Generation and Optimization of Motor Behaviors in Real and Simulated Robots
In this thesis, the problems of generating and optimizing motor behaviors for both simulated and real, physical robots have been investigated, using the paradigms of evolutionary robotics and behavior-based robotics. Specifically, three main topics have been considered: (1) On-line evolutionary optimization of hand-coded gaits for real, physical bipedal robots. The evolved gaits significantly outperformed the
hand-coded gaits, reaching up to 65% higher speed. (2) Evolution of bipedal gait controllers in simulators. First, linear genetic programming was used with two different simulated bipedal robots. In both these cases, the gait controller was evolved starting from programs consisting of random sequences of basic instructions. The best evolved programs generated stable bipedal locomotion, keeping the robot upright and moving indefinitely. However, the evolved gaits were not very human-like. Thus, a different approach, inspired by the neural mechanisms involved in the locomotion of biological organisms, was tried. Here, both the structure and parameters of a central pattern generator network, controlling the locomotion of a simulated robot, were optimized using a genetic algorithm. The evolved controllers generated a stable human-like gait and were also able to handle gait transitions. (3) Behavior selection in autonomous robots, using the utility function method. In particular, the performance of the method as a function of the polynomial degree of the utility functions was investigated. It was found that adequate behavior selection systems can be found rapidly for low polynomial degrees (1-2), but also that the best solutions can only be obtained by using a higher polynomial degree (3-4). Furthermore, the performance of different evolutionary algorithms in connection with the utility function method was also investigated and, somewhat surprisingly, it was found that the standard method, employing a simple genetic algorithm, generally outperformed the modified methods.
linear genetic programming