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.
We’re excited to announce version 1.0, which brings the GQ-CNN package up to date with recent research in Dex-Net. Version 1.0 introduces support for:
Dex-Net 4.0: Composite policies that decide whether to use a suction cup or parallel-jaw gripper.
Fully Convolutional GQ-CNNs: Fully convolutional architectures that efficiently evaluate millions of grasps faster than prior GQ-CNNs.
The goals of this project are to facilitate:
Research extensions on novel GQ-CNN architectures that have higher performance on Dex-Net 4.0 training datasets.
Our longer-term goal is to encourage development of robust GQ-CNNs that can be used to plan grasps on different hardware setups with different robots and cameras.
- GQ-CNN models are sensitive to the following parameters used during dataset generation:
The robot gripper
The depth camera
The distance between the camera and workspace.
As a result, we cannot guarantee performance of our pre-trained models on other physical setups. If you have a specific use-case in mind, please reach out to us. It might be possible to generate a custom dataset for you particular setup. We are actively researching how to generate more robust datasets that can generalize across robots, cameras, and viewpoints.
The package is currently under active development. Please raise all bugs, feature requests, and other issues under the Github Issues. For other questions or concerns, please contact Jeff Mahler (email@example.com) or Vishal Satish (firstname.lastname@example.org) with the subject line starting with “GQ-CNN Development”.
If you use the code, datasets, or models in a publication, please cite the appropriate paper:
Dex-Net 2.0 (bibtex)
Dex-Net 2.1 (bibtex)
Dex-Net 3.0 (bibtex)
Dex-Net 4.0 (bibtex)