-training, improving success rates on real robots from 35.0% to 67.5% by filtering out corrupted episodes. The approach compares gradient directions between candidate samples and clean validation data, avoiding reliance on fixed exclusion budgets.
Source: Robotics — Robot learning news discovery · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported success rate
35%
View original evidence
each training sample points in the same direction as the gradient of a clean validation sample in a similar task state. It evaluates data quality online during a one-epoch warm-up and automatically excludes bad episodes. On a real robot, the average success rate improved from 35.0% before curation to 67.5% with automatic budget selection. Title: RoboDrop: Curating VLA Post-Training Data via Local Gradient Compatibili
Open source S3Reported success rate
67.5%
View original evidence
each training sample points in the same direction as the gradient of a clean validation sample in a similar task state. It evaluates data quality online during a one-epoch warm-up and automatically excludes bad episodes. On a real robot, the average success rate improved from 35.0% before curation to 67.5% with automatic budget selection. Title: RoboDrop: Curating VLA Post-Training Data via Local Gradient Compatibili
Open source S3Reported success rate
84%
View original evidence
ts: Improved corruption detection and policy performance Figure 5: Examples of suboptimal trajectories detected by RoboDrop For temporal corruption in LIBERO, while the unfiltered average success rate was 84.0% and 83.4% with 10% random exclusion, RoboDrop achieved 91.6%. For action corruption, it improved from 91.3% and 92.0% to 95.1%. The temporal corruption detection AUROC for LIBERO-10 was 96.1%, surpassing the s
Open source S12Reported success rate
83.4%
View original evidence
ts: Improved corruption detection and policy performance Figure 5: Examples of suboptimal trajectories detected by RoboDrop For temporal corruption in LIBERO, while the unfiltered average success rate was 84.0% and 83.4% with 10% random exclusion, RoboDrop achieved 91.6%. For action corruption, it improved from 91.3% and 92.0% to 95.1%. The temporal corruption detection AUROC for LIBERO-10 was 96.1%, surpassing the s
Open source S12
- paper: Available Artifact link Open source S4
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
each training sample points in the same direction as the gradient of a clean validation sample in a similar task state. It evaluates data quality online during a one-epoch warm-up and automatically excludes bad episodes. On a real robot, the average success rate improved from 35.0% before curation to 67.5% with automatic budget selection. Title: RoboDrop: Curating VLA Post-Training Data via Local Gradient Compatibili
Open source S3
ts: Improved corruption detection and policy performance Figure 5: Examples of suboptimal trajectories detected by RoboDrop For temporal corruption in LIBERO, while the unfiltered average success rate was 84.0% and 83.4% with 10% random exclusion, RoboDrop achieved 91.6%. For action corruption, it improved from 91.3% and 92.0% to 95.1%. The temporal corruption detection AUROC for LIBERO-10 was 96.1%, surpassing the s
Open source S12
ty URL: https://arxiv.org/abs/2609.10021 Affiliation: Tsinghua University, Striding AI Authors: Runze Xu, Yuanfan Xu, Cuijie Xu, Shuang Dai, Yining Li, Yu Wang, Jincheng Yu Figure 1: RoboDrop auditing VLA data using local gradient compatibility Background: Even successful episodes contain local errors Robot demonstrations may contain unnecessary movements or misalignments in observations and actions, even if the task
Open source S4
Source：Robotics — Robot learning news discovery · note.com