The intelligence of environments 153
by the algorithms nor the specifically relevant properties of the environment nor
the behaviours of the robot itself can be fully predicted by the experimenter. Elementary cognitive abilities are part of this endeavour only to the extent that they
are part of the evolutionary story that is modelled here, and to the extent that the
general method allows for scaling up towards more complex, at least minimally
cognitive tasks – which has been the topic of later experimental work (as, e.g.
reported in Beer 2003; Beer and Williams 2015). The issue of scaling and its
possible limits has been raised both with respect to the practical difficulties of
scaling (e.g. by Nelson 2014) and as a more fundamental objection to all sorts of
behaviour-based, bottom-up, representation-free approaches to adaptive behaviour as the foundation of cognition.
A markedly different approach to building systems that use natural information
in pursuit of producing an analogue of natural adaptive behaviour is taken by what
is called “cognitive robotics” by its proponents (ICS-a/b-2015).
3
Starting, as it
were, at the other end of systemic complexity but omitting the evolutionary aspect
that would provide a foundation for genuine functional analogies, this approach
commits itself to viewing cognition from a systems perspective. This means that
cognitive systems are considered as integrated wholes, so as to include in any
model as many aspects of their embodiment as practically possible.
The rationale behind this perspective partly parallels the ecological approach to
perception discussed in Chapter 3. It commences from the critical observation that
the cognitive sciences in general, and many computational and robotic approaches
in particular, are too narrow in their focus when considering only one or a few
components of a cognitive system in isolation, in a detached and typically rather
static experimental setting. Moreover, it is maintained that the relevance of touch
and other senses and of mechanisms of bodily feedback has been underestimated
in much of (very much vision-centred) cognitive science. A plausible model of a
cognitive system, according to cognitive robotics, needs to account for its bodily
and environmental contexts and the interplay between them, in view of the integration of various sensory modalities.
Models of this kind are realised as humanoid robots that incorporate various
human-like physical traits and a broad range of sensor input, including vision,
touch and key modes of proprioception such as balance and kinaesthesia. For
example a visual system, human or robotic, will not be able to stabilise its view
on its own when moving or being moved. Stabilisation requires physiological
feedback patterns that keep the eyes or cameras fixated on the object of attention.
A robot visual system will also have to replicate the functional roles of fovea and
periphery of the visual field in human vision. In order to provide tactile feedback, the robot will be equipped with touch-sensitive artificial skin, and it will use
accelerometers to provide kinaesthetic feedback (Mittendorfer and Cheng 2011;
Wieser et al. 2011). Cognitive robots that functionally integrate these kinds of
capabilities serve as a research platform to generate plausible neurological models
of how human beings perform the same set of cognitive tasks.
Hence, the kind of humanoid robots involved here is humanoid not for the sake
of creating human-like effects on a behavioural or interactive level (as the social
by the algorithms nor the specifically relevant properties of the environment nor
the behaviours of the robot itself can be fully predicted by the experimenter. Elementary cognitive abilities are part of this endeavour only to the extent that they
are part of the evolutionary story that is modelled here, and to the extent that the
general method allows for scaling up towards more complex, at least minimally
cognitive tasks – which has been the topic of later experimental work (as, e.g.
reported in Beer 2003; Beer and Williams 2015). The issue of scaling and its
possible limits has been raised both with respect to the practical difficulties of
scaling (e.g. by Nelson 2014) and as a more fundamental objection to all sorts of
behaviour-based, bottom-up, representation-free approaches to adaptive behaviour as the foundation of cognition.
A markedly different approach to building systems that use natural information
in pursuit of producing an analogue of natural adaptive behaviour is taken by what
is called “cognitive robotics” by its proponents (ICS-a/b-2015).
3
Starting, as it
were, at the other end of systemic complexity but omitting the evolutionary aspect
that would provide a foundation for genuine functional analogies, this approach
commits itself to viewing cognition from a systems perspective. This means that
cognitive systems are considered as integrated wholes, so as to include in any
model as many aspects of their embodiment as practically possible.
The rationale behind this perspective partly parallels the ecological approach to
perception discussed in Chapter 3. It commences from the critical observation that
the cognitive sciences in general, and many computational and robotic approaches
in particular, are too narrow in their focus when considering only one or a few
components of a cognitive system in isolation, in a detached and typically rather
static experimental setting. Moreover, it is maintained that the relevance of touch
and other senses and of mechanisms of bodily feedback has been underestimated
in much of (very much vision-centred) cognitive science. A plausible model of a
cognitive system, according to cognitive robotics, needs to account for its bodily
and environmental contexts and the interplay between them, in view of the integration of various sensory modalities.
Models of this kind are realised as humanoid robots that incorporate various
human-like physical traits and a broad range of sensor input, including vision,
touch and key modes of proprioception such as balance and kinaesthesia. For
example a visual system, human or robotic, will not be able to stabilise its view
on its own when moving or being moved. Stabilisation requires physiological
feedback patterns that keep the eyes or cameras fixated on the object of attention.
A robot visual system will also have to replicate the functional roles of fovea and
periphery of the visual field in human vision. In order to provide tactile feedback, the robot will be equipped with touch-sensitive artificial skin, and it will use
accelerometers to provide kinaesthetic feedback (Mittendorfer and Cheng 2011;
Wieser et al. 2011). Cognitive robots that functionally integrate these kinds of
capabilities serve as a research platform to generate plausible neurological models
of how human beings perform the same set of cognitive tasks.
Hence, the kind of humanoid robots involved here is humanoid not for the sake
of creating human-like effects on a behavioural or interactive level (as the social