ISIS Research Generation Symposium

embarrassedlopsidedΤεχνίτη Νοημοσύνη και Ρομποτική

14 Νοε 2013 (πριν από 3 χρόνια και 9 μήνες)

95 εμφανίσεις

ISIS Research Generation Symposium

Models of Infant Development: Are We Really Serious about Environmental Interaction
and Dynamics?

Symposium Chairs: Christopher G. Prince & Lakshmi J. Gogate


Word count:

Epigenetic robotics, a recent advance in formal modeling (e.g., Lungarella et al., 2003),
shows promise for the study of infant development. Epigenet
ic robots use a collection of
techniques (e.g., connectionist models) in combination with sensors (e.g., cameras) and
effectors (e.g., robot arms) to model the behavioral development of infants. Through the
grounding provided by sensors and effectors, thes
e robots can approximate the dynamics
of a child’s interaction with the physical world. Thus, the advantage for researchers in
infant psychology in using such models lies in having access to theoretical and
methodological tools to further our understanding

of the influences of environmental
interactions and dynamics on the developing organism. In this symposium, we argue that
if we are serious about understanding the influences of environmental interaction and
dynamics on child development, more researchers

need to look to epigenetic robotics.
While models, by their nature, abstract away from the specifics of what they model, it is
also true that there is a fine line between abstraction and modeling that is
too simple
Certainly, knowledge can still be gaine
d with models that don’t explicitly interact with
the physical world. However, much of contemporary theorizing in developmental
psychology implicates the importance of the dynamics of an organism’s environmental
interaction (e.g., Thelen & Smith, 1998). Gi
ven a need for methodologies in
developmental psychology spanning empirical study, theories, and formal modeling, how
can we progress in contributions to the formal modeling of environmental interactions
and developmental dynamics by stopping short of the
physical world? For example,
connectionist models often take as inputs symbolic representations, which are highly
abstracted in their relation to the physical world (e.g., see Quinlan, 2003). However,
because psychological development involves ongoing nonl
inear interaction of the
organism’s properties with its world, even slight changes in organism or environment
properties can change the developmental trajectory. Robotic models of infant
development are promising because of their high fidelity regarding th
e dynamics of
interaction with the physical world, which should more closely approximate the
developmental trajectories of the infants being modeled. For example, models of word
learning can use the dynamics of the gaze of an adult (e.g., Yu & Ballard, 200
4), models
of swinging can use body
environment dynamics (Berthouze & Lungarella, 2004), and
models of self
other discrimination can utilize sensory
motor interaction with the world
(e.g., Edsinger
Gonzales, 2005). Because a child’s interactions with her p
hysical and
social world are crucial to her development, the place to bolster (not simplify!) our
formal models of infant development is in these dynamics. We note that epigenetic
robotics research is in its infancy, and is just starting to address these q
uestions. Many
further models and infant
model comparisons are needed.

Justification for panelist selections:

Christopher Prince has helped organize a
series of workshops on epigenetic robotics held since 2001. He publishes on sensory
oriented models of in
fant audio
visual synchrony perception, and on theoretical issues in
epigenetic robotics.

Katharina Rohlfing’s research focuses on the interface between language
acquisition (including gestural communication) and conceptual development. She
recently organi
zed a symposium on multimodal behavioral modifications in parents.
Yukie Nagai’s research is in cognitive developmental robotics. With M. Asada, she has
published robotic learning models for joint visual attention. Britta Wrede studies speech
processing in

the domain of human
robot interaction where she models prosodic aspects
of dialog and investigates multi
modal integration.

Sylvain Sirois’ research has involved neural network models of learning and
development since 1995. His collaborations with Thomas
Shultz and Denis Mareschal
have involved modeling and empirical studies of category learning and infant habituation.
His recent modeling involves robotics. He has been a Programme Committee member of
the Epigenetics Robotics workshop since 2003.

What we k

Epigenetic Robotics and Environmental Interactions and Dynamics

Christopher G. Prince, Lakshmi J. Gogate, Nathan A. Helder, & George J. Hollich

Word count:

Our thesis is that robotic models of infant development can provide developmental
logists with valuable theoretical and methodological tools to further our
understanding of the influences of environmental interactions and developmental
dynamics. In one of the first examples of this type of research, Metta et al. (1999)
constructed a rob
otic model that learned to reach. To improve its reaching behavior, the
robot used a process of visually fixating on an object, attempting to reach for that object,
and feedback from an inaccurate reach. An accurate

behavior in the robot
ed the acquisition of accurate
. Initially inaccurate reaching was
intended to approximate the asymmetric tonic neck reflex, with the robot’s arm initially
extending in roughly the direction in which the robot’s head was turned. The robot first
ated on a target object, next reached roughly towards the object, and then re
fixated on
the end of its arm. The robot used this process of reaching and re
fixation to learn a more
accurate relation between fixation and reaching. Other work on robotic mode
ls of motor
behavior development has evaluated the hypothesis that freezing and unfreezing
peripheral degrees of motor freedom (joints) enables motor control to be more easily
learned (e.g., Bernstein, 1967). While freezing and unfreezing of degrees of mot
freedom in children can be observed, the controlling variables and parameters may be
difficult to unambiguously determine via experimentation. Robotic models provide a
means by which these issues may be disambiguated. Using robotic techniques, Berthouze

and Lungarella (2004) found that a stable swinging behavior could be more readily
accomplished by phases of first freeing only one joint, and then freeing a second joint to
swing. Other examples of infant
environment dynamics include infants changing wher
they are looking based on their present interactions. Related robotic models include
habituation (Lovett & Scassellati, 2004; Chen & Weng, 2004; Sirois, 2005), caregiver
infant interaction (Breazeal & Scassellati, 2000), joint attention (Nagai et al., 20
03), and
visual discrimination (Seth et al, 2004). Another promising area of epigenetic robotics is
perceptual self
other discrimination. To learn about her own body and distinctions
between body and environment, presumably an infant relies on her interact
ions with the
world (e.g., Rochat & Striano, 2000). Initial robotics models in this area include
Gonzales (2005), Gold and Scassellati (2005), and Olsson, et al. (2005). These
models augment existing tools in developmental psychology

allowing expl
oration of
hypotheses about specific developmental mechanisms using realistic environments.

REFERENCES (Symposium & “What we know”)

Berthouze, L., & Lungarella, M. (2004). Motor skill acquisition under environmental
perturbations: On the necessity of alt
ernate freezing and freeing of degrees of freedom.
Adaptive Behavior
, 47

Bernstein, N. (1967).
The Coordination and Regulation of Movements
. Oxford:

Breazeal, C. & Scassellati, B. (2000). Infant
like social interactions between a robot an
d a
human caregiver.
Adaptive Behavior
, 49

Chen, Y. & Weng, J. (2004). Developmental learning: A case study in understanding
“object permanence.” In
Proceedings of the Fourth International Workshop on
Epigenetic Robotics: Modeling Cognitive Developm
ent in Robotic Systems
. Lund,
Sweden: Lund University Cognitive Studies.

Gonzales, A. (2005). Developmentally guided ego
exo force discrimination for
a humanoid robot.
Proceedings of the Fifth International Workshop on Epigenetic
Robotics: Modelin
g Cognitive Development in Robotic Systems

Gold, K. & Scassellati, B. (2005). Learning about the self and others through
Developmental Robotics AAAI Spring Symposium

Lovett, A., & Scassellati, B. (2004). Using a robot to reexamine looking
experiments. In
Proceedings of the Third International Conference on Development and

Lungarella, M., Metta, G., Pfeifer, R., & Sandini G. (2003). Developmental robotics: A
Connection Science
, 151

Nagai, Y., Hosoda, K., Morit
a, A., & Asada, M. (2003). A constructive model for the
development of joint attention.
Connection Science
, 211

Olsson, L., Nehaniv, C. L., & Polani, D. (2005). From unknown sensors and actuators to
visually guided movement.
Proceedings of the 4

IEEE International Conference on
Development and Learning
. Osaka, Japan, July 19

Quinlan, P. T. (2003). (Ed.)
Connectionist Models of Development: Developmental
Processes in Real and Artificial Neural Networks
. New York, NY: Psychology Press.

P. & Striano, T. (2000). Perceived self in infancy.
Infant Behavior &

, 513


Schlesinger, M., & Casey, P. (2003). Where infants look when impossible things happen:
Simulating and testing a gaze
direction model.
Connection Science

Seth, A. K., McKinstry, J. L., Edelman, G. M., & Krichmar, J. L. (2004). Visual binding
through reentrant connectivity and dynamic synchronization in a brain
based device.
Cerebral Cortex
, 1185

Sirois, S. (2005). Hebbian motor control in

a robot
embedded model of habituation.
Proceedings of the International Joint Conference on Neural Networks

(IJCNN 2005)
(pp. 2772
2777). IEEE Press.

Thelen, E. & Smith, L. B. (1998). Dynamic systems theories. R. M. Lerner (Ed.)
Handbook of Child Psycholo
, 5th Edition, Volume 1:
Theoretical Models of Human

(pp. 563
634). John Wiley: New York.

What we don’t know:

Models of Infant Development: How to make sense of environmental interaction and

Katharina J. Rohlfing, Britta Wrede and
Yukie Nagai

If we are serious about environmental interaction and dynamics, in modeling we have to
consider (1) humans’ interaction with the physical and social world [9] and (2) the
intermodality of human capabilities, i.e., their coordination across mul
tiple sensory
modalities [1].

Assuming that a newborn is bombarded with a variety of sensory stimuli [10], it is crucial
to explain how the child actively filters the information from the environment and attends
to certain sources while ignoring others. W
hat has so far been less incorporated in
algorithmic implementations [4] [11] of active filtering is the fact that the caregivers’
behavior is modified across multiple modalities such as language or hand movements
([3], [5], [13]) when directed towards a c
hild. The behavior modifications seem to be
driven by infants’ perceptual system responding especially well to them ([5], [7]). The
interplay of perceptual modalities is hypothesized to be the key in decoding meaning;
e.g., stimuli presented at the same ti
me, rhythm or intensity across modalities have the
power to guide the infant’s attention to crucial aspects of the interaction ([1], [2], [6]).

Our questions for future research are based on studies of multimodal perception and
action ([6], [10]). These
studies show that when perceptual (e.g. modified hand
movements) or social information (such as language) is redundant across modalities,
young infants selectively attend to it [9]. We ask how infants decode perceptual
information by decomposing the presen
ted behavior into meaningful units.

One solution suggested by [2] is a tight interaction between social and perceptual
information called


[8]. While considerable psychological research
focused on parental behavior being modified in dif
ferent modalities,


studies to date
have systematically investigated how parents may package their action acoustically and
how this influences infants’ action understanding and production. Additionally, an open
question is how perception within modalitie
s (e.g. visual perception of human motion or
gaze) may interact with each other, and the research on this question has just started

e.g. [11] [12]. What is lacking is concrete modification parameters and architectures of
organismic multimod
al interactions in order to apply them to robots and,
thus, to model interactive and adaptive learning processes.

By modeling a robotic system that is socially interactive we are on the way to investigate
and evaluate the interplay between cues from diff
erent modalities and to get insights into
how multimodal information enables a system to understand the basic structure of
environmental behavior.


[1] Bahrick, L. E. (2003). Development of intermodal perception. In L. Nadel (Ed.),
a of Cognitive Science (pp. 614
617). London: Nature Publishing

Brand, R. J., & Baldwin, D. A. (2005).
Motionese and Motherese: Two avenues for
supporting infant action processing.

Paper presented at the X. International
Congress for Studies in
Child Language, 25.

29. July, Berlin, Germany.

Brand, R. J., Baldwin, D. A., & Ashburn, L. A. (2002). Evidence for 'motionese':
modifications in mothers' infant
directed action. In Developmental Science, 5, 72

[4] Breazeal, C. & Scassellati, B. (
2000). Infant
like social interactions between a robot
and a human caregiver.
Adaptive Behavior
, 49

Dominey, P. F., & Dodane, C. (2004). Indeterminancy in language acquisition: the
role of child directed speech and joint attention. In Journal of

17, 121

Gogate, L. J., Bahrick, L. & Watson, J. (2000). A study of multimodal motherese: The
role of temporal synchrony between verbal labels and gestures. In Child

71 (4), 878

Gogate, L. J., & Bahrick, L.
E. (2001). Intersensory redundancy and 7
infants' memory for arbitrary syllable
object relations. In Infancy, 2, 219

[8] Hirsh
Pasek, K., & Golinkoff, R. M. (1996). The Origins of Grammar: Evidence
from Early Language Comprehension. Cambri
dge, MA: MIT Press.

[9] Hollich, G., Hirsh
Pasek, K., Tucker, M. L., & Golinkoff, R. M. (2000). The change is
afoot: Emergentist thinking in language acquisition. In P. B. Anderson, C.
Emmeche, N. O. Finnemann & P. Voetmann Christiansen (Eds.),
(pp. 143
178). Aarhus: Aarhus University Press.

Lamb, M. E., Bornstein, M. H., Teti, D. M. (2002).
Development in infancy.
Mahwah, NJ: Erlbaum.

[11] Nagai, Y (2005). Learning to Comprehend Deictic Gestures in Robots and Human
Infants. In Proc
eedings of the 14th IEEE International Workshop on Robot and
Human Interactive Communication (RO
MAN'05), pp. 217

[12] Wrede, B., Fritsch, J. & Rohlfing, K.J. (2005). How can prosody help to learn
actions? In Proceedings of the Fourth International Co
nference on Development
and Learning (ICDL 2005).

[13] Zukow
Goldring, P. (to appear). Assisted imitation: Affordances, effectivities, and
the mirror system in early language development. In M. A. Arbib (Ed.). From
action to language. Cambridge: CUP.

ere to go from here?

Taking infants and development seriously

Sylvain Sirois

School of Psychological Sciences, The University of Manchester


Word count:

Epigenetic Robotics
, which refers to embodied, situated models of cognitive development
(Berthouze & Ziemke, 2003)
, makes use of tools that allow for a
level of specification
that eludes verbal theories. In order to work, models require that theoretical ideas be
translated into explicit, formal, and complete operations. Moreover, embodied, situated
models offer a better level of interpretation than “vacuu
m” models that process
information detached from the environment (with little regard for sensorimotor
constraints on information processing). Real infants are embodied and situated. Due
consideration of infant
environment interactions and dynamics within a

framework thus promises substantial advances in understanding the emergence of
cognition (e.g.,
Schlesinger, 2003)
. Yet two separate, important difficulties face such



are crucial when considering the relative merits of embodied,
situated models over vacuum models (e.g., disembodied neural
networks). When a model
is given a body and an environment and ceases to function properly, the temptation is to
reject the model outright. But while the failure of the model could be due to the model
itself, it could also be due to decisions about input t
ransduction, motor control, and /or the
interfaces between the model and the transducers and effectors. Conversely, successful
situated and embedded models could capitalize on the increased degrees of freedom
(computational, mechanical and/or environmental
) available to them, and misrepresent
information processing in infants. In order to address these systems issues, it may prove
useful for the study of infant
environment interactions to take guidance from (but not be
reduced to) developmental cognitive ne
(Johnson, 1997)
. Known properties of
the infant brain can help design co
gnitive models and interfaces within larger
systems (e.g.,
Alexander & Sporns, 2002)
. Similarly, models that violate known
properties of the brain functions/systems involved in infants must be evaluated with

(Sirois & Mareschal, 2002)

The other issue, delineating the
mechanisms of development,

is a recurrent problem in develo
pmental psychology.
Distinguishing learning and development is typically a function of outcomes, not
(Liben, 1987)
. The majority of situated, embodied models are victims of a
lar fallacy: they are deemed developmental because they model developmental data,
yet in effect they are learning models and propose no transition mechanisms, even though
genuine computational developmental models exist
(Shultz, 2003)

In summary, the
promise of progress through models of infant
environment interactions is well founded,
but requires serious consideration of both neuroscience and development
(Quartz &
1997; Quinlan, 1998)


Alexander, W. H., & Sporns, O. (2002). An embodied model of learning,
plasticity, and reward.
Adaptive Behavior, 10
4), 143

Berthouze, L., & Ziemke, T. (2003). Epigenetic robotics


development in robotic systems.
Connection Science, 15
(4), 147

Johnson, M. H. (1997).
Developmental cognitive neuroscience : An introduction
Oxford: Blackwell.

Liben, L. S. (1987). Information processing and Piagetian theory: Conflict or

Liben, Lynn S (Ed). (1987). Development and learning: Conflict or
congruence? The Jean Piaget Symposium series.

(pp. 109
132). Hillsdale, NJ, England:
Lawrence Erlbaum Associates, Inc.

Quartz, S. R., & Sejnowski, T. J. (1997). The neural basi
s of cognitive
development: A constructivist manifesto.
Behavioral and Brain Sciences, 20
(4), 537


Quinlan, P. T. (1998). Structural change and development in real and artificial
neural networks.
Neural Networks, 11
(4), 577

Schlesinger, M. (200
3). A lesson from robotics: Modeling infants as autonomous
Adaptive Behavior, 11
(2), 97

Shultz, T. R. (2003).
Computational Developmental Psychology
. Cambridge,
MA: MIT Press.

Sirois, S., & Mareschal, D. (2002). Models of habituation in infa
Trends in
Cognitive Sciences, 6
(7), 293