Imitation of movements as a step towards the development of
artificial social intelligence for autonomous robots
Future developments in computer science and robotics technology are
expected to change our conception of daily life communication and
human, as well as with physical or virtual agents. My research direction,
which underlies the project be described in the following, aims at
outlining modes of social interaction between agents.
The project focusses on the investigation of imitation as a
prerequisite for the development of individual
interactions between autonomous robots. Imitation of movements is used as
a basic step towards acquiring new movement patterns and developing
social intelligent behavior. While the
theoretical background is rooted in the biological and social sciences the
project bridges the gap to the engineering domain as it aims at
constructing robots in order to experimentally test the theoretical
considerations. A methodology for this approach is not available but has
to be developed in parallel to the realization of the project.
Background: Social intelligence
Social skills have been mostly regarded as a `side-effect' or
neglected so far in those research areas which are dealing with the
construction of intelligent artifacts, namely
artificial intelligence, robotics and artificial life.
Approaches in artificial life on interactions in groups of physical
robots prefer the analogy of social insect societies,
anonymous organized societies
without individual relationships.
Many approaches consider other agents only
as moving obstacles or as competitors for limited
In the same way most research in robotics aims at developing robots with a
domain-specific `technical intelligence', e.g.,
building robots which can use
effectively and with high precision a manipulator or identify
In contrast, recently, different research proposals and studies have been initiated
outlining scenarios for applications of autonomous robots, mainly
focussing on service robots, i.e., robots belonging to our daily life,
and stressing the need for robots
to have complex abilities to interact and communicate with individuals.
The recognition of other individuals ('conspecifics') is
necessary as a means to control the interactions, to predict the behavior of
the `conspecific' and to develop complex social
relationships like `aversion' or `attachment'.
research direction is highly motivated by the
`social intelligence hypothesis' which derives from primatology research and
primate intelligence originally evolved to solve social problems and
was only later extended to problems outside the social domain.
I hypothesize, that this might
be one general principle for the evolution of intelligence in natural and
Therefore this project focusses on the study of `social dynamics'.
I decided to study social intelligence by using
`real robots' and, as a basis
for social interactions, to study imitation.
Project work: Imitation
Imitation is supposed to be
among the least common and most complex natural forms of learning.
It plays a crucial role in child development. The judgement
`Here is something like me' should enable the individual to distinguish
between animate and inanimate objects
in order to build
up individual contacts and
Imitation is also used to learn new movement patterns, e.g.,
ecologically relevant movements.
In technical applications movement learning by imitation can be used as a
`Programming by Demonstration' technique
which provides a means of
`implicit' knowledge transfer between different kinds of agents (robots,
`Teaching by showing'
approaches give examples for
`engineering approaches' to imitation. They use
an analytic approach of observing,
segmenting and classifying the actions of the model, producing a
high-level, i.e., symbolic
description of the observed movement as a plan to control the
movements of the imitator. In my point of view this
approach is restricted by the bottleneck of having to map actions
to and from a symbolic level to real world, continuous actions. I
suppose that this approach will be as effective for real world
applications as humans who try to imitate the movements of each other by
using only symbolic communication, namely language, via a telephone line.
Symbolic descriptions are not expected to
be a crucial prerequiste for the development of imitative skills, e.g.,
young children are experts at imitation before they fully
develop their symbolic skills. Alternatively I use
a self-observational, bottom-up approach to
imitation, i.e., the imitator does not `watch and
analyse the other, then acts itself' but rather uses a somehow opposite
direction, namely `acts and follows, then analyses its own movements'. While
following the first strategy action of the imitator only takes place
after the movements of the other have been analysed and transferred to the
`own body', namely to the movement capabilities and morphology of the
imitator, the second strategy starts with movement from the
beginning, following and synchronising the own movements with the model's
Work program, methods and intended results
The project's goal is to study how imitating behavior as described above
can be realized in a
group of autonomous, self-build fischertechnik robots
which follow, keep contact, copy movement patterns of others and
recognize other robots by
characteristic movement patterns.
For instance the robots should learn to avoid other robots with a highly
energy-consuming behavior or should learn to keept contact to robots which
already know how to approach a recharging station.
The robots are controlled using the
The project comprises the theoretical as well as practical investigation of the
following research issues:
- a) `Habitat': Developing an appropriate experimental environment for the
- b) Design and construction of a group of robots. Morphology and sensor
equipment should be adapted to the
environment and the modes of interaction with other robots.
- c) Programming of `simple behaviors' as the basic behavior repertoire for
the robots, e.g., phototaxis,
wall-following, recharging behavior and so on.
- d) Overall control architecture: A crucial point in writing
programs is to design the single behavior elements
in order to achieve emergent and observable behavior. Additional levels
of abstraction might be necessary for
implementing systematically highly complex behavior.
- e) Recognition of conspecifics: In order to recognize other robots it is
necessary to `store and retrieve' in
a sense data about characteristics of the conspecifics. So far it is not at
all clear how one could implement
memory in an `implicit' sense, in a way more relevant to a dynamic system
approach, that is,
without explicitly selecting different behaviors or using a classical
- f) Robot group structures: Using imitation and recognition of conspecifics
results in `individual
relationships'. This allows to study the development of group structures
hierarchies), depending on environmental
and social constraints.
The project's results within the time available
should comprise (1) concretization of movement
learning by imitation in a physical
realization (building and controlling autonomous robots) and (2)
the investigation of social group behavior.
Back to Kerstin's Home Page.
1996-03-20 Kerstin Dautenhahn