Invited Talk 2
Wednesday, July 20, 13:30-14:20 (International Conference Hall)
Emergence and Development of Humanoid Embodied Cognition
Yasuo Kuniyoshi
University of Tokyo, Japan
Abstract
Our brain, especially the cerebral cortex, self-organizes its
functional structure through early developmental process. Many evidences suggest
that such self-organization is driven by and reflects the information structure
inherent in the sensory-motor signals experienced during the development. And
such information structure emerges from neural-body-environment interaction
dynamics, or "embodiment". Therefore, "Body Shapes the Brain"
would be a plausible claim. Many researchers in the past pointed out the importance
of embodiment for cognition. And some quantitative experimental data were presented
for very simple robots. However, very little is known about how a human-like
body affect human cognition, particularly in terms of quantitative experiments.
Also, a mechanism for exploration, discovery and exploitation of embodied information
structure (aka. body affordances) is an open issue: Past approaches adopt search
based methods such as reinforcement learning or genetic algorithm. However,
they are not appropriate and not powerful enough for exploring complex embodiment,
because the search space is prohibitingly large for a complex body, and it is
very difficult to define appropriate and complete set of primitives. As a novel
approach to the above issues, I will first present a robotic-psychological work
which shows that a certain invariant information structure emerges from the
dynamics of interaction between a humanoid body and the environment, and that
the structure is important both in generation and perception of motor behaviors.
Then I will present experiments on a novel behavior exploration architecture
based on coupled chaotic maps. It drives a multiple degrees of freedom robot
body without any predefined primitive actions. It is capable of "instantaneously"
discovering behavior patterns consistent with the embodied constraints and the
immediate dynamic situations. This is regarded as autonomous free exploration
for the invariant information structures (affordances) revealed in the first
part of the talk. The proposed mechanism can be related to very early neural
mechanism of fetuses and new-born babies which elicit general movements. Learning
to articulate, remember and identify the discovered behavior patterns is an
open problem. Emergence of a value system and motives is another important research
issue. I will present our recent efforts towards solving these issues, including,
spatio-temporal attractor learning with competitive dynamics, hierarchical self-organization
of behavior categories, and emergence of primitive emotional states. Concluding
discussions will be on an overall framework of emergence and development of
human-like behavior concepts through embodied interactions. And how it unfolds
in development of imitation capabilities, a critical phenomenon in the development
of humnanoid cognition.