This paper explores the design requirements for integrating large language models (LLMs) with robots, focusing on human-robot interaction (HRI). The study compares LLM-powered robots with text and voice-based agents in four tasks: generate, negotiate, choose, and execute. The findings show that LLM-powered robots elevate expectations for sophisticated non-verbal cues and excel in connection-building and deliberation, but face challenges in logical communication and may induce anxiety. The study highlights the unique design needs for LLMs in robot applications, including the importance of non-verbal cues, task-specific customization, and the potential risks of hallucination errors. The research provides design implications for both robots and LLMs to improve HRI. The study also identifies that LLM-powered robots are less preferred in tasks requiring precise communication and decision-making due to the potential for communication errors and social pressure. The results suggest that LLMs should be tailored to specific tasks and contexts to optimize their performance in human-robot interactions. The study also emphasizes the need for careful implementation of LLMs in robots to ensure they align with the intended functionality and avoid unintended behaviors. The research concludes that LLMs can enhance HRI by enabling more natural and engaging interactions, but their integration requires careful design and consideration of the unique challenges posed by robot embodiment.This paper explores the design requirements for integrating large language models (LLMs) with robots, focusing on human-robot interaction (HRI). The study compares LLM-powered robots with text and voice-based agents in four tasks: generate, negotiate, choose, and execute. The findings show that LLM-powered robots elevate expectations for sophisticated non-verbal cues and excel in connection-building and deliberation, but face challenges in logical communication and may induce anxiety. The study highlights the unique design needs for LLMs in robot applications, including the importance of non-verbal cues, task-specific customization, and the potential risks of hallucination errors. The research provides design implications for both robots and LLMs to improve HRI. The study also identifies that LLM-powered robots are less preferred in tasks requiring precise communication and decision-making due to the potential for communication errors and social pressure. The results suggest that LLMs should be tailored to specific tasks and contexts to optimize their performance in human-robot interactions. The study also emphasizes the need for careful implementation of LLMs in robots to ensure they align with the intended functionality and avoid unintended behaviors. The research concludes that LLMs can enhance HRI by enabling more natural and engaging interactions, but their integration requires careful design and consideration of the unique challenges posed by robot embodiment.