COM3240 – Week 3
A Review of Children, Humanoid Robots
and Caregivers
(Arsenio, 2004)
Presented by Gizdem Akdur
A learning framework for a humanoid robot

Human-robot interactions



Inspired by cognitive development of a child




The importance of a human actor
Teaching humanoids as children
Dependence on mother
Awareness of his/her own individuality
Self-exploration of his/her surroundings
Implementation of concepts on the humanoid robot Cog
Inspiration from Mahler’s child development theory

Margaret Mahler (1897-1985) – Hungarian physician and
psychoanalyst with a main interest in mother-infant duality and
childhood development

Was influenced by Freud and Piaget

Developed the Separation-Individuation Theory of Child
Development (1979)
Mahler’s theory (1979)

Autistic Phase (from birth – 1 month old)

Symbiotic Phase (until around 5 months old)

Separation and Individuation Phase




Differentiation (5-9 months)
Practising (10-18 months)
Re-approximation (15-24 months)
Individuality and Object Constancy (24-36 months)
Learning on the Autistic and Symbiotic Phases

Autistic phase



The newborn is mostly in a sleeping state. Awakens to eat and
satisfy other necessities
Motor skills mainly consist of primitive reflexes
Symbiotic phase



Infant’s attention dropped to repeatedly moving objects and to
sudden changes of motion
Repetition helps
Motivated the design of algorithms for detection of events


Object Segmentation algorithm
extending the algorithms of previous studies – Arsenio, 2003 and
Fitzpatrick, 2003
Help from a human tutor

will guide the robot learning about its physical surroundings




Correlate data among its own senses
Control and integrate situational cues from its surrounding world
Learn about out-of-reach objects and the different representations
they might appear
therefore special emphasis will be placed on social learning along
a child’s physical topological spaces
robot executes a simple learned task (waving), and
associates the sound to the movement of its own body
Physical topological spaces
(1) the robot's personal space, consisting of itself and familiar, manipulable objects
(2) its living space, such as a bedroom or living room
(3) its outside, unreachable world, such as the image of a bear on a forest
(1) Learning about objects and itself

Strategy described for a robot to associate data from several
resources



from its own senses
from its senses and information stored on the world/robot’s
memory
3 main schemes to be implemented



Cross-modal data association
Object recognition
Educational activities
(1.1) Cross-modal data association

Extracting visual and audio features – patches of pixels and
sound frequency bands. The algorithm was therefore extended
to detect both

Identification of robot’s own acoustic rhythms and the visual
recognition of robot’s mirror image
Child and robot looking at a mirror, associating
their image to their body (image/sound
association for the robot has been amplified)
(1.2) Object recognition

A recognition scheme for objects (other than the robot’s body
part) with 3 independent algorithms




Colour
Luminance
Shape
Geometric hashing for high-speed performance

Adaptive Hash Table was implemented
Object recognition and location in a computer generated
bedroom. Scene lines matched to the train are outlined.
(1.3) Learning from educational activities

Corresponds to child’s practising (10-18 months) developmental
sub-phase towards re-approximation (15-24 months) sub-phase

Robot learns object properties not only through cross-modal data
correlations, but also by correlating human gestures and
information stored in the world structure or on its own database

Object recognition algorithm applied to extract correlations
between sensorial signals perceived from the world and geometric
shapes present in such world
(2) Learning the world structure of the robot’s physical
surroundings

Determining where objects should be stored based
on probability of finding them on that place later


If a book is placed in the fridge, the robot will hardly
find it!
The framework, developed to capture knowledge
stored in robot’s surrounding world, consists of:


(1) Learning 3D scenes from cues provided by a human
actor
(2) Learning the spatial configuration of the objects
within a scene
(2.1) Learning about scenes

The environment surrounding the robot provides additional
structure that can be learned through supervised learning
techniques

Defining scenes as a collection of objects with an uncertain
geometric configuration, each object at a minimum distance
from another
Segmentation error analysis for furniture
items on a scene – samples also shown
(2.2) Learning about objects in scenes

Humanoids (like children) need to learn the relative probability
distribution of objects in a scene


Constraining the search space is important to optimise computational
resources
Contextual features incorporate functional constraints

Wavelet transformation (Strang and Nguyen, 1996) used


Holistic representation of the scene
Main spectral characteristics of a scene encoded with a rough description
of its spatial arrangement
Reconstruction of the original image by
the Wavelet transform. An holistic
representation of the scene.
(3) Learning about the outside world through books

Books are useful to teach different object representations and to
communicate properties of unknown objects to them


Human-robot interactions are very essential at this stage. A human
tutor does the job of a mother of a child who teaches from books
by tapping on the book’s representations
Segmentation by demonstration algorithm used
(3.1) Matching multiple representations




Object representations obtained from a book are put into a
database for future recognition tasks
Methods were developed to establish a link between an object
representation and real objects from surroundings using the
object recognition technique
The framework can be applied on paintings, prints, photos and
computer generated objects
Object recognition helps with the recognition of similar shapes
with different colours but same geometric contours
Conclusion

A developmental object perception framework has been
described which aims to teach humanoids as children

The epigenetic principle taken as a foundation

Robot learned about its surrounding world by building scene
descriptions of world structures



Contextual selections by using probabilities
Storing information about object shapes for later use
The learning process with the guidance of a human tutor is
essential to help the humanoid through its cognitive development
Thanks for listening