Planning Swift Maneuvers of Quadcopter Using Motion Primitives Explored by Reinforcement Learning

Abstract

In this work, we propose a novel, learning-based approach for swift maneuver planning of unmanned aerial vehicles using motion primitives. Our approach is composed of two main stages: learning a set of motion primitives during offline training first, and utilization of them for online planning of fast maneuvers thereafter. We propose a compact disposition of motion primitives which consists of roll, pitch, and yaw motions to build up a simple yet effective representation for learning. Thanks to this compact representation, our method retains an easily transferable, reproducible, and referable knowledge which caters for real-time swift maneuver planning. We compare our approach with the current state-of-the-art methods for planning and control, and show improved navigation time performance up to 25% in challenging obstacle courses. We validate our approach through software-in-the-loop Gazebo simulations and real flight tests with Diatone FPV250 Quadcopter equipped with PX4 FMU.