Dexonomy

Project Page | Paper | Dataset | Learning

Overview

Our algorithm synthesizes contact-rich, penetration-free, and physically plausible dexterous grasps for:

• Any grasp type
• Any object
• Any articulated hand

All starting from just one human-annotated template per hand and grasp type.

Supported Features

• Hands: Shadow, Allegro, Leap, MANO, Unitree_G1
• Object Scenes: Single (floating & tabletop), Clustered (light clutter)
• Object Assets: Rigid objects (ShapeNet & Objaverse), Articulated objects (PartNet)
• Physics Simulator: MuJoCo
• Grasping Trajectory Synthesis: cuRobo integration

TODO: add scene_cfg examples for articulated objects, clutter scenes, and other tasks.

Installation

Our code is tested on Ubuntu 20.04 with NVIDIA RTX 3090 GPUs.

git submodule update --init --recursive --progress

conda create -n dexonomy python=3.10 
conda activate dexonomy
pip install -e .

(Optional) BODex is only used for collision-free grasping trajectory synthesis. Currently, we only provide support for UR10e + shadow hand.

cd third_party/BODex
git lfs pull
# install BODex
pip install -e . --no-build-isolation
# link to dexonomy's object folder
ln -s ../../../../../../assets/object src/curobo/content/assets/object  
cd ...

Quick Start

1. Prepare Object Assets

Download our pre-processed object assets DGN_5k_processed.zip from Hugging Face, and organize the unzipped folders as below.

assets/object/DGN_5k
|- processed_data
|  |- core_bottle_1a7ba1f4c892e2da30711cdbdbc73924
|  |_ ...
|- scene_cfg
|  |- core_bottle_1a7ba1f4c892e2da30711cdbdbc73924
|  |_ ...
|- valid_split
|  |- all.json
|  |_ ...

Alternatively, you can pre-process your own object assets using MeshProcess.

2. Synthesize Grasps

Run the following commands to synthesize grasps for a specific grasp type:

dexrun op=tmpl  # Convert human annotations to valid templates
dexsyn 'tmpl_name=[1_Large_Diameter]' # Run the script for complete synthesis pipeline

where dexrun is the alias for python -m dexonomy.main and dexsyn is the alias for python -m dexonomy.script.

Tutorial

For a detailed walkthrough of template annotation and code usage, please refer to getting_started.

License

This work and the dataset are licensed under CC BY-NC 4.0.

Citation

If you find this work useful for your research, please consider citing:

@article{chen2025dexonomy,
        title={Dexonomy: Synthesizing All Dexterous Grasp Types in a Grasp Taxonomy},
        author={Chen, Jiayi and Ke, Yubin and Peng, Lin and Wang, He},
        journal={Robotics: Science and Systems},
        year={2025}
      }