Robotic Grasp Detection Using Light-weight CNN Model

Abstract

Manipulating objects is the premise of many robotic applications, such as assembly, sort and palletezing. Teach programming is widely used in industrial robot, it is necessary to improve the intelligent level of robot. In this paper, we propose an accurate, real-time, light-weight convolutional neural network SqueezeNet-RCM to predict a grasp configuration from a RGB and aligned depth image of graspable object. Instead of using sliding window or regional proposal network to generate potential bounding box, we use a way of end-to-end to train and test our network. On the standard Cornell Grasping Dataset, our model achieves accuracy of 90.1% and 88.6% on image-wise split and object-wise split respectively, speed at 73 frames per second(fps) on GPU inferencing, which could meet the requirement of real-time. Moreover, our model size is 2.9MB, which is able to fit the memory to limited environment such as FPGA.