Control of a legged rover for planetary exploration using embedded and evolved dynamical recurrent artificial neural networks

This paper presents a new method for realizing the control system of a legged rover for planetary exploration. The controller is realized using a class of dynamical recurrent artificial neural networks called CTRNN, and evolutionary algorithms. The proposed approach allows realizing the design of the controller in a modular way, decomposing the global problem into a collection of low-level tasks to be reached. The embodied dynamical neural network realized has been tested on a virtual legged hexapod called N.E.Me.Sys. The neural-controller has a high degree of robustness facing sensors noises and errors, tolerates a certain amount of degradation, but above all it allows the robot performing complex reactive behaviors, as overcoming hills and narrow valleys.