1. Clone the repository: git clone --recursive [email protected]:facebookresearch/NGDF.git

2. Create a conda environment and install package dependencies. Note: mamba is highly recommended as a drop-in replacement for conda.

   cd NGDF
   bash install.sh
   

   Install PyTorch separately, based on your CUDA driver version. The command below was tested on a 3080/3090 with CUDA 11.1:

   source prepare.sh
   pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html
   

   Run source prepare.sh before running any ngdf training or evaluation code to activate the environment and set env variables.

NGDF
├── acronym                     # Submodule with utilities for ACRONYM dataset
├── contact_graspnet            # Submodule with ContactGraspnet for baselines
├── data                        # Datasets, models, and evaluation output
├── differentiable-robot-model  # Submodule for differentiable FK
├── ndf_robot                   # Submodule for pre-trained shape embedding
├── ngdf                        # Code for training and evaluating NGDF networks
├── OMG-Planner                 # Submodule with pybullet env, reach and grasp evaluation
├── scripts                     # Scripts for running training and evaluation
└── theseus                     # Submodule for differentiable FK and SE(3) ops

1. Download datasets acronym_perobj and acronym_multobj from this Google Drive link. Place the datasets in data/.

   The datasets are required to compute the closest grasp metric and are also used in training.

2. Run evaluation

   • Download pre-trained models and configs into data/models from this link
   • Download object rotations into data from this link
   • Run grasp level set evaluations:

   bash scripts/eval/grasp_level_set/perobj.sh
   bash scripts/eval/grasp_level_set/multobj.sh
   

   Results are stored in eval/ in each model dir.

   To evaluate the grasps in pybullet, you'll need to install the code in the following section, then run the above commands with a -p flag: bash scripts/eval/grasp_level_set/perobj.sh -p

1. Set up dependencies

   • OMG-Planner, follow instructions in OMG-Planner README OMG-Planner/README.md

   • pytorch3d

     pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"
     

   • differentiable-robot-model

     cd differentiable-robot-model
     git remote add parent https://github.com/facebookresearch/differentiable-robot-model.git
     git fetch parent
     python setup.py develop
     

   • Contact-GraspNet

     cd contact_graspnet
     conda env update -f contact_graspnet_env_tf25.yml
     sh compile_pointnet_tfops.sh
     pip install -e .
     

     Download trained model scene_test_2048_bs3_hor_sigma_001 from here and copy it into the checkpoints/ folder.

2. Run evaluation script

   bash scripts/eval/reach_and_grasp/perobj.sh
   

   The results are saved in data/pybullet_eval. Get summary results in jupyter notebook

   jupyter notebook --notebook-dir=scripts/eval/reach_and_grasp
   

1. Single object model training:

   bash scripts/train/perobj_Bottle.sh
   bash scripts/train/perobj_Bowl.sh
   bash scripts/train/perobj_Mug.sh
   

2. Multi-object model training;

   bash scripts/train/multobj_Bottle.sh
   

• Building docker

  cd NGDF
  docker build -t ngdf . 
  

• Run docker
  • bash docker_run.sh
  • source prepare.sh
  • Run the same commands for training in the container under root:/workspace/NGDF#

@article{weng2022ngdf,
  title={Neural Grasp Distance Fields for Robot Manipulation},
  author={Weng, Thomas and Held, David and Meier, Franziska and Mukadam, Mustafa},
  journal={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2023}
}

The majority of NGDF is licensed under MIT license, however a portion of the project is available under separate license terms: ContactGraspNet is licensed under a non-commericial NVidia License.

We actively welcome your pull requests! Please see CONTRIBUTING.md and CODE_OF_CONDUCT.md for more info.