RealTime GPU-Based Motion Planning for Task Executions

摘要

Abstract—We present a realtime GPU-based motion plan-ning algorithm for robot task executions. Many task execution strategies break down a high-level task planning problem into multiple low-level motion planning problems, and it is essential to solve those problems at interactive rates. In order to achieve high performance for the planning, our method exploits a high number of cores on commodity graphics processors (GPUs). We describe a parallel formulation of an RRT-based motion planning algorithm which is highly suited for single query motion planning. Our approach uses the properties of Poisson-disk samples to achieve a high parallelism in order to exploit the computational capabilities of GPUs. Our approach can obtain 10-20X speedup over prior CPU-based motion planning algorithms, and we demonstrate the performance on a number of benchmarks. I.