Neural approach to visual servoing for robotic hand eye coordination

Abstract

This paper describes a neural network approach to visual servoing for the control of robot arm movements. The method is based solely on visual feedback, and requires no prior information about the kinematics of the robot or the placement or calibration of the cameras. A continuous error signal (in image coordinates) is used to move the manipulator to a visually specified target. A self-learning neural controller learns a Jacobian-based mapping between the visual error signal and the joint variations of the robot. The approach is illustrated with experiments of image-controlled end-effector displacement both in simulation and in a robot-vision system implementation.