Researchers at Nankai University have developed a general navigation framework that integrates topological graphs with large language models (LLMs) to enable flexible and controllable social navigation for robots. The approach addresses the limitations of traditional model-based and learning-based methods by leveraging LLMs' world knowledge for decision-making while ensuring controllability through model-based trajectory optimization.
Key takeaways
- The framework uses obstacle clustering and graph theory to generate candidate local guiding points, simplifying environmental information for LLM processing.
- LLMs employ role-playing and few-shot closed-loop optimization to select the optimal guiding point, balancing motion efficiency and social attributes.
- Experiments across static and dynamic scenes, tested on four LLMs, show that combining guiding points with traditional trajectory optimization yields controllable navigation, with locally optimal decisions reaching 97.94%.
This work demonstrates the potential of LLMs in social navigation, offering a pathway toward more adaptable and socially aware robots in human-centric environments.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-01-13 · “基于拓扑图和大语言模型的灵活可控社交导航”