Guiding Sampling-Based Motion Planning by Forward and Backward Discrete Search

This paper shows how to effectively compute collision-free and dynamically-feasible robot motion trajectories from an initial state to a goal region by combining sampling-based motion planning over the continuous state space with forward and backward discrete search over a workspace decomposition. Backward discrete search is used to estimate the cost of reaching the goal from each workspace region. Forward discrete search provides discrete plans, i.e., sequences of neighboring regions to reach the goal starting from low-cost regions. Samplingbased motion planning uses the discrete plans as a guide to expand a tree of collision-free and dynamically-feasible motion trajectories toward the goal. The proposed approach, as shown by the experiments, offers significant computational speedups over related work in solving highdimensional motion-planning problems with dynamics.