Reducing metric sensitivity in randomized trajectory design

This paper addresses the trajectory design for generic problems that involve: (1) complicated global constraints that include nonconvex obstacles, (2) nonlinear equations of motion that involve substantial drift due to momentum, and (3) a high-dimensional state space. Our approach to these challenging problems is to develop randomized planning algorithms based on rapidly-exploring random trees (RRTs). RRTs use metric-induced heuristics to conduct a greedy exploration of the state space; however, performance substantially degrades when the chosen metric does not adequately reflect the true cost-to-go. In this paper, we present a version of the RRT that refines its exploration strategy in the presence of a poor metric. Experiments on problems in vehicle dynamics and spacecraft navigation indicate substantial performance improvement over existing techniques.