ukcastle/robotwin-8task-cosmos3-simtransfer-gr00t

RoboTwin 2.0, 8 tasks: original + Cosmos3-Nano-Sim-Transfer augmentation (GR00T N1.7 format)

Where the augmentation is: episodes 400-799, camera observation.images.cam_high only. Episode 400+i is the augmented copy of episode i. Episodes 0-399 are the unmodified originals. The wrist cameras and all tabular data (states, actions, instruction) are identical between episode i and 400+i.

This is the combined version of the eight single-task datasets ukcastle/robotwin-<task>-cosmos3-simtransfer-gr00t, for multi-task training. It holds 800 episodes and 122804 frames in LeRobot v2.1 format, the format GR00T N1.7 reads through XPolicyLab. The tasks are the eight of arXiv 2605.02757 Table 1, with 50 RoboTwin 2.0 demo_clean demonstrations each (aloha-agilex, joint actions).

task	original episodes	augmented episodes	QA Abs.Rel median	regenerated (flagged)
adjust_bottle	0-49	400-449	0.484	-
beat_block_hammer	50-99	450-499	0.474	-
pick_dual_bottles	100-149	500-549	0.206	[0, 10, 22]
place_burger_fries	150-199	550-599	0.533	-
open_laptop	200-249	600-649	0.412	[1, 17, 21]
move_can_pot	250-299	650-699	0.275	[32, 47]
rotate_qrcode	300-349	700-749	0.538	[16, 28]
grab_roller	350-399	750-799	0.386	-

How the head-camera videos were made

The pipeline follows the augmentation recipe of arXiv 2605.02757 and its official code. The video model is the one change: Cosmos3-Nano-Sim-Transfer replaces Cosmos-Transfer2.5.

1. Caption: VideoChat-Flash-Qwen2-7B_res448 captions the RoboTwin head video.
2. Rewrite: Qwen3-8B rewrites the caption with the paper's template. The template changes only the table material (wood, stainless steel or marble).
3. Depth: Video-Depth-Anything-Small, with per-video normalization.
4. Transfer: Cosmos3-Nano-Sim-Transfer generates the video at 736x544 (480p, the model's native size for 4:3) with its 4-step sampler, 30 fps and seed 2025.
5. Resize: the video is scaled to 320x240, with frames aligned 1:1 to the hdf5 episode.

Quality check

Each clip is scored with the paper's geometric-consistency metric (Sec. 4.5: Video Depth Anything on the generated clip and on the source, then Abs.Rel between them). A clip whose Abs.Rel is a within-task outlier (> Q3 + 1.5 IQR) is regenerated with seed 2026, then 2027 if needed. This step is our addition; the paper does not filter. Twelve of the 400 clips were flagged. The per-task single datasets list the regeneration details.

Known issues (qualitative)

• Arm appearance varies between episodes: some arms are white, some black or grey metal.
• Some empty table regions get objects that are not in the source.
• Some objects change color or shape.

Use

Copy this folder to $GR00T_LEROBOT_HOME/RoboTwin-clean8-paperST-aloha_agilex-joint and train with XPolicyLab GR00T_N17 train.sh RoboTwin clean8-paperST aloha_agilex joint <seed> <gpus>. stats.json and relative_stats.json cover all 800 episodes. For an original-only baseline, train on the 400 original episodes with their own stats.

License

RoboTwin 2.0 data is under the MIT license. The Cosmos3 model license, OpenMDW-1.1, places no restrictions on generated outputs.