代码主要来自此仓库，修改了ROS接口；原论文为Volumetric Grasping Network: Real-time 6 DOF Grasp Detection in Clutter。src/inference来自原仓库中的src/vgn，src/robot_helpers来自这个仓库。
conda create -c conda-forge -n vgn python=3.10 pytorch numpy scipy scikit-image catkin_pkg
使用原仓库并遵循其说明以进行训练，依赖安装方式为：
conda create -c conda-forge -n vgn python numpy=1.19 mpi4py pandas pybullet pytorch pytorch-ignite pyyaml scikit-image scipy tensorboard tqdm mayavi
pip install open3d
使用numpy=1.19以兼容原仓库代码中的np.long
需要修改源代码simulation.py中的文件名。
生成数据（一堆方块）
ln -s <path-to>/robot_helpers/robot_helpers <path-to>/vgn/scripts/robot_helpers
ln -s <path-to>/vgn/src/vgn <path-to>/vgn/scripts/vgn
LD_LIBRARY_PATH=/home/<user-name>/miniconda/envs/vgn/lib mpirun -np 6 python3 scripts/generate_data.py --root=data/grasps/blocks
生成两种数据，需要修改generate_data.py中的逻辑（添加--mode并判断调用generate_pile还是generate_packed）
LD_LIBRARY_PATH=/home/<user-name>/miniconda/envs/vgn/lib mpirun -np 6 python3 scripts/generate_data.py --root=data/grasps/pile --cfg cfg/sim/pile.yaml --count 750000 --mode 1 && LD_LIBRARY_PATH=/home/<user-name>/miniconda/envs/vgn/lib mpirun -np 6 python3 scripts/generate_data.py --root=data/grasps/packed --cfg cfg/sim/packed.yaml --count 2000000 --mode 0
packed.yaml
sim:
urdf_root: assets/urdfs
gui: False
lateral_friction: 1.0
object_urdfs: assets/urdfs/packed/train
object_count_lambda: 4
scaling:
low: 0.8
high: 1.5
scene: packed
max_view_count: 5
scene_grasp_count: 200
metric: dynamic_with_approach
使用process_data.ipynb
ln -s <path-to>/vgn/data/grasps/packed/*.npz <path-to>/vgn/data/grasps/train/ -r
ln -s <path-to>/vgn/data/grasps/pile/*.npz <path-to>/vgn/data/grasps/train/ -r
生成训练数据集
python scripts/create_dataset.py data/grasps/train data/dataset/train
训练与可视化
python scripts/train_vgn.py --dataset data/dataset/train --augment --logdir logs
tensorboard --logdir logs