A Robotic Grasping Algorithm Based On Simplified Image And Deep Convolutional Neural Network

In this paper, a grasping method based on convolutional neural network (CNN) and image simplification is proposed to solve the problem of detecting the optimal grasping pose of a parallel-jaw gripper for unknown objects in an RGB-D view. First, the outlines of objects are simplified to remove blurred details caused by the 3D-sensing device. Further, the depth data at the edge of objects is clustered to simplify the complex 3D structure of objects. Then candidates for grasp on the image are selected by force closure and is sent into the Grasp Quality Convolutional Neural Network (GQ-CNN) to estimate the best estimation of grasping pose. A variety of common objects are used for the grasp experiment and the results demonstrate the effectiveness of the proposed method.