novelty threshold, which controls how many nodes
are used (the level of granularity in the representa-
tion) .
In previous work (Marom and Hayes, 2001a),
the attention system was used only as a trigger
for learning perception-action mappings through a
feed-forward neural network with back-propagation.
When attentive, the attention system simply pro-
vided the trigger for learning, while a completely
separate system handled the perception-action asso-
ciations.
In the work presented here the attention system
forms a vital part of perception-action coupling, be-
cause the SOFM nodes are associated directly with
motor structures. When a new node is created, a
motor structure is hard-wired to it, and remains as-
sociated with that node thereafter: when the node is
updated in response to a stimulus, the motor struc-
ture is also updated in response to that stimulus, and
when the node is deleted, so is the motor structure.
Motor Schemas
Inspired from Arbib’s Schema Theory (Arbib, 1981),
we use motor schemas to represent the motor struc-
tures of the mirror system. Previous work employs
both perceptual and motor schemas (Maistros and
Hayes, 2001), here however perceptual schemas are
replaced by the SOFM nodes. As mentioned above,
the motor schemas in the current implementation are
created and updated together with SOFM nodes.
The information that is stored in the motor
schemas is the motoric representation of the recog-
nised action. One way to obtain this representa-
tion is to convert the perceptual information of the
demonstrator’s action to information that the imita-
tor’s own motor system can use. This is the diffi-
cult robotics problem of transforming the perceptual
space to the motoric space. However, in our imple-
mentation we can bypass this problem, because our
setup involves the imitator perceiving information
that is already in terms of its own body, as will be-
come evident in the experiments below. Therefore,
perceived information can be incorporated directly
into motor schemas, and we refer to this information
as targets to be achieved by the motor system.
A schema update mechanism is responsible for in-
corporating the perceived information into the mo-
tor schema. This update mechanism is very crucial
to the ability of the mirror system to generalise and
reproduce actions. According to Schema Theory, the
type of information that can go inside motor schemas
is arbitrary; for instance, sequential targets for a
chunk of a movement, parameters for force control,
etc. The update mechanism therefore needs to be
designed carefully to deal with the chosen represen-
tation (for example, processing of temporal informa-
tion). We are addressing this problem in on-going
work, for example the utilisation of a sequence of
motor targets in each schema with heuristics to up-
date them (Marom et al., 2002).
In the work presented here we have found that in-
stead of storing a sequence of motor targets in each
schema, a single representative target is sufficient for
modelling the nature and complexities of our tasks.
Such a representative is already available in the sys-
tem through the SOFM node vector, since it gener-
alises over the perceptual space which includes the
demonstrated movement. The discussion of the ex-
periments will explain why this approach works for
these tasks.
2.2 Using the Mirror System
As described above, the structures of the mirror sys-
tem are built up from experience. Once built, the
mirror system is used in a recall phase as follows:
the system receives continuous perceptual input; the
SOFM is used to recognise that input; the winning
SOFM node activates the hard-wired motor schema;
this schema provides the output of the mirror sys-
tem, i.e. a motor target; this target is then sent to
the motor system for execution.
Notice that the mirror system is only activated
perceptually, whereas real mirror neurons can be ac-
tivated both perceptually and motorically. This issue
will be discussed further in the discussion.
3. Motor System
The main component of the motor system is an in-
verse model, used to translate motor targets into
motor commands, as used in the control literature.
It is a mechanism which, given the robot’s cur-
rent state (perceptual information and propriocep-
tive feedback) and a desired target state (e.g. joint-
angle targets), calculates the motor commands that
best achieve the desired state. This implementation
of the inverse model is a simplified version of the one
used by Demiris (1999), who called his inverse mod-
els ‘behaviours’ because they were able to adapt their
parameters through a proprioceptive error signal.
In our current implementation the inverse model is
innate and remains fixed, and we believe that assum-
ing the existence of such a model is not biologically
unreasonable. Experiments on early infancy illus-
trate that prior to the development of advanced mo-
tor skills (i.e. intentional coordinated goal-directed
movements), there is already some basic knowl-
edge about fundamental motor control (Meltzoff and
Moore, 1989). One can think of the inverse model as
the information about how the robot can use its ac-
tuators, say how to use its hands, and the immediate
consequences of their use.
By coupling the mirror system with an inverse