Extended QDSEGA for controlling real robots - acquisition of locomotion patterns for snake-like robot

Reinforcement learning is very effective for robot learning. It is because it does not need prior knowledge and has higher capability of reactive and adaptive behaviors. In our previous works, we proposed new reinforce learning algorithm: "Q-learning with dynamic structuring of exploration space based on genetic algorithm (QDSEGA)". It is designed for complicated systems with large action-state space like a robot with many redundant degrees of freedom. However the application of QDSEGA is restricted to static systems. A snake-like robot has many redundant degrees of freedom and the dynamics of the system are very important to complete the locomotion task. So application of usual reinforcement learning is very difficult. In this paper, we extend layered structure of QDSEGA so that it becomes possible to apply it to real robots that have complexities and dynamics. We apply it to acquisition of locomotion pattern of the snake-like robot and demonstrate the effectiveness and the validity of QDSEGA with the extended layered structure by simulation and experiment.