ukcastle/robotwin-rotate_qrcode-cosmos3-simtransfer-gr00t

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

Where the augmentation is: episodes 50-99, camera observation.images.cam_high only, e.g. videos/chunk-000/observation.images.cam_high/episode_000050.mp4. Episodes 0-49 are the unmodified originals, and the wrist cameras and all tabular data (states, actions, instruction) are identical between episode k and 50+k.

Single-task training set in LeRobot v2.1 format (as used by GR00T N1.7 via XPolicyLab): 100 episodes, 15436 frames.

• Episodes 0-49: the 50 RoboTwin 2.0 demo_clean demonstrations of rotate_qrcode (aloha-agilex, joint actions). This is the baseline training set of arXiv 2605.02757 Table 1.
• Episodes 50-99: one augmented copy of each, with the same states, actions, instruction and wrist cameras. observation.images.cam_high is replaced by a Cosmos3-Nano-Sim-Transfer video. Episode 50+k is the copy of episode k (meta/augmentation.json).

Train on episodes 0-49 for the baseline and on all 100 for the augmented arm, with the same recipe and step budget, as the paper does.

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) and keeps the object names from the first three "seen" instructions.
3. Depth: Video-Depth-Anything-Small, with per-video normalization, the same steps Transfer2.5 runs internally.
4. Transfer: Cosmos3-Nano-Sim-Transfer generates the video from the plain-text prompt and the depth control.
   • Model settings: 480p (736x544, the model's native size for 4:3), its fixed 4-step sampler, 30 fps, seed 2025, and a 30-latent-frame history window with 3 attention sinks.
5. Resize: the video is scaled to 320x240. Frame i of the output corresponds to frame i of the hdf5 episode; the source mp4 has one extra trailing frame, which is dropped.

Quality check and regeneration

Each clip is scored with the geometric-consistency metric of the paper (Sec. 4.5). Video Depth Anything is run on the generated clip and on the source, both depth maps are normalized per video, and Abs.Rel is computed between them. Abs.Rel median over the 50 clips: 0.538.

The paper itself does no filtering. This regeneration step is our addition:

• A clip is flagged when its Abs.Rel exceeds Q3 + 1.5 IQR within the task (threshold 1.288).
• A flagged clip is regenerated with seed 2026, then 2027 if needed.
• The first retry under the threshold is kept; if none passes, the attempt with the lowest Abs.Rel is kept.

Flagged episodes: [16, 28]

episode	kept seed	Abs.Rel
16	2026	0.5329
28	2026	0.7486

Known issues (qualitative)

• Arm appearance varies between episodes: some arms are white, some black or grey metal. The paper's own Transfer2.5 examples also change the arm appearance.
• 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-single_rotate_qrcode-paperST-aloha_agilex-joint and train with XPolicyLab GR00T_N17 train.sh RoboTwin single_rotate_qrcode-paperST aloha_agilex joint <seed> <gpus>. stats.json and relative_stats.json cover all 100 episodes. For the baseline, use the 50 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.