Large Language Model as a Strategic Brain: Autonomous Selection of Path-Planning Algorithms for Mobile Robots

Abstract

Recent advancements in deep learning and transformer architectures have accelerated research on Large Language Models (LLMs), enabling multimodal reasoning across diverse domains. However, deploying LLMs in mobile robotics remains difficult due to high computational costs and output instability. This paper proposes a framework that integrates a lightweight multimodal LLM into the path-planning layer of mobile robots. Given a terrain map and mission directive, the LLM selects the most appropriate algorithm from a predefined set of classical planners (A*, Theta*, PRM, RRT*, RRT*-Smart), based on high-level reasoning. By restricting the LLM's role to algorithm selection, the framework minimizes hallucination and improves reliability. Experimental results using Gemma3 variants in obstacle-dense environments show that the LLM consistently chooses algorithms that balance path length, computation time, and trajectory smoothness. These findings demonstrate the feasibility of incorporating lightweight LLM reasoning into resource-constrained robotic platforms and highlight its potential for intelligent, cost-effective autonomous navigation.