LIMBS Seminars

July 21st 2pm "auletta di rappresentanza 1 piano"

Speaker: Yuichiro Yoshikawa

Title: Towards imitation by an internal observer

Abstract:

Many research fields focus on the capacity to imitate -- partially  because it would give an agent a powerful learning strategy for new actions, and partially because it seems to be a basis for understanding the behavior of other agents. To imitate one must know  this correspondence between the observed demonstration and one's own behavior. In previous robotics studies, the designer usually provides knowledge about the correspondence. However, it is a formidable issue to design an agent that can imitate if the agent is an "internal observer" that can obtain information only by collecting data from its sensors through its action -- in other words, if it cannot use any explicit knowledge given by the designer.

For an internal observer to find the correspondence, there are at least two points to be considered. One concerns how to construct a coordinate system to represent behaviors. This is not easy because an internal observer does not even know what its body is at the beginning. It has to solve this problem to find its body representation in its uninterpreted sensory data through internal observation. Another point concerns how to construct a mapping between the coordinate system in which the behavior of other agents is observed and the coordinate system of one's own behavior. To learn the mapping by itself, a robot needs to find references between them. This point relates to the definition of imitation.

Although these two points are related, we consider them separately as a first step. Concerning the representation of the robot's actions, we address the problem of acquiring a representation of the body, namely finding its body in its uninterpreted sensory data. A cross modal map is proposed as the learning structure based on the idea that the invariance in multi-sensory data represents the body. Concerning the construction of a mapping, we address the problem of acquiring common vowels with the caregiver. We propose an interaction model that guides a robot to acquire articulation to vocalize.

CV

Yuichiro Yoshikawa is Ph.D. student at the Dept. of Adaptive Machine Systems, Emergent Robotics Area, Graduate School of Engineering, Osaka University

-Home