Hybrid LLM Navigation System for Edge-Cloud Reasoning

This paper proposes a hybrid decision making architecture that combines an Edge LLM(large language model) and a Cloud LLM for a mobile robot performing natural language based navigation. The user’s commands are matched with predefined zone information and converted into target coordinates, while during navigation the system receives human and obstacle information from an RGB camera and LiDAR(light detection and ranging) to assess the risk level and complexity of the environment. Based on this, the router prioritizes the Edge LLM, but conditionally calls the Cloud LLM only in situations that require semantic judgment, in order to refine policy parameters such as speed adjustment and safety distance. Experiments conducted in a Gazebo warehouse environment, where the number of people was gradually increased, showed that the proposed method reduced unnecessary stops and shortened arrival time under medium congestion, and selected more stable avoidance behaviors even in highly crowded situations. These results indUicate that a hybrid approach combining the real-time responsiveness of Edge and the reasoning capability of Cloud provides effective navigation performance in complex, human centered environments.