README.md

AnyDexGrasp

Dexterous grasp pose detection network built upon MinkowskiEngine.

Requirements

• Anaconda with Python 3.8
• PyTorch 1.13 with CUDA 11.7
• MinkowskiEngine v0.5

Installation

1. Follow MinkowskiEngine instructions to install Anaconda, cudatoolkit, Pytorch and MinkowskiEngine.Note that you need export MAX_JOBS=2; before pip install due to this issue

2. Install other requirements from Pip.

    pip install -r requirements.txt

3. Install knn module.

    cd knn
    python setup.py install

4. Install pointnet2 module.

    cd pointnet2
    python setup.py install

5. Install ur toolbox.

    cd ur_toolbox
    pip install .
    cd python-urx
    pip install .
    pip install -r requirements.txt

6. Install Allegro Hand.

   Install Allegro Hand upon Allegro-Hand-Controller-DIME.

7. Download model weights and data at GoogleDrive and put it under logs/ and extract zip files in logs/data/representation_model/graspnet_v1_newformat/ to logs/data/representation_model/graspnet_v1_newformat/.
   Download the data from the Graspnet web page and extract it to logs/data/representation_model/graspnet_v1_newformat/

Generating the STL file for dexterous hand.

    python command_generate_mesh_file.sh

Training

    sh command_train_representation.sh # Representation model
    sh command_train_decision.sh # Decision model

Collecting data

    python realsense.py
    sh command_collect_multifinger_grasp_data.sh

Robot grasp

    python realsense.py
    sh command_robot_multifinger_grasp.sh

License

The code is licensed under CC BY-NC 4.0 for non-commercial purposes.

Bibtex

If you find the code useful, please consider citing our paper

        @article{fang2025anydexgrasp,
          title={AnyDexGrasp: General Dexterous Grasping for Different Hands with Human-level Learning Efficiency},
          author={Fang, Hao-Shu and Yan, Hengxu and Tang, Zhenyu and Fang, Hongjie and Wang, Chenxi and Lu, Cewu},
          journal={arXiv preprint arXiv:2502.16420},
          year={2025}
        }