Berkeley AUTOLAB’s GQCNN Package¶
The gqcnn package is a Python API for training and deploying Grasp Quality Convolutional Neural Networks (GQ-CNNs) for grasp planning using training datasets from the Dexterity Network (Dex-Net), developed by the Berkeley AUTOLAB and introduced in the Dex-Net 2.0 paper.
The goals of this project are to facilitate:
- Replicability of GQ-CNN training from the Dex-Net 2.0 paper.
- Research extensions on novel GQ-CNN architectures that have higher performance on Dex-Net 2.0 training datasets.
Our longer-term goal is to encourage development of GQ-CNNs that can be used to plan grasps on different hardware setups with different robots and cameras.
The gqcnn package currently supports only training of GQ-CNN on Dex-Net 2.0 datasets. We are working toward a grasp planning ROS service based on GQ-CNNs to work toward GQ-CNNs that work on other robot hardware setups.
Please note that performance on current datasets is not indicative of performance on other hardware setups because our datasets are specific to:
- Our ABB YuMi parallel-jaw gripper with custom fingertips due to collision geometry.
- A Primense Carmine 1.09 due to camera parameters.
- A camera positioned between 50-70cm above a table looking down due to image rendering parameters.
We are currently researching how to generate datasets that can generalize across robots, cameras, and viewpoints.
The package is currently under active development. Installation has been tested on Ubuntu 12.04, 14.04, and 16.04.
Please raise all bugs, feature requests, and other issues under the Github Issues. For other questions or concerns, please contact Jeff Mahler (firstname.lastname@example.org) with the subject line starting with “gqcnn development: “
If you use the code, datasets, or models in a publication, please cite the Dex-Net 2.0 paper.