Abstraction-Guided Sampling for Motion Planning

ion-guided Sampling for Motion Planning University of New Hampshire Department of Computer Science Technical Report 12-01 Scott Kiesel, Ethan Burns and Wheeler Ruml May 1, 2012 Abstraction-guided Sampling for Motion Planningion-guided Sampling for Motion Planning Scott Kiesel and Ethan Burns and Wheeler Ruml Abstract Motion planning in continuous space is a fundamental robotics problem that has been approached from many perspectives. Rapidly-exploring Random Trees (RRTs) use sampling to efficiently traverse the continuous and high-dimensional state space. Heuristic graph search methods use lower bounds on solution cost to focus effort on portions of the space that are likely to be traversed by low-cost solutions. In this paper, we bring these two ideas together in a technique called f -biasing: we use estimates of solution cost, computed as in heuristic search, to guide sparse sampling, as in RRTs. Estimates of solution cost are quickly computed using an abstract version of the problem, then an RRT is constructed by biasing the sampling toward areas of the space traversed by low cost solutions under the abstraction. We show that f -biasing maintains all of the desirable theoretical properties of RRT and RRT*, such as completeness and asymptotic convergence to optimality. We also present experimental results showing that f -biasing finds cheaper paths faster than previous techniques. We see this new technique as strengthening the connections between motion planning in robotics and combinatorial search in artificial intelligence.Motion planning in continuous space is a fundamental robotics problem that has been approached from many perspectives. Rapidly-exploring Random Trees (RRTs) use sampling to efficiently traverse the continuous and high-dimensional state space. Heuristic graph search methods use lower bounds on solution cost to focus effort on portions of the space that are likely to be traversed by low-cost solutions. In this paper, we bring these two ideas together in a technique called f -biasing: we use estimates of solution cost, computed as in heuristic search, to guide sparse sampling, as in RRTs. Estimates of solution cost are quickly computed using an abstract version of the problem, then an RRT is constructed by biasing the sampling toward areas of the space traversed by low cost solutions under the abstraction. We show that f -biasing maintains all of the desirable theoretical properties of RRT and RRT*, such as completeness and asymptotic convergence to optimality. We also present experimental results showing that f -biasing finds cheaper paths faster than previous techniques. We see this new technique as strengthening the connections between motion planning in robotics and combinatorial search in artificial intelligence.