Decentralized Planning-Assisted Deep Reinforcement Learning for Collision and Obstacle Avoidance in UAV Networks

Abstract

In this paper, we propose using a decentralized planning-assisted approach of deep reinforcement learning for collision and obstacle avoidance in UAV networks. We focus on a UAV network where there are multiple UAVs and multiple static obstacles. To avoid hitting obstacles without severely deviating from the ideal UAV trajectories, we propose merging adjacent obstacles based on convex hulls and design a novel trajectory planning algorithm. For UAVs to efficiently avoid collisions in a distributed manner, we propose using a decentralized multi-agent deep reinforcement learning approach based on policy gradients. In addition, we propose using a priority-based algorithm for avoiding collisions without reducing the speeds of UAVs too much. Simulation results show that the proposed decentralized planning-assisted deep reinforcement learning approach outperforms a number of baseline approaches in terms of the probability that all UAVs successfully reach their goals within the deadline.