Grasping Prediction Algorithm Based on Full Convolutional Neural Network

Abstract

Abstract Robot grasping is a very frontier and important research direction in the field of robotics. In order to solve the problem of the robot's real-time capture, reduce the time of the visual processing, we proposed a two-stage Convolutional Neural Network structure whose design is simple, with less training parameters, improving the efficiency of the visual system. Using rotation and translation to expand the Cornell fetching dataset. The best model at Cornell grasp test set has achieved 88% of forecast accuracy compared with 73% accuracy rate on one stage network. Moreover, our model size is 0.51MB, speed at 30 FPS on GPU inferencing.