Note: This is a cleaned-up, PyTorch port of the GG-CNN code. For the original Keras implementation, see the RSS2018 branch.
Main changes are major code clean-ups and documentation, an improved GG-CNN2 model, ability to use the Jacquard dataset and simpler evaluation.

Generative Grasping CNN (GG-CNN)

The GG-CNN is a lightweight, fully-convolutional network which predicts the quality and pose of antipodal grasps at every pixel in an input depth image. The lightweight and single-pass generative nature of GG-CNN allows for fast execution and closed-loop control, enabling accurate grasping in dynamic environments where objects are moved during the grasp attempt.

This repository contains the implementation of the Generative Grasping Convolutional Neural Network (GG-CNN) from the paper:

Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach

Douglas Morrison, Peter Corke, Jürgen Leitner

Robotics: Science and Systems (RSS) 2018

arXiv | Video

If you use this work, please cite:

@inproceedings{morrison2018closing,
	title={{Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach}},
	author={Morrison, Douglas and Corke, Peter and Leitner, J\"urgen},
	booktitle={Proc.\ of Robotics: Science and Systems (RSS)},
	year={2018}
}

Contact

Any questions or comments contact Doug Morrison.

Installation

This code was developed with Python 3.6 on Ubuntu 16.04. Python requirements can installed by:

pip install -r requirements.txt

Datasets

Currently, both the Cornell Grasping Dataset and Jacquard Dataset are supported.

Cornell Grasping Dataset

Download the and extract Cornell Grasping Dataset.
Convert the PCD files to depth images by running python -m utils.dataset_processing.generate_cornell_depth <Path To Dataset>

Jacquard Dataset

Download and extract the Jacquard Dataset.

Pre-trained Models

Coming Soon

Training

Training is done by the train_ggcnn.py script. Run train_ggcnn.py --help to see a full list of options, such as dataset augmentation and validation options.

Some basic examples:

# Train GG-CNN on Cornell Dataset
python train_ggcnn.py --description training_example --network ggcnn --dataset cornell --dataset-path <Path To Dataset>

# Train GG-CNN2 on Jacquard Datset
python train_ggcnn.py --description training_example2 --network ggcnn2 --dataset jacquard --dataset-path <Path To Dataset>

Trained models are saved in output/models by default, with the validation score appended.

Evaluation/Visualisation

Evaluation or visualisation of the trained networks are done using the eval_ggcnn.py script. Run eval_ggcnn.py --help for a full set of options.

Important flags are:

--iou-eval to evaluate using the IoU between grasping rectangles metric.
--jacquard-output to generate output files in the format required for simulated testing against the Jacquard dataset.
--vis to plot the network output and predicted grasping rectangles.

For example:

python eval_ggcnn.py --network <Path to Trained Network> --dataset jacquard --dataset-path <Path to Dataset> --jacquard-output --iou-eval

Running on a Robot

Our ROS implementation for running the grasping system see https://github.com/dougsm/mvp_grasp.

The original implementation for running experiments on a Kinva Mico arm can be found in the repository https://github.com/dougsm/ggcnn_kinova_grasping.

yuanyinlong/ggcnn