libero_90_clean

日本語版: README.ja.md

LIBERO's libero_90 with demonstrations that bump into non-target objects removed. The last 250 episodes are re-recorded — that is the part of this dataset that cannot be reproduced from anything else.

Episodes	3,911
Frames	560,254
Tasks	73
Format	LeRobot v3.0, 20 Hz

Why it exists

Contact cleaning had been skipped for the libero_90 portion. The 40 LIBERO-plus tasks were filtered on contact, but libero_90 had only the 10 over-300-frame episodes dropped — because the original HDF5 needed to score contact (66.7 GB) was not available locally.

That was finished on 2026-09-16. The result is this dataset, which serves as the replacement pool for libero_combined_bowl5_initpose_l90clean.

What counted as dirty

Touching or displacing a non-target object, scored from the raw MuJoCo states in LIBERO's own libero_90 HDF5 (no replay).

Reference frame	10 (chosen from 0, 5, 10, 20)
Displacement	L1 sum over the three axes
Threshold	1.0 mm (same as the grader)
Non-target	observation <name>_pos that is not robot0*, not _to_robot0_eef_pos, and not in the bddl :obj_of_interest

Across 4,500 demonstrations: 296 displaced something by more than 1 mm, 296 touched something, 363 (8.1 %) did either. Of the 3,911 libero_90 episodes in the mixture, 250 were judged dirty.

Composition

Measured, not assumed — sha256 over the six arm dimensions of the action sequence, 2026-09-19.

Part	Episodes	Origin
Head (indices 0–3660)	3,661	Straight from the distributed data — all 3,661 match libero_combined_20hz down to the arm motion
Tail (indices 3661–3910)	250	223 of them appear nowhere in the distributed data — they are re-recorded. The remaining 27 do match it

This is why the dataset cannot be deleted. Those 223 would require re-running the scripted policy to reproduce, and would not come out identical.

Gripper values are normalized to ±1 — left at the distributed {0,1}, "open" carries the opposite label.

How it is used

As the refill pool for the mixture. The 250 dirty episodes are dropped from the parent's libero_90 range (indices 17731–21641) and this dataset's last 250 are appended at the end. Per-task episode counts are unchanged; only the frame count drops (−5,803).

Provenance was established by sha256 of the action sequences, not by name or length — (task, length) does not identify the 250 uniquely.

The full rebuild bundle is in libero_plus_bowl5_initpose's rebuild/. It was run on a different machine on 2026-09-18 and all 21,642 action sequences matched.

Known limitations

• The procedure that produced the 250 re-recorded episodes is not a script. It was done by hand on 2026-09-16. The ledger of which episode went where survives (l90clean_replace.json), but the steps to generate those 250 again do not.
• libero_90 has no initial-pose-varied and no appearance-varied demonstrations. Those exist only on the LIBERO-plus side, so demonstration quality is not uniform across tasks.
• An appearance-layered sibling exists, libero_90_clean_appearance, but the refill used this one (the un-layered side).

License

Built from LIBERO (MIT, Copyright (c) 2023 Lifelong Robot Learning). The re-recorded demonstrations are released on the same terms.