Discrete Search Leading Continuous Exploration for Kinodynamic Motion Planning

This paper presents the Discrete Search Leading continuous eXploration (DSLX) planner, a multi-resolution approach to motion planning that is suitable for challenging problems involving robots with kinodynamic constraints. Initially the method decomposes the workspace to build a graph that encodes the physical adjacency of the decomposed regions. This graph is searched to obtain leads, that is, sequences of regions that can be explored with sampling-based tree methods to generate solution trajectories. Instead of treating the discrete search of the adjacency graph and the exploration of the continuous state space as separate components, DSLX passes information from one to the other in innovative ways. Each lead suggests what regions to explore and the exploration feeds back information to the discrete search to improve the quality of future leads. Information is encoded in edge weights, which indicate the importance of including the regions associated with an edge in the next exploration step. Computation of weights, leads, and the actual exploration make the core loop of the algorithm. Extensive experimentation shows that DSLX is very versatile. The discrete search can drastically change the lead to reflect new information allowing DSLX to find solutions even when sampling-based tree planners get stuck. Experimental results on a variety of challenging kinodynamic motion planning problems show computational speedups of two orders of magnitude over other widely used motion planning methods.