Techniques de contrôle bio-inspirées: application à la simulation réaliste des comportements de créatures artificielles

Abstract

Autonomous robots have always attracted attention of researchers around the world. The main challenge of traditional autonomous robotics is to design software for the given physical robot in such way, that the robot equipped with this software is able to autonomously perform given tasks. Several approaches have been introduced for designing robotic control systems. In the area of manually-designed robotic controllers, methods range from classical planning techniques to reactive planning. Another approach is to create robotic controllers automatically, training them for a given task in a simulated environment. In this thesis, we present the major research in artificial life and virtual reality inspired by the natural evolution to automatically design of artificial creature’s controller. In this way the robots can exploit the physical dynamics of their environment to generate behaviour. We begin with a set of artificial creature experiments, in which the creature's body is fixed, and only the controller is optimized using simulated evolution. We proposed two evolutionary approaches to design the controller. The first provides two key features: (a) the topology of the neural controller gradually grows in size to allow increasingly complex behaviours, and (b) the evolutionary process requires only the physical properties of the creature model and a simple fitness function. No a priori knowledge of the appropriate cycles or patterns of motion is needed. Experiments have shown that the proposed approach significantly increases the performance of the evolution on most tested tasks. The second approach is presented in the last section, in which a model of artificial development, based on genetic regulatory networks (GRNs), is introduced to evolve the controller system of artificial creatures. It is shown that gene regulatory network model may possibly be a viable solution for evolving control solutions for physical machines.