ukcastle/robotwin-cosmos3-transfer-c2

RoboTwin 2.0 x Cosmos3 Transfer (arm C2): sim-to-real video augmentation

This dataset contains 694 video clips in which RoboTwin 2.0 simulation renders were re-rendered with NVIDIA Cosmos3 Transfer (Cosmos3-Nano) under edge control. The goal was to make the clips look like real camera footage while keeping the scene geometry and the robot motion. Every augmented clip comes with the sim clip it was made from, the exact text prompt, the sampled scene description, and the generation request with its seed. The original sim clips and the outputs have the same number of frames, so the clips stay aligned with the actions in the source RoboTwin episode.

• Source: RoboTwin 2.0, embodiment aloha-agilex (AgileX Cobot Magic dual-arm), config demo_clean, 8 tasks.
• Cameras: head (head camera), left / right (wrist cameras).
• Coverage: episodes 0–28 of each task, all 3 cameras: 232 head, 231 left and 231 right clips. grab_roller episode 28 has only the head clip, because generation was stopped while that episode was still running.
• Format: H.264 mp4, 15 fps. Cosmos outputs are 320x256. The sim inputs are 320x240, and the outputs are a vertical stretch of the same field of view, with no crop. Clips are 87–314 frames long (106,497 frames in total).

Tasks

adjust_bottle, beat_block_hammer, grab_roller, move_can_pot, open_laptop, pick_dual_bottles, place_burger_fries, rotate_qrcode (87 clips each, except grab_roller with 85).

Layout

metadata.csv / metadata.jsonl          one row per augmented clip (see columns below)
cosmos/<task>/epNNN_<cam>.mp4           Cosmos3 Transfer output
sim/<task>/epNNN_<cam>.mp4              matching RoboTwin source render (input to Cosmos)
prompts/<task>/epNNN/prompt_v2_<cam>.json   structured text prompt used for that camera
prompts/<task>/epNNN/scene_v2.json      scene sampled for the episode (table, wall, lighting), shared by the 3 cameras
prompts/<task>/epNNN/meta.json          source episode, frame count, language instructions
generation/<task>/epNNN_<cam>.request.json  exact request sent to the Cosmos3 server (prompt, negative prompt, seed, params)
LICENSES/                               RoboTwin (MIT) and Cosmos (OpenMDW-1.1) license texts

metadata columns

column	meaning
task, episode, camera	RoboTwin task, episode index, camera (head/left/right)
cosmos_video, sim_video	relative paths of the output and the source clip
frames, frames_sim, fps	frame counts (always equal) and frame rate
width, height, sim_width, sim_height	resolutions
model	nvidia/Cosmos3-Nano (Cosmos3 Transfer)
control, edge_threshold	edge control computed from the sim clip; Canny preset medium (head) or very_low (wrists)
control_guidance, guidance, steps, flow_shift, seed, chunk_frames	sampler settings for this clip
prompt_file, scene_file, generation_request	relative paths of the prompt, scene and request files
table, wall, light	the sampled scene, in plain text
instruction	one RoboTwin language instruction for the episode
source_hdf5, source_config, embodiment	source RoboTwin episode
generated_at_utc, cosmos_sha256	generation time and output checksum

Generation settings

• Model: Cosmos3 Transfer, nvidia/Cosmos3-Nano, served by the Cosmos3 vLLM-Omni server. Resolution template 256, aspect 4:3 (320x256 output). 35 steps, guidance 3.0, control guidance 1.5, flow shift 10. Clips are generated autoregressively in 93-frame chunks with 1 conditional frame. Guardrails were off.
• Control: the server computes an edge map from each sim clip. The head camera uses the medium Canny preset. The wrist cameras use very_low, so that the faint table and wall boundaries survive and Cosmos does not re-invent the layout of mostly flat wrist views.
• Prompts: structured captions describe the robot (white AgileX arms with black two-finger grippers), the task objects ("keep their exact colors and shapes … do not add any objects"), the action, and the camera. The style is "live-action video, real camera footage, natural sensor noise, no CGI".
• Scene randomization: one scene per episode, sampled deterministically per task and episode. It is used for all 3 cameras and drawn from 22 table surfaces, 10 walls and 6 lighting setups. No distractor objects are added.
• Seeds: 4052 + job index, where the job list runs episode-major and covers all tasks and cameras. Each clip has its own seed (4052–4745), which is stored in metadata.csv and in the request file.

Intended use and limitations

The clips are meant for visual data augmentation, for example as replacement observations for RoboTwin episodes when training visuomotor policies, and for studying sim-to-real appearance transfer. Actions and proprioception are not included here. They live in the source RoboTwin 2.0 episodes, which source_hdf5 points to, and the frame indices line up one-to-one.

Known limitations:

• Edge control keeps the overall geometry, but fine details can drift. Spot checks found task-object colors that changed in some wrist-camera clips (for example a container rendered in a different color), even though the prompt asks for the true colors. Do not assume that object appearance is preserved exactly.
• Each clip is generated independently, so the head and wrist clips of an episode share the scene description but not a pixel-consistent scene.
• The resolution is low (320x256), matching the RoboTwin camera stream that this augmentation was built for.

License

• The RoboTwin 2.0 code and data are released under the MIT license (LICENSES/RoboTwin-MIT.txt), and so are the source sim renders in sim/.
• The Cosmos3 models are released under OpenMDW-1.1 (LICENSES/Cosmos-OpenMDW-1.1.txt). That license "does not impose any restrictions or obligations with respect to any use, modification, or sharing of any outputs generated by using the Model Materials". The augmented clips in cosmos/ are therefore released under the MIT license together with the rest of this dataset.

Citation

@article{chen2025robotwin,
  title   = {RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation},
  author  = {Chen, Tianxing and Chen, Zanxin and Chen, Baijun and Cai, Zijian and Liu, Yibin and Li, Zixuan and Liang, Qiwei and Lin, Xianliang and Ge, Yiheng and Gu, Zhenyu and others},
  journal = {arXiv preprint arXiv:2506.18088},
  year    = {2025}
}

@techreport{nvidia2026cosmos3,
  title  = {Cosmos 3: Omnimodal World Models for physical AI},
  author = {{NVIDIA Cosmos Team}},
  year   = {2026},
  url    = {https://research.nvidia.com/labs/cosmos-lab/cosmos3/technical-report.pdf}
}