AoI-Guaranteed UAV Crowdsensing: A UGV-assisted deep reinforcement learning approach

Abstract

Unmanned Aerial Vehicles (UAVs), with their excellent environmental adaptability and flexible maneuverability, are increasingly being deployed in smart city applications for data collection. The Age of Information (AoI) is essential in these applications. Prior research on AoI has predominantly focused on static task scenarios, often overlooking the dynamics of task arrivals. For this reason, we propose an unmanned ground vehicle (UGV)-assisted deep reinforcement learning approach (U-DRL), which employs key factors affecting AoI to mitigate the AoI of data in dynamic task scenarios. We use EfficientNet, a state-of-the-art neural network architecture, to effectively extract features from dynamic task arrival scenarios. Based on these features, we utilize an intrinsic reward module (IRM) to swiftly process the input encapsulating global information, optimizing flight paths and enabling the exploration of expanded areas by UAVs. In addition, we leverage the active mobility of UGVs to recharge UAVs timely, thereby maximizing the flight time of UAVs. Through an extensive series of experiments, we validate the effectiveness of U-DRL. The experimental results demonstrate that U-DRL outperforms comparative algorithms in key performance metrics, significantly reducing the AoI of data, with breakthroughs of 54.14% and 67.90% in two real-world maps.