SII-CDZ/gpt-policy-robodojo-contexts

GPT-Policy · RoboDojo contexts

Ten fixed expert demonstrations for the GPT-Policy RoboDojo simulation harness: 16 timestamps × 3 camera views per task, 480 images in total, with stage annotations, measured robot states and numeric action references.

Reproduction guide · Task/episode index · Source and terms

Download and verify

hf download SII-CDZ/gpt-policy-robodojo-contexts --repo-type dataset \
  --local-dir data/robodojo-release
python data/robodojo-release/verify.py data/robodojo-release

Load contexts/<task>/prepared/input.json using configure.py, then run the eight-GPU command in the reproduction guide. The dataset preserves the model-visible demonstration text, numeric references and original image bytes; storage paths were made portable and an internal version label was omitted. No training or re-extraction is needed.

Contents

• contexts/: ten model-ready inputs, images and recorded public prompt/tool schemas.
• protocol/: the 50-case suite, execution settings, source versions and asset hashes.
• configure.py and verify.py: portable configuration and offline validation.

Pose references use link6 in environment_origin, metres, xyzw. Recorded next-state references are distinguished from issued commands. Standard and random scene variants share the same task demonstration.

Evaluation requires the matching GPT-Policy simulation backend, Isaac Sim 5.1, the pinned RoboDojo code/assets, and an authenticated model provider. The matching simulation backend is not yet in the public GitHub default branch at this data release. Download and offline validation are self-contained.

Citation

@article{cheng2026context,
  title={In-context robot learning with VLM agents},
  author={Cheng, Dongzhou and Yi, Taoran and Fang, Ye and Zhang, Xingwu and Feng, Fan and Li, Yixuan and Zhuang, Gengxiong and Wang, Rongze and Yang, Shuai and Song, Wei and others},
  journal={arXiv preprint arXiv:2609.19138},
  year={2026}
}