Active Inference-Enhanced DRL for UAV and Integrated Sensing, Communication, and Computing Enabled Metaverse Systems

Abstract

In this article, we investigate the problem of computational task offloading and unmanned aerial vehicle (UAV) path planning within an edge computing-enabled space–ground integrated metaverse system. Specifically, we conceptualize the UAV as a “Communication – Perception – Computing” integrated device that provides real-time services within the metaverse system. The ground base station (GBS) augments UAV processing capabilities, while the edge computing platform only provides assistance for GBS, forming a two-level offloading architecture. To satisfy both user and operator requirements simultaneously, our objective is to minimize energy consumption and delay to meet user quality of service demands, and also to minimize economic costs for virtual service providers (VSPs), by jointly optimizing the sensing duration, the trajectory and the flying time of the UAV, and the data offloading ratio of the UAV and the VSP. Given the dynamic nature of the metaverse system, we reformulate the problem as a Markov decision process and employ a deep reinforcement learning algorithm based on active inference to dynamically learn the optimal joint solution. Simulation results demonstrate that our approach achieves rapid convergence and performs effectively in terms of maximizing cumulative rewards, minimizing VSP costs, reducing system delays, and optimizing UAV energy consumption.