Generative Robotic Grasping Using Depthwise Separable Convolution

Abstract In this paper, we present an end-to-end approach method using deep learning for grasp detection. Our method is a real-time processing method for discrete depth image sampling and the problems of long calculation times and difficulty in registration caused by object modelling and global searching in traditional methods. The method uses depthwise convolution and pointwise convolution to model the relations among the channels and directly parameterizes a grasp quality value for every pixel. Our method calculates a rectangular grasping box to generate a grasping pose for an input image. For the experimental evaluation on the Jacquard dataset, we compared the proposed method with other baseline methods, and the accuracy of the proposed method was improved by 5% to 7% that shows our method can effectively predict grasp points on novel class objects.