RoboDream-Data

Training data for RoboDream, built from ~41k DROID episodes (CC BY 4.0) with reliable camera calibration: robot-only renders (the kinematic anchor), scene and object priors, T5 instruction embeddings and the training splits. Model weights are in Jay-Ye/RoboDream. The raw DROID videos are not redistributed.

Path	Content
renders/robot_only_XXX.tar	Robot-only renders: the recorded end-effector trajectories replayed in Isaac Lab with calibrated cameras (black background). 1000 episodes per tar, layout <DROID rel_path>/rgb_sim_rollout_masked/{external_cam1,external_cam2,wrist_cam}/%010d.jpg (640×360, frame-aligned with the DROID MP4s)
renders/episode_to_rel_path.tsv	DROID episode id → relative path
shards/episode_priors_XXX.tar	Per-episode priors in the processed layout: <episode>/scene_prior_rgb/<cam>/0000000000_omnipaint.jpg (scene prior) and <episode>/scene_prior_object_gsam/<cam>/0000000000_gsam.jpg (object-prior canvas)
shards/scene_prior_omnipaint.tar	Scene priors: first frame with the task objects removed by OmniPaint (<episode>_<cam>.jpg)
shards/object_masks.tar	Grounded-SAM masks of the task objects on the first frame (<episode>_<cam>.png)
meta/object_segmentation_prompts.json	Task-relevant objects per episode (GPT-5-nano on the first frame + instruction)
meta/droid_t5_text_embeddings.pkl.gz	T5-11B embeddings of all DROID instructions (training and trajectory retrieval)
meta/split_train_station.txt, meta/split_val_station_first_clip.txt	Training / validation clip lists (validation = held-out lab)
meta/filtered_episode_meta25.json, meta/split_all_max512.json, meta/episodes.txt	Episode metadata and the episode split used for training
meta/pnp_cam2base_multiview.json, meta/reformatted_episodes_with_good_extrinsics.json	Camera extrinsics used for rendering and retrieval replay

Cameras: external_cam1, external_cam2, wrist_cam (the wrist camera has no priors). See data_processing/README.md in the code repository to build the training set from these files.

huggingface-cli download Jay-Ye/RoboDream-Data --repo-type dataset --local-dir RoboDream-Data
# only the metadata and T5 embeddings:
huggingface-cli download Jay-Ye/RoboDream-Data --repo-type dataset --include "meta/*" --local-dir RoboDream-Data

@article{ye2026robodream,
  title={RoboDream: Compositional World Models for Scalable Robot Data Synthesis},
  author={Ye, Junjie and Xue, Rong and Van Hoorick, Basile and Li, Runhao and Rajaprakash, Harshitha and Tokmakov, Pavel and Irshad, Muhammad Zubair and Guizilini, Vitor and Wang, Yue},
  journal={arXiv preprint arXiv:2606.02577},
  year={2026}
}