quincyu/pi05_robotwin_obb_eef

pi0.5 + Box-WAM for RoboTwin 2.0

The 35k-step end-effector pi0.5 + Box-WAM checkpoint from Box Trajectories as Compact State Representations for World Action Models (CoRL 2026), fine-tuned on 50 RoboTwin 2.0 tasks.

Code: box-wam · Dataset: quincyu/robotwin_eef_obb_27500 · Project page: https://box-wam.github.io/

Usage

# in box-wam/multitask/openpi-boxwam
uv run huggingface-cli download quincyu/pi05_robotwin_obb_eef --local-dir checkpoints/pi05_robotwin_boxwam/release
uv run scripts/serve_policy.py --port 8000 policy:checkpoint \
    --policy.config pi05_robotwin_boxwam --policy.dir checkpoints/pi05_robotwin_boxwam/release/34999

RoboTwin 2.0 evaluation is described in the box-wam README.

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}
}

License

This checkpoint is fine-tuned from Physical Intelligence's pi0.5, which contains PaliGemma weights derived from Google's Gemma, so it is a Model Derivative of Gemma. Its use and redistribution are subject to the Gemma Terms of Use (copy in LICENSE_GEMMA.txt) and the Gemma Prohibited Use Policy, and anyone you redistribute it to must receive the same terms.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms