High level decomposition for bipedal locomotion planning

Abstract

Legged robotic platforms offer an attractive potential for deployment in hazardous scenarios that would be too dangerous for human workers. Legs provide a robot with the ability to step over obstacles and traverse steep, uneven, or narrow terrain. Such conditions are common in dangerous environments, such as a collapsing building or a nuclear facility during a meltdown. However, identifying the physical motions that a legged robot needs to perform in order to move itself through such an environment is particularly challenging. A human operator may be able to manually design such a motion on a case-by-case basis, but it would be inordinately time-consuming and unsuitable for real-world deployment. This thesis presents a method to decompose challenging large-scale motion planning problems into a high-level planning problem and a set of parallel low-level planning problems. We apply the method to quasi-static bipedal locomotion planning. The method is tested in a series of simulated environments that are designed to reflect some of the challenging geometric features that a robot may face in a disaster scenario. We analyze the improvement in performance that is provided by the high- and low-level decomposition, and we show that completeness is not lost by this decomposition.