Large Language Model-Driven Autonomous UAV Systems: Technical Evolution, Core Architectures, and Critical Challenges
TL;DR
This survey reviews the technical evolution, system architectures, and deployment challenges of LLM-driven UAVs across perception, planning, control, multi-agent coordination, and edge–cloud computing, and separates semantic-reasoning latency, control timing, power, task outcomes, hardware, and validation settings to avoid misleading cross-platform comparisons.
Abstract
The rapid development of large language models (LLMs) has expanded the capabilities of autonomous unmanned aerial vehicle (UAV) systems in naturallanguage instruction understanding, multimodal perception, and decision-making. This survey reviews the technical evolution, system architectures, and deployment challenges of LLM-driven UAVs across perception, planning, control, multi-agent coordination, and edge–cloud computing. Beyond cataloguing representative systems, we separate semantic-reasoning latency, control timing, power, task outcomes, hardware, and validation settings to avoid misleading cross-platform comparisons. We further analyze Sim2Real gaps, intermittent connectivity, hallucination-induced action risk, cyberattacks, and privacy leakage. In this survey, a semantically adaptive safety certificate denotes a runtime-verifiable CBF/MPC/STL constraint whose safe set or margin is parameterized by grounded task and scene semantics but enforced by a deterministic safety layer outside the generative model. Future directions emphasize hierarchical lightweight reasoning, communication-aware autonomy, formally bounded semantic adaptation, and airworthiness-oriented assurance.