Neural Networks 18 (2005) 666–673

www.elsevier.com/locate/neunet

2005 Special Issue

Application of neural network to humanoid robots—development
of co-associative memory model
Kazuko Itoha,*, Hiroyasu Miwab,c,d, Hideaki Takanobud,e, Atsuo Takanishia,c,d,f
a
Department of Mechanical Engineering, Waseda University, #59-308, 3-4-1 Okubo, Shinjuku-ku, Tokyo 169-8555, Japan
b
Digital Human Research Center, National Institute of Advanced Industrial Science and Technology (AIST), Tokyo, Japan
c
Institute for Biomedical Engineering, ASMeW, Waseda University, Tokyo, Japan
d
Humanoid Robotics Institute (HRI), Waseda University, Tokyo, Japan
e
Department of Mechanical Systems Engineering, Kogakuin University, Tokyo, Japan
f
Advanced Research Institute for Science and Engineering, Waseda University, Tokyo, Japan

Abstract
We have been studying a system of many harmonic oscillators (neurons) interacting via a chaotic force since 2002. Each harmonic
oscillator is driven by chaotic force whose bifurcation parameter is modulated by the position of the harmonic oscillator. Moreover, a system
of mutually coupled chaotic neural networks was investigated. Different patterns were stored in each network and the associative memory
problem was discussed in these networks. Each network can retrieve the pattern stored in the other network.
On the other hand, we have been developing new mechanisms and functions for a humanoid robot with the ability to express emotions and
communicate with humans in a human-like manner. We introduced a mental model which consisted of the mental space, the mood, the
equations of emotion, the robot personality, the need model, the consciousness model and the behavior model. This type of mental model was
implemented in Emotion Expression Humanoid Robot WE-4RII (Waseda Eye No.4 Refined II).
In this paper, an associative memory model using mutually coupled chaotic neural networks is proposed for retrieving optimum memory
(recognition) in response to a stimulus. We implemented this model in Emotion Expression Humanoid Robot WE-4RII (Waseda Eye No.4
Refined II).
q 2005 Elsevier Ltd. All rights reserved.

Keywords: Neural network; Associative memory; Humanoid robot

1. Introduction neurons was described by the Brownian motion with chaotic

We know the Hopfield model as a typical model of fully
connected neural network. In the Hopfield model, the
energy function is shown to decrease monotonically and one
energy minimum corresponds to a single memory (Hopfield,
1984; Hopfield & Tank, 1985). However, if the system falls
into a spurious minimum, it cannot then escape. Recently,
many methods have been proposed to solve this problem by
using chaos. Shimizu proposed a chaotic neural network
model, which consists of N Brownian particles (neurons)
(Shimizu, 1998). The time evolution of the internal state of
* Corresponding author. Tel.: C81 3 5286 3257; fax: C81 3 5273 2209.
E-mail addresses: itoh@suou.waseda.jp (K. Itoh), takanisi@waseda.jp
(A. Takanishi).
URL: http://www.takanishi.mech.waseda.ac.jp/.
0893-6080/$ - see front matter q 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.neunet.2005.06.021
force. It was found that the network retrieves all stored and
reversed patterns in the associative memory problem. Thus,
the dynamics of one neural network have been extensively
studied, although, the study of coupled neural networks
appears far from complete.
Itoh and Shimizu studied the dynamics of a simple
harmonic oscillator driven by chaotic force. The bifurcation
parameter of the chaotic force was modulated by the
position of the harmonic oscillator (Itoh & Shimizu, 2000a,b).
We used that idea to study a neural network model, which
consists of many harmonic oscillators (neurons) interacting
via a chaotic force (Itoh & Shimizu, 2002). In addition, a
system of mutually coupled chaotic neural networks was
investigated. Different patterns were stored in each
network and the associative memory problem was
discussed therein. If two networks are not coupled and
one pattern is stored in each network, the energy of each
network has only one minimum corresponding to
 K. Itoh et al. / Neural Networks 18 (2005) 666–673
the pattern. Each network retrieves only a single pattern.
However, we found that each network can retrieve the
pattern stored in the other network if they are coupled,
meaning namely that Virtual Attractor corresponding to
the pattern stored in the other network appeared in each
network.
On the other hand, robots have become indispensable for
human life. Most popular robots are of the industrial kind
with various functions, such as assembly and conveyance.
However, operators have to define their behavior with very
complex processes and methods. We hope that personal
robots which are active in joint work and community life
with humans will become popular in the future. Such
personal robots must adapt to partners and the environment
and to communicate naturally with humans. Therefore, new
mechanisms and functions are developed in order to realize
the natural communication by expressing the emotions,
behaviors and personality in a human-like manner. Sony
Corporation developed the entertainment humanoid QRIO.
It can autonomously walk with CCD cameras information
on the head and control its behavior using the homeostasis
regulation mechanism (Fujita et al., 2003). In addition, it
realized behavior module selection and motion modulation
by emotion in Emotionally GrOunded (EGO) Architecture
for autonomous robots (Sawada et al., 2004).
We developed WE-3 (Waseda Eye No.3) series which
achieved coordinated head-eye motion with V.O.R.
(Vestibular-Ocular Reflex), depth perception using the
angle of convergence between the two eyes, adjustment to
the brightness of an object with the eyelids and four
sensations; visual, auditory, cutaneous and olfactory
sensations. In addition, we produced emotional expressions
and various kinds of behavior by Emotion Expression
Humanoid Robot WE-4 (Waseda Eye No.4) series with the
face, neck, lungs, waist, 9-DOFs emotion expression
humanoids arms and humanoid robot hands RCH-1 (Robo
Casa Hand No.1) (Miwa et al., 2004; Roccella et al., 2004).
In addition, the mental model for humanoid robots has been
under development from both robotic and psychological
perspectives in order to realize human-like motion. We
introduced a mental space with three independent par-
ameters, the mood, the second order equations of emotion,
the robot personality (Miwa et al., 2003a), the need model
(Miwa et al., 2003b), the consciousness model and the
behavior model.
However, the previous robot had just a single kind of
recognition in response to a stimulus and showed behavior
corresponding to this recognition. Humans recognize a
stimulus and generate behavior depending on their mood at
that time. We considered human-like recognition to be
realized by developing a memory model in order to retrieve
an optimum memory (recognition) in response to a stimulus.
Therefore, we proposed a co-associative memory model
using mutually coupled chaotic neural networks and applied
this model to Emotion Expression Humanoid Robot WE-4RII
(Waseda Eyes No.4 Refined II) (Itoh et al., in press).
667
We described a model of mutually coupled chaotic neural
networks in Section 2 and Emotion Expression Humanoid
Robot WE-4RII in Section 3. The relations between human
memory and their mood and between human performance and
their activation level were shown in Section 4. Subsequently,
we developed a co-associative memory model using mutually
coupled chaotic neural networks in Section 5. This model was
applied to WE-4RII in Section 6.
2. Neural network model
2.1. Chaotic neuron model
First, we discuss a neuron model that behaves as a simple
harmonic oscillator driven by a chaotic force. If the position
of the harmonic oscillator at time t is denoted by x(t),
the time evolution of the internal state of the neuron is
governed by:
_ C u20 xðtÞ Z f ðtÞ;
€ C kxðtÞ
xðtÞ (1)
where k and u0 are the damping constant and the eigen-
frequency, respectively and f(t) is the input from the
surrounding neurons. We assumed that f(t) was given by a
chaotic force, whose amplitude changes chaotically over
time interval t:
K
f ðtÞ Z pffiffiffi yðnÞ
t
for nt% t! ðn C 1Þt ðn Z 0; 1; 2; .Þ; ð2Þ
here
pffiffiffiK is the magnitude of the chaotic force f(t). The factor
1= t is required to obtain a finite diffusion constant in the
small-t limit. In Eq. (2), y(n) denotes the output of the
neuron and represents the nth iterate of a map. As an
example of this map, we employ the Logistic map:
yðn C 1Þ ZrðnÞð0:5KyðnÞÞð0:5 C yðnÞÞK0:5
ðK0:5% yðnÞ% 0:5Þ; ð3Þ
where, r(n) is the bifurcation parameter and y(n) changes
chaotically or periodically according to the bifurcation
parameter r(n). Since, the chaotic output changes stepwise
over time interval t, we can solve x(t) formally as a function
of y(n) and then observe the system stroboscopically over
time interval t. We denote the internal state and velocity at
time tZnt by x(n) and v(n), respectively. If we solve
Eq. (1), we obtain the recurrent relation:
h m  i KyðnÞ
xðn C 1Þ Z 1K sin ut C cos ut a 2 pffiffiffi
u u0 t
m 
sin ut
C sin ut C cos ut axðnÞ C avðnÞ
u u
(4)
668 K. Itoh et al. / Neural Networks 18 (2005) 666–673

sin ut KyðnÞ u20 here, Wi,j is the coupling constant from neuron j to neuron i.
vðn C 1Þ Z a pffiffiffi K axðnÞsin ut It is clear from Eqs. (9)–(13) that each neuron is driven by
u t u
m  the chaotic force yi(n) of the surrounding neurons and itself,
K sin utKcos ut avðnÞ; (5) whose bifurcation parameter ri(n) are modulated by the
u internal state xi(n). To store patterns in the network, we use
here, m, a and u are defined by: the Hebb rule to determine Wi,j as:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
k
a Z eKmt ; mZ ; u Z u20 Km2 : (6) 1 XP
2 Wi;j Z xs xs ; (14)
N sZ1 i j
If we denote u as:
u Z 2p=Tt; (7) where, xi denotes the stored pattern vector and xi takes C1
orK1. The self-coupling constant is equal to 1; Wi,iZ1 from
T denotes the period of the harmonic oscillator in units of Eq. (14). If we put TZ2.0, the neural network can retrieve
the time interval t. We proposed a model in which the the original and reverse patterns alternately, meaning we
bifurcation parameter is modulated by the internal state of can say that this neural network can perform very well.
the neuron as:
rðnÞ Z 4Kb C b cos2 bxðnÞ ð0% b% 4Þ; (8) 2.3. Mutually coupled neural networks
where b and b are the control parameters. Since, the
bifurcation parameter is modulated by the internal state of In this paper, we study a system where two neural
the neuron as in Eq. (8), chaos is controlled by the neuron networks (Networks A and B) modeled according to Eqs.
itself. Therefore, a new type of feed back mechanism is (9)–(12) are coupled mutually. We denote the internal state
included in this model. The internal state x(n) determines and the velocity of neuron i in Network a (aZA or B) at
the bifurcation parameter r(n), which, in turn, determines time tZnt by xai;j ðnÞ and vai;j ðnÞ, respectively. As a way to
the dynamics of the chaotic output y(n). The chaotic output couple two neural networks, we connected the neurons at
then affects the dynamics of the neuron. Thus, the dynamics the same location in the two networks. That is, neuron i in
of the chaotic output is changed by the system itself. Network A(B) was connected with neuron i in Network
B(A), whose coupling constant is given by ZiA;B ðZiB;A Þ, as
shown in Fig. 1. Therefore, hai ðnÞ is defined as Eq. (15)
2.2. One neural network instead of Eq. (13):

We used that idea to study a neural network model which X

N

consists of many harmonic oscillators (neurons) interacting hAi ðnÞ Z Wi;jA yAj ðnÞ C ZiA;B yBi ðnÞ
jZ1
via chaotic force. We denote the internal state of neuron i at
(15)
time tZnt by xi(n)(iZ1,2,.,N). The time evolution of a X
N
system is defined by: hBi ðtÞ Z Wi;jB yBj ðtÞ C ZiB;A yAi ðtÞ:
h m  i Kh ðnÞ jZ1
i
xi ðn C 1Þ Z 1K sin ut C cos ut a pffiffiffi
u u20 t If two networks are not coupled and a different pattern
m  was stored in each network, the energy of each network has
sin ut
C sin ut C cos ut axi ðnÞ C avi ðnÞ only one minimum corresponding to the pattern. Each
u u network retrieves only one pattern. In this model, however,
(9) each network can retrieve the pattern stored in the other
network if they are coupled.
sin ut Khi ðnÞ u20
vi ðn C 1Þ Z a pffiffiffi K axi ðnÞsin ut
u t u
m 
K sin utKcos ut avi ðnÞ (10)
u

yi ðn C 1Þ Z ri ðnÞð0:5Kyi ðnÞÞð0:5 C yi ðnÞÞK0:5 (11)

ri ðnÞ Z 4Kb C b cos2 bxi ðnÞ; (12)

hi(n) denotes the interaction among neurons:
X
N
hi ðnÞ Z Wi;j yi ðnÞ; (13)
jZ1 Fig. 1. Coupling of two neural networks.
 K. Itoh et al. / Neural Networks 18 (2005) 666–673 669

3. Emotion expression humanoid robot WE-4RII Table 1

DOFs configuration of WE-4RII

Part DOF
3.1. Mechanical hardware
Neck 4
Eyes 3
Fig. 2 presents the hardware overview of Emotion Eyeilds 6
Expression Humanoid Robot WE-4RII (Waseda Eye No.4 Eyebrows 8
Refined II) developed in 2004. It has 59 DOFs for Lips 4
expressing motions and emotions as shown in Table 1, Jaw 1
Lung 1
and four of the five human senses: visual, auditory, olfactory
Waist 2
and tactile senses for detecting external stimuli. The hand is Arms 18
Humanoid Robot Hand RCH-1 (RoboCasa Hand No. 1) Hands 12
which is designed and developed by ARTS Lab, Scuola Total 59
Superiore Sant’Anna in order to express not only emotional
expressions but also active behavior. clarify the object of the robot’s behavior, and the behavior
model in order to generate various kinds of behavior.
3.2. Previous mental model The 3D mental space is shown in Fig. 3. Emotion Vector
E was defined in the mental space as the robot’s mental
The mental model was developed with a 3D mental space state as:
consisting of pleasantness, activation, and certainty, the E Z ðEP ; EA ; EC Þ; (16)
mood, the second order equations of emotion, the robot
personality (Miwa et al., 2003a), the need model in order to where EP is the pleasantness component of the emotion, EA
realize not only passive motion but also active motion is the activation component of the emotion and EC is the
(Miwa et al., 2003b), the consciousness model in order to certainty component of the emotion. Seven different
emotions, namely happiness, anger, surprise, sadness, fear,
disgust, and neutral, are mapped into the space as shown in
Fig. 4. The robot determines its emotion depending on the
emotional region traversed by Emotion Vector E.
The mental state is affected by not only the emotion but
also mood. Therefore, Mood Vector M, consisting of a
pleasantness component and activation component was
introduced:

M Z ðMP ; MA ; 0Þ; (17)

where MP is the pleasantness component of the mood and

MA is the activation component of the mood, respectively.
Since, we considered the current mental state to influence
the pleasantness of mood, MP was defined as the integral of
the pleasantness component of Emotion Vector as Eq. (18).
On the other hand, since the activation component of Mood
Vector is similar to the human biological rhythm such as the
internal clock, Van del Pol equation was applied to define

certain
Emotion Vector E
arousal

Pleasantness
unpleasant n pleasant
tio
va
Certainty

t i
Ac Trajectory of E
asleep
uncertain

Fig. 2. Emotion expression humanoid robot WE-4RII. Fig. 3. 3D mental space and emotion vector E to define robot mental state.
670 K. Itoh et al. / Neural Networks 18 (2005) 666–673

Certainty auto-associative memory: e.g. humans retrieve all of a

face from a part of a face, and co-associative memory: e.g.
Surprise humans retrieve ‘Apple’ from ‘Red’ or retrieve ‘Lemon’
Happines from ‘Sour’. Human memory relates to their mood, that is
s
Ple mood state-dependency and mood congruency (Takano,
asa al Anger
ntne e ur
t 1995). If humans store a memory when in certain mood,
ss N
they easily retrieve the same memory when in the same
Sleep mood. This is known as mood state-dependency. On the
other hand, mood helps to retrieve a memory corresponding
Disgust Fear
to the same mood, known as mood congruency. Basically,
tion
t iva humans tend to retrieve pleasant memories if they are
Ac al
e utr feeling pleasant and conversely, unpleasant memories if
N Sadness they are feeling unpleasant.
Sleep
Moreover, human performance is related to their
activation level (Hebb, 1975). If an activation level becomes
too high or too low, active human performance becomes
Fig. 4. Seven different emotions mapped into 3D mental space.
impossible. The best human performance comes at a
MA as Eq. (19). medium activation level.
ð In this paper, we proposed a model of co-associative
MdP Z EP dt (18) memory with mood state-dependency, mood congruency
and the relation between performance and activation level.
The robot retrieves a pleasant memory if its mood is
2
€ dA C ð1KMdA _ dA C MdA Z 0
M ÞM (19) pleasant and an unpleasant memory if its mood is unpleasant
In addition, the need model was also developed, for realizing the former two of the three. In addition, we
consisting of Appetite, Need for Security and Need for controlled the time of memory retrieval depending on the
Exploration as: activation component of the robot emotion for realizing the
2 3 last one of the three.
NA
6 7
N Z 4 NS 5: (20)
NE 5. Co-associative memory model
where, NA is the appetite component of the need, NS is the
component of the need for security and NE is the component Next, a different memory was stored in each network and
of the need for exploration. The appetite depends on the co-associative memory was discussed in these networks.
total spent energy, derived from the sum of the basal ‘Tomato’ was stored in Network A as the unpleasant
metabolism energy and output energy: memory for ‘Red’ and ‘Apple’ was stored in Network B as
the pleasant memory for ‘Red’. Of course, we can select
DNA Z f ðDAÞ DA Z DABM C DAEA ; (21) other memories as the robot personality. Fig. 5 shows the
where, DA means the small difference in spent energy of the number with which Network A retrieved each memory to
robot, ABM means the basal metabolism energy, and AEA the coupling constant from Network B to Network A ZiA;B if
means the output energy. the coupling constant from Network A to Network B ZiB;A
However, the robot with the previous mental model had was fixed. For ZiB;A Z 0:0, we have found that the number of
just a single kind of recognition in response to a stimulus and ‘Tomato’ retrieval decreased and that of ‘Apple’ retrieval
showed behavior corresponding to this recognition. Humans increased if ZiA;B becomes large as shown in Fig. 5(a).
recognize a stimulus and generate behavior depending on Therefore, mood state-dependency and mood congruency
their mood at that time. We considered development of a were introduced whereby ZiA;B is modulated by the
memory model for humanoid robots to be possible using pleasantness component of the mood. The robot recognizes
mutually coupled chaotic neural networks and that optimum ‘Red’ as ‘Apple’ for a pleasant mood and as ‘Tomato’ for an
memory retrieval (recognition) in response to a stimulus unpleasant mood. However, humans retrieve their favorite
could be achieved by that memory model. food, even if they are feeling unpleasant due to hunger.
Therefore, ZiA;B was denoted as the sum of the pleasantness
component of the mood and appetite component of the need:
4. Human memory ZiA;B Z Mp C NA : (22)
Associative memory denotes a memory for links Then, the robot recognizes ‘Red’ as ‘Apple’ even if feeling
or connections between items. This is divided into unpleasant due to hunger. In addition, the region of ZiA;B in
 K. Itoh et al. / Neural Networks 18 (2005) 666–673 671

(a) External Environment

60000
Number of Retrieving times

Apple
Apple Robot
50000
Toma
Tomatoto Sensing Motion Reflex Motion
40000 Recognition Intelligence Behavior
Memory
30000
Sensing Emotion Expression
20000 Personality Mood Personality
10000 Emotion Emotion
Consciousness
Need Need
0 Need
0 15 30 ZiA,B 45 60 75
B,A
Zi =0.0 Se
Sensing Autonomic Reflex Motion
(b) Internal Environment
60000
Number of Retrieving times

50000
40000
30000
20000
10000
0 for ZiB;A ZK4:0. Network A retrieved ‘Tomato’ in the
0 15 30 Z A,B 45
i
Fig. 5. Number of retrieving each memory to ZiA;B .
which the robot recognizes ‘Red’ as ‘Red’ itself was
defined. Accordingly, human-like recognition is realized.
On the other hand, we controlled the time required for
memory retrieval depending on the activation component of
the robot emotion in order to realize the relation between
performance and activation level. A robot with a suitable
activation level can retrieve a memory corresponding to its
mood. However, the robot needs time for memory retrieval
if the activation component Ea of the emotion becomes too
Fig. 6. Relation between performance and activation level.
Apple
Apple
Fig. 7. New mental model with co-associative memory model for humanoid
Toma
Tomatoto robot control.
high or too low. The relation between performance and
activation level was assumed as sown in Fig. 6. Moreover,
we considered that the wrong memory was sometimes
retrieved for too high or too low activation level. Fig. 5(b)
shows the number with Network A retrieved each memory
60 75
B,A
smaller half region of ZiA;B and ‘Apple’ in the larger half
Zi =–4.0
region of ZiA;B , mostly. However, it sometimes retrieved
another memory. Therefore, we defined ZiB;A as equal to 0.0
for a suitable activation level and K4.0 for too high or too
low activation level. Fig. 7 shows a new mental model for
humanoid robots, including a co-associative memory
model.
6. Experimental result
We evaluated a new co-associative memory model
through implementation in Emotion Expression Humanoid
Robot WE-4RII. We set the parameter values as KZ14.0,
mZ0.05, TZ2.0, bZ0.05, tZ0.1 and bZ1.2, and the
initial values as xi(0)Z25.0, vi(0)Z0.0 and yi(0)Z0.2.
Figs. 8 and 9 show the time evaluation of the robot emotion,
the robot mood and the retrieved memory.
At first, we describe the results of mood state-
dependency and mood congruency. In Fig. 8(a), we showed
the red ball to the robot. Since, the robot started to feel
pleasant when stroked, it retrieved the pleasant memory
‘Apple’ and exhibited ‘Happiness’ emotional expression.
The pleasant level Ep became higher once ‘Apple’ was
retrieved. Next, the robot became unpleasant by recognizing
ammonia. The robot retrieved the unpleasant memory
‘Tomato’ and exhibited ‘Disgust’ emotional expression.
Ep dropped when retrieving ‘Tomato’. In Fig. 8(b), the robot
started feeling unpleasant after being hit. Therefore, it
retrieved the unpleasant memory ‘Tomato’. If the robot
moved considerably, it felt hungry. Then it retrieved the
pleasant memory ‘Apple’ in spite of unpleasant mood due to
672 K. Itoh et al. / Neural Networks 18 (2005) 666–673

Hit Retrieve “Tomato” Retrieve “Apple”

(a) Stroke Retrieve “Apple” Retrieve (b) 2000
- - -
“Tomato”
Ammonia
10000 1000

Emotion and Mood Vector

Emotion and Mood Vector

0
–1000 20 40 60 80 100
5000
–2000
–3000
0
0 10 20 30 40 50 60 70 –4000
–5000 Ep
–5000
Ep –6000 Mp
Mp
Recalling –7000 NA
A
–10000 –8000
time s time s

Fig. 8. Experimental results of mood state-dependency and mood congruency by implementing in WE-4RII.

Hit Retrieve Retrieve

(a) Hitt Red Retrieve “Tomato” (b) Red “Apple” “Tomato”
10000 12000
Ea

Emotion and Mood Vectors

10000
Emotion and Mood Vectors

8000 Mp
Ea
8000
6000 Mp
6000
4000
4000
2000
2000
0 0
0 5 10 15 20
–2000 0 5 10 15 20 25 30 35
–2000
–4000 –4000
–6000 –6000
time s time s

Fig. 9. Experimental results of the relation between performance and activation level by implementing in WE-4RII.

hunger. The robot takes the apple and mimics eating it. If the order to realize human-like recognition. We used mutually
robot’s hunger is satisfied by breathing its favorite smell, the coupled chaotic neural networks for this model, which
robot becomes happy. consists of many harmonic oscillators (neurons) interacting
Next, we considered the relation between performance via a chaotic force. This memory model was implemented in
and activation level. Fig. 9(a) shows the result for the Emotion Expression Humanoid Robot WE-4RII. We
medium activation level and Fig. 9(b) shows the result for confirmed the ability of the robot to retrieve optimum
an excessively high activation level. It is clear from Fig. 9(a) memory according to the mood and appetite in response to a
that it took only about 2[s] for the robot from looking at the stimulus. Moreover, the time taken to retrieve a memory was
red ball until retrieving ‘Tomato’. On the other hand, the controlled by the activation level. The robot could retrieve
robot needed time for memory retrieval since the activation the memory corresponding to its mood at the time for the
level became very high by receiving strong stimuli ‘Hit’ in medium activation level. However, it sometimes mistook
Fig. 9(b). It took about 9[s]. In addition, the robot retrieved retrieving memory for an overly high or low activation level.
the wrong memory ‘Apple’, in spite of an unpleasant mood. Subsequently, the robot shows optimum behavior according
Thus, we confirmed the ability of the robot to recognize a to the mood, appetite and activation level.
stimulus depending on the mood and appetite and show In this paper, the robot’s mood was divided into two
optimum behavior. In addition, the time changed according cases of pleasantness and unpleasantness. However, we
to the activation level and the robot sometimes mistook the consider that it is necessary for natural communication with
recognition for an overly high or low activation level. humans to express the robot mood continuously. In addition,
we will study a method to retrieve far more memories.

7. Conclusions and future work

Acknowledgements
In this paper, we developed a co-associative memory
model with mood state-dependency, mood congruency and Part of this research was conducted at the Humanoid
the relation between performance and activation level in Robotics Institute (HRI), Waseda University. The authors
 K. Itoh et al. / Neural Networks 18 (2005) 666–673
would like to express their thanks to Okino Industries LTD,
OSADA ELECTRIC CO. LTD, SHARP CORPORATION,
Sony Corporation, Tomy Company LTD and ZMP INC. for
their financial support for HRI. The authors would also like
to thank Italian Ministry of Foreign Affairs, General
Directorate for Cultural Promotion and Cooperation, for
its support to the establishment of the ROBOCASA
laboratory and for the realization of the two artificial
hands. In addition, this research was supported by a Grant-
in-Aid for the WABOT-HOUSE Project by Gifu Prefecture.
Finally, the authors would like to express thanks to ARTS
Lab, NTT Docomo, SolidWorks Corp., Advanced Research
Institute for Science and Engineering of Waseda University,
and Prof. Hiroshi Kimura for their support to our research.
References
Fujita, M., Kuroki, Y., Ishida, T., & Doi, T. (2003). Autonomous behavior
control architecture of entertainment Humanoid Robot SDR-4X.
Proceedings of the IROS2003, (pp.960–967).
Hebb, D.O. (1975). Koudougaku Nyumon (in Japanese). Kinokuniya
Syoten. (pp.257-262).
Hopfield, J. J. (1984). Neurons with graded response have collective
computational properties like those of two-state neurons. Proceedings
of National Academy of Sciences USA, 81, 3088–3092.
Hopfield, J. J., & Tank, D. W. (1985). Neural computation of decisions in
optimization problems. Biological Cybernetics, 52, 141–152.
Itoh, K., & Shimizu, T. (2000a). New type of feed-back mechanism with
bifurcation parameter modulation. American Institute of Physics
Conference Proceedings, 519, 649–651.
673
Itoh, K., Shimizu, T. (2000). Cooperative phenomena between the chaotic
force and the harmonic oscillator. Proceeding of NOLTA2000, (pp.645–
648).
Itoh, K., & Shimizu, T. (2002). The virtual attractor in mutually
coupled networks. Journal of the Korean Physical Society, 40(6),
1018–1022.
Itoh, K., Miwa, H., Nukariya, Y., Zecca, M., Takanobu, H., Dario, P., &
Takanishi, A. (in press). New memory model for humanoid robots—
introduction of co-associative memory using mutually coupled chaotic
neural networks. Proceedings of the IJCNN2005.
Miwa, H., Okuchi, T., Itoh, K., Takanobu, H., & Takanishi, A. (2003). A
new mental model for humanoid robots for human friendly
communication—introduction of learning system, mood vector and
second order equations of emotion. Proccedings of the ICRA2003,
(pp.3588–3593).
Miwa, H., Itoh, K., Ito, D., Takanobu, H., & Takanishi, A. (2003).
Introduction of the need model for humanoid robots to generate active
behavior. Proceedings of the IROS2003, (pp.1400–1406).
Miwa, H., Itoh, K., Matsumoto, M., Zecca, M., Takanobu, H., Roccella, S.,
Carrozza, C.M., Dario, P., & Takanishi, A. (2004). Effective emotional
expressions with emotion expression humanoid robot WE-4RII.
Proceedings of the IROS2004, (pp.2203–2208).
Roccella, S., Carrozza, C.M., Cappiello, G., Dario, P., Cabibhan, J., Zecca,
M., Miwa, H., Itoh, K., Matsumoto, M., & Takanishi, A. (2004).
Design, fabrication and preliminary results of a novel anthropomorphic
hand for humanoid robotics: RCH-1. Proceedings of the IROS2004,
(pp.266–271).
Sawada, T., Takagi, T., & Fujita, M. (2004). Behavior selection and motion
modulation in emotionally grounded architecture for QRIO SDR-4X II.
Proceedings of the IROS2004, (pp.2514–2519).
Shimizu, T. (1998). Chaotic brownian network. Physica A, 256,
163–177.
Takano, Y. (1995). Ninchi Shinrigaku 2 Kioku (in Japanese). Tokyo
Daigaku Shuppan Kai. (pp. 11-13, 240-241).