Deep Reinforcement Learning Approaches for Sensor Data Collection by a Swarm of UAVs

The number of applications that use IoT devices as a sensing platform is ever‐increasing. A major obstacle to the flexibility of these systems is the requirement for constant connectivity to the sensing devices, constraining the applications to scenarios where network infrastructure is present. Autonomous unmanned aerial vehicles (UAVs) can serve as data harvesters, flying to the device locations, collecting their data using local network connectivity, and transporting this information to a data sink. Applications benefit from up‐to‐date data; thus, UAV data collection should strive for speed. A single UAV's capabilities are limited. A swarm has the benefit of allowing collaborative work, potentially decreasing collection times. This work proposes four decentralized deep reinforcement learning approaches to the problem of coordinating intelligent UAV agents. The approaches investigate noncollaborative and collaborative solutions and propose ways to increase the scalability of RL systems. Results on simulated scenarios show that all proposed algorithms can perform the task of data collection. Collaborative algorithms suffer from higher training times but offer better results. Training time can be reduced by training algorithms that are more general than variations in the data collection scenario. Results also show that there is a slight trade‐off in performance for algorithm generality.