Development of Microcontroller-Based AI Robot Tour Guide Utilizing Custom Language Models

This paper presents the design and implementation of an AI-powered robotic tour guide system for the West Virginia University Lane Department of Computer Science and Electrical Engineering (LCSEE). The system enables prospective students to engage in real-time, natural language conversations with a robot equipped with department-specific knowledge. Two methodologies were explored: (1) a DirectLLM approach that fine-tunes a large language model on a custom dataset, and (2) a Classify-Retrieve-Generate (CRG) pipeline that modularizes classification, answer retrieval, and natural response generation. A custom SQuAD-style dataset was developed using LCSEE data, supporting both pipelines. The system was deployed on a Raspberry Pi 4 integrated with a MangDang Mini Pupper robot. Evaluation results show that while DirectLLM excels in fluency, it is hindered by retraining requirements and limited scalability. In contrast, the CRG method provides modularity and easier maintenance, with strong performance on classification and retrieval but pending refinement in response generation. Future work includes dataset expansion, CRG generator improvements, DirectLLM fine-tuning, and exploration of hybrid fusion models. This research demonstrates the feasibility of deploying compact, domain-aware AI agents for real-time educational engagement.