Perceptive patterns for mobile robots via RD-CNN and reinforcement learning

In this paper we present a bio-inspired framework for sensing-perception-action of a roving robot, in a random foraging task. The core of this framework is the exploitation of Turing patterns to build a set of internal perceptive states, from sensorial inputs, to generate proper actions. To this aim a reaction diffusion cellular neural network (RD-CNN) is used. The basins of attraction of the Turing patterns are dynamically tuned by unsupervised learning in order to best match the sensor dynamics to the geometry of the pattern basins. Each pattern is associated with an action through reinforcement learning. The system is also provided with a contextual layer to realize a higher level control.