GPT-Driven Gestures: Leveraging Large Language Models to Generate Expressive Robot Motion for Enhanced Human-Robot Interaction

Abstract

Expressive robot motion is a form of nonverbal communication that enables robots to convey their internal states, fostering effective human-robot interaction. A key step in designing expressive robot motions is developing a mapping from the desired states the robot will express to the robot's hardware and available degrees of freedom (design space). This letter introduces a novel framework to autonomously generate this mapping by leveraging a large language model (LLM) to select motion parameters and their values for target robot states. We evaluate expressive robot body language displayed on a Unitree Go1 quadruped as generated by a Generative Pre-trained Transformer (GPT) provided with a set of adjustable motion parameters. Through a two-part study (N = 120), we compared LLM-generated expressive motions with both randomly selected and human-selected expressions. Our results show that participants viewing LLM-generated expressions achieve a significantly higher state classification accuracy over random baselines and perform comparably with human-generated expressions. Additionally, in our post-hoc analysis we find that the Earth Movers Distance provides a useful metric for identifying similar expressions in the design space that lead to classification confusion.