Collision- and Freezing-Free Navigation in Dynamic Environments Using Learning to Search

Abstract

While collision-free navigation could be done using existing rule-based approaches, it becomes more attractive to use learning from demonstration (LfD) approaches to ease the burden of tedious rule designing and parameter tuning procedures. In addition, in the freezing robot problem, once the environment surpasses a certain level of complexity, there may be no sufficient space for a robot to navigate using these planning or navigation approaches even with perfect predictions of moving entities. In this paper, it is argued that collision-free navigation in dynamic environments is learnable from demonstrations with proper feature sets without the use of a path planner. It is feasible to solve the freezing robot problem using the policies learned from demonstration. The simulation results demonstrate that the Learning to Search (LEARCH) approach with the proposed modification is capable of achieving collision- and freezing-free navigation in dynamic environments.