What are GQ-CNNs?

GQ-CNNs are neural network architectures that take as input a depth image and grasp, and output the predicted probability that the grasp will successfully hold the object while lifting, transporting, and shaking the object.

../_images/gqcnn1.png

Original GQ-CNN architecture from Dex-Net 2.0.

../_images/fcgqcnn_arch_diagram.png

Alternate faster GQ-CNN architecture from FC-GQ-CNN.

The GQ-CNN weights are trained on datasets of synthetic point clouds, parallel jaw grasps, and grasp metrics generated from physics-based models with domain randomization for sim-to-real transfer. See the ongoing Dexterity Network (Dex-Net) project for more information.