The Repetition Roadmap for Repetitive Constrained Motion Planning

Abstract

We present the Repetition Roadmap, a motion planner that effectively exploits the repetitiveness of a set of tasks with small variations to efficiently compute new motions. The method learns an abstract roadmap of probability distributions for the configuration space of a particular task set from previous solution paths. We show how to construct the Repetition Roadmap by learning a Gaussian mixture model and connecting the distribution components based on the connectivity information of the prior paths. We present an algorithm that exploits the information in the Repetition Roadmap to guide the search for solutions of similar tasks. We illustrate our method in a maze, which explains the construction of the Repetition Roadmap and how the method can generalize over different environments. We show how to apply the Repetition Roadmap to similar constrained manipulation tasks and present our results including significant speedup in computation time when compared to uniform and adaptive sampling.