LeBoard scans a LeRobot dataset and finds its defects; harm-v0, our harm model, predicts how much they will lower the success of the policy trained on it, with an honest margin of error.
Every dataset below was scanned by LeBoard v2. The prediction comes from harm-v0, our harm model: effect sizes from the robot-learning literature, applied to what the scan measured.
| Defect | In the data | Predicted cost (points) |
|---|
robomimic recorded the same task with worse, okay and better operators and published how well a robot trained on each did. harm-v0 never saw these numbers: its predictions come from LeBoard's scan of the demonstrations and the literature priors alone.
| Check on unseen dataset families | Error |
|---|
Error: RMSE of the predicted differences between datasets of the same task, in log-odds, when each family of datasets is held out. Lower is better.
Each pair averages robomimic's 3 policy types. The 18 results also include mixed-operator subsets, all inside the predicted range.
Counted on the Hugging Face Hub, October 2026.
harm-v0 limits, said plainly: effect sizes come from the literature and are checked on 7 families of datasets, not yet learned beyond them; defects LeBoard does not measure (failed or inconsistent demonstrations) are not in the prediction; the static-action effect was calibrated with our own LIBERO-10 run, so that example is a calibration point, not a proof; the direction of the operator-skill effects (shorter, smoother demonstrations help) comes from papers that include robomimic's own, but none of robomimic's success numbers was used.