Guidelines Proposed for Grounded Robot Personas in LLM-Based HRI

The increasing use of Large Language Models (LLMs) in social robots for verbal interaction has introduced challenges, including robots exhibiting hallucinated capabilities, unclear behavioral boundaries, and inconsistent personas. This lack of clear prompt design in Human-Robot Interaction (HRI) can lead to unpredictable and potentially unsafe robot behavior. To address these issues, this paper develops a comprehensive framework for prompt design in LLM-based robots. It introduces a structured prompt template comprising eight functional components, which allow for precise specification, bounding, and adaptation of robot behavior. The framework is built upon a review of existing LLM-based HRI work and incorporates insights from HRI experts. Qualitative findings from expert surveys highlighted concerns about the legibility of robot personality, the need for user adaptation, and significant ethical considerations regarding safety, deception, and governance. Based on these insights, the proposed prompting guidelines and template serve as a structured aid for HRI research, advocating for prompt design to be treated as a socio-technical problem requiring explicit capability boundaries, transparent assumptions, and context-sensitive safeguards to ensure reliable and interpretable human-robot interactions.