DexJoCo/DexJoCo-Pi05

Model Card for DexJoCo Pi-0.5

This repository provides the Pi-0.5 models trained on the DexJoCo datasets.

Model Summary

• Base Models: pi05_base, pi05_base_action_dim_44
• Finetuned Checkpoints:
  • pi05_dexjoco_ckpt ("rand-obj" tasks)
  • pi05_dexjoco_rand_full_ckpt ("rand-full" tasks)
  • pi05_dexjoco_multi_task ("multi-task" setup, "rand-obj")

Training Details

All models are finetuned using LoRA on the DexJoCo datasets with LeRobot format:

• Single-arm checkpoints: LoRA finetuned for 30,000 steps.
• Bimanual & Multi-task checkpoints: LoRA finetuned for 60,000 steps.

Links & Resources

• Hugging Face Paper: https://huggingface.co/papers/2605.16257
• arXiv Paper: https://arxiv.org/abs/2605.16257
• Project Page: https://dexjoco.github.io
• GitHub Repository: https://github.com/brave-eai/dexjoco
• Dataset: https://huggingface.co/datasets/DexJoCo/DexJoCo-Datasets-LeRobot
• Raw Dataset: https://huggingface.co/datasets/DexJoCo/DexJoCo-Datasets-Raw

BibTeX:

@misc{wang2026dexjocobenchmarktoolkittaskoriented,
      title={DexJoCo: A Benchmark and Toolkit for Task-Oriented Dexterous Manipulation on MuJoCo}, 
      author={Hanwen Wang and Weizhi Zhao and Xiangyu Wang and Siyuan Huang and He Lin and Boyuan Zheng and Rongtao Xu and Gang Wang and Yao Mu and He Wang and Lue Fan and Hongsheng Li and Zhaoxiang Zhang and Tieniu Tan},
      year={2026},
      eprint={2605.16257},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2605.16257}, 
}

License

This repository is derived from OpenPI and is distributed under the Apache 2.0 License.

Gemma-based model and checkpoints are subject to the Gemma Terms of Use. See LICENSE_GEMMA.txt