Collision Detection and Avoidance for Multi-UAV based on Deep Reinforcement Learning

In recent years, the demand for improving the autonomy of UAVs has continued to increase in the civilian and military fields. Collision detection and avoidance is one of the key technologies to this end. In this paper, we propose a fully-distributed collision detection and avoidance method for multi-UAV based on deep reinforcement learning. Different from traditional methods, we implement an end-to-end control, which takes the information of sensor, UAV status and destination as inputs, and directly outputs the control references. Besides, based on the PPO algorithm and the paradigm of centralized training and decentralized execution, we provide the design of the deep reinforcement leaning network and the policy update strategy. In addition, we build a series of training and verification environments, including 2D Stage scenes and 3D Gazebo scenes. The experiment results show that our method can successfully avoid the obstacles and achieve no collision between UAVs.