Provided by Cognitive Sciences ePrint Archive
Towards Learning Affective Body Gesture
Andrea Kleiiismitli
University of Aizu, Database Systems Lab
Aizu Wakamatsu, 965-8582 Japan
[email protected]
Nadia Bianclii-Bertlioiize
University of Aizu, Database Systems Lab
Aizu Wakamatsu, 965-8582 Japan
nadia⅛⅛u-aizu. ac. j p
Abstract
Robots arc assuming an increasingly impor-
tant- role in our society. They now become
pets and help support children healing. In
other words, they arc now trying to entertain
an active and affective communication with
human agents. However, up to now, such
systems have primarily relied on the human
agents’ ability to empathize with the system.
Changes in the behavior of the system could
therefore result- in changes of mood or behav-
ior in the human partner. This paper de-
scribes experiments we carried out- to study
the importance of body language in affective
communication. The results of the experi-
ments led us to develop a system that can in-
crement-ally learns to recognize affective mes-
sages conveyed by body postures.
1. Introduction
Many researchers have been trying to exam-
ine the relation between emotion and nonver-
bal cues. In the same vein of work as
(Brcazcal and Scasscllati, 1999), researchers arc ex-
plicitly exploiting the empathy of human caretakers
for the human-like characteristics of the system and
arc trying to identify the causes for this empathy.
This is achieved by codifying expressions to repre-
sent- some emotional state. The robots often do not
have any real learning capability, besides hard-wired
evolution. These studies also investigate the effects
of changes in the system on the behavior of the hu-
man caretaker. In those systems, affective behaviors
arc hcncc mainly passive, and the system’s ability
to react- to its human partner’s emotion is generally
very limited or missing, resulting in a social interac-
tion that fails over time. Indeed, over the long run,
habituation takes place - the human partner gets
bored - because the system docs not- seem to react
to the user, or only in clearly stereotypical ways. In
that sense, current- systems miss the b і-directionality
and/or unpredictability typical of human social in-
teraction (Bianchi-Bcrthouzc and Lisctti, 2002).
Suggested by (Picard, 1997), a different- line of re-
search exploits information from physiological cues
to detect- and react- to emotions. For example pulse,
galvanic skin response, temperature, and blood pres-
sure can be measured by sensors to understand
changes in the affective state of the human.
Meanwhile, little attention has been placed on the
visual signals to be extracted from body posture.
Dance and choreographic studies have shown that
it- is a powerful and frequently used means of com-
munication.
2. Movement Features and their Af-
fective Messages
An interesting aspect- of body language is that it
docs not- take body structural similarities between
agents for successful interaction to occur. Humans
can read affective body language in bodies that are
not- human-like. Accordingly, the modeling of af-
fective body language in robots is very interesting
because it- requires going beyond pure recognition
of the posture to explore a more general mapping.
Studies, on dancing motions (von Laban, 1988) in
particular, have shown that a movement- conveys a
different- affective messages when its features, e.g.
its spatial dimensions, are modified. Other stud-
ies, such as described in (Nakata, 2002), have shown
that the same motion was accepted with different-
degrees of naturalness if its features, such as speed,
were changed. These studies have been performed on
non-anthropomorphic bodies, demonstrating the ex-
tent- of humans’ abilities to empathize to non-human
systems.
In this study, we carried out- preliminary experi-
ments to understand how humans recognize affective
postures. Specifically, we aimed at identifying the
features of body language that lead an observer to
associate an emotional state to a given posture or
motion. To this end, we created an avatar with a
human-like body. Facial expressions were not- used
so as to focus only on body motion. The avatar was
animated using key-framing and inverse kinematics.
Table 1 summarizes the observations made by
seven Japanese students on 5 types of affective ani-
mation. From the users’ observations, we identified
as relevant- the following motion features: Complex-
ity (C), Irregularity (I), Symmetry (S), Speed (Sp),