Assessing Human Likeness by Eye Contact in an Android Testbed

Assessing Human Likeness by Eye Contact in an Android Testbed Karl F. MacDorman, 1,2 Takashi Minato, 1 Michihiro Shimada, 1 Shoji Itakura, 3 Stephen Cowley, 4 and Hiroshi Ishiguro 1,5 Department of Adaptive Machine Systems and 2 Frontier Research Center Osaka University, Suita, Osaka 565-0871 Japan Department of Psychology, Kyoto University Yoshida Honmachi, Sakyo-ku, Kyoto 606-8501 Japan School of Psychology, University of Hertfordshire, College Lane, Hertfordshire AL10 9AB United Kingdom Intelligent Robotics and Communication Laboratories, ATR 2-2-2 Hikaridai, Keihanna Science City, Kyoto, 619-0288, Japan Abstract uncanny valley humanoid robot familiarity The development of robots that closely resemble human beings enables us to investigate many phenomena re- lated to human interactions that could not otherwise be investigated with mechanical-looking robots. This is be- cause more humanlike devices are in a better position to elicit the kinds of responses people direct at each other. In particular, we cannot ignore the role of appearance in giving us a subjective impression of social presence or intelligence. However, this impression is influenced by behavior and the complex relationship between it and appearance. As Masahiro Mori observed, a humanlike appearance does not necessarily give a positive impres- sion. We propose a hypothesis as to how appearance and behavior are related and map out a plan for an- droid research to investigate this hypothesis. We then examine a study that evaluates the behavior of androids according to the patterns of gaze fixations they elicit in human subjects. Studies such as these, which integrate the development of androids with the investigation of human behavior, constitute a new research area fusing engineering and science. healthy person toy industrial robot human likeness moving corpse prosthetic hand Figure 1: Mori’s uncanny valley for animated objects [Mori, 1970]. Introduction Progress is underway to develop humanoid robots that can support rich, multimodal interaction [Kanda et al., 2004], and we may expect to see ad- equate competencies within the next decade for brief exchanges in stereotyped situations. However, these robots will be of substantially less value if because of their appearance, ordinary people are unable to accept them as a social presence. Studies of person-to-person interaction in psychology and other fields generally take our human form for granted. This leaves us to assume that our everyday impressions of sociality are a subjective phenomenon arising from our interactions with other people. However, the importance of a humanlike appearance has yet to be discounted, and there are a number of reasons why it might be significant. We have a range of biomechanical structures that have evolved or been adapted to express volition, intention, and emotion: Our eyes indicate the direction of gaze, which supports joint attention and other interactive responses; our faces and vocal tract are populated by scores of muscles in- volved in controlling facial expressions and the voice; and our bodies are animated by gestures and other meaningful acts. In addition, we are highly sensitized to these biomechanical structures and have developed specialized brain centers to interpret them, including those implicated in identifying faces [Farah et al., 2000], detecting faces [Kanwisher et al., 1997] and hands [Downing et al., 2001], and recognizing emotion. Honed by evolution and experience, our most highly developed model of a social other is our model of other people. If we cannot accept humanoid robots as a so- cial presence—even socially “competent” ones—because they do not look human, this is something robotics engi- neers need to know and plan for accordingly. This need has strongly motivated robotics engineers to learn some- thing about us as people and how the human form—and deviations from it—affect our perceptions and reactions. Simply put, what makes something a social presence? Is it mainly its behavior, or is there instead some complex interplay between appearance and behavior? Running counter to the view that we should build robots that look like people—what we call androids— is Masahiro Mori’s hunch that our goal should instead be stylishly designed robots, because robots that look too human might be disturbing [Mori, 1970]. Mori pro- posed that our sense of familiarity increases as robots appear more human until an uncanny valley is reached at which subtle defects in human likeness appear repul- sive (Fig. 1). The impression would not be unlike that of a moving corpse. Only recently is Mori’s hunch materializing into a re- search program for understanding the uncanny valley [Minato et al., 2004]. The effect of similarity can be separated into the effects of appearance and behavior,