M2HRI: An LLM-Driven Multimodal Multi-Agent Framework for Personalized Human-Robot Interaction

A research team from the University of Texas at Austin and the University of Colorado Boulder has introduced M2HRI, a novel framework designed to make teams of robots more socially intelligent. The system uses large language models (LLMs) as the core reasoning engine, enabling two key innovations for each robot: a distinct, LLM-generated personality (like 'helpful' or 'curious') and a long-term memory that stores past interactions with users. This moves beyond current multi-robot systems, which typically treat all units as functionally identical.

Crucially, M2HRI includes a centralized coordination mechanism that considers these individual robot personalities when assigning tasks and managing group behavior. In a controlled user study involving 105 participants, the framework proved highly effective. Users could significantly distinguish between robot personalities, and the presence of these traits, combined with memory for personalization, enhanced the overall quality of interaction. The coordinated system also successfully reduced task overlap among robots, creating a more coherent and efficient team dynamic.

The results, published on arXiv, demonstrate that both agent individuality and structured coordination are essential for effective human-robot interaction (HRI) in social environments like homes or hospitals. The project's code is publicly available, providing a foundational framework for developing robot teams that can build long-term, personalized relationships with users, a significant step toward practical social robotics.