Generalization of optimal motion trajectories for a biped walking machine based on machine learning

Abstract

Optimal walking trajectories are essential for bipedal walking. Recent developments enabled the use of motion planer based on nonlinear optimization. They are however, not directly applicable in real-time tasks due to a high computation time. Therefore, a task space consisting of the stride-length and a step time is used to precompute corresponding trajectory parameters in a certain range regarding a cost function. The resulting trajectories define an optimal joint space motion. The mapping from the task space to the joint space is generalized with multiple machine learning methods. A parametrization of every method was determined to represent the underlying model of the data as good as possible, without over-fitting. Finally, the performance regarding the accuracy and runtime and the evaluation of the cost value and constraint violation in the walking tasks is discussed.