Official Implementation for paper "Hybrid Physical Metric For 6-DoF Grasp Pose Detection" ICRA 2022.

arxiv

Download the GraspNet-1Billion dataset from graspnet. In this paper, we use a new evaluation metric to generate grasp confidence scores for grasp poses. You can get the new score labels under hybrid physical metric from here. The data directories should like this:

FGC_GraspNet/
├── grasp_data/
|       ├── scenes
|       ├── models
|       ├── dex_models
│       ├── FGC_label
│       ├── grasp_label
│       └── collision_label

• Python 3.6.13
• PyTorch 1.8.1+cu111
• Open3d 0.8
• TensorBoard 2.3
• NumPy
• SciPy
• Pillow
• tqdm

Get the code.

git clone https://github.com/luyh20/FGC-GraspNet.git

Install packages via Pip.

pip install -r requirements.txt

Compile and install pointnet2 operators (code adapted from votenet).

cd pointnet2
python setup.py install

Compile and install knn operator (code adapted from pytorch_knn_cuda).

cd knn
python setup.py install

Install graspnetAPI.

git clone https://github.com/graspnet/graspnetAPI.git
cd graspnetAPI
pip install .

Training:

sh train.sh

Testing:

sh test.sh

Realsense model's link，https://drive.google.com/file/d/1Y-CWHr_eZDoZm3XJocrUJq1SA5tfrONX/view?usp=sharing

The demo uses the RGBD data collected in real time from the Realsense D435i camera as input, and predicts the grasp poses results by the FGC_GraspNet. If you want to use it in your own experiment, you need to check the camera model or change the camera intrinsic for your camera model.

Run Demo:

python demo.py

@inproceedings{lu2022hybrid,
  title={Hybrid Physical Metric For 6-DoF Grasp Pose Detection},
  author={Lu, Yuhao and Deng, Beixing and Wang, Zhenyu and Zhi, Peiyuan and Li, Yali and Wang, Shengjin},
  booktitle={2022 International Conference on Robotics and Automation (ICRA)},
  pages={8238--8244},
  year={2022},
  organization={IEEE}
}