demonstrated task. A mirror system is built up from
experience using an on-line machine learning ap-
proach (variation of the SOFM) that self-organises,
and in effect temporally segments the experiences of
the robot. We have seen that different SOFM net-
work sizes are needed to represent each of the two
tasks presented, and this is because they have differ-
ing complexities; the object-interactions experiment
requires more nodes than the wall-following one.
Our mirror system represents how to use basic in-
nate skills such as moving a hand to particular posi-
tions, attaching a weightless glass to the wrist, etc.
in order to perform the task of the first experiment;
or turning to face the wall, turning to be parallel to
the wall, keeping parallel to the wall, etc., in order
to perform the task of the second experiment.
The nature of both tasks involves only observable
information, i.e. there are no physical constraints.
Recall that motor schemas hold the SOFM node vec-
tor information, which is of a purely perceptual na-
ture. Therefore perceptual information in the struc-
tures of the mirror system is alone sufficient to recog-
nise and reproduce both tasks. In future work we
intend to devise new tasks that will require the ar-
chitecture to be extended, so that motor schemas in-
clude factors of a more motoric nature; for example,
weight, force control, somatosensory feedback, etc.
As briefly discussed in Section 2, our system is
only an attempt towards an implementation of the
biological mirror system, since it is only perceptually
activated. A more complete implementation would
involve the system being activated both perceptually
and motorically. This would involve extending our
current architecture to include external factors to ac-
tivate the mirror system motorically; this is currently
beyond the scope of our work.
Our architecture relies on the existence of an in-
verse model (a set of innate skills). This is not bi-
ologically unreasonable, because there are low-level
motor skills already present at birth, prior to ad-
vanced motor control. Further, these skills need not
necessarily be hand-coded prior to implementation
of the mirror system; they can in fact be acquired
and/or modified through self-exploration, as we have
done in other experiments (Marom et al., 2002).
We present a system that attempts to model the
functional role of mirror neurons, namely the activa-
tion of structures in response to both the observation
of a demonstrated task, and its generation. Through
social situatedness and a set of innate skills, percep-
tual and motor structures develop for the recogni-
tion and reproduction of demonstrated actions. We
believe this is an implementation towards a mirror
system, and we have tested it on two different plat-
forms.
7. Related Work
Pomplun and Mataric (2000) and Fod et al. (2000)
have developed other methods to segment and clus-
ter data from demonstrated movements, using tech-
niques such as Principal Component Analysis and K-
Means clustering; these approaches differ from our
work in that they operate on a batch of data and
hence need to be re-trained to handle novel move-
ments.
Andry et al. (2002) use a similar approach to ours,
however they also self-organise the robot’s proprio-
ception; a simple winner-take-all mechanism is used
to segment the visual input into ‘perceptual struc-
tures’, and then each perceptual structure is associ-
ated with a separate SOFM that maps the robot’s
proprioception. This means that a one-to-many re-
lationship between perception and action is built, as
we mention in this paper; i.e. there are different
ways of achieving each perceptual situation.
There are several examples of implementations
that rely on the pre-existence of perceptual-motor
structures, i.e. where the movement segmentation is
performed either off-line, such as the examples men-
tioned above, or hand-coded arbitrarily by the de-
signer, such as our earlier work (Maistros and Hayes,
2000). Such perceptual-motor structures are often
referred to as primitives (Demiris and Mataric, 1998;
Schaal, 1999; Mataric, 2000).
Although our architecture would need to be ex-
tended to include motoric activation of the mirror
system, it is generally the case also with other imple-
mentations that they lack both types of activation.
In contrast, Fagg and Arbib (1998) and Arbib et al.
(2000) model the mirror system together with other
brain areas including the ones responsible for action
selection, constraints, motor plans, etc., which they
use to trigger their mirror system motorically.
References
Andry, P., Gaussier, P., and Nadel, J. (2002). From
visuo-motor development to low-level imitation. In
Proceedings of the Second International Workshop
on Epigenetic Robotics: Modeling Cognitive Devel-
opment in Robotic Systems, Edinburgh, UK.
Arbib, M. A. (1981). Perceptual structures and dis-
tributed motor control. In Brooks, V. B., (Ed.),
Handbook of Physiology - The Nervous System II.
Motor Control, pages 1449-1480. Amer. Physiol.
Soc.
Arbib, M. A., Billard, A., Iacoboni, M., and Oztop,
E. (2000). Synthetic brain imaging: grasping,
mirror neurons and imitation. Neural Networks,
13(8-9):975-997.