MENTOR: A Mission-Level ExecutioN ArchiTecture fOR LLM-Guided UAVs

Jinze Yang1, Yi Cui1, Yuhang Liu1, Zhen Li1, Haikuo Liu1, Changkun Du2

1 FALCON Lab, Beijing Institute of Technology
2 Intelligent Science & Technology Academy Limited of CASIC

Project Demonstration

A real-world demonstration of language-guided multi-stage UAV mission execution.

In complex real-world environments, enabling unmanned aerial vehicles (UAVs) to understand natural-language instructions and execute them safely remains a central challenge.

To address this problem, we redefine language-guided UAV autonomy as a Mission-Level Execution problem and formally introduce the MENTOR architecture.

MENTOR is an innovative UAV execution framework built around the principle of decoupling reasoning from control while preserving closed-loop safety. It separates high-level semantic reasoning from flight-critical low-level control, while maintaining a tightly coupled closed-loop feedback interface between them.

✨ How it Works

🧠 Structured Mission States: The system parses complex natural-language instructions and converts them into highly structured mission states.

🛡️ Guarded Mission Executive: During execution, MENTOR continuously integrates online visual evidence and low-level skill feedback, providing reliable decision safeguards for mission progress.

🚁 Typed UAV Skills: Validated stage-level decisions are precisely grounded into a standardized library of UAV action skills.

🌟 Why MENTOR?

Through this architectural design, MENTOR establishes a clear division of labor between the "brain" and the "cerebellum": it gives large language models (LLMs) substantial freedom to perform adaptive, context-aware reasoning over evolving mission contexts, while returning flight-critical responsibilities to reliable low-level autonomy modules. This fundamentally improves the safety and consistency of UAV mission execution in the real world.

在复杂的现实环境中,如何让 UAV 听懂人类的自然语言指令并安全执行,一直是一项核心挑战。

为了解决这一问题,我们将语言引导的无人机自主控制重新定义为一个 Mission-Level Execution 问题,并正式推出 MENTOR 架构。

MENTOR 是一种创新的无人机执行框架,其核心理念是 “解耦推理与控制,保持闭环与安全”。它成功地将高层级语义推理与涉及飞行安全的底层控制分离开来,同时维持了一个紧密耦合的闭环反馈接口。

✨ 核心机制

🧠 Structured Mission States: 系统能够解析复杂的自然语言指令,并将其转换为高度结构化的任务状态。

🛡️ Guarded Mission Executive: 在执行过程中,持续整合在线视觉证据与底层技能反馈,为任务的推进提供可靠的决策守护。

🚁 Typed UAV Skills: 将经过验证的阶段级决策,精准向下映射并落实到一组标准化的无人机动作技能库中。

🌟 为什么选择 MENTOR?

通过这种架构设计,MENTOR 实现了“大脑”与“小脑”的完美分工:赋予了 LLMs 在不断演进的任务上下文中,进行自适应、情境感知推理的极大自由度;同时,将飞行的重任交还给可靠的底层自主控制模块,从根本上确保了无人机在现实世界中执行任务的安全性和一致性。