Robot and locomotion-controller design optimization for a reconfigurable quadruped

We present an automated approach to robot and locomotion-controller design optimization, using reinforcement learning methods that have been successfully demonstrated to teach a real prototype quadruped various walking gaits. The same machine learning methods are used here for a different purpose: to optimize robot and locomotion-controller design. Optimization can be used before or after building the robot, to reconfigure it for improved performance in a specific terrain or learning task. We propose measures for evaluating: (1) different controller symmetry assignments – couplings among the robot’s motors; (2) different lower-leg lengths. The measures are constructed with respect to the reward function of a simple learning task – better designs are those that achieve higher rewards. Averaged learning curves of each design are also used as an evaluation measure. We also present manual design considerations of the quadruped, which was mainly created as a test bed for machine learning methods for rugged terrain locomotion.