In this paper, we propose a new approach to evolve controllers of autonomous robots, and experimental results of its application to a real mobile robot are described as well. It is based on two concepts: Firstly, behavior of a system in environment is generated by combinations of multiple sensory-motor reflexes, and secondly, the system behaves and evolves under the direct influence of its environment, thus the system is expected to adapt well for its environmental situations with flexibility.
The sensory-motor reflexes were realized by logic circuits connecting sensors to motors, which are composed of AND, OR and NOR elements. The genetic evolution in the environment was employed to determine the connections among the elements, thus types of reflexes and the ways of their combinations being obtained automatically by the interactions between the environment and the system itself.
This algorithm was implemented in a real miniature mobile robot, Khepera, and several experiments were performed. Khepera successfully learned to navigate and avoid obstacles in test fields. In comparison to a conventional algorithm, the acquired behavior scored higher in the values of fitness functions.