UltraDexGrasp: Learning Universal Dexterous Grasping for Bimanual Robots with Synthetic Data

Sizhe Yang, Yiman Xie, Zhixuan Liang, Yang Tian, Jia Zeng, Dahua Lin, Jiangmiao Pang
Shanghai AI Laboratory, The Chinese University of Hong Kong, Zhejiang University, The University of Hong Kong, Peking University

ICRA 2026

📋 Contents

• 🔥 Highlight
• 🛠️ Getting Started
• 🔗 Citation
• 📄 License
• 👏 Acknowledgements

🔥 Highlight

UltraDexGrasp is a framework for universal dexterous grasping with bimanual robots.

The proposed data generation pipeline integrates an optimization-based grasp synthesizer with a planning-based demonstration generation module, and supports multiple grasp strategies, including two-finger pinch, three-finger tripod, whole-hand grasp, and bimanual grasp.

Trained on data produced by this pipeline, the policy demonstrates robust zero-shot sim-to-real transfer and strong generalization to novel objects with varied shapes, sizes, and weights.

🛠️ Getting Started

Installation

First, clone this repository.

git clone https://github.com/InternRobotics/UltraDexGrasp.git

(Optional) Use conda to manage the python environment.

conda create -n ultradexgrasp python=3.10 -y
conda activate ultradexgrasp

Install dependencies.

# Install PyTorch according to your CUDA version. For example, if your CUDA version is 11.8:
pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu118

# Install other dependencies
pip install -r requirements.txt

mkdir third_party; cd third_party

# Install PyTorch3D. If you encounter any problems, please refer to the detailed installation instructions at https://github.com/facebookresearch/pytorch3d.
git clone https://github.com/facebookresearch/pytorch3d.git
cd pytorch3d
pip install -e . --no-build-isolation
cd ..

# Install cuRobo
sudo apt install git-lfs
git clone https://github.com/NVlabs/curobo.git
cd curobo
pip install -e . --no-build-isolation
cd ..

# Install BODex_api (adapted from https://github.com/JYChen18/BODex)
git clone https://github.com/yangsizhe/BODex_api.git
cd BODex_api
pip install -e . --no-build-isolation
pip install torch-scatter -f https://data.pyg.org/whl/torch-2.4.1+cu118.html  # for torch-2.4.1 and cuda-11.8
conda install coal -c conda-forge -y
conda install boost=1.84.0 -y
cd src/bodex/geom/cpp; python setup.py install
cd ../..

Usage

Generate trajectories for various grasp strategies by running:

# left whole hand grasp
python rollout.py \
    --hand 0 \
    --object_scale_list '[0.08]'

# right whole hand grasp
python rollout.py \
    --hand 1 \
    --object_scale_list '[0.08]'

# bimanual grasp
python rollout.py \
    --hand 2 \
    --object_scale_list '[0.25]'

# left three-finger tripod
python rollout.py \
    --hand 3 \
    --object_scale_list '[0.04]'

# right three-finger tripod
python rollout.py \
    --hand 4 \
    --object_scale_list '[0.04]'

You can change the object_mesh_path in env/config/env.yaml to the mesh path of another object. For mesh processing, please refer to BODex.

🔗 Citation

If you find our work helpful, please cite:

@article{yang2026ultradexgrasp,
  title={UltraDexGrasp: Learning Universal Dexterous Grasping for Bimanual Robots with Synthetic Data},
  author={Yang, Sizhe and Xie, Yiman and Liang, Zhixuan and Tian, Yang and Zeng, Jia and Lin, Dahua and Pang, Jiangmiao},
  journal={arXiv preprint arXiv:2603.05312},
  year={2026}
}

📄 License

This repository is released under the Apache 2.0 license.

👏 Acknowledgements

Our code is built upon BODex. We thank the authors for open-sourcing their code and for their significant contributions to the community.