Behavior Generation of Humanoid Robots Depending on Mood

Transcription

1 Behavor Generaton of Humanod Robots Dependng on Mood Kazuko ITOH a,1, Hroyasu MIW b,c,d, Yuko NUKRIY e, Massmlano ZECC f,g, Hdeak TKNOBU d,h, Stefano ROCCELL f, Mara Chara CRROZZ f, Paolo DRIO f, and tsuo TKNISHI a,c,d,g a Department of Mechancal Engneerng, Waseda Unversty, Tokyo, Japan b Dgtal Human Research Center, Natonal Insttute of dvanced Industral Scence and Technology (IST), Tokyo, Japan c Insttute for Bomedcal Engneerng, SMeW, Waseda Unversty, Tokyo, Japan d Humanod Robotcs Insttute (HRI), Waseda Unversty, Tokyo, Japan e Graduate School of Scence and Engneerng, Waseda Unversty, Tokyo, Japan f RTS Lab, Scuola Superore Sant nna, Pontedera (PI), Italy g ROBOCS, Waseda Unversty, Tokyo, Japan h Department of Mechancal Systems Engneerng, Kogakun Unversty, Tokyo, Japan bstract. Personal robots, whch are expected to become popular n the future, are requred to be actve n jont work and communty lfe wth human. Therefore, the objectve of ths study s the development of new mechansms and functons for a humanod robot to express emotons and to communcate naturally wth human. We developed both the mental model from psychologcal pont of vew and the Emoton Expresson Humanod Robot WE-4RII (Waseda Eye No.4 Refned II) from engneered pont of vew. In ths paper, a co-assocatve memory model usng mutually coupled chaotc neural networks was proposed and mplemented n WE-4RII as ts mental model. We confrmed that the robot could generate the behavor dependng on ts mood n response to a stmulus. Keywords. Humanod Robot, Neural Network, Mental Model, Memory Introducton Industral robots have varous functons, such as assembly and conveyance. However, an operator has to defne the robot s behavor wth very complex processes or methods. On the contrary, personal robots, whch are expected to become popular n the future, have to be actve n jont work and communty lfe wth human. daptaton to partners or the envronment and communcaton wth partners are necessary for them. Therefore, the objectve of ths study s the development of new mechansms and functons for the natural blateral nteracton between a robot and a human, e.g. expressng the emotons and personalty, and generatng actve behavors. 1 Correspondng uthor: Research ssocate at Waseda Unversty; ddress: #59-308, Okubo, Shnjuku-ku, Tokyo, Japan; Tel: ; Fax: ; URL: E-mal: toh@suou.waseda.jp (K. Itoh), takans@waseda.jp (. Takansh).

2 Varous communcaton robots are researched n robotcs. Brooks at MIT has developed an expressve robotc creature that expresses facal expressons usng ts eyes, eyelds, eyebrows and mouth. It communcates wth human by usng vsual nformaton from CCD cameras [1]. Sony Corporaton has developed the entertanment humanod QRIO, whch s 50 [cm] tall and has 50-DOFs. It can autonomously walk by usng CCD cameras nformaton on the head, and t can control ts behavor wth the homeostass regulaton mechansm [2][3]. We have produced emotonal expressons and actve behavor wth the Emoton Expresson Humanod Robot WE-4 (Waseda Eye No.4) seres, whch have the face, neck, lungs, wast, 9-DOFs emoton expresson humanods arms and humanod robot hands RCH-1 (Robo Casa Hand No.1) [4]. In addton, we have developed the mental model for humanod robots n order to realze the communcaton wth humans. mental space wth three ndependent parameters, mood, second order equatons of emoton, robot personalty, need model [5], conscousness model and behavor model have been ntroduced nto the mental model. On the other hand, many authors research new models for neural networks. In the Hopfeld model, whch s the most typcal neural network, t s shown that the energy functon decreases monotoncally, and one energy mnmum corresponds to one memory [6][7]. However, f the system falls nto a spurous mnmum, t cannot then escape from t. Therefore, several methods have been proposed to solve ths problem by usng chaos. Shmzu has proposed a chaotc neural network model, whch conssts of N Brownan partcles [8]. It was found that the network retreves all stored patterns and reversed patterns n the assocatve memory problem. We have been researchng a system of multple harmonc oscllators nteractng va chaotc force as a model of neural network [9]. We have nvestgated a system of mutually coupled neural networks, n whch each neuron s connected only wth the correspondent neuron n the coupled network. Storng the dfferent pattern n two networks, we have found that each network retreves not only the pattern stored n t but also the pattern stored n the coupled network. In the prevous mental model, the robot has showed just one behavor n response to a stmulus, though human behavor n response to a stmulus depends on the mood at the tme. Therefore, we proposed a co-assocatve memory model usng mutually coupled chaotc neural networks for generatng the behavor to a stmulus dependng on the mood. We mplemented new mental model wth memory model n the Emotonal Expresson Humanod Robot WE-4RII (Waseda Eyes No.4 Refned II). In ths paper, we descrbe n detal the new memory model. 1. Prevous Mental Model We have developed the mental model wth the mental space wth three ndependent parameters, mood, second order equatons of emoton, robot personalty, need model, conscousness model and behavor model for a humanod robot to nteract blaterally wth a human. Frst, we have defned the mental space consstng of the pleasantness, actvaton and certanty axes shown n Fgure 1. The Emoton Vector E has been defned n the mental space as the robot s mental state as: E = E, E, E ), (1) ( p a c

3 certan Emoton Vector E Pleasantness unpleasant asleep Certanty ctvaton uncertan arousal pleasant Trajectory of E Certanty Surprse Happness nger Pleasantness Sleep ctvaton Sleep Neutral Dsgust Sadness Neutral Fear Fgure 1. 3D Mental Space and Emoton Vector E. Fgure 2. Mappng of Seven Dfferent Emotons. where E p s the pleasantness component of emoton, E a s the actvaton component of emoton, and E c s the certanty component of emoton. We have mapped out seven dfferent emotons n the 3D mental space as shown n Fgure 2. The robot expresses the emoton correspondng to the regon traversed by the Emoton Vector E. We have consdered that the transton of a human mental state s expressed by smlar equatons to the equaton of moton. Therefore, we have expanded the equatons of emoton nto the second order dfferental equaton as shown n Eq.(2). lso three Emotonal Coeffcent Matrxes, the Emotonal Inerta, Emotonal Vscosty and Emotonal Elastcty Matrxes have been ntroduced. M E & + ΓE & + KE = F E, (2) M : Emotonal Inerta Matrx Γ : Emotonal Vs costy Matrx K : Emotonal Elastcty Matrx F E : Emotonal pprasal where the Emotonal pprasal F E stands for the total effects of nternal and external stmul on the mental state. The robot expresses dfferent reactons to a stmulus by changng the Emotonal Coeffcent Matrxes. The mental state s affected not only by emoton, but also by mood. Therefore, we have defned the Mood Vector M, consstng of a pleasantness component and actvaton component: ( M, M,0 p a ) = E pdt 2 ( 1 M ) M& + M = 0 M =, (3) M p, (4) M &, (5) a + a a a where M p s the pleasantness component of the mood and M a s the actvaton component of the mood, respectvely. Snce we have consdered the current mental state to nfluence the pleasantness of mood, M p has been defned as the ntegral of the pleasantness component of the emoton (Eq.(4)). On the other hand, snce the actvaton

4 component of Mood Vector s smlar to the human bologcal rhythm such as the nternal clock, the Van del Pol equaton has been appled to defne M a (Eq.(5)). In addton, we developed the need model, whch conssts of the ppette, the Need for Securty, and the Need for Exploraton, n order to generate actvely a behavor for blateral nteracton wth a human [5]. The need matrces N t at tme t and N t+ t at t+ t are descrbed by the Equaton of Need: N = N + P N t+ t t N, (6) P N : Personalty Matrx :Need N Small dfferences between two need states where N s determned by the stmul from the nternal and external envronment, P N s a 3*3 matrx and stands of the personalty for the need. In psychology, each need s ndependent, so P N s a dagonal matrx. Especally, the appette depends on the total spent energy (descrbed as the sum of the basal metabolsm energy and output energy): N = f ( ) = BM + E, (7) where means the varaton n the energy spent by the robot, BM s the basal metabolsm energy, and E s the output energy. However, the robot wth the prevous mental model has been able to determne just a sngle knd of recognton n response to a stmulus and has showed behavor correspondng to the recognton. On the contrary, humans recognze a stmulus and generate a behavor dependng on ther mood at the tme. Therefore, we proposed a new memory model for recognton (memory retreval) dependng on mood n response to a stmulus, usng mutually coupled chaotc neural networks. 2. Human Memory Human memory s related to ther mood by mood state-dependency and mood congruency [10]. Humans store Memory when n a certan mood. They can easly retreve Memory f ther mood becomes the same mood agan. Ths s known as the mood state-dependency. On the other hand, a certan mood helps retrevng a memory correspondng to that mood (mood congruency). Bascally, humans tend to retreve pleasant memores f they are pleasant and conversely, unpleasant memores f they are unpleasant. Moreover, human performance s related to ther actvaton level [11]. If an actvaton level becomes too hgh or too low (.e. superexctaton or blearness), actve human performance becomes mpossble. The best human performance comes at a medum actvaton level. In ths paper, we proposed a co-assocatve memory model for generatng varous behavors to a stmulus, whereby the robot retreves a pleasant memory f t s pleasant and an unpleasant memory f t s unpleasant. In addton, we controlled the tme for retrevng a memory by the actvaton component of the robot s emoton, n order to realze the connecton between performance and actvaton level.

5 Z B, Z,B B B W, j W, j Network Network B Fgure 3. Couplng of Two Neural Networks. 3. Co-assocatve Memory Model 3.1. Mutually Coupled Chaotc Neural Networks We proposed a system of mutually coupled chaotc neural network, whch conssts of many harmonc oscllators (neurons), as a co-assocatve memory model. The nternal state of each neuron s represented as the poston of the harmonc oscllator drven by the chaotc force. If the nternal state of neuron (=1,2,,N) n Network (= or B) at tme t=nτ s denoted by x (t), the tme evoluton of the neuron s provded by Eq.(8). f(t) s the nput from the surroundng neurons and tself, whose ampltude changes chaotcally at tme nterval τ: 2 ( t) + kx& ( t) + ω x ( t) = f ( t) & x, (8) f () t h 0 K = for n τ t < ( n+ 1) τ ( n= 0,1,2, L), (9) τ where, k, ω 0 and K are the dampng constant, the egen-frequency and the magntude of f(t), respectvely. The factor 1/ τ s needed to obtan a fnte dffuson constant n the small τ lmt [12]. h(n) denotes the nteracton among neurons. We coupled two networks by connectng neuron n Network (B) wth neuron n Network B() as shown n Fgure 3: N = j= 1 N = j= 1, B B B B B B, h W, j yj + Z y, h W, jyj + Z y, (10) W, j = 1 N P s = 1 ξ ξ, (11) s s j where W,j s the couplng constant from neuron j to neuron and Z,B (Z B, ) s that from neuron n Network B() to neuron n Network (B). ξ denotes the stored pattern vector and ξ takes +1 or -1. To store patterns n the network, we used the Hebb rule to determne W,j as shown n Eq.(11). The self-couplng constant s equal to 1; W, =1. In Eq.(10), y(n) denotes the output of the neuron and represents the n th terate of a map. s an example of the map, we employed the Logstc map, whose bfurcaton parameter r(n) s modulated by the nternal state of the neuron:

6 y ( n+ 1) = r (0.5 y )(0.5 + y ) 0.5 r ( ) y (12) 2 = 4 b + b cos β x (0 b 4) (13) where b and β are control parameters. y(n) changes chaotcally or perodcally accordng to the bfurcaton parameter. Snce the bfurcaton parameter r(n) s modulated by the nternal state of the neuron as shown n Eq.(13), the chaos s controlled by the neuron tself. new type of feed-back mechansm s ncluded n ths model. The nternal state x(n) determnes the bfurcaton parameter r(n), whch, n turn, determnes the dynamcs of the chaotc output y(n). The chaotc output then affects the dynamcs of the neuron. Thus, the dynamcs of the chaos s changed by the system tself. In partcular, f we put the perod of the harmonc oscllator 2.0, the neural network can retreve the orgnal and reverse patterns alternately, meanng ths neural network can perform very well. 3.2 Co-assocatve Memory Model In ths paper, we defned pple as the pleasant memory for Red and Tomato as the unpleasant memory for Red. Of course, we can select other memores accordng to the robot personalty. In the case of Z B, =0.0, f Z,B becomes large the number of tmes n whch Network retreves pple ncreases and the number of tmes n whch Network retreves Tomato decreases. Therefore, we ntroduced the mood state-dependency and the mood congruency by modulatng Z,B by the pleasantness component of the mood M p. The robot retreves pple for pleasant mood and Tomato for unpleasant mood when t recognzes Red. However, human retreve the favorte food even f they are unpleasant by feelng hungry. We solved ths problem by defnng Z,B as the sum of the pleasantness component of mood M p and the ppette N :, Z = M + N. (14) B p,, Z B = 4.0 B = 0.0 Sleep Surprse Z n tmes Ea Fgure 4. Relaton between Performance and ctvaton Level.

7 In addton, we controlled the tme for retrevng a memory by the actvaton component of the robot emoton E a, n order to realze the connecton between performance and actvaton level, as shown n Fgure 4. The robot wth sutable actvaton level can retreve the memory correspondng to ts mood soon. However, the robot needs tme to retreve the memory f E a becomes too hgh or too low. Moreover, we consdered that a wrong memory s sometmes retreved n case of too hgh or too low actvaton level. For Z B, =-4.0, Network retreves Tomato n almost smaller half regon of Z,B but sometmes retreved pple. Reversely, Network retreves pple n almost larger half regon of Z,B but sometmes retreves Tomato. Therefore, we defned that Z B, s equal to 0.0 for sutable actvaton levels and to -4.0 for too hgh or too low actvaton levels. 4. Expermental Results We evaluated a new co-assocatve memory model through mplementaton n the Emoton Expresson Humanod Robot WE-4RII as shown n Fgure 5. We set the parameter values as K=14.0, k=0.1, ω 0 =31.4, β=0.05, τ=0.1 and b=1.2, and the ntal values x& as x (0)=25.0, (0 (0)=0.0 and y (0)=0.2. Fgure 6 shows the tme evaluaton of the emoton, the mood and the retreved memory of WE-4RII. t frst, we showed the red ball to the robot and the robot became unpleasant by beng ht. It retreved the unpleasant memory Tomato and shows the Dsgust emotonal expresson. Next, the robot felt hungry snce t moved consderably. Due to hunger, t retreved the pleasant memory pple n spte of unpleasant mood. fter that, the robot can take the apple usng ts hand, and generate the behavor such as eatng. If ts hunger s satsfed, the robot becomes happy. Moreover, we confrmed that t took only about 2[s] for the robot from lookng at the red ball untl retrevng a memory. On the other hand, the robot needed about 9[s] for memory retreval, snce the actvaton level became very hgh by beng ht many tmes. t ths tme, sometmes the robot could not retreve the memory, dependng on mood. Emoton and Mood Vectors Ht Retreve Tomato Retreve pple Ep Mp N tme s Fgure 5. WE-4RII. Fgure 6. Expermental Results of Mood State-Dependency and Mood Congruency

8 5. Conclusons and Future Work In ths paper, we nvestgated the co-assocatve memory model ncludng the mood state-dependency, the mood congruency and the connecton between performance and actvaton level n order to generate a behavor dependng on mood n response to a stmulus. Ths memory model was realzed usng mutually coupled chaotc neural networks, whch consst of many harmonc oscllators (neurons) nteractng va the chaotc force. We confrmed that the robot can retreve the memory dependng on mood n response to a stmulus and show the behavor correspondng to the memory by mplementng ths memory model n the Emoton Expresson Humanod Robot WE-4RII. In the future, we wll ncrease a number of retrevable memores. Furthermore, we wll study a method for storng memores dependng on mood. cknowledgment Part of ths research was conducted at the Humanod Robotcs Insttute (HRI), Waseda Unversty. The authors would lke to express ther thanks to Okno Industres LTD, OSD ELECTRIC CO. LTD, SHRP CORPORTION, Sony Corporaton, Tomy Company LTD and ZMP INC. for ther fnancal support to HRI. The authors would lke to thank Italan Mnstry of Foregn ffars, General Drectorate for Cultural Promoton and Cooperaton, for ts support to the establshment of the ROBOCS laboratory and for the realzaton of the two artfcal hands. In addton, ths study was supported n part by the Mnstry of Educaton, Scence, Sports and Culture, Grant-n-d for Young Scentsts (B), , lso, ths research was supported by a Grant-n-d for the WBOT-HOUSE Project by Gfu Prefecture. Fnally, the authors would lke to express thanks to RTS Lab, NTT Docomo, SoldWorks Corp., dvanced Research Insttute for Scence and Engneerng of Waseda Unversty, Prof. Hrosh Kmura for ther supports to our research. References [1] C. Breazeal and B. Scassellat: How to buld robots that make frends and nfluence people, Proceedngs of the IROS1999, pp , [2] M. Fujta, Y. Kurok, et al.: utonomous Behavor Control rchtecture of Entertanment Humanod Robot SDR-4X, Proceedngs of the IROS2003, pp , [3] T. Ishda, Y. Kurok, et al.: Mechancal System of a Small Bped Entertanment Robot, Proceedngs of the IROS2003, pp , [4] H. Mwa, K. Itoh, et al.: Effectve Emotonal Expressons wth Emoton Expresson Humanod Robot WE-4RII, Proceedngs of the IROS2004, pp , [5] H. Mwa, K. Itoh, et al.: Introducton of the Need Model for Humanod Robots to Generate ctve Behavor, Proceedngs of the IROS2003, pp , [6] J. J. Hopfeld: Neurons wth graded response have collectve computatonal propertes lke those of two-state neurons, Proceedngs of Natl. cad. Sc. US 81, pp , [7] J. J. Hopfeld and D. W. Tank: Neural Computaton of Decsons n Optmzaton Problems, Bol. Cybern 52, pp , [8] T. Shmzu: Chaotc Brownan Network, Physca 256, pp , [9] K. Itoh and T. Shmzu: The Vrtual ttractor n Mutually Coupled Networks, Journal of the Korean Physcal Socety, Vol.40, No.6, pp , [10] Y. Takano: Nnch Shnrgaku 2 Koku (n Japanese), Tokyo Dagaku Shuppan Ka, pp.11-13, , [11] D. O. Hebb: Koudougaku Nyumon (n Japanese), Knokunya Shoten, pp , [12] T. Shmzu: Relaxaton and bfurcaton n brownan moton drven by a chaotc force, Physca 164, pp , 1990.