Abstract:
This paper examines the use of large language models (LLM) to control unmanned aerial vehicles (UAVs) using natural language commands. The research is aimed at solving a key problem – the discrepancy between the high computing requirements of LLM and the limited resources of on-board computers. The main focus is on optimizing LLM for operation on energy-efficient single-board computers with neuroprocessors, such as OrangePi 5B based on Rockchip RK3588S. The paper presents the results of testing Qwen2.5-Coder quantized models, demonstrating the preservation of code generation quality at processing speeds of up to 17.8 tokens/s. A specialized test (benchmark) was developed to evaluate the effectiveness of LLM integration into UAV autonomous control architecture and the correctness of code generation, including 125 scenarios. The results confirm the feasibility of LLM in autonomous drone control systems, although they reveal typical errors associated with sensor data processing and coordinate systems. The study proposes a promising direction for the development of intelligent UAV control systems with a natural language interface (NLP). The study included both a scientific approach (development of a specialized test) and a technological innovation (performance analysis on single-board computers) aimed at integrating LLM into UAV autonomous control architecture.
Keywords:unmanned aerial vehicle, large language model, quantization, single-board computers, autonomous control, natural language processing.
UDC:
004.93 + 004.8 BBK:
39.52
Received: May 13, 2025 Published: September 30, 2025
Fulltext:
PDF file (906 kB)