Mimesis Embodiment and Proto-symbol Acquisition for Humanoids

Mimesis is the primitive skill of imitative learning, one of the methods for recognition of others' behavior and construction of self behavior. Mimesis is thought an origin of human intelligence because this function is observed not only at humans but also at animals. When the mimesis is adopted as learning method for humanoids, convenience for designing full body behavior increase because bottom-up learning approaches from robot side and top-down teaching approaches from user side involved each other. Therefore, we have developed a behavior acquisition and understanding system for humanoids based on the mimesis. This system is able to abstract observed others' behaviors into conceptual symbols, to recognize others' behavior using proto-symbols, and to generate self motion patterns using the proto-symbols. In this paper, we mention the integration of mimesis loop, and confirmation of the feasibility on virtual humanoids. 1 I n t r o d u c t i o n The research of humanoid robots has a long history and has accumulated a substantial amount of literature. The focus of early efforts was mostly on the dynamics and control of biped walk. Although it has not yet reached the level of complete solution with full of liability and adaptability, the hardware technology has been established for building autonomous humanoids. The demonstrations of HONDA P2 in the end of 1996 and the other models were widely publicized and accepted with a surprise even in the research community of robotics. It was the time when the door opened to the human behavioral science and the human intelligence, as a real research issue of robotics. Although the motivation of the artificial intelligence originated there, the physical limitations have forced or justified the researchers to carry on their research in a limited scope and scale of complexity. It ought to be the major challenge of contemporary robotics to study robotic behaviors and intelligence in the full scale of complexity mutually sharing research outcomes and hypotheses with the human behavioral science and human intelligence. -7803-6736-7/01/510.00 © 2001 IEEE 1 5 Mimesis is the primitive skill of imitative learning and forms the foundation of skills understanding others' behaviors and constructing self behaviors. It is commonly observed for humans and the giant apes[l]. The group of animals coincides with those who live with full of social communications and manipulate symbols for the purpose. It is an recent hypothesis by Deacon[2] that the humans have developed the brains through co-evolution with the development of language, the ultimate symbol system. Therefore, the mimesis could be hypothetically thought the origin of intelligence of such animals. The discovery of "mirror neurons" [3] ,that activate not only when one observes an specific behavior of others, but also when one intends to act it, was suggestive. It implies that the basic functions for imitation was locally integrated in the hardware of the brain. The contemporary technology, on the other hand, poses a grand challenge, namely to design the architecture of intelligence for complex machines such as the humanoids. The role of visual observation was studied by Kuniyoshi et al[4], to extract reusable knowledge of tasks. The behavior pattern acquisition through the observation based learning was studied by Kawato et al[5], for robotic manipulators and more recently for a humanoid. Schaal[6] and Mataric[7] discussed the issues between the imitation learning and the humanoids. The authors' research goal is to bridge from the mimesis to the symbol manipulation through synthetic intelligence design for the humanoid robots. In this paper, we propose an architecture to realize the function of mimesis. A sequence of other's motion was cut off and segmented from the ever-lasting sequence of motion with the help of the trainer. It is them decomposed using the probability neural network (PNN) such that its approximate is represented by a sequence of the innate primitive motions of the humanoid. The sequence of the primitives were modeled by the dynamics of a hidden Markov model (HMM). An acquired HMM is used for the dual purposes. It can compute the likelihood when the humanoid observes a new sequence of motion, namely, how close it is to the one he had acquired. The HMM is also used to generate the acquired motion. However, the observed motion would be inappropriate for a self motion unless it is modified to be compatible with the