Application of deep reinforcement learning to uav fleet control

Abstract

The increasing presence of robots and unmanned systems as a result of technological breakthroughs and falling costs has increased the demand for robust and scalable multi-agent control systems. We developed a multi-UAV fleet control system based on recent findings in the deep reinforcement learning literature. A deep convolutional neural network with a linear output layer is chosen as control policy, due to its wide spread applicability, and is trained, in simulation, for two tasks: aerial surveillance and base defense, with five UAVs. The generalization power of the architecture with respect to different fleet sizes was evaluated. For both tasks, at test time, we varied the number of UAVs from one to ten and we found that for all settings the policy was able to accomplish both tasks robustly. We deployed the control policy on a fleet of five DJI Mavic Pro drones and found that it performed well.