Prompt-Based Lightweight LLM Navigation for GPS-Denied UAVs with Simulated Visual-Inertial Fusion

Abstract

Autonomous navigation in GPS-denied environments remains a major challenge for unmanned aerial vehicles (UAVs), especially in cluttered scenarios requiring both high-level decision-making and real-time control. This paper presents an LLM-guided navigation framework for a single UAV that combines prompt-based semantic action generation with visual-inertial odometry, front-depth sensing, online SLAM, and obstacle-aware motion planning. To improve real-time applicability, the framework is extended with a drone emulation platform implemented on the FPGA-based Kria KR260, enabling hardware-oriented evaluation of navigation and control functions. Experimental results in indoor-like GPS-denied environments demonstrate accurate path tracking, reliable obstacle avoidance, and robust overall navigation performance, underscoring the promise of FPGA-assisted, LLM-driven UAV navigation.