People typically expect social agents' verbal utterances to be congruent with their physical appearances. However, given the vast diversity in robot appearances, expectations for each robot can vary widely, making it impractical to constrain a robot's topics of conversation. We propose a solution that allows building “body images” for robots using a free word association task (Li et al., 2023). By feeding words associated with a robot into a large language model (LLM), we created unique body images for two robots, Buddy and Atlas. These body images affected the way the robot interacted with a human. Our method provided each robot with a verbal interaction style that aligned with their physical appearances and human social expectations.
We present results showing how extended verbal interaction with these robots changes people's ratings of the robots on several anthropomorphic dimensions. Hence, expectations for how a robot will act can affect the way people respond to the robot. For instance, a body image based on words associated with the large humanoid resulted in higher perceived humanness scores (see Ho & MacDorman, 2010) than a child-like body image - regardless of the robot’s physical appearance. Counterintuitively, when there is a mismatch between body image and appearance, people tend to attribute higher animacy (see Spatola, Kühnlenz, & Cheng, 2021) to the robot. These results demonstrate the potential for improving human-robot interaction by integrating a body image empowered by LLM and real human data.