DA2 Dataset: Toward Dexterity-Aware Dual-Arm Grasping

The project website is https://sites.google.com/view/da2dataset. This repo contains the code for DA² dataset generation and some scripts that can visualize grasp pairs and render virtual scenes. The paper is available here.

Installation

Basic installation

conda create -n DA2 python=3.8
conda activate DA2
git clone https://github.com/ymxlzgy/DA2.git
cd path/to/DA2
mkdir grasp test_tmp
pip install -r requirements.txt

Meshpy installation

cd path/to/DA2/meshpy
python setup.py develop

Pytorch installation

Please refer to pytorch official website to find the best version in your case, e.g.,:

pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

Mayavi installation

conda install mayavi -c conda-forge

Mesh Download

Download meshes here. After download, unzip them under path/to/DA2 make it like path/to/DA2/simplified.

Toy generation

This is used to individual generation. Need to modify the path OBJ_FILENAME inside posetest.py to your customized path, and run:

cd path/to/DA2/scripts
python posetest.py

The generated grasp pairs will be saved under test_tmp.

Whole dataset generation

Need to modify file_dir to the customized mesh path.

cd path/to/DA2/scripts
python generate_dual_dataset.py

The generated grasp pairs will be stored under grasp. generate_dual_dataset2.py is used for a parallel generation. If use this script, you need to modify the len(file_list) in generate_dual_dataset.py to a customized number.

Visualize

To visualize the mesh with accompanying grasp pairs, run:

cd path/to/DA2/scripts
python dual_grasps_visualization.py absolute_path/to/grasp_file --mesh_root path/to/meshes

Render scenes

After download or generating the dataset, run:

cd path/to/DA2/scripts
python render_point_dex.py path/to/dataset

path/to/dataset in our case is path/to/DA2. You may need to change the path under the function load_grasp_path to a customized defined path. The generated scenes will be under path/to/dataset/table_scene_stand_all. Simulated point clouds will be under path/to/dataset/pc_two_view_all, Virtual camera info will be under path/to/dataset/cam_pose_all.

Scene rendering is time-consuming. To render in parallel, just run multiple scripts in the same time.

If you want to visualize the generated scenes, run:

python render_point_dex.py path/to/dataset --load_existing number_of_the_scene --vis

Acknowledgement

This work is based on Dex-Net, Acronym, Contact-GraspNet, and diverse-and-stable-grasp.

Many functions under dexnet are from Dex-Net. Didn't remove them in case anyone can notice them and may facilitate one's research.

If you think this repo can help your research, please consider citing:

@article{da2dataset,
  author={Zhai, Guangyao and Zheng, Yu and Xu, Ziwei and Kong, Xin and Liu, Yong and Busam, Benjamin and Ren, Yi and Navab, Nassir and Zhang, Zhengyou},
  journal={IEEE Robotics and Automation Letters},
  title={DA$^2$ Dataset: Toward Dexterity-Aware Dual-Arm Grasping},
  year={2022},
  volume={7},
  number={4},
  pages={8941-8948},
  doi={10.1109/LRA.2022.3189959}}