Episodes Preview ALOHA Visualizer
27.5k episodes · 50 fps · 3 cameras · 640×480 av1

RoboTwin 2.0 demonstrations with object box trajectories

Training data of Box Trajectories as Compact State Representations for World Action Models (CoRL 2026): 27,500 RoboTwin 2.0 demonstrations (50 tasks × 50 clean + 500 randomized scenes, 6.4 M frames) in LeRobot v2.1 format, with the projected 3-D bounding boxes of the manipulated objects recorded at every step.

Feature Shape Content
observation.images.cam_high, cam_left_wrist, cam_right_wrist 480 × 640 video, 50 fps head and wrist cameras
observation.state (16,) [left x, y, z, qw, qx, qy, qz, gripper, right ...] end-effector pose
action (16,) end-effector pose of the next step
observation.obb_head (32,) `[left slot (16)
task text RoboTwin instruction of the episode

Collected with RoboTwin-boxwam (demo_clean_eef, demo_randomized_eef) and converted with scripts/robotwin/ of openpi-boxwam.

Usage

# in box-wam/multitask/openpi-boxwam
uv run huggingface-cli download quincyu/robotwin_eef_obb_27500 --repo-type dataset \
    --local-dir ~/.cache/huggingface/lerobot/quincyu/robotwin_eef_obb_27500
uv run scripts/compute_norm_stats.py --config-name pi05_robotwin_boxwam
XLA_PYTHON_CLIENT_MEM_FRACTION=0.9 uv run scripts/train.py pi05_robotwin_boxwam --exp-name boxwam

Trained model: quincyu/pi05_robotwin_obb_eef. Code and paper: box-wam · https://box-wam.github.io/

Citation

@inproceedings{yu2026boxwam,
  title     = {Box Trajectories as Compact State Representations for World Action Models},
  author    = {Yu, Qinxi and Ding, Wenzhou and Wang, Gloria X. and Zhai, Albert J. and Wang, Shenlong},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}

Simulation assets and tasks are from RoboTwin 2.0 (MIT license).

Downloads last month
301

Models trained or fine-tuned on quincyu/robotwin_eef_obb_27500