《通用大语言模型：无人机指挥与控制接口》最新

Javier Ramos-Silva1,2*and Peter Burke2,31*Department of Electrical Engineering and Computer Science,University of California, Irvine, MS 2625, Irvine, 92697, CA, USA.2*BME, MSE, CBEMS, Physics, University of California, Irvine, MS2625, Irvine, 92697, CA, USA. *Corresponding author(s). E-mail(s): pburke@uci.edu;Contributing authors: pburke@uci.edu; Abstract The use of artificial intelligence (AI) for drone control can have a transforma-tive impact on drone capabilities, especially when integrated with drone sensing,command, and control, part of a growing field of physical AI. Large languagemodels (LLMs) for agentic drone control provide a new opportunity for droneautonomy. However, challenges remain in the interface, with each use case requir-ing a tedious, labor intensive effort to connect the LLM to the drone. Here, usingthe new model context protocol (MCP), we solve that problem, by developingand demonstrating in flight and simulation an LLM agnostic and drone agnos-tic approach, providing the first universal, versatile, comprehensive and easy touse drone control interface. This provides a universal way for AI systems toaccess drone data, tools, and services. Our cloud based MCP server supports theMavlink protocol, an ubiquitous drone control language used almost universally Keywords:Large language model (LLM), drone, unmanned aerial vehicles (UAVs),unmanned aircraft systems (UAS) 1 Introduction Autonomy in drones is usually based on onboard image and state processing fromonboard sensors[1]. An outstanding example is the 2023 AI controlled racing drone[2]that beat a human drone pilot, similar to the IBM supercomputer beating the worldchess master in the 1990s. However, challenges remain in the amount of compute poweronboard due to weight and energy restrictions for all drone classes and sizes, espe-cially small drones. With the coming of age of internet connected drones, using AI inthe cloud opens new opportunities to harness the virtually unlimited compute powerof data centers deployed globally. LLMs are one such technology, trained at scale. Todate, the interface between the AI LLM and the drone has been an unsolved challengefor general use cases. Here we show, using the model context protocol (MCP) standard, 2 Architecture Fig. 1 shows the architecture, and Fig. 2 the concept of operations, designed to bridgethe virtual world of AI with the material (physical) world of drones. 2.1 Concept of operations Any LLM (private proprietary or open source, cloud or locally hosted) that supportsthe MCP standard can connect to the MCP server developed in this work, which we A list of LLMs as of the writing of this paper that support the MCP serverarchitecture is: •Anthropic Claude– Native support, including Claude 3.5 Sonnet, Claude Opus4, and Claude Sonnet 4.5. Claude Desktop app has built-in MCP client support. •OpenAI– Adopted MCP in March 2025. Works with GPT-4, GPT-4 Turbo, andthe GPT-5 series models. The OpenAI Agents SDK includes MCP support.•Google Gemini– Announced support in April 2025. Works with Gemini 2.0 Flash,Gemini 2.5 Pro, and Gemini 2.5 Flash. Google launched production-ready managedMCP servers in December 2025. – Llama 3.2 and Llama 3.3 70B– Qwen 2.5 72B In turn, the MCP server can connect to any drone that supports the Mavlinkprotocol and has an internet connection. This includes PX4 and Ardupilot, the two largest drone software packages, already used on millions of drones. The MCP serverhandles all drone communications and provides LLMs with a description of capabilities 2.2 Mavlink protocol Mavlink[3] is the protocol used by open source autonomous drones running Ardupilotand PX4. This includes millions of drones (land, air, sea) and the vast majority of theglobal fleet of open source drones. supported by libraries in most common modern high level programming languagessuch as C, Python, etc. 2.3 Model context protocol: From manual tools to universalstandard LLM interface 2.3.1 Tools: Tedious and manual agentic AI programming A large language model (LLM) supports inference and text prediction based on tokens,but does not otherwise perform any function. Recently, LLMs have been programmedwith the ability to call ”tools” in order to interact outside the context of pure text. Thisenables LLMs to perform as agents, accessing internet content, reading and modifyingfiles and databases, and even writing code. Typically, each tool is hand coded or 2.3.2 Model Context Protocol In order to enable more seamless connection between LLMs and tools, the modelcontext protocol concept was introduced by Anthropic in Nov. 2024[4]. In December2025, Anthropic donated the MCP standard the Linux Foundation, in order to secureit as an open, industry wide standard[5]. The MCP serves as a layer between LLMsand the rest of the virtual world by standardizing the LLM communication protocol.An MCP server can connect any LLM to any tool. MCP is now supported by allthe major LLM providers (OpenAI, Anth