Teaching robot navigation behaviors to optimal RRT planners

Abstract

This work presents an approach for learning navigation behaviors for robots using Optimal Rapidly-exploring Random Trees (RRT*) as the main planner. A new learning algorithm combining both Inverse Reinforcement Learning (IRL) and RRT* is developed to learn the RRT* ’s cost function from demonstrations. A comparison with other state-of-the-art algorithms shows how the method can recover the behavior from the demonstrations. Finally, a learned cost function for social navigation is tested in real experiments with a robot in the laboratory.