Embodied and Evolved Dynamical Neural Networks for Robust Planetary Navigation

The N.E.Me.Sys project has the aim of controlling a legged rover for planetary exploration using dynamical recurrent neural networks and evolutionary algorithms. This paper describes the realization of the navigation module of such a rover using a 2D chemiotaxis scenario, in which the agent must reach the source of a chemical signal. The analyses carried out in this work show the high degree of robustness of the neuro-controller versus uncertainties, noise, errors, or unpredicted situations. Moreover an analysis of the topology of the network has been realized in order to find the reasons of the good performances of the proposed methodology: it is possible to prove that different individuals share the same topology, i.e. the evolutionary process looks for the same feedback paths more than for the optimal set of parameters.