Reinforcement learning approach to acquisition of stable gaits for locomotion robots

Emergence of motion patterns in locomotion robots is studied. Acquisition of stable periodical gaits can be organized by learning how to reach a goal position. Classifier systems are used for sensory motor control of individual legs. During the learning process, the classifiers are implicitly coordinated by sharing the total sensor space of the robot. The proposed approach is tested under simulation and experiment on a special four-legged robot. It is shown that periodical gaits emerge as a result of interaction between the four classifier systems.