Feelix Growing Portal

Gaussier, Ph., Andry, P., Boucenna, S. (2008) 'Dynamic fields and Interactive Systems'. In: Proceedings of Conference on Dynamics & Applications. Braga, Portugal. September 2008

Abstract:
The dynamical system approach is an interesting framework to analyse and design complex control architectures
[7, 6]. Focusing on the dynamics allows to overstep some limitations of functional approaches and
to enlight possible emergent properties. For instance, in previous works, using the perception ambiguity, we
have shown that a simple visuo-motor homeostat can be used to trigger low level imitation capabilities [5, 4].
Moreover, dynamical neural fields allow to combine easily in a single system different control strategies (different
motor commands obtained from different neural networks working at different frequencies can be easily merged
in a single neural field allowing the control of several degrees of freedom). Yet, in these systems performances
directly depend on the human capabilities to maintain the interaction. To allow turn taking or simply long term
interactions the robot must not be only a reactive system but must be endowed with some ”will” to interact.
In recent works, we have shown a simple internal oscillator can be used to maintain low level interactions. To
go one step further, we try to address the question of predicting what could be the stable states of a system
interacting with its environment [2, 3]. As a toy problem, we have analysed how an expressive robot head could
learn to associate the facial expression of a human or another robot with its own internal emotional state. We
have shown in the case of a simple reactive architecture that a solution to obtain a stable state of interaction is
that the human teacher mimics the robot facial expressions. This idea has been successfully tested with a real
robot head. Moreover, we have shown that the robot head can learn through the interaction game to perform
in an unsupervised manner a face / non face discrimination by using the capability to predict the rhythm of
the interaction [1] as a learning modulation to decide whether some visual features can belong to a face or not.
At last, we propose a definition of the shared perception as a dynamical system and question the possibility to
develop mathematical tools allowing to predict and study how interacting systems can develop more and more
complex skills through the interactions.