Adaptive Hierarchical RAG-Empowered LLM for UAV Scheduling in Multi-Access Edge Computing

Abstract

The proliferation of Multi-Access Edge Computing (MEC) has led to massive data generation. This imposes complex and dynamic requirements on task scheduling. Conventional Uncrewed Aerial Vehicle (UAV) scheduling methods struggle with task diversity. They lack adaptability to stochastic environments. Large Language Models (LLMs) provide powerful reasoning capabilities, while Retrieval-Augmented Generation (RAG) facilitates external knowledge integration. In this paper, we propose a RAG-empowered LLM framework for UAV scheduling. Proposed framework aims to minimize energy consumption and completion steps while maximizing rewards. We formulate the problem as a Mixed-Integer Non-Linear Program (MINLP). It is decoupled into discrete task allocation and continuous motion control. Our proposed framework parses natural language instructions via LLMs. To reduce latency, we design an adaptive hierarchical retrieval mechanism. This mechanism adjusts retrieval quantity based on task novelty. Finally, we introduce a dynamic management mechanism for long-term operations. It uses knowledge sparsity and policy contribution to prevent knowledge base bloat. Experiments show that proposed framework significantly outperforms baselines in energy consumption, scheduling efficiency, and rewards. It achieves up to 31.0% lower energy consumption, 13.5% lower decision latency, and 20% lower knowledge storage. Notably, it exhibits superior robustness and generalization under stochastic environments.