Datasets:
Behavior-layer audit of xarm_lift / xarm_push / jaco_play — same-platform idle spread
Sharing an objective, freshly re-run audit of three datasets in the lerobot org (tool: RDA v0.5.4, default thresholds, offline; fresh HF download on 2026-08-25). All integrity layers (NaN/Inf, timestamps, missing frames, schema) were clean; these are behavior-layer signals only.
xarm_lift_medium(800 ep): PASS 767 / REVIEW 33 / 0 · spikes 6 · median idle 20.8%xarm_push_medium(800 ep): PASS 238 / REVIEW 562 / 0 · spikes 845 (500 ep) · median idle 83.3%jaco_play(1,085 ep): PASS 390 / REVIEW 695 / 0 · spikes 11,958 (837 ep) · median idle 74.1%
Takeaway worth your attention: if you train a policy on a ~75–83% idle distribution without weighting/curriculum, the loss is structurally biased toward predicting "do nothing." For anyone combining ALOHA/xArm-style arms, this also shows the same xArm platform spans 20.8%–83.3% idle purely from task type — a useful calibration point, not a quality judgment.
For data owners: if the high idle in xarm_push_medium reflects the genuine contact/hold rhythm of pushing, that's expected and worth documenting in the card so downstream users don't over-trim.
Repro: pip install robot-data-audit && rda audit <path>