Content uploaded by Min-Gyu Kim
Author content
All content in this area was uploaded by Min-Gyu Kim on Jun 29, 2015
Content may be subject to copyright.
Int J Soc Robot (2014) 6:5–15
DOI 10.1007/s12369-013-0184-0
Comparative Study of Human Behavior in Card Playing
with a Humanoid Playmate
Min-Gyu Kim ·Kenji Suzuki
Accepted: 27 February 2013 / Published online: 15 March 2013
© Springer Science+Business Media Dordrecht 2013
Abstract This paper describes the study of human behav-
iors in a poker game with the game playing humanoid robot.
Betting decision and nonverbal behaviors of human play-
ers were analyzed between human–human and the human–
humanoid poker game. It was found that card hand strength
is related to the betting strategy and nonverbal interaction.
Moreover, engagement in the poker game with the humanoid
was assessed through questionnaire and by measuring the
nonverbal behaviors between playtime and breaktime.
The findings of this study contribute to not only design of
socially interactive game playing robot, but also the theoret-
ical approach on the realization of the robot that behaves in
the way of human doing in game playing.
Keywords Human–robot interaction ·Social playmate ·
Poker game ·Humanoid
1 Introduction
Current advanced technologies offer many chances in a
number of different ways to play games with each other.
In modern life, people enjoy games through video game
and social network game and now even play games with
robots. The people factor of game that creates opportuni-
ties of social experience such as competition, teamwork,
M.-G. Kim ()
Department of Intelligent Interaction Technologies, University
of Tsukuba, Tennodai 1-1-1, Tsukuba, Ibaraki, Japan
e-mail: mingyu@ai.iit.tsukuba.ac.jp
K. Suzuki
Faculty of Engineering, Information and Systems, University
of Tsukuba, Tennodai 1-1-1, Tsukuba, Ibaraki, Japan
e-mail: kenji@ieee.org
social bonding and personal recognition [1] plays an im-
portant role in either developing game playing robots as
a social partner in the domain of human–robot interac-
tion.
Many researchers have made efforts to exploit game play-
ing robots. Neil et al. have successfully implemented phys-
ical chess playing robot capable of the motion grammar
for the manipulation of chess pieces [2]. Frank et al. intro-
duced that an industrial robot arm system with vision cam-
era plays German board game, Mensch ärgere dich nicht
while projecting board game field onto the workbench [3].
On the other hand, there are game playing robots with com-
municative abilities. A poker playing agent that exhibits
emotions to communicate with human player through mul-
timodal channels [4]. A virtual character plays chess by
physically manipulating chessman while speaking and ex-
pressing several emotions [5]. These game playing robots
have been designed from robot-centric interaction point of
view.
In this study, we focus on poker game that is the op-
timal social setting to observe human complex behav-
iors and social interactions. During the game, player usu-
ally needs to gain advantage with hiding intention. They
attempt to play carefully because they need to see the
game pertaining to a financial exchange. Besides, peo-
ple engage in social communication and interaction while
deciphering other’s intention from verbal and nonverbal
behaviors or sometimes behaving deceptively for strate-
gic purpose. Furthermore, people can make new social
contact and build social relationship with fellow players
through the poker game. In such a social, friendly game
environment, a poker playing humanoid robot to under-
stand human interactions and socially respond to human
in the game is essential as well as establishing computa-
tional model of game strategies and opponent decisions.
6Int J Soc Robot (2014) 6:5–15
In order to direct the goals, the following studies should
be investigated; (1) what kind of human behaviors the
robot has to perceive, and (2) how the robot creates so-
cial response on the basis of perceived human behav-
iors.
The first study examined several comparative effects be-
tween human–human poker game and human–humanoid
poker game. Considering human–centric interaction design
of poker playing robot, it is important to study the human
behaviors in human–robot poker game as compared with
human–human poker game. Many researchers in humani-
ties and social sciences have presented their studies about
the human behaviors related to poker game setting for a
long time [6–10]. Applying the discoveries of the human
studies to designing socially interactive poker playing robot
is highly questionable because it is unsure whether the same
human behaviors would be observed in the human–robot
poker game. In order to clarify this kind of vague of whether
human–human communication can be directly transferred
into human–robot interaction, Walters et al. compared the
human–robot personal spatial zones with the aspect of
human–human personal spatial zones [11]. Mutlu et al. at-
tempted to explain the task structure and user attributes of
human–robot interaction design through comparative study
of human–human interaction [20].
The purpose of the follow-up study is to analyze re-
lationship between human current situation (player’s own
hand strength) and behaviors (his nonverbal behavior and
decision-making) in human–humanoid poker game. As
usual in human–human poker game, it is seen that player’s
psychological state is affected by his own hand strength
and shows up through face, body and decision in any way.
Hence, a player who holds strong hand tends to try to dis-
guise his hand as being weak for strategic purpose. If the
same relationship in human–humanoid poker game is found
also, it can be critical for robot to understand human behav-
iors, which means the robot predicts human fellow player’s
own hand strength by observing his nonverbal behaviors and
analyzing game strategic decisions. Such robots that under-
stand and imitate human behaviors are distinguished from
usual poker agents capable of computing winning strate-
gies with efficient decision-making algorithm designed to
cope with imperfect information [12]. The agents accurately
evaluate the game rather than understand complex dynam-
ics of human interactions. From the second study, we expect
that the human-like characteristics of poker game robot suc-
cessfully will encourage a human social behavior and keep
people motivated to play the game.
In Sect. 2, we establish theoretical frameworks by re-
viewing relevant works. Section 3describes the methodol-
ogy of this study in detail. We introduce our humanoid plat-
form, experimental procedure and measured human behav-
iors. Sections 4and 5show the experimental results and gen-
eral discussion.
2 Theoretical Background
2.1 Competitive Card Playing with Humanoid Robot
Preceding studies have examined effects of robot on human
behavior and perception based on cooperative type of inter-
actions. For instance, it has reported how the robot’s physi-
cal presence influences a person’s perception in the cooper-
ative task and the differences between a robot and an an-
imated character in terms of engagement and perceptions
[16,17]. The effect of physical embodiment on the feeling
of an agent’s social presence has been evaluated by letting
the participants interacting with the agent like touching [15].
The result showed that lonely people provide more positive
responses to the agents than non-lonely people. Kanda et al.
have measured the body movements of human who inter-
acts with the physical robot and found the positive correla-
tions between cooperative body movements and subjective
evaluations [18]. Moriguchi et al. have examined in the card
sorting task with an android whether it triggers young chil-
dren’s perseveration and compared the effects of the android
with those of a human model [19].
Meanwhile, Mutlu et al. have researched on user at-
tributes in competitive task as compared with cooperative
tasks [20]. They have introduced significant differences of
task structure such as that involvement in the human–robot
interaction was significantly higher in the competitive task
than in the cooperative task. The differences might have
relevance to the findings that people in competitive group
display more rejection to the other and believe other per-
sons are uniformly competitive while people in coopera-
tive group evaluate their group members more positively
[13,14]. In our study, it is assumed that people will see robot
as competitor in competitive game task like people usually
does in human–human competitive game. More specifically,
it is expected that people will show different type of re-
sponse toward human competitor and robot competitor be-
cause robotic appearance, mechanical gestures and speech
synthesis would make people less absorbed in the competi-
tive task itself and more interested in interacting with robot.
Therefore, we hypothesized that in poker game with robot
opponent, people will show different aspect of betting deci-
sion and nonverbal interaction as compared with poker game
with human opponent.
•Hypothesis 1: When participants play poker game with
robot opponent, they will show different betting decision
and nonverbal interaction as compared with poker game
with human opponent.
2.2 Poker Game
In the domain of artificial intelligence, poker game is an in-
teresting test bed because it includes imperfect information,
Int J Soc Robot (2014) 6:5–15 7
competitive agents, risk management, agent modeling, de-
ception and dealing with unreliable information [12]. In or-
der to confront the professional human poker players, solv-
ing these difficulties is the main issue in developing poker
agent.
However, poker in real life has two not mutually exclu-
sive reasons why the poker game we play cannot be a com-
plete zero-sum game [21]. (a) “Some players make wrong
decisions most of the time because their mental model of
the mathematical structure of the game and the psychologi-
cal nature of the opponents is false or incomplete”. (b) “So-
cial reasons, enjoyment, and all sorts of personal considera-
tions may contribute to the utility value of particular actions
in the game”. Therefore, it is not easy for players to use a
hypothetical normative strategy based on money.
This study pays attention to the game playing humanoid
as a social, friendly partner rather than gambling machine
to deal with regular or professional players. If we obtain
the mental model of human players as applied to the robot
by combining the mathematical structure of the poker game
with the psychological model of the opponent, it will be
likely to perform as human does in poker game of real life.
We first evaluate the bluffing strategy on the basis of money,
and then look into whether card hand strength is associated
with bluff decision and nonverbal interaction.
•Hypothesis 2: Participants who play with the robot op-
ponent will be affected in their decision and nonverbal
interaction by their given hand strength as much as the
participants who play with the human opponent is done.
•Hypothesis 3: Given hand strength to participants is re-
lated to their betting decision and nonverbal interaction.
2.3 Engagement in Game with Humanoid
Engagement is an important matter of successfully attracting
people into a human–robot interaction. According to [22],
when a sense is more immersed in the interaction, the ex-
perience is likely to be more compelling. Karolien et al.
established a categorization of digital game experience in
terms of in-game and post-game using focus group method
[23]. Regarding suspense dimension, they categorized chal-
lenge, tension and pressure into in-game experience and re-
lease, relief and exhausted into post-game experiences. In a
poker game, players also similarly go through mutually ex-
clusive states in every turn such as in-game state (playtime)
and post-game state (breaktime). When poker players expe-
rience in tension and release alternately, a pattern of their
nonverbal interaction is changed.
The use of eye gaze has been discussed to evaluate user’s
engagement in a conversation task with virtual agent and
robot [24,25]. This study bases on the eye gaze to measure
how people are engaged in the poker game with our robot.
Fig. 1 Humanoid robot platform used in this study, named Genie
Abdullah et al. have already compared the frequency of oc-
currence for three events of eye gaze in a social game; ‘look-
ing at opponent’, ‘looking at team member’ and ‘looking at
game board’ to identify social interaction trends and game
involvement [26].
•Hypothesis 4: Participants who play with the robot oppo-
nent will show a different interaction pattern of eye gaze
between playtime and breaktime.
•Hypothesis 5: Subjectively evaluated engagement will
show a difference between human–human and human–
humanoid poker game.
3 Overview of Experimental Setup
3.1 Humanoid Platform
As can be seen from Fig. 1, the humanoid robot used in this
study has been built to investigate human behavioral analy-
sis [27]. The humanoid robot with the upper torso that in-
cludes 3DOF waist mounted on the wall has totally 32DOF
(8DOF for the head, 3DOF for the waist, 7DOF for the right
arm, 4DOF for the right hand, 5DOF for the left arm and
5DOF for the left hand). The humanoid platform has anthro-
pomorphic size and shape. It does not have a face to exhibit
emotions, but two eyes equipped with vision camera can im-
itate human gaze.
3.2 Participants
Ten participants (aged 22–29, 6 males and 4 females) and
a paid dealer were recruited from local communities in
University of Tsukuba. The participants are undergraduate
and graduate students with a wide academic background
in education, physics, nursing, computer science, etc., and
computer-friendly as mostly use a computer 8 or more times
a week. They are non-professional, beginner-leveled play-
ers who barely play a poker game such as playing less than
once a month. Some of them have never played poker game
8Int J Soc Robot (2014) 6:5–15
Fig. 2 Experimental setup: participants played Texas hold’em both
with (a) human and (b) humanoid
before. The participants has never seen and also interacted
with real robots. Since non-experience of the participants
in both poker and robot prohibits from the biased response,
the experimental results would fairly confirm the effect of
opponent presence, the effect of card hand strength and en-
gagement.
The dealer is a graduate student who has many experi-
ences in psychological studies. She was trained in advance
to master the rule of Texas hold’em and the purpose of the
experiment. Also, the dealer was carefully instructed not
to talk and smile a lot, in order to minimize the effect of
dealer’s speech and behaviors on the participant’s decision
and interaction. The dealer spoke only to notify the betting
turn to the players.
3.3 Procedure
Each participant played Texas hold’em first with the human
experimenter and then with the humanoid one on one as
showninFig.2. The dealer explained how to play the sim-
plified Texas hold’em to the participants before the experi-
ment and then the participants practiced the game until they
got a hold of the game. The participants played five rounds
of the poker game with play bills. The entire episode was
recorded. The video cameras in a human–humanoid poker
game were installed in the humanoid’s body and right side
of the humanoid whereas a video camera was set in right
side of the experimenter in the human–human poker game.
The experimenter in the human–humanoid poker game
operated the humanoid’s whole motions and made speech
synthesis behind the experimental space according to Wiz-
ard of Oz approach [28]. The humanoid was manually con-
trolled to display informative behaviors such as affirma-
tive head nods, negative head shakes and head orientation
(i.e., gazing at human player, looking at table and looking
at a card on the hand) as well as to play Texas hold’em
with its arm and hand. For time-saving, the dealer assisted
the humanoid to easily handle the cards. The play bill was
also handled by the dealer. Instead, the humanoid raised the
stakes or folded by saying “bet one thousand yen” or “fold”.
3.4 Simplified Rule for Two-Person Texas Hold’em
The original rule of Texas hold’em was modified to sim-
plify the betting structure. Each betting round, the partici-
pants can bet from a thousand yen to three thousand yen.
As usual, in the game, the players share five community
cards that are opened on the table and have two own cards.
To win the game, the players should find the best combina-
tion among five community cards and two own cards during
the game. The players included a blind bet before the card
distribution. After the blind bet, the players place their bet
with the first three community cards (flop cards) and their
cards (pocket cards). When the tokens which a player bet
would be equally matched by his opponent, the first bet-
ting round ends and the fourth community card (turn card)
is opened. Then the players begin to bet money again. The
round closes when the bet is matched equally by the oppo-
nent. The final community card (river card) is opened and
final bet is placed.
After the end of the final betting round, the players open
their pocket cards for a showdown. If one of the players folds
in the middle of the game, the game is over and the pocket
cards are not revealed. Each participant played the poker
game five times with the experimenter and the humanoid
respectively.
3.5 Robot Movements
The robot movements in the experiment are categorized into
two different motions as shown in Fig. 3, manipulative and
expressive motions. The manipulative motion is to pick and
lift up its own pocket card to check the card suits. The hu-
manoid performs the ‘Checking own Card Suits’ action us-
ing its left arm and rotating head only when the pocket card
is distributed to the players in the beginning of the game.
The expressive head movements create a sense of looking at
table or its opponent. It executes the ‘Looking at Table’ ac-
tion when the participants bet. ‘Looking at its Opponent’ ac-
tion is done when the robot bets. The profile of robot motion
and speech used in the experiment is described in Table 1.
3.6 Manipulation of Card Hand Strength
Poker player’s nonverbal behaviors are related to his own
card hand [29]. Since the participants were not professional,
it was assumed that they would show appreciably different
responses according to the card hand strength.
The card hand strength was established as a primary in-
dependent variable in this experiment, and the poker game
of five rounds was manipulated prior to the experiment to
observe an effect of card hand strength on human nonverbal
response. Given card hands were divided into two groups;
strong hand and weak hand as shown in Table 2. The par-
ticipants played Texas hold’em with the strong card hands
Int J Soc Robot (2014) 6:5–15 9
Fig. 3 Snapshots in the demonstration of Genie’s Texas hold’em playing: (a) is the manipulative arm and head movements to display ‘Checking
own Card Suits’ and (b) is the expressive head movements to create a sense of ‘Looking at Table’ and ‘Looking at its Opponent’
Table 1 Robot action and speech profiles according to game sequence during the experiment
Betting structure Game sequence description Robot action Robot speech
Blind bet Betting without any cards Looking at opponent (expressive) Saying ‘Bet XXXX yen’
Dealer distributes cards to the players Looking at table (expressive)
Robot picks up its own poker cards Cheking own card suits (manipulative)
Flop (first) Dealder asks the players to bet Looking at opponent (expressive)
When robot bets Looking at opponent (expressive) Saying ‘Bet XXXX yen’
When human bets Looking at table (expressive)
Turn (second) Dealder asks the players to bet Looking at opponent (expressive)
When robot bets Looking at opponent (expressive) Saying ‘Bet XXXX yen’
When human bets Looking at table (expressive)
River (final) Dealder asks the players to bet Looking at opponent (expressive)
When robot bets Looking at opponent (expressive) Saying ‘Bet XXXX yen’
When human bets Looking at table (expressive)
Table 2 Manipulation of card hand strength
1st
round
2nd
round
3rd
round
4th
round
5th
round
(a) Human–human poker game
AQ♣4♠6♦10♦4♦A♦K♣Q♦5♣5♦
Tab l e J ♠3♦10♠
3♣8♥
7♦K♠9♦
8♦6♣
2♦8♣9♣
3♥Q♥
7♠4♥2♠
10♥6♥
3♠5♥5♠
A♣K♥
BJ♥10♠J♣8♠6♠Q♠2♥7♣A♠A♥
(b) Human–humanoid poker game
A8♣A♥2♥3♠6♠Q♦8♥Q♠7♦7♣
Tab l e 1 0♦3♥5♦
9♠K♣
6♦5♠4♦
Q♣2♠
10♠9♦4♣
A♦5♥
J♣10♣4♥
9♥A♠
J♦7♠7♥
2♦A♣
B10♥5♣6♥K♥4♠8♦J♠K♠2♣9♣
Note: A is participant’s pocket card, B is the opponent’s pocket card
and table is community cards
in 2nd and 5th rounds and with the weak card hands in 1st,
3rd and 4th rounds. The strong card hands were straight and
four of a kind, which means five consecutive cards and four
cards of the same rank, respectively. The weak card hands
indicated no pair to make them lose.
3.7 Measurements
3.7.1 Nonverbal Behaviors
Since the studies on nonverbal behaviors in human–robot
poker game have not been sufficiently researched, we re-
ferred to the psychological research findings in the field of
a human nonverbal communication related to deception. Es-
pecially, we paid attention to smile, hand movements and
eye blink that have been discussed about human nonverbal
behaviors in a high stake, deceptive situation that also can
be seen in poker game.
In this study, smile was defined as smiling as perceived
by the coders, who were given no specific definition or were
given a definition not involving specific AUs (e.g. corners of
the mouth are pulled up’; laughing is also included) [30].
Regarding hand movement, adaptors that are movements
indicating a low level of awareness such as self-touching
were chosen because we assumed that self-touching will be
mostly detected in poker game more than other gestures as-
sisting speech like illustrator [31]. Self-touching was cate-
gorized as three different hand movements that are touch-
ing face, touching head and touching arm. Eye blink was
10 Int J Soc Robot (2014) 6:5–15
Fig. 4 Betting decision tree
defined as eyes opening and closing quickly, and eye gaze
was defined as facing the other person/objects and gazing
at the person/objects [30]. Especially, the eye gaze was di-
vided into three groups; gazing at opponent, gazing at table
and gazing at other.
We coded the amount of four nonverbal behaviors from
recorded video respectively and then calculated frequency
that is a total amount of each nonverbal behavior divided by
a minute.
3.7.2 Bluff Decision
Betting strategies of players rely upon given circumstances
such as card hand strength [32]. Hence, bluffing strategy is
used according to player’s own hand strength. For instance, a
player with a winning hand (strong card hand), would try to
bet carefully in a non-threatening way to keep the pot grow-
ing and at the same time keep the opponent from folding
early. And on the other hand, the participant with a losing
hand (weak card hand) would rather try to bet in a way that
the other players would assume otherwise and raise the bet
taking high risks. Based on these assumptions, the player
with a strong hand would avoid maximum betting and the
player with a weak hand would avoid minimum betting, is
considered as the bluff strategies.
Figure 4shows the all instances with betting turn in three
columns and possible choices following opponent’s decision
in twelve rows. We could decide whether the participant’s
betting strategy is bluffing or not with the judging table as
showninTable3. It was confirmed that all betting decisions
of the participants were included in the table.
Table 3 Result table of judging bluff and truth
Consequence No. Strong hand Weak hand
Subject Opponent Subject Opponent
1 Bluff Bluff Truth Truth
2 Truth Bluff Bluff Bluff
3 Bluff Bluff Truth Bluff
4 Truth Truth Bluff Bluff
5 Bluff Truth Truth Bluff
6 Bluff Bluff Truth Truth
7 Bluff Bluff Bluff Bluff
8 Truth Truth Bluff Bluff
9 Bluff Truth Truth Bluff
10 Bluff Bluff Bluff Truth
11 Truth Truth Bluff Bluff
12 Truth Bluff Bluff Truth
3.7.3 Questionnaire
We analyzed the participants’ written responses to see how
engaging the poker game with the humanoid and human ex-
perimenter was. A questionnaire was developed based on
the presence as immersion, one of the six parts of Lombard
and Ditton’s work [22]. This questionnaire consisted of ob-
jective type questions as well as descriptive questions that
the participants answered in addition to other individual in-
terviews.
Q1 To what extent did you feel mentally immersed in the
experience with Human/Genie?
Q2 How involving was the experience with Human/Genie?
Q3 How completely were your senses engaged in playing
with Human/Genie?
Q4 How relaxing or exciting was the experience with Hu-
man/Genie?
Q5 How engaging was the game played with Human/Genie?
4 Experimental Results
The assessment of inter-rater reliability for a total amount
of the nonverbal behaviors coded by two experimenters was
executed. As shown in Table 4, the calculated results using
Cohen’s Kappa statistics indicated a reasonable agreement
in the good to moderate or substantial ranges between two
raters.
In this study, because participants spent difference time
of playing a poker game, all the analysis was performed in
a way to transform the quantity of nonverbal behaviors into
the unit of frequency. For the calculation of frequency, we
Int J Soc Robot (2014) 6:5–15 11
Table 4 The inter-rater reliability for measured nonverbal behaviors
(Kappa coefficient statistics)
Cohen’s Kappa value
H-H Poker Game H-R Poker Game
Smile 0.67 0.68
Touching Face 0.76 0.85
Touching Head 0.73 0.68
Touching Arm 0.80 0.70
Eye blink 0.53 0.42
Gazing at Opponent 0.59 0.61
Gazing at Table 0.66 0.54
Gazing at Other 0.68 0.58
required to separate each game round from the recorded en-
tire episode, and used the dealer’s comments noticing begin-
ning and end of each game round. For instance, the dealer
said both “2nd round of game is started, you take first turn”
at the beginning of game round and “you have straight hand,
so you win!” at the end of game round. We defined playtime
as a time spent for playing a poker game and breaktime as a
time between rounds taken to stand by for next round. The
calculated frequencies of nonverbal behavior during play-
time or breaktime were utilized to analyze effects of card
hand strength and the presence of an opponent, and also to
examine relationship between card hand strength and human
behaviors and engagement between human–human poker
game and human–humanoid poker game.
4.1 Effect of the Presence of an Opponent
We used Kruskal-Wallis test to evaluate the first hypothesis
(see effect of opponent in Tables 5and 6). Hypothesis 1 was
supported. In Table 5, smile (χ(1)=3.23, p<.10 in strong
hand condition, and χ(1)=3.57, p<.10 in weak hand
condition) shows marginally significant difference in both
hand conditions. Also, eye gaze in strong hand shows highly
significant difference (Gazing at Opponent: χ(1)=10.35,
p<.01, Gazing at Table: χ(1)=10.62, p<.01, Gazing at
Other: χ(1)=14.42, p<.001), and in weak hand condition
indicates highly significant difference (Gazing at Opponent:
χ(1)=12.24, p<.001, Gazing at Table: χ(1)=13.21,
p<.001, Gazing at Other: χ(1)=5.60, p<.05).
The frequency of smile was decreased when the partic-
ipants played with the humanoid. According to the follow-
up interview, some of the participants told that they con-
trolled over their facial expressions because they felt that
the expressive head movements of the humanoid sometimes
seems to observe themselves. The participants intended not
to smile both in human–human and the human–humanoid
poker game. This result confirms other research that peo-
ple try to manage facial expressions in a deceptive situation
Fig. 5 Average percentage of bluff with standard error
in which the poker game also involves deceptive interaction
[33]. It is considered that the participants seriously accepted
the poker game with the humanoid as much as a real poker
game with human in which deceptive interaction is inherent.
On the contrary, the frequency of eye gaze was increased
when playing with the humanoid. The increase means that
the humanoid movements draw participants’ interest and at-
tention rather than the human opponent. This result is as-
sociated with the significant difference in eye gaze between
playtime and breaktime in Table 6. According to the result
in Table 6, the frequency of eye gaze in the playtime with
the humanoid was significantly increased than that in the
playtime with the human opponent whereas the frequency of
eye gaze in the breaktime with the humanoid is significantly
lower than that in the breaktime with the human opponent
(Gazing at Table: χ(1)=4.50, p<.05). This presents that
the participants surely did not pay attention to the humanoid
in the breaktime when the humanoid did not show any move-
ments. Therefore, we could say that the humanoid motions
fascinated the participants during the poker game.
Regarding the bluff analysis as shown in Fig. 5, in weak
hand condition, the bluff rate in human–human poker game
(M=67.78, S.D.=9.23) is significantly higher than the
bluff rate in human–humanoid poker game (M=47.78,
S.D.=27.24). Only in the weak hand condition, the bluff
rate shows significant difference (χ(1)=3.99, p<.05). In
this analysis, it was particularly found that while playing
with the humanoid, the participants with weak hand used
bluff strategies at the rate of about 50%. According to the
follow-up interview, the participants assumed that the robot
would not have an ability to play Texas hold’em very well
after initially facing its body and movements. Therefore,
they did not play carefully in the beginning of the poker
game. This kind of awareness seemed to encourage the par-
ticipants to play loosely in the beginning of the game. Their
answers proved the assumption that the robotic appearance
and mechanical movements would keep people from being
involved into the game itself.
12 Int J Soc Robot (2014) 6:5–15
Table 5 The difference in human behaviors between human–human poker game and human–humanoid poker game: Kruskal-Wallis test results
Mean and standard deviations Chi-Square
Human–Human poker game Human–Humanoid poker game Effect of hand condition Effect of opponent
Strong Weak Strong Weak in H-H in H-R Strong Weak
Smile:
Smile 1.36 (1.14) 1.82 (1.58) 0.62 (0.91) 0.86 (1.23) 0.37 1.53 3.23+3.57+
Hand movement:
Touching Face 0.60 (1.13) 0.83 (1.63) 0.30 (0.35) 0.40 (0.66) 1.27 0.55 0.01 0.36
Touching Head 0.10 (0.21) 0.03 (0.10) 0.25 (0.49) 0.17 (0.32) 3.32+0.96 0.30 0.25
Touching Arm 0.10 (0.32) 0.00 (0.00) 0.15 (0.34) 0.03 (0.10) 1.00 1.30 1.30 1.00
Eye blink:
Blink 13.21 (7.50) 11.37 (8.12) 17.35 (6.44) 14.88 (6.15) 0.37 0.57 1.46 1.65
Eye gaze:
Gazing at Opponent 2.40 (2.07) 1.17 (1.20) 8.35 (3.30) 6.92 (2.96) 1.06 0.58 10.35∗∗ 12.24∗∗∗
Gazing at Table 3.30 (1.99) 2.30 (1.01) 9.20 (2.66) 7.80 (2.77) 3.93∗9.64∗∗ 10.62∗∗ 13.21∗∗∗
Gazing at Other 0.25 (0.42) 0.50 (0.53) 2.20 (1.49) 1.67 (1.23) 12.05∗∗ 14.04∗∗∗ 14.42∗∗∗ 5.60∗
Note:∗∗∗p<.001, ∗∗p<.01, ∗p<.05, +p<.10
Table 6 The difference in human nonverbal behaviors between playtime and breaktime: Kruskal-Wallis test results
Mean and standard deviations Chi-Square
Human–Human poker game Human–Humanoid poker game Effect of game state Effect of opponent
Playtime Breaktime Playtime Breaktime in H-H in H-R Playtime Breaktime
Smile:
Smile 1.64 (1.37) 4.09 (2.71) 0.78 (1.15) 2.89 (1.99) 4.48∗5.51∗3.72+1.04
Hand movement:
Touching Face 0.69 (1.18) 1.16 (2.00) 0.18 (0.24) 0.31 (0.50) 0.15 0.05 0.72 0.18
Touching Head 0.06 (0.14) 0.43 (0.98) 0.12 (0.19) 0.13 (0.40) 0.05 1.67 0.87 0.53
Touching Arm 0.04 (0.14) 0.29 (0.60) 0.04 (0.08) 0.55 (1.15) 0.53 0.18 0.85 0.16
Eye blink:
Blink 12.11 (7.68) 16.15 (8.30) 16.03 (6.07) 23.55 (10.89) 1.85 3.02+1.65 2.29
Eye gaze:
Gazing at Opponent 1.83 (1.60) 3.27 (2.13) 4.04 (1.73) 2.15 (1.22) 2.06 4.81∗5.86∗0.97
Gazing at Table 2.84 (1.41) 7.71 (2.51) 4.50 (1.56) 5.44 (0.99) 13.72∗∗∗ 1.66 4.81∗4.50∗
Gazing at Other 0.37 (0.37) 3.60 (2.56) 1.05 (0.81) 2.71 (1.18) 14.34∗∗∗ 7.62∗∗ 5.01∗0.05
Note:∗∗∗p<.001, ∗∗p<.01, ∗p<.05, +p<.10
4.2 Effect of Card Hand Strength
Hypothesis 2 was supported. Touching face in human–
human poker game shows marginally significant difference
between strong and weak hand conditions (χ(1)=3.32,
p<.10). Gazing at table (human–human poker game:
χ(1)=3.93, p<.05, human–humanoid poker game:
χ(1)=9.64, p<.01) indicates significant difference. Also,
Gazing at other (human–human poker game: χ(1)=12.05,
p<.01, human–humanoid poker game: χ(1)=14.04,
p<.001) shows highly significant difference (see effect of
hand condition in Table 5).
In case of the hand movement, when the participants had
weak hands, the frequency of touching face was increased
and the frequency of touching head and arm was decreased.
Although there is no significant effect of the hand strength,
Int J Soc Robot (2014) 6:5–15 13
the increased frequency of touching face partially supports
the study that self-touching is increased in deceptive com-
munication [29]. In the poker game, the participants seemed
to consciously hide their facial expressions by touching their
face with hand. The frequency of touching arm, mostly
cross arms, shows that the participants sometimes controlled
their hand movements by fixing hand position like crossing
arms [34].
The frequency of eye gaze in strong hand condition was
significantly increased, meaning that the time of fixed gaze
was decreased. Especially, the frequency of gazing at other
shows the most significantly increase than gazing at oppo-
nent and gazing at table. It presents that the participants with
their undoubted winning hand felt so relaxed and seemed
to concentrate on other things, rather than focusing on the
game playing.
The another result was obtained from the bluff decision
that is dependent on the card hand strength (human–human
poker game: χ(1)=14.65, p<.001, human–humanoid
poker game: χ(1)=3.80, p<.10). The bluff decision was
affected by card hand strength except for the initial round
when the participants did not try to bet deceptively. Mean-
while, it is expected that if the subjects would get hold of
Texas hold’em much more with technical skills, the effect
of card hand strength would be accordingly decreased over
time.
4.3 Relationship Between Card Hand Strength and Human
Behaviors
The binomial logistic regression analysis with backward
stepwise was applied to see which of predictors would best
account for the relationship between hand strength and hu-
man behaviors. Independent and dependent variables were
set as follow; dependent variable (given situation) is the
card hand strength of the participants, and independent vari-
ables (decision and interaction) are the bluff decision and
nonverbal behaviors of the participants. From this analysis,
we learned that the bluff rate (β=−0.86, p<.05), smile
(β=−0.62, p<.10), blink (β=0.76, p<.05) and gazing
at other (β=0.60, p<.10) are the best predicted variables.
Since the result in Table 7supported Hypothesis 3, we could
say that the given situation of the participants has relation to
their decision and nonverbal interaction.
4.4 Effect of Game State: Tension vs. Release
The effect of game state in terms of tension and release
was examined using Kruskal-Wallis test. This analysis, in
particular, focuses on the frequency of eye gaze between
playtime and breaktime. In human–human poker game, gaz-
ingattable(χ(1)=13.72, p<.001) and gazing at other
Table 7 Result table of binomial logistic regression analysis
Coefficient OR p-value 95% C.I for OR
Lower Upper
Bluff −0.86 0.43 0.02∗0.21 0.88
Smile −0.62 0.54 0.06+0.28 1.03
Blink 0.76 2.13 0.03∗1.07 4.23
Gazing at Other 0.60 1.82 0.08+0.94 3.52
Constant −0.94 0.39 0.53
Note:∗p<.05, +p<.10
Model χtest: p<0.01
(χ(1)=14.34, p<.001) show highly significant differ-
ence. In the human–humanoid poker game, gazing at op-
ponent (χ(1)=4.81, p<.05) and gazing at other (χ(1)=
7.62, p<.01) show significant difference. Considering that
the increase in frequency of eye gaze means that the time
of fixed gaze is decreased, the participants in the human–
human poker game fixed their gaze on the table to observe
the game situation while frequently looking at the humanoid
in the human–humanoid poker game. Hypothesis 4 was sup-
ported.
The result derives another interaction pattern change
from the smile between playtime and breaktime. The fre-
quency of smile between playtime and breaktime shows a
significant difference (human–human poker game: χ(1)=
4.48, p<.05, human–humanoid poker game: χ(1)=5.51,
p<.05). It is supposed that the participants controlled their
facial expression more while engaged in the game playing
with tension.
4.5 Measurement of Engagement
The qualitative study about engagement was performed us-
ing Mann-Whitney U test. The result shows no significant
difference between human–human and human–humanoid
poker game (Q1: z=−0.87, Q2: z=−0.91, Q3: z=
−0.08, Q4: z=−0.27, Q5: z=−0.32) as shown in Fig. 6.
Hypothesis 5 was not supported. Although the humanoid
performed simple head and arm movements, the partici-
pants evaluated the similar level of engagement. This result
presents that the humanoid kept the participants engaged in
the poker game in a different way of interaction pattern with
the human–human poker game.
5 General Discussion and Conclusion
5.1 Contributions
It has been found that the opponent presence and card hand
strength were significantly effective on the participant’s bet-
14 Int J Soc Robot (2014) 6:5–15
Fig. 6 The average ratings with standard error intervals with respect
to the question items for engagement
ting decision and nonverbal behaviors. Besides, the relation-
ship between hand strength and human behaviors has been
analyzed and the human opponent model that incorporates
both mathematical and psychological terms has been discov-
ered. Also, the engagement was evaluated through the quali-
tative study and analyzed with the frequency of eye gaze and
smile.
Through our study, two prominent considerations in de-
signing a socially interactive poker playing humanoid were
drawn. The first one is to observe human smile and eye gaze
in the human–humanoid poker game. These can be used for
measuring how engaged human player is in socially inter-
acting with the robot. The participants in human–humanoid
poker game preferred to interact with the robot more than in
the human–human poker game. It does not mean that they
did not play the game with the robot attentively because they
attempted to hide intentions by controlling over their facial
expressions. On the other hand, the participants paid atten-
tion to play the game more than in the human–humanoid
poker game. We found that the effect of hand strength in
the human–human poker game was highly significant while
being marginally significant in the human–humanoid poker
game.
It is inferred that our humanoid significantly evoked dif-
ferent nonverbal interaction and difference of betting strat-
egy pattern. Our results embrace the importance of physical
embodiment that has been explored in game tasks [35]. Even
machine-like appearance facilitated social interaction with
the poker playing humanoid through employing anthropo-
morphism that is motivated to design a system that functions
in physical and social space and make its mechanism to fa-
cilitate social interaction with people [36]. From social inter-
action perspective, our poker-playing robot can be utilized
as a social partner. Although the benefit of attracting peo-
ple to socially interact with it differentiates itself from other
virtual agents in video game and networked game, it will
successfully encourage a human social behavior and keep
people motivated to play games with it.
The second consideration is to model human opponent
based on betting decision and the nonverbal interaction al-
together. We expect that this model will enable the poker
playing robot to perform in a same way of understanding
human player as discussed. Consider the following exam-
ple: the robot assesses its opponent hand strength in order
to make its betting decision. It often makes wrong deci-
sion because of imperfect information of the game the in-
tentional nonverbal behaviors of the opponent. To resist be-
ing defeated in succession, the robot modifies the opponent
model parameter by learning the game situation, opponent’s
intentions, etc. Similar research modeled robot deception to
capacitate human-like survival skills in situations involving
conflict [37]. This robot recognizes a situation and selects
deceptive strategy to reduce the chance of being found. Like-
wise, the model of human opponent that we investigated in
this study will provide the potential of adapting the game
playing humanoid to human social poker game.
As a final remark, discovering commonalities and differ-
ences of the human perceptions on human and robot guides
us to learn how to make the robot to be perceived as it acts
like a human. Likewise, in order to be a social partner, the
game playing humanoid must be perceived as it interacts
like human. Therefore, we conducted comparative study to
see how differently people respond toward human and robot.
We hope that this study contributes to not only the practical
design of socially interactive game playing robots, but also
the theoretical approach on the realization of behaving in a
similar way of human doing in a game playing.
5.2 Future Works
In the future, further psychological research should be con-
ducted with real money in long-term in order to investigate
the effects of opponent presence and hand strength and en-
gagement in more real situation. Furthermore, it should be
investigated how robot behavior affects people to interpret
it in the poker game with the developed humanoid plat-
form. And the proposed decision making structure of the
humanoid playmate will be verified to be able to understand
and learn human behaviors while playing the poker game
autonomously.
Acknowledgements This work is partially supported by Grand-in-
Aid for Scientific Research and Global COE Program on “Cybernetics:
fusion of human, machine, and information systems” by MEXT, Japan.
References
1. Lazzaro N (2004) Why we play games: four keys to more emotion
without story. In: Game developers conference
2. Dantam N, Kolhe P, Stilman M (2011) The motion grammar for
physical human–robot games. In: Int conf on robotics and automa-
tion, pp 5463–5469
Int J Soc Robot (2014) 6:5–15 15
3. Wallhoff F, Bannat A, Gast J, Rehrl T, Dausinger M, Rigoll
G (2009) Statistics-based cognitive human–robot interfaces for
board games—let’s play! In: Human interface part II. Lecture note
in computer science, vol 5618. Springer, Berlin, pp 108–715
4. Marquis S, Elliott C (1994) Emotionally responsive poker playing
agents. In: Notes for the 12th national conf on artificial intelli-
gence workshop on artificial intelligence, artificial life, and enter-
tainment, pp 11–15
5. Kovács G, Ruttkay Z, Fazekas A (2007) Virtual chess player with
emotions. In: 4th Hungarian conf on computer graphics and ge-
ometry, pp 182–188
6. Caro M (2003) Caro’s book of poker tells. Cardoza Publishing,
Las Vegas
7. Berne E (1996) Games people play. the basic hand book of trans-
actional analysis. Ballantine, New York
8. Siler K (2010) Social and psychological challenges of poker
source. J Gambl Stud 26(3):401–420
9. Seale DA, Phelan SE (2010) Bluffing and betting behavior in a
simplified poker game. J Behav Decis Mak 23(4):335–352
10. Sung M, Pentland AP (2005) PokerMetrics: stress and lie de-
tection through non-invasive physiological sensing. Ph.D. thesis,
MIT Media Laboratory
11. Walters ML, Dautenhahn K, te Boekhorst R, Koay KL, Kaouri
C, Woods S, Nehaniv C, Lee D, Werry I (2005) The influence of
subjects’ personality traits on personal spatial zones in a human–
robot interaction experiment. In: IEEE int workshop on robot and
human interactive communication, pp 347–352
12. Billings D, Papp D, Schaeffer J, Szafron D (1998) Opponent mod-
eling in poker. In: Proc AAAI-98, pp 493–499
13. King GA, Sorrentino RM (1983) Psychological dimensions of
goal-oriented interpersonal situations. Pers Soc Psychol 44:140–
162
14. Kelley HH, Stahelski AJ (1970) Social interaction basis of coop-
erators’ and competitors’ beliefs about others. J Pers Soc Psychol
16(1):66–91
15. Lee KM, Jung Y, Kim J, Kim SR (2006) Are physically em-
bodied social agents better than disembodied social agents?: The
effects of physical embodiment, tactile interaction, and people’s
loneliness in human–robot interaction. Int J Hum-Comput Stud
64(10):962–973
16. Bainbridge WA, Hart JW, Kim ES, Scassellati B (2008) The effect
of presence on human–robot interaction. In: 17th IEEE int sym on
robot and human interactive communication (ROMAN)
17. Bainbridge WA, Hart JW, Kim ES, Scassellati B (2011) The ben-
efits of interactions with physically present robots over video-
displayed agents. Int J Soc Robot 3(1):41–52
18. Kanda T, Ishiguro H, Imai M, Ono T (2003) Body movement
analysis of human–robot interaction. Int Joint Conf Artif Intell
18:177–182
19. Moriguchi Y, Minato T, Ishiguro H, Shinohara I, Itakura S (2010)
Cues that trigger social transmission of disinhibition in young chil-
dren. J Exp Child Psychol 107:181–187
20. Mutlu B, Osman S, Forlizzi J, Hodgins J, Kiesler S (2006) Task
structure and user attributes as elements of human–robot interac-
tion design. In: The 15th IEEE int symp on robot and human in-
teractive communication, pp 74–79
21. Findler NV (1977) Studies in machine cognition using the game
of poker. Commun ACM 20(4):230–245
22. Lombard M, Ditton T, Crane D, Davis B, Gil-Egul G, Horvath
K, Rossman J (2000) Measuring presence: a literature-based ap-
proach to the development of a standardized paper and pencil in-
strument. In: The third international workshop on presence, Delft,
The Netherlands
23. Poels K, de Kort Y, IJsselsteijn W (2007) “It is always a lot of
fun!!”: exploring dimensions of digital game experience using fo-
cus group methodology. In: Proc of the conf on future play, pp
83–89
24. Gu E, Badler NI (2006) Visual attention and eye gaze during mul-
tiparty conversations with distractions. In: Lecture notes in com-
puter science, vol 4133. Springer, Berlin, pp 193–204
25. Sidne CL, Kidd CD, Lee C, Lesh N (2004) Where to look: a study
of human–robot engagement. In: Proc of intelligent user interaces,
pp 78–84
26. Mahmud AA, Mubin O, Shahid S, Martens JB (2010) Designing
social games for children and older adults: two related case stud-
ies. Comput Entertain 1(3):147–156
27. Kim MG, Suzuki K (2010) A card playing humanoid for un-
derstanding socio-emotional interaction. In: Entertainment com-
puting ICEC’10. Lecture notes in computer science, vol 6243.
Springer, Berlin, pp 9–19
28. Kelley JF (1984) An iterative design methodology for user-
friendly natural language office information applications. ACM
Trans Off Inf Syst 2(1):26–41
29. Hayano DM (1980) Communicative competency among poker
players. J Commun 30(2):99–104
30. DePaulo BM, Lindsay JJ, Malone BE, Muhlenbruck L, Carlton K,
Cooper H (2003) Cues to deception. Psychol Bull 129(1):74–118
31. Ekman P (1985) Telling lies: clues to deceit in the marketplace,
politics and marriage. W.W. Norton, New York
32. Ponse M, Tuyls K, Jong SD, Ramon J, Croonenborghs T,
Driessens K (2008) The dynamics of human behaviour in poker.
In: Proc of the 20th Belgium-Netherlands conf on artificial intelli-
gence (BNAIC 2008), pp 225–232
33. Eckman P (2003) Darwin, deception, and facial expression. Ann
NY Acad Sci 1000:205–221
34. Vrij A, Edward K, Roberts KP, Bull R (2004) Detecting deceit
via analysis of verbal and nonverbal behavior. J Nonverbal Behav
24(4):239–263
35. Pereira A, Martinho C, Leite I, Paiva A (2008) ICat, the chess
player: the influence of embodiment in the enjoyment of a game.
In: Proc of the 7th int joint conf on autonomous agents and multi-
agent systems, pp 1253–1256
36. Duffy BR (2003) Anthropomorphism and the social robot. Robot
Auton Syst 42(3–4):177–190
37. Wagner AR, Arkin RC (2009) Robot deception: recognizing when
a robot should deceive. In: IEEE int symp on computational intel-
ligence in robotics and automation, pp 46–54
Min-Gyu Kim He is currently a Senior Researcher of Interaction
Science Research Center at Sungkyunkwan University, Korea. He re-
ceived the B.E. in Electronics Engineering and M.E. in Mechanical En-
gineering from Korea Aerospace University, Korea, in 2003 and 2005
respectively. He also received Ph.D. in Intelligent Interaction Tech-
nologies from University of Tsukuba, Japan in 2012. He was a Post-
doctoral Researcher in Artificial Intelligence Laboratory at University
of Tsukuba, Japan in 2012. He is interested in Robotics, Human Robot
Interaction and Cognitive Engineering.
Kenji Suzuki He is currently an Associate Professor of the Faculty
of Engineering, Information and Systems, and Principal Investigator
of Artificial Intelligence Laboratory, University of Tsukuba, Japan. He
also belongs to the Center for Cybernetics Research, and he is a JST
PRESTO Researcher, Japan Science and Technology Agency, Japan.
He received the B.S. in Physics, M.E. and Dr. Eng. in Pure and Ap-
plied Physics from Waseda University, Tokyo, Japan, in 1997, 2000
and 2003 respectively. He was a visiting researcher at the Laboratory
of Musical Information, University of Genoa, Italy from 1997 to 1999,
and also at LPPA, Laboratory of Physiology of Perception and Action,
at the College de France, Paris in 2009. His research interests include
Robotics, Assistive and Rehabilitation Robotics, Human Robot Inter-
action, Augmented Human, and Affective computing. He is currently
a member of IEEE and ACM.
