Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SAGE-LLM: Towards Safe and Generalizable LLM Controller with Fuzzy-CBF Verification and Graph-Structured Knowledge Retrieval for UAV Decision

Wenzhe Zhao, Yang Zhao, Ganchao Liu, Zhiyu Jiang, Dandan Ma, Zihao Li, Xuelong Li

TL;DR

A train-free two-layer decision architecture based on LLMs, integrating high-level safety planning with low-level precise control is proposed, suggesting its potential for generalization to broader embodied intelligence systems and safety-critical control domains.

Abstract

In UAV dynamic decision, complex and variable hazardous factors pose severe challenges to the generalization capability of algorithms. Despite offering semantic understanding and scene generalization, Large Language Models (LLM) lack domain-specific UAV control knowledge and formal safety assurances, restricting their direct applicability. To bridge this gap, this paper proposes a train-free two-layer decision architecture based on LLMs, integrating high-level safety planning with low-level precise control. The framework introduces three key contributions: 1) A fuzzy Control Barrier Function verification mechanism for semantically-augmented actions, providing provable safety certification for LLM outputs. 2) A star-hierarchical graph-based retrieval-augmented generation system, enabling efficient, elastic, and interpretable scene adaptation. 3) Systematic experimental validation in pursuit-evasion scenarios with unknown obstacles and emergent threats, demonstrating that our SAGE-LLM maintains performance while significantly enhancing safety and generalization without online training. The proposed framework demonstrates strong extensibility, suggesting its potential for generalization to broader embodied intelligence systems and safety-critical control domains.

SAGE-LLM: Towards Safe and Generalizable LLM Controller with Fuzzy-CBF Verification and Graph-Structured Knowledge Retrieval for UAV Decision

TL;DR

A train-free two-layer decision architecture based on LLMs, integrating high-level safety planning with low-level precise control is proposed, suggesting its potential for generalization to broader embodied intelligence systems and safety-critical control domains.

Abstract

In UAV dynamic decision, complex and variable hazardous factors pose severe challenges to the generalization capability of algorithms. Despite offering semantic understanding and scene generalization, Large Language Models (LLM) lack domain-specific UAV control knowledge and formal safety assurances, restricting their direct applicability. To bridge this gap, this paper proposes a train-free two-layer decision architecture based on LLMs, integrating high-level safety planning with low-level precise control. The framework introduces three key contributions: 1) A fuzzy Control Barrier Function verification mechanism for semantically-augmented actions, providing provable safety certification for LLM outputs. 2) A star-hierarchical graph-based retrieval-augmented generation system, enabling efficient, elastic, and interpretable scene adaptation. 3) Systematic experimental validation in pursuit-evasion scenarios with unknown obstacles and emergent threats, demonstrating that our SAGE-LLM maintains performance while significantly enhancing safety and generalization without online training. The proposed framework demonstrates strong extensibility, suggesting its potential for generalization to broader embodied intelligence systems and safety-critical control domains.
Paper Structure (33 sections, 12 equations, 3 figures, 3 tables)

This paper contains 33 sections, 12 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Safety Research in UAV Adversarial Games
  4. Large Language Models in Embodied Intelligence and UAV Control
  5. Preliminaries
  6. Control Barrier Function
  7. Method
  8. Overall Framework
  9. Fuzzy CBF Safety Verification Theory
  10. Directional Cones for Semantic Actions
  11. Fuzzy CBF Theory Based on Directional Cones
  12. Star-Hierarchical Graph Based RAG
  13. Construction of the Star-Hierarchical Graph
  14. Hierarchical Recursive Elastic Retrieval
  15. Experience
  16. ...and 18 more sections

Figures (3)

  • Figure 1: System Framework
  • Figure 2: Generalization Experience Visualization. From left to right, the scenarios are: 5 spherical obstacles at random positions, 8 spherical obstacles at random positions, and 5 cylindrical obstacles with infinite height at fixed positions. The results show that SAGE-LLM successfully pursues the target while avoiding all obstacles in unseen, challenging environments, showcasing its zero-shot generalization capability.
  • Figure 3: Generalization performance scatter plot with error bars. Proximity to the top-right signifies superior overall performance, while smaller error bars reflect higher stability. SAGE-LLM demonstrates a leading and robust position.

Theorems & Definitions (1)

  • Definition 1: Formal Representation of a Semantic Action