MetaComm: Multi-Agent Reinforcement Learning Based Trajectory Planning in Multi-UAV Assisted Crowd Sensing Systems

In the context of emerging metaverse applications, unmanned aerial vehicle (UAV)-assisted mobile crowdsensing (MCS) systems have proven effective for high-quality data collection. These systems are capable of meeting strict quality-of-information (QoI) requirements, including criteria such as throughput and age-of-information (AoI). However, existing UAV trajectory planning methods that rely on predetermined or complete communication between UAVs often lead to unnecessary redundancy and struggle to achieve optimal performance in scenarios with partial observability. To address these challenges, this paper proposes a metaverse-oriented communication-assisted multi-agent deep reinforcement learning-based UAV trajectory planning method (MetaComm). First, a communication graph is established to describe the connectivity relationships among UAVs. This graph is parameterized by an attention mechanism and optimized through gradient-based bi-level learning to retain task-relevant inter-UAV links while reducing redundant communication. Under the centralized training and decentralized execution paradigm, a bi-level optimization approach is adopted: the inner loop focuses on training actor–critic policies for a given communication topology, while the outer loop adjusts the topology in response to QoI feedback and communication cost constraints. Extensive experiments on two real-world mobility datasets from NCSU and KAIST show that MetaComm consistently outperforms five baselines, achieving an average QoI improvement of 5.27% and a maximum improvement of 19.47% over the best-performing baseline.