6-DoF GraspNet: Variational Grasp Generation for Object Manipulation

This is a PyTorch implementation of 6-DoF GraspNet. The original Tensorflow implementation can be found here https://github.com/NVlabs/6dof-graspnet.

License

The source code is released under MIT License and the trained weights are released under CC-BY-NC-SA 2.0.

Installation

This code has been tested with python 3.6, PyTorch 1.4 and CUDA 10.0 on Ubuntu 18.04. To install do

pip3 install torch==1.4.0+cu100 torchvision==0.5.0+cu100 -f <https://download.pytorch.org/whl/torch_stable.html>
Clone this repository: git clone git@github.com:jsll/pytorch_6dof-graspnet.git.
Clone pointnet++: git@github.com:erikwijmans/Pointnet2_PyTorch.git.
Run cd Pointnet2_PyTorch && pip3 install -r requirements.txt
cd pytorch_6dof-graspnet
Run pip3 install -r requirements.txt to install necessary python libraries.
(Optional) Download the trained models either by running sh checkpoints/download_models.sh or manually from here. Trained models are released under CC-BY-NC-SA 2.0.

Disclaimer

The pre-trained models released in this repo are retrained from scratch and not converted from the original ones https://github.com/NVlabs/6dof-graspnet trained in Tensorflow. I tried to convert the Tensorflow models but with no luck. Although I trained the new models for a substantial amount of time on the same training data, no guarantees to their performance compared to the original work can be given.

Updates

In the paper, the authors only used gradient-based refinement. Recently, they released a Metropolis-Hastings sampling method which they found to give better results in shorter time. As a result, I keep the Metropolis-Hastings sampling as the default for the demo.

This repository also includes an improved grasp sampling network which was proposed here https://github.com/NVlabs/6dof-graspnet. The new grasp sampling network is trained with Implicit Maximum Likelihood Estimation.

Update 9th June 2020

I have now uploaded new models that are trained for longer and until the test loss flattened. The new models can be downloaded in the same way as detailed in step 7 above.

Demo

Run the demo using the command below

python -m demo.main

Per default, the demo script runs the GAN sampler with sampling based refinement. To use the VAE sampler and/or gradient refinement run:

python -m demo.main --grasp_sampler_folder checkpoints/vae_pretrained/ --refinement_method gradient

Dataset

Get ShapeNet Models

Download the meshes with ids written in shapenet_ids.txt from https://www.shapenet.org/. Some of the objects are in ShapenetCore and ShapenetSem.

Prepare ShapeNet Models

Clone and build: https://github.com/hjwdzh/Manifold
Create a watertight mesh version assuming the object path is model.obj: manifold model.obj temp.watertight.obj -s
Simplify it: simplify -i temp.watertight.obj -o model.obj -m -r 0.02

Download the dataset

The dataset can be downloaded from here. The dataset has 3 folders:

grasps folder: contains all the grasps for each object.
meshes folder: has the folder for all the meshes used. Except cylinder and box the rest of the folders are empty and need to be populated by the downloaded meshes from shapenet.
splits folder: contains the train/test split for each of the categories.

Verify the dataset by running python grasp_data_reader.py to visualize the evaluator data and python grasp_data_reader.py --vae-mode to visualize only the positive grasps.

Training

To train the grasp sampler (vae or gan) or the evaluator with bare minimum configurations run:

python3 train.py  --arch {vae,gan,evaluator}  --dataset_root_folder $DATASET_ROOT_FOLDER

where the $DATASET_ROOT_FOLDER is the path to the dataset you downloaded.

To monitor the training, run tensorboard --logdir checkpoints/ and click http://localhost:6006/.

For more training options run GAN Training Example Command:

python3 train.py  --help

Quantitative Evaluation

I have not converted the code for doing quantitative evaluation https://github.com/NVlabs/6dof-graspnet/blob/master/eval.py to PyTorch. I would appreciate it if someone could convert it and send in a pull request.

Citation

If you find this work useful in your research, please consider citing the original authors' work:

inproceedings{mousavian2019graspnet,
  title={6-DOF GraspNet: Variational Grasp Generation for Object Manipulation},
  author={Arsalan Mousavian and Clemens Eppner and Dieter Fox},
  booktitle={International Conference on Computer Vision (ICCV)},
  year={2019}
}

as well as my implementation

@article{pytorch_6dof-graspnet,
      Author = {Jens Lundell},
      Title = {6-DOF GraspNet Pytorch},
      Journal = {<https://github.com/jsll/pytorch_6dof-graspnet},>
      Year = {2020}
}

mrkulk / pytorch_6dof-graspnet