-- This paper examines how functional hierarchy can self-organize
through sensory-motor interactions, without assuming predefined
level-structured functions. A humanoid robot was implemented with
so-called the multiple timescales recurrent neural network (MTRNN).
The MTRNN consists of the fast neurons part and the slow neurons one
which are interconnected each other within a single network.
The results of the robot learning experiments showed that functional
hierarchy emerges with accompanying a compositional structure such that
the continuous sensory-motor flow is segmented into reusable behavior 
primitives in the fast neurons part and those primitives are integrated
into specified goal-directed actions in the slow neuron part.

-- This paper investigates how manipulable representations with
compositionality can be acquired without losing fluidicity, 
generalization and context-dependency.
The comparative reviews among our prior neuro-robotics experiments address
the issues of local vs distributed representations in representing behavior
and the effectiveness of level structures associated with different
sensory-motor articulation mechanisms. It is concluded that the compositional
structures can be acquired 'organically'by achieving generalization in
learning and by capturing the contextual nature provided that open-ended
sensory-motor interactions with internal neuronal dynamics are sustained.

-- This paper shows how meaning space can be self-organized
through dynamic interactions between a linguistic module and
a behavior module which are implemented by RNNPB models. 
The robotic learning experiment showed that compositional structure
represented by combinations of verbs and objective nouns appear as 
generalized with forming distributed representations in the network.

-- This paper describes a mirror neuron model from dynamical
systems perspective. Forward predictive models are learned for multiple
goal-directed actions distributedly in a single RNN. In action generation,
units called parametric bias (PB) play role of bifurcation parameter for the
RNN dynamics to generate multiple goal-directed actions. On the other hand
in recognizing actions, the best PB values to fit with perceived sensory
sequences are identifies through inverse computation. A set of robotics
experiments evaluate how generalization by learning can be achieved with
this distributed representation scheme.

-- This paper attempts to describe possible correspondences of
(Tani & Nolfi, 1999) to phenomenology of immanent time of Husserl.
It is argued that time passing can be consciously perceived when continuous
flow of sensory-motor flow is segmented from one chunk to another accompanying
switching of coherences among local modules.

-- This paper shows how continuous sensory-motor flow can be
segmented into hierarchically organized chunks through anticipatory
learning of local mixture of RNN experts with multiple levels.
The study addresses the issue of how compositional representation
can emerge solely through row sensory-motor experiences using a
localist neural network model.

-- This study attempts to describe the notion of the "self" from
dynamical systems perspective based on our robot experiments. 
A vision-based mobile robot implemented with an RNN model learns to predict
landmark sequences experienced during its dynamic exploration of environment.
It was shown that the learning process switches spontaneously between coherent
phases in which the top-down prediction agrees with the bottom-up sensation
and incoherent phases in which conflicts appear between the two.
By investigating possible analogies between this result and the phenomenological
literature on the "self", we draw the conclusions that (1) the structure of
the "self" corresponds to the "open dynamic structure" which is characterized by
co-existence of stability in terms of goal-directedness and instability
caused by embodiment; (2) the open dynamic structure causes the system's
spontaneous transition to the unsteady phase where the "self" becomes aware.

-- This paper describes a neuro-robotics experiment to show how
'symbolic structure' emerges as embedded in neuronal dynamics as
the results of internalizing experiences of combinatorial sensory-motor
interactions of robots. The action-sensation causality is learned
as a forward model by using a Jordan-type recurrent neural net (RNN)
which is implemented in a mobile robot. After the learning, the RNN
generated on-line prediction of next sensation for given action as well as
mental simulations for combinatorial action sequences without actual movements.
Our dynamical system analysis showed that a finite state machine like symbolic
structure emerges in a fractal-like global attractor of the RNN dynamics
which is naturally situated with sensory-motor interactions with environment.

-- This paper describes how symbolic dynamics can be learned in RNN.
A Jordan type RNN was trained with stochastic symbolic sequences with a
grammar. The learning result showed that the stochastic symbolic sequences
are reconstructed by self-organizing deterministic chaos in RNN.