In Proceedings of the International Symposium on Computational Intelligence in Robotics and Automation (CIRA) Banff, Alberta, Canada. Jul 2001. Evaluation Metrics and Results of Human Arm Movement Imitation Marc Pomplun and Maja Matari´ c University of Toronto, Ontario, Canada M5S 3G3, marc@psych.utoronto.ca <mailto:marc@psych.utoronto.ca> University of Southern California, Los Angeles, CA 90089-0781, USA, mataric@cs.usc.edu <mailto:mataric@cs.usc.edu> WWW home page: <http://www-robotics.usc.edu/˜maja> Abstract. We present a psychophysical study of human arm movement imitation, and an approach to analyzing the resulting data, which is general enough to be applied to human or humanoid movement analysis. We describe a joint-space based segmentation and comparison algorithm that allows us to evaluate the performance of 11 different subjects performing a series of arm movement imitation tasks. The results provide analytical evidence for the strong interference effects of simultaneous rehearsal during observation. Additionally, the results also demonstrate that repeated imitation in these tasks did not affect the subjects’ performance. 1 Introduction Our interest is in gaining insight into the mechanisms behind imitation, the ability to repeat an observed behavior as well as to learn arbitrary new skills, i.e., skills that are irrelevant to the goal, by observation [2]. So called “true imitation” is thought to be a complex mechanism, since it is found in very few species, while mimicry is more common [17]. We have proposed a model of imitation [14,9] based on evolutionarily older substrates including motor primitives as the basis for motor control and the mirror system [8, 19, 20] as the basis for a direct sensory- motor mapping for mimicry. Our other work [14, 9, 4, 23, 1] is aimed at developing and implementing this model on a variety of synthetic humanoid platforms, in order to validate the model on real-world tasks. In this paper, we focus on a set of psychophysical imitation experiments we performed in order to gain further insight and constraints for our imitation model. We gathered arm movement data from a collection of normal subjects imitating video demonstrations, and have developed a set of analysis tools for evaluating the quality of the resulting imitation. Evaluation of imitation is an open problem, and our evaluation metrics are meant as a general tool for both human and robot imitation and general motor control evaluation. The rest of this paper is organized as follows. Section 2 summarizes previous research and develops the ques- tions investigated by the present study. The resulting experimental design is described in Section 3. In Section 4, the components of an appropriate metric for measuring distances between pairs of trajectories are specified, and in Section 5 the contribution of each component to the performance of the metric is analyzed. Section 6 provides the experimental results, including a statistical analysis of the factors influencing the subjects’ imitation performance. In Section 7, these results are discussed, and the paper is summarized in Section 8. 2 Motivation Imitation is a powerful tool of cultural skill transfer and interaction [2]. Thus, it is potentially of great value to robotics, where control of and interaction with highly-complex humanoid robots that are currently becoming available (Honda P3, NASA Robonaut, Sarcos full-body, etc.) are open problems. Our earlier work [12, 15] in the area of psychophysics of imitation involved studying the attentional behavior of people watching movements for subsequent imitation. Specifically, we recorded the visual fixations of 40 subjects (half male, half female), wearing an eye-tracker and watching videos of finger, hand, and arm movements. We constructed the experiments so as to address the following questions: 1. Is there a difference between watching a movement with the intention to imitate and just watching? 2. When watching with the intention to imitate, what features are fixated on? Our data demonstrated conclusively that the answer to the first question was negative, and the answer to the second was revealing. Subjects fixated at the end-point (finger or hand, or if a pointer was present, the tip of the pointer), regardless of whether they were just watching or intending to imitate. The only difference we observed was pupil dilation, which was larger in the imitation condition, presumably indicating cognitive processing.