Realizing the quadruped robot walking by using reinforcement learning

We apply the reinforcement leashing to acquire a gait pattern of a quadruped locomotive robot. Where, no prescribed teaching signals as for supervised learnings but only scalar reinforcement signals are assumed. Tile advantage of tile reinforcement learning for such a problem is that we need no exact robot models for calculating prescribed teaching signals, but we simply need to evaluate results of trials and generate reinforcement signals. It is expected as a result that tile robot can acquire a walking pattern suitable to its structure, dynamics and environment by itself. We use here a well-known Actor-Critic learning method. It is applied to determine the value of amplitude and argument of assumed sinusoidal curves for angles of every joints. The computer simulations showed that it could generate various stable walking pattern suitable to the environment and dynamics of the robot. We also applied the proposed method to an experimental real robot and could ascertain the learning process for getting the walking pattern. As a conclusion, we have proposed a method of the reinforcement learning scheme which assumed the sinusoidal curve for each joint angle trajectory, and we have showed, by simulation and experiment, its effectiveness for generating the stable walking patterns for the quadruped robot.