IEEE Consumer Electronics Magazine - July 2017 - 28

A few neurorobotics models
have used more realistic
implementations of brain-like
circuitry and neurons to take direct
inspiration from neuroscience
findings on the distributed
control of behavior.
which uses artificial neural networks (and brain-machine
interfaces) for robot control and learning.
All of these approaches share four key principles in the
design of social robots interacting with people:
1) embodiment, i.e., the robot's sensors and actuators shape
the type of behavior and cognitive control strategy developed by the robot [29], [30]
2) situatedness, i.e., where cognition is situated in the agent's
interaction history in its physical and social environment [31]
3) grounding, where the agent's internal representations, in
regard to language, are directly and intrinsically grounded
in its experience of the world [32], [33]
4) learning, as the agent's behavioral and cognitive strategies
are autonomously acquired during its lifetime through
evolutionary or developmental learning, e.g., implemented
with neural networks, reinforcement learning, or other
machine-learning techniques.
In the next section, we will look at two examples of cognitive robotics applications exploiting the four key principles.
The first approach shows the advantages of using a developmental robotics approach that implements the four design
principles: this consists of an experiment where the robot uses
its body posture (embodiment principle) to acquire language
via a situated human-robot interaction (situatedness principle)
to ground the language (grounding principle) in the robot's
task and representation. Language is acquired (learning principle) during development. The second example focuses on
neurorobotics, with a specific emphasis on the implementation
for neural network controllers in neuromorphic systems. This
puts the main emphasis on the learning principle, though
within a situated and embodied interaction with its world.

INTERACTING WITH ROBOTS: A DEVELOPMENTAL
LANGUAGE-LEARNING APPROACH
Developmental robotics is the "interdisciplinary approach to
the autonomous design of behavioral and cognitive capabilities in artificial agents (robots) that takes direct inspiration
from the developmental principles and mechanisms observed
in natural cognitive systems (children)" [27]. As such, this
approach puts a strong emphasis on the embodied and situated interaction between the (baby) robot and its caregiver or
human tutor, taking inspiration from child development. For
example, recent developmental psychology studies have
28 IEEE Consumer Electronics Magazine

july 2017

investigated how body posture might be playing a critical role
in early word learning [34]. To learn anything at all from realtime experiences, a physical learner must be able to orient its
sensors, and thereby its physical body, to attend to the referred
object. Part of the learning challenge then is to react appropriately, e.g., orienting to the spatial locations of objects. Here,
we briefly present an embodied developmental approach,
mapping the body posture to expected sensory experience for
the learning of object names.
Morse et al. [35] use the humanoid robot iCub specifically
to show the role of embodiment and body posture in supporting early word learning both in human infants and in baby
robots. The robot's control system uses the epigenetic robotics architecture (ERA), which is a connectionist model combining self-organizing maps (for the robot's categorization
abilities) and Hebbian learning (for the learning of objectname associations) [Figure 4(a)]. To highlight the relevance
of this approach, we show a short example using epigenetic
[36] architecture as a metaphor in the construction of learning systems. Using this paradigm, we can compare the epigenetic approach to both robotics and neuroprosthetics
learning processing.
The term epigenetics, originally used by Jean Piaget,
emphasizes the role of both the environment and of genetics
in development. The ERA robotics example also refers to the
concept of a Hebbian network [37]: a simulation based on a
concept closely connected to epigenetics, where the connective model becomes predominant, expressing itself to build
simulations of brain plasticity and connections learned in an
epigenetic way. So, a set of Kohonen self-organized maps
[38] are pretrained for object classification according to their
color and shape (color and space maps). These maps tend to
data reduction, with the consequence of a specialization of
neural networks, rather than to its expansion. This neural
model increases the connectivity of the system by reducing
the number of artificial connections (nodes). In this case, the
network models are based on a competitive learning algorithm. In addition, in ERA an extra Kohonen map is pretrained to recognize the robot's body posture (posture map).
A set of Hebbian associative learning weights, connecting the
nodes between different Kohonen maps, are trained during
language training sessions with a human teacher. These are
directly inspired by the child psychology studies of Samuelson et al. [34].
During these language-learning sessions [Figure 4(b)], the
iCub changes its own posture to attend to different parts of the
scene (i.e., to the left and the right side of the robot's peripersonal space) where two objects are shown. The Hebbian
weights are then adjusted to create new associations between
the visual color and shape maps, the posture map, and the nodes
representing words. This is the key mechanism implementing
language grounding in perception and action.
The same modeling setup has also been used not only to
replicate previous child psychology experiments, but also to
make predictions on additional phenomena. This, for example, includes the prediction of the role of body posture in

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - July 2017