Dialogue-Based Intention Detection in Human-Robot Interaction using Large Language Models

Abstract

This thesis addresses the development of a modular human-robot interaction system focused exclusively on voice-based communication. While current dialogue systems often rely on multimodal cues, this work explores intention detection through spoken language alone, aiming to enhance robots' ability to interpret and respond to user commands in natural dialogue. Utilizing the Robotic Operating System 2 (ROS 2) framework, the system integrates speech recognition, synthesis, and large language models with prompt engineering techniques to improve accuracy in understanding explicit user intentions. The resulting architecture enables robots to perform tasks based solely on verbal input, even within complex and evolving dialogues. Simulation scenarios demonstrated the system's ability to handle uncertainty, clarify ambiguous requests, and adapt to changes in user goals, showcasing robust intention classification, reclassification, and task execution. This voice-only interaction model offers a strong foundation for accessible and natural communication in real-world environments, such as education and customer service. Although real-world deployment was not conducted due to hardware limitations, future work will focus on real-world testing with physical robots and human participants. A detailed experimental design has been outlined to evaluate user satisfaction, perceived intelligence, and communication effectiveness, using both standardized metrics and open-ended feedback. These future studies aim to refine intention inference, expand task diversity, and enhance the overall naturalness and reliability of verbal human-robot communication.