ArticlePDF Available

Gestural Turing Test A Motion-Capture Experiment for Exploring Believability In Artificial Nonverbal Communication

Authors:

Abstract and Figures

One of the open problems in creating believable characters in computer games and collaborative virtual environments is simulating adaptive human-like motion. Classical artificial intelligence (AI) research places an emphasis on verbal language. In response to the limitations of classical AI, many researchers have turned their attention to embodied communication and situated intelligence. Inspired by Gestural Theory, which claims that speech emerged from visual, bodily gestures in primates, we implemented a variation of the Turing Test, using motion instead of text for messaging between agents. In doing this, we attempt to understand the qualities of motion that seem human-like to people. We designed two gestural AI algorithms that simulate or mimic communicative human motion using the positions of the head and the hands to determine three moving points as the signal. To run experiments, we implemented a networked-based architecture for a Vicon motion capture studio. Subjects were shown both artificial and human gestures, and were told to declare whether it was real or fake. Techniques such as simple gesture imitation were found to increase believability. While we require many such experiments to understand the perception of human-ness in movement, we believe this research is essential to developing a truly believable character.
Content may be subject to copyright.
Gestural Turing Test
A Motion-Capture Experiment for Exploring
Believability In Artificial Nonverbal Communication
Jeffrey Ventrella
Simon Fraser University, School of Interactive
Art and Technology, Vancouver BC
jeffrey@ventrella.com
Magy Seif El-Nasr
Simon Fraser University, School of Interactive
Art and Technology, Vancouver, BC
magy@sfu.ca
Bardia Aghabeigi
Simon Fraser University, School of Interactive
Art and Technology, Vancouver BC
b.aghabeigi@gmail.com
Richard Overington
Emily Carr University of Art and Design,
Vancouver, BC
roverington@ecuad.ca
ABSTRACT
One of the open problems in creating believable characters in
computer games and collaborative virtual environments is
simulating adaptive human-like motion. Classical artificial
intelligence (AI) research places an emphasis on verbal language.
In response to the limitations of classical AI, many researchers
have turned their attention to embodied communication and
situated intelligence. Inspired by Gestural Theory, which claims
that speech emerged from visual, bodily gestures in primates, we
implemented a variation of the Turing Test, using motion instead
of text for messaging between agents. In doing this, we attempt to
understand the qualities of motion that seem human-like to
people. We designed two gestural AI algorithms that simulate or
mimic communicative human motion using the positions of the
head and the hands to determine three moving points as the signal.
To run experiments, we implemented a networked-based
architecture for a Vicon motion capture studio. Subjects were
shown both artificial and human gestures, and were told to declare
whether it was real or fake. Techniques such as simple gesture
imitation were found to increase believability. While we require
many such experiments to understand the perception of human-
ness in movement, we believe this research is essential to
developing a truly believable character.
Categories and Subject Descriptors
I.2.0 [Artificial Intelligence]: General
I.2.m [Artificial Intelligence]: Miscellaneous
J.4 [Social and Behavioral Sciences]: Psychology
General Terms
Algorithms, Measurement, Documentation, Performance, Design,
Experimentation, Human Factors, Languages, Theory.
Keywords
Turing test, believability, gestural theory, virtual agent, motion
capture, nonverbal, point light displays
1. INTRODUCTION
Alan Turing’s thought-experiment of the 1950’s was proposed as
a way to test a machine’s ability to demonstrate intelligent
behavior. Turing had been exploring the question of whether
machines can think. To avoid the difficulty of defining
“intelligence”, he proposed taking a behaviorist stance, and to ask:
can machines do what we humans do? [19] In this thought
experiment, a human observer engages in a conversation (using
text-chat only) with two hidden agents one of them is a real
human and the other is an AI program. Both the human and the AI
program try to appear convincingly human. If the observer
believes that the AI program is a real human, then it passes the
Turing Test.
The focus on verbal language in this and other
explorations of intelligence is characteristic of classical AI
research. Verbal language may indeed be the ultimate indicator of
human intelligence, but it may not be the most representative
indicator of intelligence in the broadest sense. Inventor/thinkers
such as Rodney Brooks remind us that intelligence might be best
understood, not as something based on a system of abstract
symbols and logical decisions, but as something that emerges
within an embodied, situated agent that must adapt within an
environment [2]. If we can simulate at least some basic aspects of
the embodied foundations of intelligence, we may be better
prepared to then understand higher intelligence, and thus model
and simulate believable behaviors in computer games and
collaborative virtual environments. Justine Cassell said it well:
We need to locate intelligence, and this need poses problems for
the invisible computer. The best example of located intelligence,
of course, is the body.” [3].
Gestural Theory [8] claims that speech emerged out of
the more primal communicative energy of gesture. If this theory is
correct, then perhaps we should explore this gestural energy as a
© AAMAS This is the author's version of the work. It is posted here for your personal use. The definitive
version was published in Jeffery Ventrella, Magy Seif El-Nasr, Bardia Aghabeigi, and Richard Overington.
Gestural Turing Test: A Motion-Capture Experiment for Exploring Nonverbal Communication. AAMAS 2010
International Workshop on Interacting with ECAs as Virtual Characters, May 10, 2010. Retrieved from the Simon
Fraser University Institutional Repository.
viable indicator of intelligence. We have set up an experiment to
run a Turing Test using an “alphabet” of three moving dots
instead of the alphabet of text characters. In this experiment, the
agents (one of then might be non-human) interact, and generate
spontaneous body language through their ongoing interactions.
One may ask: what is there to discuss if you only have a few
points to wave around in the air? In the classic Turing Test, you
can bring up any subject and discuss it endlessly. But remember
the goal of the Turing Test: to fool a human subject into believing
that an AI program is a human. However that is accomplished is
up to the subject and the AI program. Turing chose the medium of
text chat, which is devoid of any visual or audible queues. Body
language was thus not an option for Turing. In contrast, we are
using a small set of moving dots, and no verbal communication.
Moving dots are abstracted visual elements (like the alphabet of
written language), however, they are situated in time, and more
intimately tied to the energy of natural language.
1.1 Prior Work, and Stated Contribution
Several variations of the Turing test based on simulated human
motion have been implemented [20], [18], [11]. Imitation of
human behavior has been shown to be effective in creating
believability in virtual agents, such as work by Kipp [10]
describing a system that uses imitation of human gesture to
generate conversational gestures for animated embodied agents.
Gorman [6] shows that imitation of the behaviors of a computer
game player creates enhanced believability in artificial agents.
Stone, et al [17] describe a technique for reproducing the structure
of speech and gesture in new conversational contexts. Neff, et al,
show how the gestural styles of individual speakers can be
reconstructed, focusing on arm gestures [12]. The emotional and
narrative content that emerges through extended interaction
between a virtual agent and a human can be used for simulating
memory and emotional states, thus increasing believability, as
indicated by work by Seif El-Nasr [16]. Modeling the affective
dimensions of characters and their personalities, as demonstrated
by Gebhard [5], provides more robust, consistent behavior in an
agent over extended time. While we have not developed such
components, we have developed a scheme by which believability
over extended interaction time can be measured.
Studies in using point light displays have shown that
humans are sensitive to the perception of human movement, such
as detecting human gait [1][9], and there are findings of distinct
patterns of neural activity associated with the perception of
human-made movement [15][14], indicating that a small number
of visual elements can be used, not only for testing perception of
believable motion, but also for use as control points in an
animated character, using inverse kinematics (IK). IK is
commonly used in computer animation to determine the joint
rotations of a character, based on goal positions.
In our research we have chosen to reduce the visual
aspect to a minimum, as a way to work with first principles of
motion behavior and also to establish an efficient and manageable
set of controllers for animating a character. This paper
demonstrates a scheme for testing the believability in this highly-
reduced set of primary motion features, using an established, well-
studied method: the Turing Test.
We do not address the issue of coverbal gesture or ways
to add a nonverbal layer to an existing verbal layer for
conversational agents. This is basic research focusing on “silent
copresence” and primitive communication through motion only.
1.2 Graphical Representation
The plastic human brain routinely adapts to new communication
media and user interfaces. If a human communicator spends
enough time “being” three dots, and communicating with three
dots that behave in a similar way, then the body map quickly
adapts to that schema. We have designed AI algorithms that
“know” they exist as 3 dots, and must use a three-dot interface to
communicate. Consider that an AI used for the classic Turing Test
does not require the simulation of a mouth, tongue, lungs,
diaphragm, or any apparatus used to generate verbal language,
and nor does it require the simulation of fingers tapping on
computer keyboards. Similarly, the Gestural Turing Test AI does
not need to simulate the entire muscular and skeletal apparatus
required for moving these points around. Semiosis is confined to
the points only they are the locus of communication.
How many points are needed to detect communicative
motion? We had originally considered two dots even one dot, as
comprising the gestural alphabet. Our hypothesis was that, given
enough time for a subject to interact with the dot(s), the
intelligence behind it (or lack thereof) would eventually be
revealed. With one dot, there would be very little indication of a
physical human puppeteer however, the existence of a human
mind might become apparent over time, due to the spontaneous
visual language that naturally would emerge, given enough time
and interaction.
Figure 1. The human visual system can detect living motion from only
a few dots.
Several points of light (dozens) would make it easier for the
subject to discern between artificial and human (as illustrated in
Figure 1). But this would require more sophisticated physical
modeling of the human body, as well as a more sophisticated AI.
For our experiment, we chose three points, because we believe the
head and hands to be the most motion-expressive points of the
body. The majority of gestural emblems originate in the head and
hands.
2. EXPERIMENTS
Figure 2 shows a schematic of the studio setup. A human observer
(the subject, shown at right) sits in a chair in front of a screen
projected with two sets of three white dots.
The subject wears motion-capture markers (attached to
a hat and two gloves), which are used to move the three dots on
the right side of the screen. The three dots on the left side of the
screen are moved by a hidden agent obscured by a room divider.
This hidden agent is either another human with similar motion
capture markers, or a software program that simulates human-
made motion of the dots. No sounds or text can be exchanged
between the subject and the hidden agent. Sign language is not
possible, due to the limited number of dots.
Figure 2. The human subject (right) interacts with the moving dots on
the left and must decide if they are created by a hidden human or an
AI program.
To generate the points from the human subjects we used the
Vicon motion capture studio at Emily Carr University of Art and
Design in Vancouver, BC. Figure 3 shows a screenshot of the
Vicon interface (top). In order for the Vicon system to
differentiate between the various objects in the scene, one of the
hats used four markers, and each glove on the opposite side used
two markers. Everything else used only one. This explains the
linear-connected figures in the screenshot. This is only for
purposes of calibration and disambiguation for the Vicon system,
and makes no difference to the subject’s view. Twenty cameras
are deployed in the studio (six of them can be seen represented in
wireframe at the top). Because of the room divider, some camera
views of the markers are obscured, which accounts for occasional
drop-out and flickering in the resulting points.
The stream of 3D positional data generated by the
Vicon system while the subjects moved was distributed via a local
area network to a laptop running the Unity game engine. We
formatted the data using XML, which was also used for recording
motions and archiving results from the experiments. A 3D scene
consisting of six small white spheres three on either side of a
black divider are animated by this data stream at 30 Hz. We
used the Unity engine because we intend to extend this research to
drive realistic avatars in a subsequent version of the project. An
example display from Unity is shown at the bottom of Figure 3.
2.1 Artificial Gesture Algorithms
To generate the artificial gestures, we designed two algorithms.
Both algorithms relied on the detection of the energy of the
human’s motions to trigger responsive gestures. We calculated
energy from continually measuring the sum of the instantaneous
speeds of the three points. If at any time energy changed from a
value below a specified threshold to greater than that threshold, a
response could be triggered, which depended on what kinds of
gestures were playing at the time.
The first algorithm (AI1) employed a state-machine that
chose among a set of short, pre-recorded gestures made by a
human. These pre-recoded gestures included “ambient” motions
(shifting in the chair, scratching, etc.) and a set of more dynamic
gestures (emblems and communicative gestures such as waving,
pointing, drawing shapes,chair-dancing”, etc.). When triggered,
it played gestures form the set of dynamic gestures, and when it
detected smaller movements, it responded by playing smaller,
ambient gestures. AI1 did not include a sophisticated blending
scheme for smooth transitions between gestures, and so it was
often apparent to the subjects that it was not human. This was
intentional: we wanted to expose the subjects to less-believable
behaviors so that they could establish a base-level of non-
believability on which to judge other motions.
The second algorithm (AI2) used a combination of
procedurally-generated motions and imitative motions created
while the experiment was being done. The procedurally-generated
gestures were continuous (no explicit beginning, middle or end)
and so any one of them could be blended in or out at any time, or
layered together. These were constructed through combinations of
several sine and cosine oscillations, with carefully-chosen phase
offsets and frequencies. This included slight motions using a
technique similar to Perlin Noise [13]. The imitative gestures were
created by recording the positions of the subject’s motions,
translating them to the location where the AI wassitting, and
playing them back after about a second, with some variation,
when a critical increase in energy was detected. AI2 used a more
sophisticated blending technique, achieved by allowing multiple
gestures (each with varying weights) to play simultaneously, such
that the sum of the weights always equals 1. A cosine function
was used for blending transitions to create an ease-in/ease-out
effect, which helped to smooth transitions.
Figure 3. ‘Skeleton template’ of the Vicon motion capture system,
required for labeling and calibration (top). Rendering with Unity
engine, used in experiments (bottom)
Our rationale for recording the subject’s gestures and playing
them back, using AI2, was that it would not only appear human,
but that it would also be imitative. Imitation is one of the most
primary and universal aspects of communication especially
when there is a desire for rapport and emotional connection.
Gratch, et. al, report that virtual agents that exhibit postural
mirroring, and imitation of head gestures enhance the sense of
rapport within subjects [7].
3. Results
There were 17 subjects. We ran 6 to 12 tests on each subject. A
total of 168 tests were done. Figure 4 shows the results in
chronological order from top to bottom. In this graph, the set of
tests per subject is delineated by a gray horizontal line. The length
of the line is proportional to the duration it took for the subject to
make a response. The longest duration was just over 95 seconds.
If the response was “false”, a black dot is shown at the right end
of the line. Wrong guesses are indicated by black rectangles at the
right-side of the graph.
Figure 4. Test results displayed chronologically for all subjects
This graph reveals some differences in subjects’ abilities to make
correct guesses, and also differences in duration before subjects
made a response. But we are more interested in how well the two
AI algorithms performed against the human. This can be shown
by separating out the tests according to which hidden agent was
used (Human, AI1, or AI2). Figure 5 shows the percentages of
wrong vs. right responses in the subjects for each of the three
agents. As expected, the human had the most guesses of real. Also
as expected, AI2 scored better than AI1 in terms of fooling
subjects into thinking it was real.
Figure 5. Percentages of right and wrong responses for each agent.
Believability in virtual agents can be measured in many ways it
doesn’t have to be a binary choice, and in fact it has been
suggested by critics of the classic Turing Test that its all-or-
nothing test criterion may be a problem, and that a graded
assessment might be more appropriate and practical [4]. One
approach might be to measure how quickly a subject is convinced
that a virtual agent is real. We calculated the average durations for
each case of right and wrong responses, as shown in Table 1.
Table 1. Average durations for both right and wrong responses.
The average durations before responding for the human agent are
less than the average durations for the other agents, except for
when subjects guessed wrongly for AI2. It may be that the
authenticity of the human agent is easily and quickly determined,
on average, which accounts for the slightly shorter average
durations. But this is not conclusive. We ran a t-test and did not
find any significant differences. In future experiments we would
need more experimental data and more thorough analyses.
4. OBSERVATIONS
We selected both males and females, with the majority of the
subjects being females. Some of the subjects displayed great
confidence, and made quick decisions (which were not necessarily
more correct). Some subjects took very long to decide (over a
minute). We also found large variations among the kinds of
gestures that the subjects made. Some of them were very reserved,
holding their hands close together, and making small motions,
while others made large motions. Some subjects gestured broadly
and stopped to wait for a response, while others appeared to be
swimming in place with no breaks or pauses. These variations in
subject gesturing had a pronounced effect on AI response. Both
AI algorithms relied on the gestures to be fast enough to trigger a
response (specifically, the sum of the instantaneous speeds of the
three points had to be greater than a certain threshold).
Consequently, the subjects who made small, slow gestures were
met with fairly uncommunicative artificial agents, and this had the
compounding effect of less activity from the subject. It was not
unlike two shy people who are unable to get a conversation going.
This suggests to us that a more sophisticated AI would need to be
designed that is able to gauge the overall energy of the subject’s
gestures and adjust its gesture-detection threshold accordingly.
4.1 Signal versus Noise
The Vicon motion capture system relies on multiple markers on
the body to construct a reliable 3D representation. Since we used
so few points, the system sometimes lost the labeling of points,
and as a consequence, there was occasional drop-out and
swapping between markers. This kind of problem is typically
cleaned up in post-processing before motion capture data are used
in a film, for instance. In our case, we were streaming the data in
realtime, and had to manage this problem on the fly. At first we
spent a bit of effort trying to remove these visual artifacts. But
later we realized that the human subject may forgive the noise and
still appreciate the signal, similar to the way that some static is
tolerated in telephony. So, instead we decided to add artificial
artifacts to the simulated output! The subject can easily discover
that these glitches occur in his/her own two dots, and will quickly
forgive them seeing them in the artificial agent actually could
actually enhance its believability. (Recall that in the Turing Test,
any trick that can fool the human is fair game).
4.2 Problems
For subsequent tests, we would like to improve a number of
things. For instance, the Vicon motion capture system is not set up
to deal with very small numbers of markers. It relies on several
markers, with many of them being a fixed distance apart, in order
to keep track of the labeling of markers. Often, when the subject’s
hands were held together, the Vicon system would swap the
hands, which required a recalibration in the middle of the test (this
only took a second each time, but still we would prefer not to have
to do it).
We cannot be 100% sure that a subject did not hear or
“sense” that the human on the other side of the divider was the
one gesturing. This is why we had originally planned on
conducting the test in two remote locations (the rationale for
developing a flexible network architecture). This setup however
would require considerable technical work. Also, internet latency
would introduce new problems that would have to be dealt with.
5. CONCLUSIONS
The results of these experiments show that when an artificial
agent imitates the gestures of a human subject, there is more
acceptance of that agent as being alive. We also show that this can
be tested with a small set of visual indicators. This is consistent
with research in studying human vision with point-light displays,
and it suggests that believability need not be supported by visual
realism.
The scope of this project did not permit the design of an extended
AI with the ability to build on a collaborative semiotic process.
But we believe that layering more sophisticated algorithms on top
of the base behaviors we have implemented would create
sustained believability over longer durations of time.
The contribution of the paper is twofold. First: the
Gestural Turing Test itself can act as a methodology for validating
gestural AI algorithms. It can easily be extended to include
motion specified by many more control points, as well as any
level of sophistication of intelligence, emotion, memory, natural
language, and physical modeling. Secondly: the developed
imitation algorithm can be analyzed to define a model or set of
design lessons for creating better believable characters. The dots
used in this experiment (which are actually projected 3D
positions) are ultimately intended to become the control points for
a fully-rendered avatar using IK. In a subsequent experiment, we
intend to replace the graphical representation of dots with 3D
avatars. It does not take a lot of control points to achieve
reasonable motion, especially if the human model has a well-
crafted constraints system to generate natural poses, given the
pushing and pulling of the control points. One reason we feel that
IK is a reasonable technique is because communicative motion
often takes the form of hands moving in complicated paths,
whereby the relative positioning of the hands (less so the elbows
and shoulders) constitute the informational content. Thus, a
gesturing system that is based on head and hand positioning (and
rotation) over time is a valid scheme to use. One could even claim
that communicative motion is IK-based at the neurological level.
In future work, we aim to address several key questions related to
the semantics of believable motions towards the development of
techniques for designing believable characters.
6. ACKNOWLEDGMENTS
We would like to acknowledge the NSERC - NCE (Networks of
Centres of Excellence of Canada), GRAND (graphics, animation
and new media) for providing funds to support this project. We
wish to thank Professor Leslie Bishko of the Emily Carr
University of Art and Design for initial discussions and ideas, and
for helping with access to the motion capture studios.
7. REFERENCES
[1] Blake, et al., Perception of Human Motion. 2007
[2] Brooks, Rodney A. Elephants Don't Play Chess. Robotics
and Autonomous Systems 6, 1990 3-15
[3] Cassell, J. "Embodied Converstational Agents:
Representation and Intelligence in User Interfaces," AI
Magazine, vol. 22, 2001.
[4] French, R. If it walks like a duck and quacks like a duck...
The Turing Test, Intelligence and Consciousness. Oxford
Companion to Consciousness. Wilken, Bayne, Cleeremans
(eds.) Oxford Univ. Press. 461-463. 2009.
[5] Gebhard, P. ALMA a Layered Model of Affect.
Proceedings of the Fourth International Joint Conference on
Autonomous Agents and Multiagent Systems. 29-36
[6] Gorman, B., Thurau, C., Bauckhage, C., and Humphrys, M.
Believability Testing and Bayesian Imitation in Interactive
Computer Games. From Animals to Animats 9. Springer,
2006.
[7] Gratch, J., Wang, N., Gerten, J., Fast, E., and Duffy, R.
Creating Rapport with Virtual Agents. Proceedings of the 7th
international conference on Intelligent Virtual Agents.
Springer, 2007.
[8] Hewes, Gordon W. 1973. Primate Communication and the
Gestural Origins of Language. Current Anthropology 14:5-
24.
[9] Jacobs, A., and Pinto, J. Experience, Context, and the Visual
Perception of Human Movement. Journal of Experimental
Psychology. 2004, Vol. 30, No. 5, 822835
[10] Kipp, M. Gesture Generation by Imitation. Dissertation for
Doctor of Engineering. Saarland University, Saarbruecken,
Germany. Published by Dissertation.com. 2003
[11] Livingstone. Turing's Test and Believable AI in Games. 2006
[12] Neff, M. et al. Gesture modeling and animation based on a
probabilistic re-creation of speaker style. ACM Transactions
on Graphics. Volume 27, Issue 1. 2008
[13] Perlin, K. Real Time Responsive Animation with
Personality. IEEE Transactions on Visualization and
Computer Graphics. Volume 1, Issue 1, March, 1995
[14] Pinto, J., and Shiffrar, M. The visual perception of human
and animal motion in point-light displays. Social
NeuroScience Volume 4, Issue 4 August 2009 , pages 332
346
[15] Saygin, A.P., Wilson, S., Hagler, D., Bates, E., and Sereno,
M. I.. Point-Light Biological Motion Perception Activates
Human Premotor Cortex. The Journal of Neuroscience, July
7, 2004.
[16] M. Seif El-Nasr, T. Ioerger, and J. Yen, "FLAME - Fuzzy
Logic Adaptive Model of Emotions," Autonomous Agents
and Multi-Agent Systems, vol. 3, pp. 219-257, 2000.
[17] Stone, M., et al. Speaking with Hands, Speaking with Hands:
Creating Animated Conversational Characters from
Recordings of Human Performance. ACM Transactions on
Graphics Volume 23, Issue 3. 2004
[18] Stuart, J. et al. Generating Novel, Stylistically Consonant
Variations on Human Movement Sequences. 2007
[19] Turing, Alan. Computing Machinery and Intelligence, 1950
[20] Van Welbergen. Informed Use of Motion Synthesis Models.
Motion in Games. Springer, 2008.
... Dynamic figures with little anthropomorphic appearance can be perceived as human movement when certain kinematic features (speed, stiffness, and phase etc.) approach those of natural human movements (Morewedge et al., 2007; Thompson et al., 2011; Blake and Shiffrar, 2007; Johansson, 1973; Heider and Simmel, 1944). Similarity, the contingencies between a human and a virtual agent's movement also lead to humanness attribution to the virtual agent following their interaction (Pfeiffer et al., 2011; Ventrella et al., 2010). In the current study, VP's hand has anthropomorphic Table.3 ...
Article
This FDG 2011 Doctoral Consortium paper gives an overview of my Computer Science research approach and initial results in applying principles and theories from the Performing Arts to problems related to embodied agent design and animation for games.
Data
Full-text available
In this paper we introduce ALMA – A Layered Model of Affect. It integrates three major affective characteristics: emotions, moods and personality that cover short, medium, and long term affect. The use of this model consists of two phases: In the preparation phase appraisal rules and personality profiles for characters must be specified with the help of AffectML – our XML based affect modeling language. In the runtime phase, the specified appraisal rules are used to compute real-time emotions and moods as results of a subjective appraisal of relevant input. The computed affective characteristics are represented in AffectML and can be processed by sub-sequent modules that control the cognitive processes and physical behavior of embodied conversational characters. ALMA is part of the VirtualHuman project which develops interactive virtual characters that serve as dialog partners with human-like conversational skills. ALMA provides our virtual humans with a personality profile and with real-time emotions and moods. These are used by the multimodal behavior generation module to enrich the lifelike and believable qualities.
Article
Full-text available
Abstract A common task in dance, martial arts, animation, and many other movement genres is for the character to move in an innovative and yet stylistically consonant fashion. This
Article
Full-text available
If it walks like a duck and quacks like a duck, it probably is a duck." So goes a pithy observation, undoubtedly as old as English itself, that provides a quick-and-dirty way to recognize ducks. This little maxim constitutes an operational means of duck identification that, for better or for worse, sidesteps all the thorny issues associated with actually explicitly defining a set of features allowing us to identify a duck (e.g., has feathers, can fly, weighs less than 10 pounds, has webbed feet, swims well, has nucleated red blood cells, has a four-chambered heart, has a flat bill, etc.). What folk wisdom did for ducks, Alan Turing did for intelligence. He was the first to suggest an operational means of identifying intelligence that has come to be called the Turing Test (Turing, 1950). The underlying idea of his Test is the same as our folk means of duck identification, viz., whatever acts sufficiently intelligent is intelligent. Translated into the vernacular of modern electronic communication, the Turing Test says that if, by means of an extended e-chat alone, you cannot tell whether you are chatting with a machine or a person, then whomever or whatever you are chatting with is intelligent. Since it first appeared nearly six decades ago, Turing's article has become the single most cited article in artificial intelligence. In fact, few articles in any field have caused so much ink to flow. References to the Turing Test still appear regularly in artificial intelligence journals, philosophy journals, technical treatises, novels and the popular press. Type "Turing Test" into any internet search engine and there will be, literally, thousands of hits. How the Turing Test Works Turing's original description of his "Imitation Game" was somewhat more complicated than the simpler version that we describe below. However, there is essentially universal agreement that the additional complexity of the original version adds nothing of substance to its slightly simplified reformulation that we refer to today as the Turing Test. The Turing Test starts by supposing that there are two rooms, one of which contains a person, the other a computer. Both rooms are linked by means of text-only communication to an Interrogator whose job it is, by means of questioning the entities in each room, to determine which room contains the computer and which the person. Any and all questions that can be transmitted via typed text are fair game. If after a lengthy session of no-holds-barred questions, the Interrogator cannot distinguish the computer from the person, then the computer is declared to be intelligent (i.e., to be thinking). It is important to note that failing the Turing Test proves nothing. It is designed to be a sufficient, but not necessary, condition for intelligence.
Conference Paper
Full-text available
Recent research has established the potential for virtual characters to estab- lish rapport with humans through simple contingent nonverbal behaviors. We hy- pothesized that the contingency, not just the frequency of positive feedback is crucial when it comes to creating rapport. The primary goal in this study was evaluative: can an agent generate behavior that engenders feelings of rapport in human speakers and how does this compare to human generated feedback? A secondary goal was to an- swer the question: Is contingency (as opposed to frequency) of agent feedback crucial when it comes to creating feelings of rapport? Results suggest that contingency mat- ters when it comes to creating rapport and that agent generated behavior was as good as human listeners in creating rapport. A "virtual human listener" condition per- formed worse than other conditions.
Conference Paper
Full-text available
In this paper we introduce ALMA - A Layered Model of Affect. It integrates three major affective characteristics: emotions, moods and personality that cover short, medium, and long term affect. The use of this model consists of two phases: In the preparation phase appraisal rules and personality profiles for characters must be specified with the help of AffectML - our XML based affect modeling language. In the runtime phase, the specified appraisal rules are used to compute real-time emotions and moods as results of a subjective appraisal of relevant input. The computed affective characteristics are represented in AffectML and can be processed by sub-sequent modules that control the cognitive processes and physical behavior of embodied conversational characters. ALMA is part of the VirtualHuman project which develops interactive virtual characters that serve as dialog partners with human-like conversational skills. ALMA provides our virtual humans with a personality profile and with real-time emotions and moods. These are used by the multimodal behavior generation module to enrich the lifelike and believable qualities.
Article
The abstract for this document is available on CSA Illumina.To view the Abstract, click the Abstract button above the document title.
Book
This dissertation shows how to generate conversational gestures for an animated agent based on annotated text input. The central idea is to imitate the gestural behavior of human individuals. Using TV show recordings as empirical data, gestural key parameters are extracted for the generation of natural and individual gestures. For each of the three tasks in the generation pipeline a software was developed. The generic ANVIL annotation tool allows to transcribe gesture and speech in the empirical data. The NOVALIS module uses the annotations to compute individual gesture profiles with statistical methods. The NOVA generator creates gestures based on these profiles and heuristic rules, and outputs them in a linear script. In all, this work presents a complete work pipeline from collecting empirical data to obtaining an executable script and provides the necessary software, too.
Article
Suppose that sometimes he found it impossible to tell the difference between the real men and those which had only the shape of men, and had learned by experience that there were only two ways of telling them apart: first, that these automata never answered in word or sign, except by ch ance, to questions put to them; and second, that though their movements were often more regular and certain than those of the wisest men, yet in many things which they would have to do to imitate us, they failed more disastrously than the greatest fools.
Chapter
I propose to consider the question, “Can machines think?”♣ This should begin with definitions of the meaning of the terms “machine” and “think”. The definitions might be framed so as to reflect so far as possible the normal use of the words, but this attitude is dangerous. If the meaning of the words “machine” and “think” are to be found by examining how they are commonly used it is difficult to escape the conclusion that the meaning and the answer to the question, “Can machines think?” is to be sought in a statistical survey such as a Gallup poll.
Article
There is an alternative route to Artificial Intelligence that diverges from the directions pursued under that banner for the last thirty some years. The traditional approach has emphasized the abstract manipulation of symbols, whose grounding in physical reality has rarely been achieved. We explore a research methodology which emphasizes ongoing physical interaction with the environment as the primary source of constraint on the design of intelligent systems. We show how this methodology has recently had significant successes on a par with the most successful classical efforts. We outline plausible future work along these lines which can lead to vastly more ambitious systems.