Hierarchical LLM framework boosts UAV navigation in 3D aerial highways

A team of researchers (Zijiang Yan, Hao Zhou, Wael Jaafar, et al.) have introduced a novel Hierarchical LLM-driven control framework for navigating multiple high-speed UAVs in 3D aerial highways within Integrated Terrestrial and Non-Terrestrial Networks (ITNTNs). The approach addresses the limitations of both Deep Reinforcement Learning (DRL)—which lacks zero-shot strategic reasoning—and Large Language Models (LLMs), which suffer from high inference latency unsuitable for real-time aerodynamic control. The framework uses a massive cloud-based LLM deployed on a High-Altitude Platform Station (HAPS) to manage slow-timescale global load balancing, while lightweight edge-LLMs on individual UAVs convert local observations into tactical sub-goals. A fast-timescale DRL controller then executes collision-free, handover-aware trajectories based on these sub-goals.

Simulation results show that this agentic architecture significantly reduces collision rates and improves aggregate system throughput compared to existing baselines. The paper has been accepted by IEEE GLOBECOM 2026 and is available on arXiv (2607.18604). This work represents a promising step toward practical, intelligent coordination of UAV swarms in complex, dynamic 3D airspace, combining the strategic reasoning of LLMs with the real-time control capabilities of DRL.