Semantic-Aware Model Cache and Resource Allocation for Air-Ground Embodied Intelligence: A Two-Timescale M2ADRL Algorithm

In emergency rescue missions, the perception of embodied agents (e.g., rescue robots) rely heavily on high-quality semantic information. However, deploying complex embodied semantic models within constrained Unmanned Aerial Vehicle (UAV) networks introduces severe resource conflicts. We propose a semantic-aware framework tailored for embodied intelligence networks, jointly optimizing model caching, user association, semantic compression factors, and computational resources to minimize total system delay while satisfying energy and semantic fidelity constraints. To solve this multi-timescale mixed-integer problem, we develop a Two-Timescale integrated Matching and Multi-Agent Deep Reinforcement Learning (2T-M2ADRL) algorithm. It utilizes Hierarchical Bipartite Matching (HBM) for discrete large timescale topologies and a Multi-Head Attention based MADDPG (MHA-MADDPG) for small timescale resource orchestration. Furthermore, we construct an offline surrogate model to bypass evaluating semantic similarity via Large Language Models (LLMs) during training. Extensive simulations confirm our framework outperforms baselines, ensuring optimal latency and semantic accuracy.