Multi-UAV trajectory optimizer: A sustainable system for wireless data harvesting with deep reinforcement learning

A wireless sensor network assisted by multiple autonomous unmanned aerial vehicles (UAVs) is a promising solution for harvesting data and monitoring the circumstance in various applications. However, the complicated path planning problem of each UAV is still problematic. In this paper, we propose an optimal operation strategy based on multi-agent reinforcement learning (MARL) to tackle these hurdles. Various parameters such as the number of deployed UAVs, charging start capacity, and charging complete capacity define a multi-UAV system. This approach is applicable without a time-consuming and costly policy control. We also describe how to balance multiple objectives, such as data harvesting, charging, and collision avoidance, using transfer learning. Finally, learning a policy control that generalizes multiple scenario parameters allows us to analyze the performance of individual parameters in a specific scenario, which helps find the macro-level optimal parameter within a particular scenario. Videos are available at https://github.com/mincheolseong/UAV-Trajectory-Optimizer.