harpreetsahota/lerobot-curation-metrics-figures

LeRobot curation metrics: blog figures and plugin screenshots

The ten figures from A Practical Guide to LeRobot Dataset Quality Metrics (a Hugging Face article) and its longer companion post, You Can't Watch 10,000 Episodes. Measure These Instead. Both are guides to the metrics worth computing on a LeRobot v3 dataset before you train a policy or fine-tune a VLA.

All signals are synthetic and illustrative. Every number printed on a figure is computed from its synthetic signal with the definition the post describes, except figure 10, whose per-metric z-scores are made-up example values.

File	What it shows
01_integrity.png	A v3 episode is a row range in a shared Parquet file plus a time window in a shared MP4 per camera, and the integrity checks that catch a pointer that stopped matching the data
02_sparc_ldlj.png	The same two reaches moved smooth, stop-and-go and jittery, with SPARC and LDLJ values
03_time_efficiency.png	One episode's arm speed with its idle lead, mid-episode pause and idle trail, and length_z against a task group
04_tracking.png	Action versus state for one joint with a stall, the residual after lag alignment, and the cross-correlation that estimates lag
05_gripper.png	A clean grasp, chatter and a regrasp recovery, with flips per second
06_action_divergence.png	Neighbouring episodes' actions from a similar state, agreeing and contradicting each other
07_camera_frames.png	Reference, blur, exposure, clipping and frozen-feed frames with their metric values
08_video_action_lag.png	Arm speed and video motion energy offset in time, the cross-correlation peak, and the lag across episodes
09_normalization.png	Pooled versus per-task warn lines on the same two tasks
10_scoring.png	Group values for an example episode, with the camera group setting the score

Plugin screenshots

Screenshots of the LeRobot curation panel and its compute form in the FiftyOne App, in the panel/ folder. They are real (not synthetic): 23 episodes of a bimanual robot with dexterous hands (dexmate_vega1_and_sharpa_wave), scored with a dual-arm layout, hands on, and three cameras. The "low-confidence" banner is expected here: each of the 23 episodes has its own task string, so none of them reaches the 20-episode group size and all are scored against the whole view.

File	What it shows
panel/01_overview.png	Overview: score histogram with warn and fail lines, verdicts, episodes per task, outlier scatter, and the worst-first ranking with Score, Flags and Driver
panel/02_motion_action.png	Motion & Action: one histogram per metric, with arm and hand signal chips
panel/03_vision.png	Vision: the five camera metrics with camera chips, and a worst-first table
panel/04_integrity.png	Integrity & Coverage: verdicts, episodes per task, per-episode integrity
panel/05_inspector.png	Inspector: score, driver, verdicts, flagged spans, every metric with its per-arm breakdown, and joint traces
panel/06_language.png	Language: task-string checks
panel/07_form_data_top.png	Compute form, Data tab: arms, arrays, and the joint-positions question
panel/08_form_data_joints.png	Compute form, Data tab: joints as chips, and the list of what will be scored
panel/09_form_metrics.png	Compute form, Metrics tab: collapsible families with counts
panel/10_form_camera.png	Compute form, Camera tab: pick cameras, the camera metrics, and a decode-time estimate

Gallery

Citing and crediting these figures

If you use any of these figures in a post, paper, slide deck, course or video, please credit and cite them. A caption or a slide footnote is enough:

Figure from Harpreet Sahota, "A Practical Guide to LeRobot Dataset Quality Metrics" (2026). https://huggingface.co/datasets/harpreetsahota/lerobot-curation-metrics-figures

BibTeX:

@misc{sahota2026lerobotcurationfigures,
  author       = {Sahota, Harpreet},
  title        = {A Practical Guide to LeRobot Dataset Quality Metrics: figures},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/harpreetsahota/lerobot-curation-metrics-figures}},
  note         = {Hugging Face dataset}
}

Please keep the "Synthetic, illustrative signals" footnote that appears on the figures, so the values are not mistaken for measurements on a real dataset.

License

These figures, GIFs and screenshots are released under CC BY 4.0. You may share and adapt them, including commercially, as long as you give appropriate credit (author, source link, and a note if you changed them) as shown in the citation block above.