An Efficient Robotic Grasping Pipeline Base on Fully Convolutional Neural Network

Abstract

In this paper, we present a pipeline for grasping unknown objects based on RGB-D images from a stereo camera mounted on the wrist of the robot arm. The proposed grasping pipeline composes of a fully convolutional neural network (FCNN) and a Simplified Grasp Pose Detection (S-GPD). The FCNN is used to generate grasping candidates accurately from a four channel image synthesized by RGB and depth image, and the S-GPD method is responsible for scoring the candidates in the point cloud. To evaluate the performance of our proposed grasping pipeline, we demonstrate the proposed pipeline in Kinova Mico <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> robotic arm to execute grasping tasks for a single and multiple objects. The experimental results show that our grasp pipeline reaches a grasping success rate of 85.8% for a single object and 84.3% for multi-objects. Furthermore, the proposed pipeline achieves a balance between effectiveness and efficiency compared with other advanced grasping methods.