A Hybrid LLM/Model-Based Architecture for Flexible and Adaptive Robot Coaching

Human-Robot Interaction (HRI) presents significant challenges in accurately assessing situations, adapting robotic behavior to human intentions, ensuring explainability, pertinence, and acceptability, and effectively managing uncertainty. Traditional model-based approaches provide reliability but struggle with human unpredictability, often approximating human behavior through specific models that fail to encompass all possible scenarios. Conversely, statistical approaches, such as Large Language Models (LLMs), offer greater adaptability but lack deterministic guarantees. This paper introduces a hybrid architecture that integrates structured methodologies with the flexibility of LLMs to enhance robotic coaching in dynamic environments. The proposed architecture leverages deterministic modules for critical constraints and safety guarantees while employing LLMs for contextual understanding, natural language interaction, and adaptation to unpredictable human behaviors. Through a healthcare robotic coach scenario implementation and the conduction of a user study, we aim to demonstrate how this balanced approach enables effective monitoring of task execution, dynamic adaptation to human states, and seamless verbal interaction while maintaining system reliability. By bridging deterministic and statistics-based techniques, the proposed architecture aims to advance HRI toward safer, more transparent, flexible, and adaptive interactions.