The aim of this dissertation is to show that dynamic control of quadruped locomotion is achievable through the use of genetically evolved central pattern generators. This strategy is tested both in simulation and on a walking robot. The design of the walker has been chosen to be statically unstable, so that during motion less than three supporting feet may be in contact with the ground.

The control strategy adopted is capable of propelling the artificial walker at a forward locomotion speed of ~1.5 Km/h on rugged terrain and provides for stability of motion. The learning of walking, based on simulated genetic evolution, is carried out in simulation to speed up the process and reduce the amount of damage to the hardware of the walking robot. For this reason a general-purpose fast dynamic simulator has been developed, able to efficiently compute the forward dynamics of tree-like robotic mechanisms.

An optimization process to select stable walking patterns is implemented through a purposely designed genetic algorithm, which implements stochastic mutation and cross-over operators. The algorithm has been tailored to address the high cost of evaluation of the optimization function, as well as the characteristics of the parameter space chosen to represent controllers.

Experiments carried out on different conditions give clear indications on the potential of the approach adopted. A proof of concept is achieved, that stable dynamic walking can be obtained through a search process which identifies attractors in the dynamics of the motor-control system of an artificial walker.