Object recognition and robot grasping technology based on RGB-D data

Abstract

In this paper, a robot grasping method with object recognition and autonomous grasping ability based on RGB-D camera is designed. For object recognition, foreground extraction and point cloud clustering are proposed to realize object segmentation based on point cloud data and a set of object dataset. A kind of multiple modal convolution neural network model with dual channel is designed based on VGG and tested with homemade training dataset. For grasping planning, a heuristic nonuniform random grasp sample algorithm is presented according to the local reference frame and the local mean curvature of point clouds. The grasp candidates are scaled up from sample grasp pose by grid searching. The internal points in close region of every grasp hypothesis are encoded to an image and then the image is inputted into a simple convolutional neural network to evaluate the grasp success rate to rank the candidate set. The experimental results show that the proposed robot grasping method can realize object recognition and grasp objects accurately.