IEEE Robotics & Automation Magazine - December 2012 - 51

engage with humans [17]. Their robot exhibited different
emotions on the basis of its success at recruiting a passerby to
take a poll. The results showed that having an expressive face
and appropriate movement (body language) increases a
robot's chance of successful interactions.
The work of Delaunay et al. [31] focused on understanding
how a person in a human-robot interaction can read gaze
direction from a robot. Results of their experiments indicate
that although it is hard to recreate human-human interaction
performance, robot faces having a humanlike physiognomy
perform equally well, which seems to suggest that these are the
preferred candidates to implement joint attention in humanrobot interaction [31]. Saerbeck and Bartneck [92] analyzed the
relationship between the motion of a robot and perceived effect
on human beings. Acceleration and curvature appear to be the
most influential factors in how the motion is perceived. Experimental results suggest a strong relationship between motion
parameters and attribution of affect, while the type of embodiment had no effect.
Riek et al. [90] showed that people cooperate with abrupt
gestures quicker than with smooth gestures. A person's speed
at decoding robot gestures is correlated with his or her ability
to decode human gestures, and negative attitudes toward
robots are strongly correlated with a decreased ability in
decoding human gestures [90]. Chaminade et al. [21] gave
instructions to volunteers to explicitly attend to the emotion
shown by human and robot subjects. A significant increase
was observed in response to robot, but not human facial
expressions in the anterior part of the left inferior frontal
gyrus, a neural marker of motor resonance [21].
The research on robots exhibiting emotions has a rich history. The following is a short list of robots that demonstrate
different levels of facial expressions [18].
Affective Tiger is a toy robot developed as a tool for social
and emotional awareness education of young children. Tiger's
face has two degrees of freedom (2 DoF) (mouth and eyes) [57].
Minerva is an interactive tour guide robot that displays
four emotions: neutral, happy, sad, and angry. It has 4 DoF:
two for mouth control and two for the eyebrows [23].
Sparky is a teleoperated robot, which uses facial expressions, gestures, motions, and sounds to interact with people.
His face has 4 DoF to control three expressive features-eyebrows, eyelids, and lips [98].
Kismet is a famous MIT robot developed as a testbed for
learning social interactions between robots and people. Kismet's face has 18 DoF that allow the robot to express a full
range of emotions, which have been shown to be correctly
interpreted by people [16].
iCub was designed by a consortium of European institutions to simulate the perceptual system and articulation of a
small child and to interact with the world in the same way
that a child does. The robot has 53 DoF distributed between
its arms (seven each), hands (nine each), head (six), torso
(three), and legs (six each) [77].
In addition to experiments on understanding the
emotional states of robots, some work has been started on the

general analysis of avatar behavior [15]. Another novel
research direction is known as Avatar DNA, a patent-pending
technology from Raytheon [110]. A recently published article
demonstrates the feasibility of applying strategy-based purely
behavioral biometrics developed for the recognition of
human beings to the recognition of intelligent software agents
[122]. The article lays the theoretical groundwork for research
in the authentication of nonbiological entities. Specifically, it
is demonstrated that behavioral biometrics is a sound
approach to intelligent robot authentication.
Artimetrics
Database Generation and Availability
In well-established fields such as biometrics, numerous standardized and publicly available datasets exist [66], making it
possible to compare different algorithms and to test developed systems. Labeled public datasets of robot faces, avatars,
or attributed conversations from artificially intelligent agents
are currently unavailable. A visual survey of robots [38] contains images of various artificial entities but not a standardized database suitable for subsequent research. Methods for
synthetic iris, face, and fingerprint database generation for
biometric research were recently surveyed [119], [120]. However, for the robot domain, this is mainly an unexplored area
of research. Techniques for the creation of such standardized
datasets that are consistent with real-world datasets can be
imitated by examining the approaches to the generation and
evaluation of facial datasets [59], [37] utilized by biometric
systems or from chat mining research applied to gender attribution and human versus bot classification [27], [39].
The authors have begun to work on the generation of a
publicly available avatar face dataset [82], and on the collection of speech corpora from intelligent agents-two types of
data, which are of specific interest in the early artimetrics
research. One database consists of a set of high-resolution
facial images of avatars collected from two of the most popular virtual worlds: SecondLife.com and EntropiaUniverse.
com. The other database consists of a text corpus from intelligent agents who have performed extremely well in the recent
Loebner Prize in Artificial Intelligence (AI) competitions
(Loebner.net). We have developed automated tools utilizing
the power of AutoItScript.com and the Linden scripting language (LindenLab.com) for the creation of customized datasets of both kinds. With the assistance of the developed tools,
researchers in the field can effortlessly generate virtually
unlimited amount of data for visual and stylometric robot
authentication experiments. Additional work is still necessary
to make it possible to generate data with specific characteristics. Currently, it is only possible to specify the desired
amount of data, the gender of the avatars' faces, and the overall area of knowledge over which the intelligent agents communicate. It is, however, already possible to generate multiple
samples for each nonbiological entity, making it easy to perform training and testing on disjoint datasets. In parallel, we
are working on assembling a dataset of hardware robots' faces,
december 2012

IEEE ROBOTICS & AUTOMATION MAGAZINE

http://www.SecondLife.com http://www.Loebner.net http://www.AutoItScript.com http://www.LindenLab.com

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2012