Autonomous Vision-Guided Robot Manipulation Control

This paper presents a hand-eye-head system which learns to perform temporal actions. In this system, the robot learns how to control its hand according to what is seen and the speciic mission. The learning process is interactive and on-line. Several networks which store the learned information are automatically generated through learning processes. The hierarchical structure of the network allows each associative recall to be completed in O(log n) time, where n is the number of cases learned. A visual attention selection system is also implemented by using a recurrent network and an autonomous learning scheme. The recurrent network enables the learning process of the visual attention selection system to draw information not only from the current input image but also from the history. The new autonomous learning scheme combines both supervised and reinforcement learning modes so that the system gradually improves its performance through a process of trial and error.