High-Level Semantic Instruction-Guided UAV Area Search Using LLM Agents

Abstract

Area search is a fundamental capability for unmanned aerial vehicles, but practical missions often begin with high-level semantic instructions and incomplete environmental information rather than a fully specified geometric search region. Large language model agents can interpret such instructions and interact with external tools, yet they are unreliable when directly responsible for long-horizon geometric coverage planning. This paper presents SAGE-UAV, a semantic agent-guided exploration framework for UAV area search that separates semantic decision making from low-level coverage path generation. The agent interprets mission intent, maintains task-level environment memory, invokes a deterministic area coverage planner, receives execution feedback, and iteratively refines the search strategy. The coverage planner supports circular and polygonal regions, single- and multi-UAV execution, and workload-aware lane partitioning for cooperative search. We evaluate SAGE-UAV on the Area Search subset of MultiUAV-Plat, covering 500 tasks across five difficulty levels. SAGE-UAV achieves a 69.4% task pass rate, outperforming the evaluated ReAct baseline by 32.4 percentage points. It also reaches an 85.0% average check pass rate. These results demonstrate the end-to-end effectiveness of the complete SAGE-UAV system in the evaluated partially observable benchmark.