A CNN approach for controlling a roving robot

In this paper an approach based on Cellular Nonlinear Networks (CNNs) to control a roving robot is presented. The control system consists of two levels: a visual feedback plans the robot trajectory to reach a given target, and a low level controls the obstacle avoidance. Two different CNN structures are used for this purpose: visual feedback is implemented by a programmable 64x64 one-layer CNN, while the low level is implemented by a 3x3 two-layer CNN able to generate Turing patterns. Thanks to this vertical distribution of tasks, the algorithm for visual feedback is quite simple. At the same time the CNN for obstacle avoidance can be simply extended (by enlarging its size and keeping the same approach) to include many sensors and to perform an analogue data fusion of sensors in order to implement generic reactive behaviours.