DGGCNN: An Improved Generative Grasping Convolutional Neural Networks

The traditional robot grasping detection methods suffer from unstable grasping accuracy and slow convergence rate of training. In this paper, a depth generative grasping convolutional neural networks (DGGCNN) is proposed. A modified convolutional neural network architecture is designed to output the grasp quality, angle and width of the target. A novel loss function is also defined to further optimize the training quality of the network. The Cornell dataset is then used to train the network. The results of the simulation show that the proposed method has a superior success rate of grasping compared with original generative grasping convolutional neural networks (GGCNN).