Predictive robot control with neural networks

Abstract

. For a target tracking task, the hand-held camera of the anthropomorphic OSCAR-robot manipulator has to track an object which moves arbitrarily on a table. The desired camera-joint mapping is approximated by a feedforward neural network. Through the use of time derivatives of the position of the object and of the manipulator, the controller can inherently predict the next position of the moving target object. In this paper several `predictive' controllers are proposed, and successfully applied to track a moving object. Key Words. Robotics, Neural Networks, Visual Tracking 1. Introduction In [1] a real-time learning neural controller is described for the control of the anthropomorphic OSCAR-robot manipulator. An object, which is placed arbitrarily on a table, is observed by a camera in the end-effector of the manipulator. In order to position the camera above the object at a fixed distance, a feedforward network learns the camera-joint mapping. During operation, learning sample...