SAHABAT: An Interactive LLM-Powered Tour Guide Robot for IIUM Gallery

Abstract

Autonomous robots can improve access to guided tours, but conversational interaction and physical navigation are often implemented as separate functions. This paper presents SAHABAT, an interactive large language model powered tour guide robot developed for the International Islamic University Malaysia Gallery. The platform integrates a touchscreen and voice interface, an NVIDIA Jetson Orin Nano, LiDAR and stereo vision sensors, a differential drive base, and a Robot Operating System 2 navigation stack. A FastAPI backend coordinates speech recognition, language model inference, speech synthesis, and a lightweight retrieval augmented generation mechanism that retrieves curated gallery records through deterministic station and keyword matching. Detected station references are also translated by the custom interface into navigation waypoints, allowing a spoken request to trigger both a grounded explanation and autonomous movement. Preliminary testing across ten consecutive voice interactions produced an average end-of-speech to audible-response latency of 1.24 seconds, with a mean language model time to first token of 434 milliseconds and text-to-speech time to first audio of 460 milliseconds. The robot consumed approximately 27 watts while idle and 35 watts during navigation, corresponding to an estimated practical mixed-operation runtime of 13 hours. These results demonstrate the feasibility of integrating responsive conversational guidance, curated domain knowledge, and autonomous indoor navigation within a single mobile gallery robot.