Grasp as You Say: Language-guided Dexterous Grasp Generation

Yi-Lin Wei, Jian-Jian Jiang, Chengyi Xing, Xiantuo Tan, Xiao-Ming Wu, Hao Li,
Mark Cutkosky, Wei-Shi Zheng

[Paper]     [Project]

(NeurIPS 2024) Official repository of paper "Grasp as You Say: Language-guided Dexterous Grasp Generation"

Install

• Create a new conda environemnt and activate it.

conda create -n dexgys python=3.8
conda activate dexgys

• Install the dependencies.

conda install -y pytorch==1.10.0 torchvision==0.11.0 cudatoolkit=11.3 -c pytorch -c conda-forge
pip install -r requirements.txt

• Build the pakage.

Note: The CUDA enviroment should be consistent in the phase of building and running (Recommendation: cuda11 or higher).

cd thirdparty/pytorch_kinematics
pip install -e .

cd ../pointnet2
python setup.py install 

cd ../
git clone https://github.com/wrc042/CSDF.git
cd CSDF
pip install -e .
cd ../../

Data Preparation

1. Download dexterous grap label and language label of DexGYS from here ["coming soon"], and put in the "dexgys" in the path of "./data".

2. Download ShadowHand model mjcf from here, and put the "mjcf" in the path of "./data".

3. Download 3D mesh of object from here, and put the "oakink" in the path of "./data".

4. Finally, the directory should as follow:

.data/
├── dexgys/ 
│ ├── train_with_guide_v2.1.json
│ ├── test_with_guide_v2.1.json 
├── oakink/ 
│ ├── shape/
└── mjcf/

Usage

Train

1. Train Intention and Diversity Grasp Component (IDGC)

python train.py -t "./config/idgc.yaml"

2. Infer IDGC on train and test set to obatin training and testing pairs for QGC.

python ./test.py \
--train_cfg ./config/idgc.yaml \
--test_cfg ./config/infer_idgc_train.yaml \
--override model.checkpoint_path \"<checkpoint-path>\"

python ./test.py \
--train_cfg ./config/idgc.yaml \
--test_cfg ./config/infer_idgc_test.yaml \
--override model.checkpoint_path \"<checkpoint-path>\"

3. Train Quality Grasp Component (QGC).

• Set the "data.train.pose_path" and "data.test.pose_path" of "./config/qgc.yaml" to the <matched_results.json> of the outcome of step2.
• For example:

data:
  name: refinement
  train:
    data_root: &data_root "./data/oakink"
    pose_path: ./Experiments/idgc/test_results/epoch_<the epoch number>_train/matched_results.json
    ...
  val:
    data_root: *data_root
    pose_path: ./Experiments/idgc/test_results/epoch_<the epoch number>_test/matched_results.json

• Then run:

python train.py -t "./config/qgc.yaml"

Test

• Infer QGC to refine the coarse outcome of IDGC.
• Set "data.test.pose_path" of "./config/infer_qgc_test.yaml" to the <matched_results.json> of the outcome of LDGC.

data:
  name: refinement
  train:
    data_root: &data_root "./data/oakink"
    pose_path: ./Experiments/idgc/test_results/epoch_<the epoch number>_train/matched_results.json
    sample_in_pose: &sample_in_pose True

• Then run:

python ./test.py \
--train_cfg ./config/qgc.yaml \
--test_cfg ./config/infer_qgc_test.yaml \
--override model.checkpoint_path \"<checkpoint-path>\"

TODO

• Release the datasets of GraspGYSNet
• Release the visualization code of GraspGYS framework
• Release the evaluation code of GraspGYS framework
• Release the training code of GraspGYS framework
• Release the inference code of GraspGYS framework

Acknowledgements

The code of this repository is based on the following repositories. We would like to thank the authors for sharing their works.

• UniDexGrasp

• Scene-Diffuser

• DGTR

Contact

• Email: [email protected]

Citation

Please cite it if you find this work useful.

@article{wei2024grasp,
  title={Grasp as you say: language-guided dexterous grasp generation},
  author={Wei, Yi-Lin and Jiang, Jian-Jian and Xing, Chengyi and Tan, Xian-Tuo and Wu, Xiao-Ming and Li, Hao and Cutkosky, Mark and Zheng, Wei-Shi},
  journal={arXiv preprint arXiv:2405.19291},
  year={2024}
}