reverse_lerobot_v2

日本語版: README.ja.md

LeRobot-format demonstrations for "reverse-direction" tasks — taking an object out rather than putting it in. Every episode was produced by a scripted policy inside the simulator; none of it is a subset of distributed data.

Episodes	397
Frames	61,430
Tasks	3
Format	LeRobot v3.0, 20 Hz

What it is for

PARC2026's Track 3 asked for the reverse of the training tasks (organizer disclosure, 2026-08-28). Where the forward data says "put the soup can in the tray", Track 3 asks "take the can out of the tray and put it on the table".

Forward-trained weights scored 0 of 8 on every Track 3 submission. This dataset exists to close that gap.

Task	What it does
Yellow book	take it off the shelf
Alphabet-soup can	take it out of the tray
Red mug	take it off the left plate

All 397 instruction strings are byte-identical to the Track 3 bddl text. Track 3's fourth task — the two mugs — has no episodes here.

How it was generated

These are not forward demonstrations played backwards. They are episodes recorded by stepping the simulator normally, forward in time.

Time reversal is used only to produce the reference path for the carry segment — after the object is grasped. The approach-and-grasp segment is synthesized, because a forward demonstration contains only "how to approach before grasping" and says nothing about where to put the object once it is out.

1. gripper_shift_sweep.py sweeps every (forward demo × distributed initial state) pair. The best combination was --mode phased --init-source t3 --grasp-frame clamped --standoff 0.06
2. A dry run first, then a min-cost assignment (scipy.optimize.linear_sum_assignment), so each of the 50 distributed initial-state rows is used once, with the best-quality demo
3. Episodes that reached the reverse goal are written as npz and converted with npz_to_lerobot.py

What changed from v1

Three changes from v1 (reverse_lerobot_389, 389 episodes, 67,775 frames):

• Rotation limit applied by scaling the whole rotation vector, not by clipping per component (ROT_LIMIT=0.375)
• Dropped the book task's horizontal-levelling motion
• Settle for 80 frames after the lift

	v1	v2
Frames with |z| > 10	15.2 %	0.0 %
Max |z|	26.1	9.8
Drop	40 mm / 37 mm	1.0 mm / 2.9 mm
Red mug	139	147
Total	389	397
Median length	173	152

Known limitations

• It damages forward tasks that share objects. Training with 49.5 % reverse data dropped the object-sharing rows by 22 to 25 points, while the rest stayed flat. The mixing ratio is a trade, not a free addition.
• The Track 3 submission score stayed 0.000 throughout. Local evaluation became non-zero for the first time with this lineage (0.000 → 0.0235), but that never reached the leaderboard.
• No demonstrations for the two-mugs task.
• The arm's initial pose is identical across every episode.
• Colour-only disambiguation is not learned.

Mixing

siruku6/jaist_mix_v2 (6,176 episodes) is libero_plus thinned to 75 episodes per task (3,000) plus this dataset duplicated eight times (397 × 8 = 3,176). The reverse share is 49.5 %.

COPIES=8 PER_TASK=75 REV_DIR=work/artifacts/reverse_lerobot_v2 \
MIX=jaist_mix_v2 bash solution/train/jaist_dataset_build.sh

Normalization statistics are frozen to the forward side's (difference 0.000e+00).

3,176 of that mixture's episodes are eight copies of these 397. Stored as the component it is one-eighth the size.

License

Built from LIBERO (MIT, Copyright (c) 2023 Lifelong Robot Learning). The scripted demonstrations here were produced in the same simulator and are released on the same terms.