DexGraspBench

A standard and unified simulation benchmark in MuJoCo for dexterous grasping, aimed at enabling a fair comparison across different grasp synthesis methods, proposed in BODex: Scalable and Efficient Robotic Dexterous Grasp Synthesis Using Bilevel Optimization [ICRA 2025].

Project page ｜ Paper | Grasp synthesis code

Introduction

Main Usage

• Replay and test open-loop grasping poses/trajectories in parallel.

• Each grasping data point only needs to include:

  • Object (must be pre-processed by MeshProcess): obj_scale, obj_pose, obj_path.
  • Hand: approach_qpos (optional), pregrasp_qpos, grasp_qpos, squeeze_qpos.

Highlight

• Comprehensive Evaluation Metrics: Includes simulation success rate, analytic force closure metrics, penetration depth, contact quality, data diversity, and more.
• Diverse Experimental Settings: Covers various robotic hands (e.g., Allegro, Shadow, Leap, UR10e+Shadow), data formats (e.g., motion sequences, static poses), and scenarios (e.g., tabletop lifting, force-closure testing).
• Multiple Baseline Methods: Includes optimization-based grasp synthesis approaches (e.g., DexGraspNet, FRoGGeR, SpringGrasp, BODex) and data-driven baselines (e.g., CVAE, Diffusion Model, Normalizing Flow).
• Reproducible and Standardized Testing: The hand assets are sourced from MuJoCo_Menagerie, with modification details provided in the assets/hand directory.

Getting Started

Installation

1. Clone the third-party library MuJoCo Menagerie.

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

2. Install the python environment via Anaconda.

conda create -n DGBench python=3.10 
conda activate DGBench
pip install numpy==1.26.4
conda install pytorch==2.2.2 pytorch-cuda=12.1 -c pytorch -c nvidia 
pip install mujoco==3.3.2
pip install trimesh
pip install hydra-core
pip install transforms3d
pip install matplotlib
pip install scikit-learn
pip install usd-core
pip install imageio
pip install 'qpsolvers[clarabel]'

3. you may need to run the following command to avoid potential errors related to MKL such as undefined symbol: iJIT_NotifyEvent

conda install -c conda-forge mkl=2020.2 -y

(Optional) Object Preparation

If you need the object assets used in BODex, please download our pre-processed object assets DGN_2k_processed.zip from here and organize the unzipped folders as below.

assets/object/DGN_2k
|- processed_data
|  |- core_bottle_1a7ba1f4c892e2da30711cdbdbc73924
|  |_ ...
|- scene_cfg
|  |- core_bottle_1a7ba1f4c892e2da30711cdbdbc73924
|  |_ ...
|- valid_split
|  |- all.json
|  |_ ...

If you need the object assets used in Dexonomy, please download and organize DGN_5k and objaverse_5k from here.

Running

We have provided several scripts, which optionally includes format conversion, evaluation, statistic calculation, and visualization with OpenUSD or OBJ files.

For a quick start, some example data is provided in the output/example_shadow directory, which can be directly evaluated by

bash script/example.sh

To evaluate the synthesized grasps of BODex,

bash script/test_BODex_shadow.sh

To evaluate the synthesized grasps of DexLearn,

bash script/test_learning_shadow.sh

To visualize the synthesized grasps of Dexonomy,

bash script/vis_Dexonomy.sh

To evaluate the conditional synthesized grasps of DexLearn,

bash script/test_learning_conditional.sh

Changelog

The main branch serves as our standard benchmark, with some adjustments to the settings compared to the BODex paper, aimed at improving the practicality. Key changes include increasing the object mass from 30g to 100g, raising the hand's kp from 1 to 5, and supporting more diverse object assets. One can further reduce friction coefficients miu_coef (currently 0.6 for tangential and 0.02 for torsional) to increase difficulty.

The original benchmark version is available in the baseline branch. This branch also includes code to test other grasp synthesis baselines, such as DexGraspNet, FRoGGeR, SpringGrasp.

Citation

If you find this project useful, please consider citing:

@article{chen2024bodex,
  title={BODex: Scalable and Efficient Robotic Dexterous Grasp Synthesis Using Bilevel Optimization},
  author={Chen, Jiayi and Ke, Yubin and Wang, He},
  journal={arXiv preprint arXiv:2412.16490},
  year={2024}
}