Visually guided movements: Learning with modular neural maps in robotics

Abstract

Robotics involves complex processing and requires modular controllers. For the connectionist approach, the adaptation of each module within the global system remains a major problem to be solved. This paper proposes the idea that biological learning can take advantage of the structures of the modules and the nature of modular decomposition. Therefore, we address this problem starting with the architecture of the system. We illustrate this approach using a robotic application: the visual servoing of the arm's end- effector. The on-line adaptation of a simple controller permits excellent results. To process several variables, and to limit the size of the memory required, this controller is decomposed into modules, in the image of sensorial or motor processing centers. The learning of the modules is realized on-line, a bi-directional architecture permits the adaptation of each module using a simple algorithm. The results obtained with various modular arrangements, both during intensive computer simulations and on our robotic platform, confirm the practical interest of this approach.