AoI-minimal trajectory planning and data collection in UAVs crowdsensing by graph reinforcement learning

Abstract

In multi-Unmanned Aerial Vehicles (UAVs) data collection scenarios, it remains a critical challenge to jointly optimize time-sensitive responsiveness and maximize the volume of collected data. To address this issue, a graph reinforcement learning (RL) framework termed STIG-RL is proposed. The framework integrates an attention-based spatio-temporal interactive graph (ASTIG) network to extract efficient graph representations of the UAV-user state. Additionally, a hybrid decision-making strategy that combines model-based and model-free RL is employed to enhance policy efficiency. To further improve learning performance and reduce reliance on real-world experience, a simulated trajectory generation module (SRM) with a state prediction mechanism is incorporated. Extensive experimental evaluations demonstrate that the proposed approach outperforms five baseline algorithms in terms of data collection efficiency and responsiveness.