In order to further close the gap between robots and their natural counterparts, current research is driving towards exploiting the natural dynamics of robots. This requires a rethinking about the way we design and control robots for moving in a more dynamic, efficient and natural way.
Dynamic bipedal locomotion is a challenging problem and remains an open area of research. A key characteristic is the decoupling between the center of mass and the multibody dynamics. Particular difficulties arise from effective underactuation, the mechanism complexity, as well as nonlinear and hybrid dynamics. While model-free approaches, like reinforcement learning, made significant progress in simulated environments in recent years, most of the remarkable improvements achieved on real humanoids are devoted to model-based optimal control methods. A direct approach to generate dynamic walking gaits for underactuated systems is to design controllers that produce trajectories equivalent to those of passive dynamic walkers. Consequently, trajectory optimization is considered a promising approach for generating dynamic and energy efficient trajectories for high-dimensional robotic systems.
This master’s thesis aims to contribute to the research field of dynamic bipedal locomotion by applying, evaluating and extending recently presented optimal control approaches on a new humanoid robot with series-parallel mechanisms. In a first phase, we aim to generate feasible trajectories for bipedal walking using trajectory optimization. This involves analyzing and identifying suitable state of the art methods, applying them to the RH5 robot and evaluating the offline-generated trajectories on the real system. In the second phase, we extend the analysis to more complex, highly-dynamic maneuvers. By this, we want to identify the shortcomings of the approach in terms of control performance and system design, and derive appropriate suggestions to guide future investigations.