Faster Sample-Based Motion Planning Using Instance-Based Learning

摘要: We present a novel approach to improve the performance of sample-based motion planners by learning from prior instances. Our formulation stores results collision and local planning queries. This information is used accelerate based on probabilistic checking, select new paths in free space, compute an efficient order perform queries along search path graph. fast algorithms k-NN (k-nearest neighbor) high dimensional configuration spaces locality-sensitive hashing derive tight bounds their accuracy. The are instance-based have sub-linear time complexity. general, makes no assumption about sampling scheme, can be with various planners, including PRM, Lazy-PRM, RRT RRT*, making small changes these planners.We observe up 100% improvement rigid articulated robots.