Efficient Extreme Motion Planning by Learning from Experience

Abstract

In this paper, we propose the segment-based roadmap (SRM) method for extreme motion planning. Unlike existing roadmap-based approaches, each vertex in the SRM contains a sequence of configurations. This segment-based motion planning can effectively handle the narrow passage problem caused by stability constraints in a high dimensional space. The SRM is generated from trajectory examples and trajectory optimization. We extract motion segments from the trajectory examples. The extracted motion segments and its connection is stored in vertex set and edge set, respectively. Furthermore, a trajectory optimization method is used to increase the con-nectivity of the SRM. In particular, a Gaussian random path (GRP) is used to initialize the trajectory optimization problem and shown to be more effective in terms of final cost as well as the running time. In simulation study, the average final cost using the GRP initialization shows 96.7% improvements compared to the initialization with linear interpolation which is often used in practice. In experiment study, we conducted experiments on NAO in order to verify the proposed motion planner using the SRM.