Abstract
In the sixth-generation (6G) era, wireless networks need to support a large number of ultra-low latency and high-reliability applications. However, conventional bit-level communication paradigms fail to capture the intrinsic meaning of multi-modal data, leading to inefficiencies in both communication and computation for downstream tasks. To address this limitation, we propose a semantic-driven task offloading framework in low-altitude wireless networks (LAWNs), where multiple unmanned aerial vehicles (UAVs) provide on-demand edge computing services to ground terminals (GTs). Specifically, we employ a vector quantized-variational autoencoder (VQ-VAE) to enable joint coding and modulation (JCM) of cross-modal data. Then, we formulate an optimization problem that simultaneously determines UAV deployment, task offloading decisions, transmit power allocation, and computational resource scheduling, with the objective of maximizing the quality of experience (QoE) for GTs. To solve this problem, we employ the Karush-Kuhn-Tucker (KKT) conditions to address the UAV deployment subproblem, and utilize a multi-agent proximal policy optimization (MAPPO) approach to tackle the task offloading and resource allocation subproblem. Simulation results demonstrate that the proposed method significantly enhances QoE, achieving more than 5.87% improvement over representative benchmarks, while reduces task latency by more than 4.89% and improves energy efficiency by more than 3.76%.