A DEEP REINFORCEMENT LEARNING APPROACH TO FLOCKING AND NAVIGATION OF UAVS IN LARGE-SCALE COMPLEX ENVIRONMENTS

摘要

This paper aims at enabling unmanned aerial vehicles (UAV) to flock and meanwhile perform navigation tasks in large-scale complex environments in a fully decentralized manner. By incorporating the insights of flocking control inspired by bird flocking in nature, the problem is structured as a Markov decision process and solved by deep reinforcement learning. In particular, coordination among agents is achieved by following a local interaction protocol that each agent only considers the relative position of the nearest two neighbors on its left side and right side. In addition, a flocking control-inspired reward scheme is designed for the emergence of flocking and navigation behaviors. Simulation results demonstrate that by training with three UAVs, the learned policy, shared across all agents, can enable a larger number of UAVs to perform navigation tasks as a group in large-scale complex environments.