Prompting Guidelines for Grounded Personas in LLM-Based Robots

Summary

The paper presents a structured eight-component prompt template for specifying and adapting behavior in LLM-based social robots. Drawing on prior HRI research and input from 27 experts, it addresses unclear personalities, hallucinated capabilities, user adaptation, deception, safety, and governance. The authors argue that explicit capability boundaries and transparent behavioral assumptions are essential for reliable and interpretable human-robot interaction.