gives the output for motor sub-system i (e.g., a head
control motor), at time t.
∑ak (t) ak (t) vk (t)
U (t) = k V zλ----- [1]
2 ak (t)
k
In equation [1], vk (t) is a vector giving the current
sensory input from sensory subsystem k (e.g., a joint
angle or a camera) at time t, ak (t) is a measure of the
reliability (confidence) of sensory subsystem k (a scalar),
and ak(t), defines the strength (e.g., priority) of a
particular sensory input (also a scalar).
In a perceptual context, the weighted component
integration (or democratic integration) algorithm
forwarded by Triesch and von der Malsburg (2001) offers
similar ideas, and presents a more detailed investigation
of the integration concept than Cheng et al. (2001). In
Triesch and von der Malsburg (2001), a group of
perceptual components such as motion, color, and shape
detection are adapted both in terms of the weighting of
the components and in terms of prototypes for the
perceptual components.
Unfortunately, the integration mechanisms proposed
by both Cheng et al. (2001) and Triesch and von der
Malsburg (2001) do not focus on or provide specific
means of incorporating skills that result from the process
of integration—leaving Criterion 2 unsatisfied. Two
additional computational mechanisms would seem
needed past Cheng et al. (2001) and Triesch and von der
Malsburg (2001) in order to provide skill incorporation.
First, the system needs a (at least implicit) means of
determining that a soft-assembled skill is re-occurring.
That is, the system needs a way to determine when that
skill should be considered “stable” (Criterion 5). Second,
once stable, the skill needs to be represented in a manner
similar to the existing skills. This last part, at least in
terms of this present scenario seems particularly difficult.
We have been working from the design premise of
utilizing current results from epigenetic robotics to form
the bootstrapping components of a robot, intended to
display ongoing emergence. But, in this case, there is no
particular means to add to this static collection of
bootstrapping skills. Presumably, a computational
mechanism would be needed to learn the aspects of the
new, now-stable soft-assembled skill. This new skill
would, after this learning, be part of the repertoire of the
system and with the other skills would form the basis for
further development (i.e., it would then be termed
incorporated; Criterion 2).
5.2. Bootstrapping Ongoing Emergence
With Primitive Ongoing Emergence
A possible limitation of adopting the strategy proposed
above is that one may miss common underlying
mechanisms that helped create the individual skills in the
first place. That is, in the ideal case, the goal would be to
create a robot that exhibits ongoing emergence, where the
bootstrapping primitives themselves are emergent. Thus,
68
in this ideal case the bootstrapping primitives are
generated by the same processes that underlie subsequent
skill development.
This presents a rather different problem than in
Section 5.1. On the one hand, a robot that has a pre-
programmed set of behaviors can presumably exhibit
those behaviors (e.g., in a sequence, or through a
blending of behaviors, such as shown in Breazeal,
Buchsbaum, Gray, Gatenby, & Blumberg, 2005), but is in
need of mechanisms to incorporate stabilized soft-
assembled behaviors into its skill repertoire. On the other
hand, a robot without a pre-programmed set of behaviors,
in addition to needing mechanisms to provide ongoing
emergence itself, is in need of an initial set of skills—it
needs initial means of perceiving, representing, and
behaving.
One conceptual way that such initial—and
emergent—skills might be created is through self-
exploration. A number of authors in epigenetic robotics
have suggested the need for some form of “self” in these
robotic systems. For example, Weng (2004) proposes that
developing robots must be SASE—Self-Aware and Self-
Effecting agents, Blank et al. (2005) talk about self-
motivation and self-organization, and Steels (2004)
suggests that robotic agents should self-regulate their
build-up of skills and knowledge as a way to increase
their rate of learning. In the present context of
bootstrapping a developing agent without pre-
programmed skills, self-exploration could be used to
facilitate differentiation between self and other (e.g., see
Michel, Gold, & Scassellati, 2004), which is important
because such a developing agent would likely need to
figure out what parts of its “environment” are part of the
agent (e.g., its own limbs) versus part of the external
world. We hypothesize that the basic properties of
ongoing emergence (i.e., Criterion 1 through 6) could
also provide the basis for these self-other discrimination
skills, and hence can provide the means to bootstrap the
skills of a developing robot.
6. Discussion
The foregoing has been a largely theoretical discussion of
ongoing emergence. Ongoing emergence describes, in
brief, behavioral growth in humans and (hopefully, in the
future) in robots. If we have achieved our goal, this paper
will stimulate further theoretical and empirical research
towards these ends. We hope that this is but one of many
entries to follow in the continuing discussion of
behavioral growth in robots. In closing this paper, we
want to argue for a relation between ongoing emergence
and theorizing in cognitive science, we discuss additional
means by which ongoing emergence may be achieved
incrementally in epigenetic robotics research, and we
close with a view to the future.
In Section 5 we raised a distinction between using
pre-programmed initial skills (Section 5.1) and not using
pre-programmed initial skills but instead relying purely
on the properties of ongoing emergence (Section 5.2). In
the pre-programmed initial skills case, we take this to be
analogous to Fodor’s concept of modularity (Fodor,