Deep regression of convolutional neural network applied to resolved acceleration control for a robot manipulator

Abstract

This paper presents a modified resolved acceleration control scheme based on deep regression of the convolutional neural network. The resolved acceleration control scheme can achieve precise motion control of robot manipulators by regulating the accelerations of the end-effector, and the conventional scheme needs the position and orientation of the end-effector, which are obtained through the direct kinematics of the robot manipulator. This scheme increases the computational loads and might obtain inaccurate position and orientation due to mechanical errors. To overcome the drawbacks, a camera is used to capture the images of the robot manipulator, and then a deep regression of convolutional neural network is imposed into the resolved acceleration control to obtain the position and orientation of the end-effector. The proposed approach aims to enhance the positioning accuracy, to reduce the computational loads, and to facilitate the deep regression in real-time control. In this study, the proposed approach is applied to a 3-DOF planar parallel robot manipulator, and the results are compared with those by the conventional resolved acceleration control and a visual servo-based control. The results show that those objectives are achieved. Furthermore, the robustness of the proposed approach is tested through only the partial image of the end-effector available, and the proposed approach still works functionally and effectively.