Team-RAMEN/ramen-g1-ikea-assembly-curation

RAMEN G1 IKEA Assembly — segment curation labels

Per-segment quality labels for all 533 episodes of BitRobot/G1_WBT_Dex1_Building-Children-Table (revision 4b5961c6f8b97ececa75ee73041b30bcbf463ae2), made by Team RAMEN for the IROS 2026 Humanoid IKEA Assembly Challenge (Unitree G1 + Dex1-1 assembling a children's table).

Each label marks a frame range of one episode, says which skill the robot performs there, and grades how well it is done. We used the labels to cut the episodes into clean per-skill training sets. This repository holds labels only, no images or robot data.

Files

file	content
labels.parquet	The final labels: 15,928 segments, one row each
labels.jsonl	The edit log behind them (16,958 lines). A later line with the same key replaces an earlier one; a line with "__delete__": true removes the segment. Folding the log this way gives exactly labels.parquet
failure_categories.json	The failure categories (keys and their original Japanese labels)

Columns

The key of a segment is (episode_id, frame_start).

column	content
episode_id	Episode index in the BitRobot dataset (0–532)
frame_start, frame_end	Frame indices within the episode at 30 fps; both ends included
task_index	Skill, using the BitRobot dataset's task indices (below)
verdict	optimal, success, ambiguous or failure (below)
failure_category	For failure only (below); empty otherwise
reviewer	Who labelled it (one reviewer for all segments)
reviewed_at	When it was labelled (ISO 8601, 2026-07-19 to 2026-08-01)
schema_version	4

Skills and verdicts

task_index	BitRobot task	optimal	success	ambiguous	failure
0	insert table leg to table base	1,664	466	576	725
1	move to table	452	53	108	898
2	flip table	5	945	26	218
3	rotate leg to tighten	439	291	1,331	1,726
4	pick table leg	786	433	1,083	220
5	rotate table base	1,430	61	81	1,390
6	building children table	0	0	2	36
7	move table base	387	34	28	34
	total	5,163	2,283	3,235	5,247

• optimal: the skill is done cleanly.
• success: the skill is done, but not cleanly (for example it needed two adjustments, or the result is slightly off).
• ambiguous: hard to judge.
• failure: the skill fails or the segment is unusable.

Failure categories (failure segments only):

key	original label	meaning	segments
unknown	フレーム落ちかつありえない変な行動	dropped frames together with implausible, odd motion	4,133
fail_insert	挿入失敗	the insertion fails	616
hold_stack	固まった	the robot freezes	322
lazy_buffer	いらない行動	unnecessary motion	125
side_flip	横型のまま回す	the table is turned while still on its side	25
fail_env	環境が違う	a different setup	22
slip	滑り落ち	the part slips out of the hand	4

How we used it

We built one training set per skill from these labels with data/bitrobot_lerobot_subtask_datasets/scripts/build_curated_task_dataset.py in the code repository:

skill (task_index)	segments used	episodes	dataset
insert_table_leg (0)	optimal	1,664	ramen-g1-ikea-assembly-ramen-ori-insert-table-leg
flip_table (2)	success, only the last one in each episode (earlier ones are practice or a turn of the table)	429	ramen-g1-ikea-assembly-ramen-ori-flip-table
rotate_leg_to_tighten (3)	success (tight enough without over-tightening, and short)	290	ramen-g1-ikea-assembly-ramen-ori-rotate-leg-to-tighten
pick_table_leg (4)	optimal	786	ramen-g1-ikea-assembly-ramen-ori-pick-table-leg
rotate_table_base (5)	optimal, only after the first pick (or the first insert / tighten when there is no pick)	1,401	ramen-g1-ikea-assembly-ramen-ori-rotate-table-base
move_table_base (7)	optimal, only before the first pick	384	ramen-g1-ikea-assembly-ramen-ori-move-table-base

Each segment becomes one episode of the training set. Where the counts are a little lower than in the table above, the order and overlap filters (see Known issues) dropped segments. We did not use move to table (1; the robot walks with the SDK instead) or building children table (6).

Known issues

• One reviewer. All labels come from one person watching the videos; the verdicts are a judgement call.
• A few overlapping segments. Sometimes an old label was not deleted after a correction, so two segments of the same episode overlap. Our build script keeps the one with the later reviewed_at.
• move table base (7) and rotate table base (5) can be mixed up. The two motions look alike and differ mainly in when they happen (before or after the legs are picked). Our build script filters them by that order.

License and attribution

These labels are released under CC BY 4.0. They describe episodes of BitRobot/G1_WBT_Dex1_Building-Children-Table by BitRobot, licensed under CC BY 4.0; the skill names in task_index are that dataset's task names.