Robot vision system with self-learning mechanism

This article describes a self-learning robot vision system to identify the target position. We try to use a neural network as a decision maker that determines how to move the robot to reach the exact target on the base of the image acquired by the robot vision. Moreover, we try to teach the above function automatically to the neural network: We set the target at the known fixed position, and the robot moves randomly and acquires the target images from each position where it stands. In this situation, the relative locations between the target and the robot are known; thus, these data are fed to the neural network as input images and teaching signals. The total system works as follows: (1) we set a target object at a known position, and teach the position to the system; (2) the robot moves randomly around the target and the neural network learns the relation between the relative positions and images; and (3) after enough learning, the robot can catch the target located at an arbitrary position.