A Multi-Agent Deep Reinforcement Learning Approach for Practical Decentralized UAV Collision Avoidance

Abstract

This letter proposes an improved deep reinforcement learning controller for the decentralized collision avoidance problem. Using a unique architecture incorporating ‘long-short term memory cells’ and a reward function inspired from gradient-based approaches, the controller outperforms existing techniques in environments with variable numbers of agents. The design of the reward function is also evaluated to critique existing deep reinforcement learning approaches and used to establish a better result. The proposed controller is subsequently tested in simulation and a real-world 3-dimensional drone environment: outperforming the predominant classical approach in the literature by a significant margin.

Topics & Concepts