Modeling and Learning Walking Gaits of Biped Robots

Abstract

This paper describes an open loop modeling of a walking gait by mimicking the human walking style. A parameterizable model for the leg and arm movement will be developed. For finding the parameters of these problem classes often machine learning approaches are used. Thus, several optimization techniques are discussed and finally Evolution Strategies chosen for the optimization process. The best fitting parameters like population size or the selection operator are then found out by doing walk evolution with different configurations of the strategy in a robot simulator. Finally the best performing strategy is used to evolve a forward walk on a real robot.