On Global Optimization of Walking Gaits for the Compliant Humanoid Robot, COMAN Using Reinforcement Learning

In ZMP trajectory generation using simple models, often a considerable amount of trials and errors are involved to obtain locally stable gaits by manually tuning the gait parameters. In this paper a 15 degrees of Freedom dynamic model of a compliant humanoid robot is used, combined with reinforcement learning to perform global search in the parameter space to produce stable gaits. It is shown that for a given speed, multiple sets of parameters, namely step sizes and lateral sways, are obtained by the learning algorithm which can lead to stable walking. The resulting set of gaits can be further studied in terms of parameter sensitivity and also to include additional optimization criteria to narrow down the chosen walking trajectories for the humanoid robot.