TAME: Time-varying Affective Response for Humanoid Robots

Transcription

1 TAME: Tme-varyng Affectve Response for Humanod Robots Lla Moshkna, Sunghyun Park, Ronald C. Arkn, Jamee K. Lee, HyunRyong Jung Lla Moshkna, Sunghyun Park, Ronald C. Arkn Georga Tech Moble Robot Laboratory, Atlanta, GA, USA E-mal: {lla, ts2883, Jamee K. Lee, HyunRyong Jung Samsung Advanced Insttute of Technology, Kheung, South Korea E-mal: {jamee.lee, Abstract. Ths paper descrbes the desgn of a complex tme-varyng affectve software archtecture. It s an expanson of the TAME archtecture (Trats, Atttudes, Moods, and Emotons) as appled to humanod robotcs. In partcular t s ntended to promote effectve human-robot nteracton by conveyng the robot s affectve state to the user n an easy-to-nterpret manner. Keywords: Humanods, emotons, affectve phenomena, robot archtectures Ths research s funded under a grant from Samsung Electroncs. Portons of ths paper are modfed from content appearng n Moshkna et al.[1-3]. 1

2 1 Introducton Wth every advance n robotcs: vsual percepton workng more precsely, communcatons becomng more relable, moblty mprovng, we are gettng ever closer to robots becomng a part of our everyday lves, movng from factores nto our homes and workplace. As robots gan more autonomy and start nteractng wth people not specally traned n robotcs, t becomes ncreasngly mportant for them to be able to communcate n a way easly understandable to nurses and patents n a hosptal settng, or elderly n ther homes, or vstors at museums and exhbtons. The development of humanod robots s certanly a step n the rght drecton, as they, due to ther embodment, wll be more lkely to be accepted and used n everyday and long-term stuatons. However, human-lke appearance alone would not guarantee smooth, natural and enjoyable nteracton. In partcular, humans employ one form of affect or another n almost every step of ther lves, and are qute capable of decpherng others affectve nonverbal behavor [4]. Gven that people also tend to treat computers as socal actors [5] and thus expect them, even f subconscously, to behave n a socally approprate manner, we beleve that endowng humanods wth affectve capabltes would be advantageous for the successful human-robot nteracton. Our research group has had extensve pror experence mplementng motvatonal and affectve phenomena n robotc systems. Some of our earler research ncluded: addng motvatonal behavors to a robotc Sowbug and a prayng mants; an mplementaton of an emotonal attachment mechansm n smulaton and on real robots; an ethologcally-nspred archtecture for a robotc dog Abo wth ncluded a number of drves and emotons; and developng emotonallygrounded symbols wthn EGO archtecture on a humanod robot Sony QRIO (see Arkn [6] for a more detaled summary). Based on our pror experence, we are developng an applcaton of cogntve and psychologcal models of human Trats, Atttudes, Moods, and Emotons (TAME) for use n humanod robots. These affectve states are embedded nto an ntegrated archtecture and desgned to nfluence the percepton of a user regardng the robot s nternal state and the humanrobot relatonshp tself. Recent work by Arkn et al n non-verbal communcaton [7] and emotonal state for the AIBO [8] addressed powerful yet less complex means for accomplshng these tasks. Introducng tme-varyng affectve states that range over multple tme scales spannng from an agent s lfetme to mere seconds wth orentaton towards specfc objects or the world n general provdes the power to generate heretofore unobtaned rchness of affectve expresson. Ths paper descrbes the cogntve and psychologcal underpnnngs of ths work n the context of humanod robots and affectve software archtecture, and presents the drectons beng taken n ths ongong project to mplement and test t on a small humanod robot. 2 Related Work Although most research on humanods focuses on the physcal aspects (e.g., perfectng walkng gats, sensors or appearance), there are some who also explore affectve nteracton. For example, humanod Waseda Eye No. 4 Refned [9] combnes emotons, moods, and personalty. The overall goal of the system s to acheve smooth and effectve communcaton for a humanod robot. Although many elements of ths system are not psychologcally or bologcally founded, t provdes a few nterestng mechansms, such as modelng personalty s nfluence on emoton va a varety of coeffcent matrces and usng nternal-clock actvaton component n moods. Another example of a robotc system that combnes multple affectve aspects s the Roboceptonst [10]. In ths system, emotons and moods are generated as a response to the robot s nteracton wth passerbys and events n an magnary story lne; atttudes are treated as long-term moods assocated wth a certan person or thng. Although the Roboceptonst s not a humanod, t s equpped wth a flat-screen montor dsplayng an anmated character face through whch affect s expressed. Fukuda et al. [11] also nclude the notons of emotons and moods n ther Character Robot Face; emotons are represented as semantc networks, and the combnaton of currently actve emotons s deemed as mood. Two other humanod robotc head robots, Ksmet [12] and MEXI [13] have emoton and drve systems. Ksmet s modeled after an nfant, and s capable of protosocal responses, ncludng emotonal expressons, whch are based on ts affectve state. In MEXI, the Emoton Engne s composed of a set of basc emotons (postve that t strves to acheve and negatve t tres to avod) and homeostatc drves. In ERWIN, yet another socally nteractve robot head, fve basc emotons are generated through modulaton of hormonal-lke parameters [14]. 2

3 Fnally, LEGO robot Feelx [15] s capable of expressng a subset of basc emotons elcted through tactle stmulaton. Other work nto humanod affect prmarly focuses on producng recognzable emotonal facal and bodly expressons, rather than affect generaton. In partcular, Nakagawa et al. [16] propose a method to control affectve nuances by mappng dmensons of valence and arousal onto velocty and extensveness of moton and body posture; ths method was tested to produce subtle affectve expressons on two humanod robots, Robove-mn R2 (an upper body humanod) and Robove-M (a bped). As another example, Hanson Robotcs androd head Ensten [17] s capable of learnng and producng a large number of realstc facal expressons based on Ekman s Facal Acton Codng System, FACS [18]. Fnally, the research nto robotc affect on non-humanod platforms ncludes: producng emotonal expressons based on the crcumplex model of affect on a huggable robot Probo by Salden et al. [19]; dsplay of affect on expressve robotc head EDDIE [20], based on the crcumplex model and Ekman s FACS; expresson of Extraverson and Introverson on robotc dog AIBO [21]; and a number of other related endeavors. 3 Cogntve Bass of TAME In comparson wth the aforementoned systems, the TAME framework encompasses a wder range of affectve phenomena, and provdes psychologcal groundng for each. It has been ntally tested on the entertanment robot dog AIBO [3], and ts applcaton to humanod robots s farly straghtforward n prncple. Moreover, humanod platforms provde certan benefcal affordances for the use of the framework. The frst one s ther expressve potental, exhbted not only n facal and bodly expressons (e.g., a smle, a shoulder shrug, a handshake), but also n a varety of tasks they could perform for whch human-lke personaltes are readly applcable. Another affordance les n the emphass on long-term nteracton, and the potental for humanods to act as partners or companons, rather than bystanders; as some of the components of the framework are subtle n expresson and would be notced best over multple nteractons wth the same person. The TAME framework tself takes nspraton from a large number of theores and fndngs from personalty, emoton, mood and atttude areas of psychology [22-24, 26-28, 30-33], whch are specfcally adapted to the needs of enhancng human-robot nteracton. As multple defntons of affectve phenomena exst, here we present those we use for our work: 1. Affect s an emboded reacton of pleasure or dspleasure sgnfyng the goodness or badness of somethng; 2. Personalty trats dentfy the consstent, coherent patterns of behavor and affect that characterze ndvduals; although not affectve per se, they provde a profound nfluence on generaton and applcaton of affectve phenomena; 3. Affectve atttudes are general and endurng, postve or negatve, feelngs about an object, a person or a ssue; 4. Moods are a low-actvaton, slowly-varyng dffuse affectve state; 5. Emotons are a hgh-actvaton short-term affectve state and provde a fast, flexble response to envronmental contngences n the form of relevant stmul. 3.1 Overvew The Affectve Module, the core of TAME, s subdvded nto Dspostons and Affectve State. Dspostons nclude personalty trats and affectve atttudes, and represent a propensty to behave n a certan way. They are more or less persstent, long-lastng, and ether slowly changng (atttudes) or permanent (trats) throughout a robot s lfe. Affectve state conssts of emotons and moods, that are more fleetng and transent affects, and manfest as ether hgh-ntensty, shortduraton peaks (emotons) or slow smooth undulatons (moods). Table 1 summarzes the dfferences n duraton and temporal changes of these four components. Another drecton along whch these components dffer s object specfcty: emotons and atttudes appear and change n response to partcular stmul (such as fear n the presence of an attacker or dslke towards an unfrendly person), whereas trats and moods are dffuse and not object-specfc they manfest regardless of the presence or absence of objects. Each component can be postoned n the two-dmensonal space defned by duraton and specfcty [22-24] (Fgure 1). Trats and emotons are at the opposte ends of ths spectrum: trats are tme-nvarant and global, whereas emotons are short-term, dynamcally changng and focused. Although all the components can be categorzed dfferently and each can have a dstnct functon and purpose, these 3

4 phenomena cannot be regarded as ndependent, as they strongly nfluence each other and nterweave to create a greater lluson of lfe. Trats Atttudes Moods Emotons Duraton Lfe-long A few days to a few years A few hours to a few weeks A few seconds to a few mnutes Change n Tme Tmenvarant Persstent across tme; change slowly wth the number of tmes an object of atttude s encountered. Change cyclcally as a varable of underlyng envronmental and nternal nfluences; any drastc changes are smoothed across prevous mood states Intensty changes n short-term peaks as elctng stmul appear, dsappear, and change dstance; habtuaton effects descrbe decay of emoton even n the presence of stmul. Table 1. Summary of Tme-varyng Aspects of TAME Components Moods Global/nonspecfc Trats Instantaneous Emotons Fgure 1: Relatve Poston of Types of Affect Atttude Focused/specfc Lfe-tme The Affectve Module fts wthn the behavor-based robotc control paradgm [25] by frst processng relevant perceptual nput (be t color and dstance to certan emoton-elctng objects or level of lght affectng moods) and then drectly nfluencng behavoral parameters of affected lowlevel behavors and/or behavor coordnaton gans as they are composed nto behavoral assemblages (Fgure 2). Fgure 2: Conceptual Vew of TAME 3.2 Psychologcal and Mathematcal Foundatons Personalty Trats Personalty defnes an organsm s recurrent patterns of behavor and emotonalty. The Fve Factor Model of personalty trats [26] was chosen as the model for ths component for ts unversalty: t s consstent over tme, cultures, ages, and even applcable to non-human anmals. 4

5 To a large extent, trats are nherted or mprnted by early experence, therefore n TAME we treat them as nvarable (the excepton s that an operator can specfy a dfferent personalty confguraton dependng on a task at hand, but t would reman the same durng that task). The taxonomy has fve broad dmensons, each of whch s further subdvded nto facets; therefore a robot s personalty can be as smple or as complex as desred. Trats provde a two-fold advantage for humanod robots: frst, they serve a predctve purpose, allowng humans to understand and nfer the robot s behavor better; second, they allow adaptaton to dfferent tasks and envronments, where certan trat confguratons are better suted for one or another task or envronment. The fve global dmensons are Openness, Conscentousness, Extraverson, Agreeableness and Neurotcsm. Openness refers to actve magnaton, preference for varety and curosty; Conscentousness descrbes socally desred mpulse control that facltates task- and goal-drected behavor; Extraverson refers to lkng people and preferrng large groups and gatherngs, and also affect postve emotonalty; Agreeableness s a dmenson of nterpersonal tendences, and refers to beng sympathetc to others, cooperatve and eager to help; fnally, Neurotcsm s the general tendency to experence negatve affect, such as fear, sadness, anxety, etc and be more senstve to sgns of danger. Each of them has ts own effect on robot behavor: for example, n a humanod, extraverson could be expressed by keepng a closer dstance to the human, frequent smles, more gestures, etc. Ths trat would be approprate for tasks requrng engagement and entertanment from a robot, e.g., a museum gude or a play partner for kds. Another example of a useful trat s Neurotcsm: a humanod can, through correspondngly neurotc behavor, suggest to an accompanyng human to pay more attenton to potentally dangerous surroundngs. The trats are modeled as vectors of ntensty, where ntensty refers to the extent to whch a trat s represented. In the robot, these ntenstes: are defned a pror by a human; don t change throughout the robot s lfe (ths could be a sngle run, an nteracton wth a person, or the robot s entre physcal lfe-span); and are not nfluenced by any other affectve phenomena. We provde a functonal mappng from the trat space onto behavoral parameter space as a 2 nd degree polynomal, where 3 pars of correspondng data ponts are mnmum trat/parameter, maxmum, and default/average (the values are taken from the normally dstrbuted human psychologcal data [27]). Trats can have a drect or an nverse nfluence on partcular behavors and ths relatonshp s defned n a matrx beforehand. Fgure 3 presents a 2 nd degree polynomal mappng from the trat of Neurotcsm onto two behavoral parameters: drectly to obstacle avodance gan (a degree to whch an agent should avod obstacles) and nversely to wander gan (related to exploraton). Hghest parameter value at the lowest trat value Wander Gan Hghest parameter value at the hghest trat value Default = 1 Obstacle Avodance Gan Mean=79.1 Fgure 3: Comparson of Drect and Inverse Influences of Trats on Behavor In cases where multple trats affect the same behavor (e.g., Neurotcsm may push the robot away from the obstacles whle Conscentousness could make t go closer for a faster route), frst a trat/parameter mappng s calculated, accordng to the chosen functon f j (p j ), where trat nfluences behavor j, a polynomal n ths case. Then, the results are averaged across all nfluencng personalty trats to produce the fnal parameter value used thereafter: 5

6 B j N 1 1 pb j N 1 f ( p ) j (1) where B j s a partcular behavoral parameter, f j (p ) s the functon that maps personalty trat p to B j, N s the total number of trats, and pb s personalty/behavor dependency matrx; f there s no nfluence, the result of f j = 0. As the trats are relatvely tme-nvarant, the obtaned trat-based behavor parameters serve as default behavors for the robot Emotons From an evolutonary pont of vew, emotons provde a fast, flexble, adaptve response to envronmental contngences. They appear as short-term, hgh-ntensty peaks n response to relevant stmul (we don t usually lve n a constant flux of emotons), and serve a number of functons, of whch most applcable for humanods are communcatve, expressve and afflatve, e.g., fear communcates danger and a request for help, whle joy n response to a brght smle helps forge trust and camaradere. The prmary, reactve emotons of fear, anger, dsgust, sadness, joy and nterest were chosen, n part because these basc emotons have unversal, well-defned facal expressons [28], are straghtforwardly elcted, and would be expected, perhaps subconscously, on a humanod s face, as appearance does affect expectatons. Each emoton s ntensty s stored n the emoton ntensty matrx E [ E ], where 0 E g, the value E represents the ntensty of a currently actve emoton, 0 sgnfes the absence of emoton, and g s the upper bound for emoton. From an emoton generaton pont of vew, Pcard [29] suggests a number of propertes are desrable n an affectve system: 1. Actvaton. Refers to certan stmulus strength below whch the emoton s not actvated. 2. Saturaton. Refers to the upper bound of an emoton, after whch, regardless of the ncreasng stmulus strength, the emoton doesn t rse any more. 3. Response decay. States that emotons decay naturally over tme unless they are restmulated. 4. Lnearty. Emotons can be modeled as lnear under certan condtons; due to the propertes of actvaton and saturaton, the emotons wll approxmate lnearty only for certan stmulus strength range, and wll approach a sgmod at ts edges. Takng these propertes nto consderaton, the resultng functon for emoton generaton (based on stmulus strength) resembles a sgmod, n whch the left sde corresponds to actvaton, the rght sde corresponds to saturaton (ampltude), and the mddle models the actual response. The elctng stmulus strength for each emoton s calculated by takng nto account a number of object propertes, such as ts physcal propertes (sze, shape, etc.), ts poston (dstance to the object, ts velocty, etc.), and any exstng atttude of the agent towards the elctng stmulus. Then, the base emoton level s calculated as follows: E, base e ( s g a )/ d e ( s e b )/ d,, ( a )/ d f f a s s ( a ( a b ) / 2 b ) / 2 where b 2 d ln( g e ( a )/ d ) / 2 a (2) where E,base s the base emoton value for emoton, s s the strength of stmulus elctng emoton, a s the varable that controls the actvaton pont for emoton, d s the varable that controls the maxmum slope for emoton, g s the ampltude of emoton, and b s the break-pont, at whch the emoton reverses ts rate of growth. Fgure 4 presents the resultng curve graphcally. 6

7 Ampltude (g) Pont of maxmum slope (controlled by d) Actvaton Pont (controlled by a) Fgure 4: Emoton Generaton Based on Stmulus Strength Emotons are also hghly dependent on trats and moods: personalty may nfluence the threshold of elctng stmulus (actvaton pont), peak (ampltude) and rse tme to peak (affectng the slope of the generaton curve) [30]; and moods can vary the threshold of experencng a specfc emoton [31]. For example, Extraverson s correlated wth postve emotons, therefore a humanod robot hgh n ths dmenson would dsplay more smles, excted gestures and other expressons of joy. Atttude also has an effect on emoton the object of lke or dslke may serve as a stmulus for emoton generaton. A lnear mappng from trats to ampltude, actvaton pont, and maxmum slope s used to obtan personalty nfluence on emoton generaton. For example, the trat of Extraverson provdes a drect nfluence on the ampltude, actvaton pont, and slope of postve emotons (joy and nterest), therefore a robot wth a hgher level of Extraverson wll have a stronger postve emoton that wll be actvated at weaker stmulus strength and wll rse faster than that of an ntroverted robot. Smlarly, current mood wll nfluence the actvaton pont, where the negatve mood wll make t easer for an agent to experence negatve emotons, and postve mood postve emotons. Fgure 5 presents combned nfluence of trats and mood on emoton generaton. Hgher Mood Value Neutral Mood Value Hgher Trat Value Lower Mood Value Lower Trat Value Fgure 5: Combned Mood and Trat Influences on Emoton Generaton To account for the short-term duraton of emotons and habtuaton to prolonged stmulus, emoton decay s modeled as a slowly decreasng exponental: E, t, decay E t, base e ( t t0 )* d where E,t,decay s the ntensty of emoton at tme t, t o s the tme at whch emoton s actvated (becomes greater than 0), and d s a varable that controls the rate of decay. Ths ensures that a hgh-actvaton emotonal state s not mantaned beyond the ntal epsode, and, provded the stmulus doesn t change, the emoton t nvoked wll eventually dsspate. Fnally, n order to smooth the emoton change n cases of sudden appearance and dsappearance of elctng stmul, a weghted averagng flter can be used: (3) 7

8 E w * E w * E ) /( w w ) (4), t, fltered ( current, t, decay pror, t 1, fltered current pror where E,t,fltered s the fnal ntensty of emoton at tme t after flterng, w current and w pror are weghtng varables controllng the relatve mportance of current and prevous emotonal states. Ths flterng functon wll help to account for short-term lngerng emotons even after the elctng stmulus has dsappeared. Emotons can have a vared mpact on behavor, from a subtle slowng to avod a dsgustful object to a drastc flght n response to extreme fear. Ths effect can be modeled by lnear mappng from emoton strength to relevant behavoral parameters, and Fgure 6 provdes a comparatve vew across tme of stmulus strength (an object appears, comes closer, and then s gone), correspondng emoton actvaton (after response decay and smoothng), and the Object Avodance Gan (whch causes an avodance response to Fear); duraton s plotted along the x axs, and normalzed values for stmulus strength, fear and object avodance gan along the y axs. If the object contnued to be present and unchanged, then Fear would eventually be brought down to 0. Fgure 6: Example of Fear to Object Avodance Gan Mappng In a humanod, dsplay of fear may sgnal mmnent danger to nearby people, and be more persuasve than words alone, n case an evacuaton s requred. Expressons of dsgust, smlarly, may alert a human to the presence of some noxous stmulus, whch, though not necessarly hazardous, may stll be best avoded Moods Unlke emotons, moods represent a global, contnuous affectve state, cyclcally changng and subtle n expresson. Mood can be represented along two dmensons, Postve Affect and Negatve Affect [24], where Negatve Affect refers to the extent to whch an ndvdual s presently upset or dstressed and descrbes level of stress and tenson, and Postve Affect generally refers to one s current level of energy, enthusasm, and pleasure. The level of arousal for both categores can vary from low to hgh; a low postve mood value has a negatve connotaton ( sluggsh, dsnterested ) and refers to nsuffcent level of energy, pleasure and enthusasm, rather than just low. One advantage of provdng a humanod wth expresson of mood would be to let nearby humans know when the system needs attenton, be t a low battery resultng n low energy level, or nsuffcent amount of lght resultng n poor sensor relablty. There are two broad types of mood change: envronmental/external (lght or nose level, external temperature, amount of recent nteracton, etc.) and nternal (e.g., battery level n case of a robot, nternal temperature); and short-term stuatonal varables, ncludng emotonal epsodes. The current base level of mood of a robot s defned as a weghted summaton of varous external and nternal varables. Assumng that the same varables affect both postve and negatve moods, strengths of envronmental and nternal nfluences can be represented n a matrx l l ], where 0 l b, where b s the hardware-dependent upper bound (e.g., lght can only be detected up to a certan level, etc.). The relatve weghts for each varable are stored n the mood generaton matrx mg mg ]. The values n ths matrx are unt converson factors, to convert the varous [ j mood generaton varables (whch may correspond to raw sensor data) to the same unt, and are [ 8

9 found expermentally for each varable. In addton, negatve mg stands for nverse nfluence of the varable on the mood, and postve mg stands for drect nfluence. Accordng to Set Pont theory [24], a certan base level of mood s mantaned at all tmes, and, though events and changes n the envronment cause t to fluctuate, t tends to return to the same level over tme. As moods are contnuous, always present streams of affect, the base mood s contnuously generated based on the current envronmental and nternal nfluences as follows: m base m m postve negatve N 1 mg mg, postve, negatve ( l l, neutral ) (5) where mg s the mood generaton matrx, l s the mood generaton varable strength matrx, l, neutral s the set pont for mood, and N s number of mood generaton varables. Fgure 7 llustrates the effect of an envronmental varable on mood. Fgure 7 presents an example of the nfluence of lght on mood generaton. Addtonally, smlarly valenced emotons can affect the correspondng mood ntenstes addtvely, and change the exstng base mood level n the followng manner: m m postve negatve m m postve, base negatve, base E, k f E E, f E k k k { Interest, Joy} { Fear, Anger, Dsgust, Sadness} (6) where m postve s the emoton-based ntensty of postve mood, m postve ntensty of postve mood, and E s the emoton ntensty matrx. s the emoton-based Fgure 7: Dfferent Effect of Lght on Postve and Negatve Mood As mood s a low-actvaton, slow-varyng affectve state, sudden changes are smoothed out by takng nto consderaton pror mood states. Flterng over a longer perod of tme results n slower and smaller mood changes and helps tone down any drastc spkes due to emotons. Addtonally, n humanods that are desgned for sharng lvng condtons wth humans for a prolonged tme, crcadan varatons n mood may be ntroduced to provde mood congruency wth the human, where user-defned cyclcal daly, weekly and seasonal hgh and low ponts would be supermposed onto the base mood values. Moods are mld by defnton, and would only produce a small, ncremental effect, or a slght bas, on the currently actve behavors. Moods can have a drect or nverse nfluence on a behavoral parameter. A behavor-mood dependency matrx mb mb ] s defned, where mb j [ j { 1,0,1} s defned, where 1 corresponds to nverse nfluence, +1 to drect nfluence, and 0 to absence of mood nfluence on behavor. Postve and negatve moods may nfluence the same behavoral parameters, and ths nfluence s treated as addtve. As moods are updated contnuously, new mood-based values of behavoral gans/parameters replace the exstng tratbased values n the followng manner: 9

10 B N mood B trat K mb j m (7),, j j 1 where B, s the updated behavoral parameter, mb j s the mood-behavor dependency mood matrx value for mood j, m j s the current value of mood j, N s the total number of mood categores (2), and K s a scalng factor to ensure that the moods produce only ncremental effect as opposed to overpowerng any of the parameters. Fgure 8 shows an example of ncremental effects of moods on behavor. Suppose that mood can bas robot s obstacle avodance behavor. For example, f vsblty s poor, t may be advantageous to stay farther away from obstacles to accommodate sensor error, and vce versa, n good vsblty t may be better to concentrate on task performance. Thus, negatve mood can bas the obstacle avodance gan by rasng t, and postve mood by lowerng t. Neurotcsm also affects t by settng the default parameters to be used throughout the lfe-cycle, and the ncremental effect of moods s shown aganst the space of trat-based defaults (plotted n sold blue center lne). Fgure 8: Drect/Inverse Mood Effects on Behavor at Dfferent Neurotcsm Values For human-robot nteracton, expressve manfestaton of mood can alert a person to favorable or unfavorable changes n the envronment or n the robot tself, especally f percepton of these changes s based on sensor nput not avalable through human senses. Consder the followng scenaro. A humanod s gudng a human nspector through a partally secured search-and-rescue ste, when the lghts become dm. Although no mmedate danger s vsble, the robot s negatve mood rses, and t dsplays the sgns of anxety and nervousness; no acton per se s warranted yet, but the nspector, pckng up the cues from the robot, becomes more alert and ready for acton. We are currently usng ths scenaro to formally evaluate, through human subject experments, the effect of robotc mood dsplay on complance wth and persuasveness of a robot s request Affectve Atttudes From a multtude of defntons of atttudes, the followng was adopted as the workng defnton for TAME: a general and endurng postve or negatve feelng about some person, object or ssue [32]. It stresses relatve tme-nvarance ( endurng ), object/stuaton specfcty, and the role of affect/affectve evaluaton n the atttude concept. We propose two methods for atttude formaton: one s more general and does not requre any nput from an nteractng human, and the other s more experental, and requres ntal human nput. In the general method, atttudes are descrbed by valence and ntensty, and are represented as a sngle value A, rangng from to, where 0 sgnfes a neutral (or absence of) atttude, negatve values represent ncreasngly strong negatvely-valenced atttude (rangng from a mld dslke to hatred), and postve values refer to ncreasngly strong postvely-valenced atttude (e.g., from a subtle lke to adoraton). Atttudes are object-specfc, and an ntal atttude for a partcular object (y) would consst of a combnaton of postve or negatve attrbutes of ths object (as a facetous example, a robot may develop a dslke to a man wth a mustache), represented as a matrx o y [ o y ], where o y. Such attrbutes are not lmted to propertes of the object only; for example, an emoton nvoked by the object and any actons taken by the object 10

11 may be consdered attrbutes. The ntal value of the atttude for object y (A y,nt ) s calculated as follows: N A y o (8), nt y 1 where A y,nt s the newly-formed atttude for object y, o y s an attrbute of object y that s nvolved n the atttude formaton, and N s the number of attrbutes for object y. Assumng that an ntal mpresson s the strongest, substantal changes n atttude are farly hard to acheve, therefore any subsequent exposure to the same object would result only n ncremental change. Ths s done by dscountng any addtonal postve or negatve object attrbutes to a certan extent. The updated atttude value for object y for n-th encounter (A y,n ) would then be calculated as follows: A y, n A y, n 1 n ( N 1 o y ) (9) where A y,n-1 s the atttude towards object y at encounter n-1, n s the total number of encounters up to date, s the dscount factor. o s the matrx of attrbutes for object y, and y Fnally, consstent wth the fndng on mood-congruent judgment, postve mood ncreases the value of the atttude ( a ) towards an object y, and negatve mood decreases t as follows: y A A K( m m y, mood y postve negatve ) (10) where A y,mood s mood-enhanced value of agent s atttude towards object y, A y s the orgnal value of agent s atttude towards object y, m postve s the current postve mood value, m negatve s the current negatve mood value, and K s a scalng factor to brng moods and atttudes to the same unts. In the experence-based method, robotc atttudes are based on those held by people commonly nteractng wth robot. In ths method, a Case-Based Reasonng approach [33] s used, where an ntal set of cases s provded by each nteractng human, and the resultng atttudes are expressed through correspondng emotons by the robot. Each case contans a set of object propertes (ndexed by these propertes and a user ID, to dfferentate between dfferent people) and correspondng atttudes for a varety of objects. When a robot encounters a new object, the most smlar case s retreved from the case lbrary and appled. As affectve atttudes are closely related to emotons (n fact, some psychologsts even descrbe love and hate as emotons, albet long-term and persstent), the output of both methods produces stmulus strengths to generate correspondng emotons. Thus, atttudes are not expressed n behavoral changes per se, but rather through the emotons they nvoke. In the course of long-term nteractons wth people, t may be easer for humanods that share atttudes wth ther human companons to establsh rapport and understandng. For example, a chld playng wth a robot nanny or tutor may feel greater affnty towards the robot that acknowledges the chld s lkes and dslkes n toys and games. 3.3 Exploratory Expermental Study In order to explore the ssues of feasblty and potental usefulness of the TAME framework, a prelmnary exploratory study was conducted (please refer to Moshkna et al [3] for a detaled report). For ths between-subject longtudnal human-robot nteracton study a partal mplementaton of Emoton and Trat modules was performed on a robotc dog AIBO. The emotons of Interest, Joy, Anger and Fear were expressed va head, ears and tal postons and movement, a varety of gats, and LED dsplay; and Extraverted and Agreeable personalty was acheved by modfyng the percentage of tme the robot spent walkng around and waggng ts tal. The two condtons used n the study were Emotonal (wth the aforementoned affectve expressons) and Control (wthout affect). The study was set as robot as a personal pet and protector scenaro, n whch 20 people nteracted wth the robot for a total of four sessons, rangng from 20 to 60 mnutes each. For each sesson, the partcpants were gven one or two new commands to ntroduce to the robot (7 total: Stop, Go Play, Follow the Ball, Kck the 11

12 Ball, Follow Me, Come to Me, and Sc em ), and the last sesson was cumulatve. The measures used for ths study were: PANAS (mood) questonnare [35] to assess partcpants Negatve and Postve mood at the end of each sesson; Mn-Markers Bg-Fve personalty questonnare [34] to assess the subjects personalty n the begnnng and the robot s personalty at the end of the study; and a post questonnare, to evaluate ease of use and pleasantness of nteracton. The post questonnare conssted of sx 5-pont Lkert scale questons wth three subquestons, wth Strongly Agree anchored at 5, and Strongly Dsagree anchored at 1. The questons were as follows: 1) It was easy to get the robotc dog perform the commands; 2) It was easy to understand whether the robotc dog was performng the command or not; 3) The robotc dog showed emotonal expressons; 4) The robotc dog had a personalty of ts own; 5) Wth every sesson, I was gettng more attached to the dog; 6) Overall, I enjoyed the nteracton wth the robotc dog. If the partcpants answered Agree or Strongly Agree to questons 3 or 4, they were also asked to answer questons 3a,b and 4a, respectvely. The subquestons were as follows: 3a) Emotonal expressons exhbted by the dog made the nteracton more enjoyable; 3b) Emotonal expressons exhbted by the dog made the nteracton easer; 4a) I enjoyed nteractng wth the robot, partly because t possessed some personalty. A total of 20 people partcpated n the study, 10 males and 10 females, dstrbuted equally between the two condtons. The subjects were recruted va flyers posted on and around the Georga Insttute of Technology campus, and they vared wdely n the demographcs accordng to age (from between 20 and 30 to over 50 years old), ther educatonal level and backgrounds (from Hgh School dploma to workng on a Ph.D., wth majorty havng ether a Bachelor s or Master s degrees), and computer experence. Most of the partcpants had owned pets at some pont n ther lves (18 out of 20), and had ether no or very lmted robot nteracton experence (only 2 out of 20 had nteracted wth moble or entertanment robots pror to the study). A number of nterestng and encouragng observatons were made n ths study, as reflected n the results of 1-taled Independent Samples T-tests and Pearson Correlatons (unless specfed otherwse). Frst, those partcpants who beleved that the robot dsplayed emotons (5 out of 10 n the Control condton, and 8 out of 10 n the Emotonal condton) and/or personalty (6 out of 10 n both condtons) also beleved that these features made ther nteracton more pleasant: the average answer for queston 3a was 4.46, and for queston 4a was There was also a number of sgnfcant correlatons between questonnare responses regardng the pleasantness of the nteracton: 1) the response to queston 4 (robotc dog had a personalty) had a sgnfcant postve correlaton (r =.66, p=0.02) wth the response to queston 5 (partcpant got more attached to the dog); 2) the response to queston 3a (emotonal expressons made the nteracton more enjoyable) had a sgnfcant postve correlaton wth response to questons 5 (r=617, p=.025) and 6 (r=.749, p=.003, enjoyed nteracton wth the robotc dog ). Although there was no sgnfcant statstcal dfference between the two condtons n emoton dsplay recognton (queston 3), possbly due to small sample sze (M non-emotonal =2.7, M emotonal =3.6, F=.693, p<0.088, eta 2 =.1), the perceptons of the robot s emotonalty and personalty seem to make the nteracton more enjoyable and result n greater attachment. Ths was encouragng, as t suggested that people value expresson of emoton and personalty n ther nteracton wth an autonomous entertanment robot. Second, we observed a reduced level of Negatve Mood n the subjects n the Emotonal condton (M non-emotonal =13.9, M emotonal =12.125, F=6.462, p<0.048, eta 2 =.146), whch suggests that affectve behavor contrbutes to the qualty of nteracton. Addtonally, a sgnfcant postve correlaton (r=.598, p=.007) between average Postve Mood and the response to queston 4 (robot dsplayed personalty) was observed, thus provdng a lnk between perceved robotc personalty and users mproved mood. Fnally, women were found to be more attuned to emotonal expressons and more ready to attrbute emotons to the robot than men: 2-factor ANOVA on Gender and Emotonalty resulted n a sgnfcant man effect of Gender on the answer to queston 3: dsplay of emotons (M female =3.8, M male =2.5,F =4.829, p<0.043, partal eta 2 =.232). Ths should be taken nto consderaton for systems adapted to groups wth gender-based compostons. Other lessons learned from ths study ncluded the followng: 1) the physcal platform affects the percepton of emotonalty; 2) for between-subject experments, affectve expressons may need to be exaggerated to be conscously acknowledged; 3) there s a great need for sutable repeatable metrcs to evaluate usefulness of robotc affect va human-robot nteracton studes. 12

13 4 Software Archtecture In ths secton, we ntroduce the overall software archtecture for our affectve system, also referred to as TAME Module henceforth (Fgure 9). The system s desgned to be a stand-alone process to acheve maxmum platform-ndependence. Wth an nterface to connect to the system s TAME Communcaton Manager (to supply sensory data), and approprate confguraton fles, ths software can potentally be ntegrated nto any robotc platform wth ease and flexblty. The archtecture tself s farly straghtforward, and conssts of: TAME Manager (the man module of the system), TAME Communcaton Manager (receves sensor data and passes the updated affectve values to the robot), a module for each of the affectve components, and Stmul Interpreter. Confguraton Fle Intal confguraton Trat values Trat Emoton Mood Atttude Updated values Updated values Updated values Stmul Interpreter TAME Manager Updated TAME varables Raw sensor data Processed stmul Processed stmul / TAME varables Processed stmul TAME Comm. Manager Confguraton Fle Updated TAME varables Raw Sensor data Robot Fgure 9: Hgh-level archtectural vew of the TAME system. 4.1 Affectve Modules These are comprsed of four dfferent affectve components of TAME (namely Trat, Atttude, Mood, and Emoton), and each module processes sensory and nternal nformaton (current values of other affectve components) and calculates the updated affectve varables, passng them along to TAME Manager. In order to provde flexblty and adaptaton to ndvdual users and stuatons, each component s loaded wth some ntal default values from a confguraton fle. For the Trat component, a default value can be specfed for each of the fve personalty dmensons: Openness, Agreeableness, Conscentousness, Extraverson, and Neurotcsm. Once the values are specfed, they reman unchanged throughout the lfetme of the system executon snce personalty s generally regarded to be tme-nvarant. For the Emoton component, confguraton settngs nclude specfcatons prescrbng how each personalty dmenson may nfluence each emoton (e.g., drectly or ndrectly), as well as mnmum and maxmum values for a number of varables used for emoton generaton. Although t would be possble for an advanced user to select these values to sut a partcular task, n general, settng these defaults would be best left to the desgner or admnstrator, as they would nfluence complex nteractons wthn the module. For the Mood component, confguraton settngs nclude settngs for hgh and low ponts for crcadan changes, as well as mappng between sensor data types and nfluence on mood generaton. Agan, some of these settngs are best left for desgner/admnstrator, except for 13

14 specfyng crcadan changes to match user s mood. Renforcement learnng s beng nvestgated as the bass for determnng these parameters. Fnally, for the Atttude component, the confguraton fle can specfy the types and propertes of objects towards whch the robot can have postve or negatve atttudes. Case-based learnng s beng developed for ths component as a means to both set the parameters and generalze over broad classes of objects. 4.2 TAME Manager and TAME Communcaton Manager TAME Manager s the man module n the system that runs as a threaded process to manage all the affectve components. It supples each of them wth relevant sensor data (processed as stmul) or necessary values of certan varables from other affectve components. The affectve varables of all the affectve modules such as the Joy varable n the Emoton component or the Extraverson varable n the Trat component are comprehensvely called the TAME varables n the system. Then, TAME Manager receves the updated values of the TAME varables for each affectve component after approprate calculatons. TAME Communcaton Manager s a separate thread that s responsble for recevng sensor data from the robot and relayng them to Stmul Interpreter, and then passng approprately processed stmul nto TAME Manager. It also receves the most up-to-date values of the TAME varables from TAME Manager and communcates the nformaton to the robot controller. Behavoral arbtraton or the changes to behavoral parameters accordng to dfferent affectve states can then be acheved on the robot controller sde. By avodng drect manpulaton of behavoral parameters, the desgn of our affectve system allows for hgh portablty and scalablty. On the robot sde, dependng on the capabltes of a partcular platform, correspondng affect can be mplemented n ether contnuous or dscrete manner. For example, n the contnuous case, an emoton can be expressed n accordance wth Ekman s FACS on an anmated robot face, or mapped onto velocty and expensveness of gestures and posture, n a manner smlar to that proposed by the desgners of Robove [16] or through behavoral overlay method proposed by Brooks et al. [7]. In the dscrete case, a number of affectve expressons (facal and/or bodly), perhaps dfferng n ntensty, can be desgned on the robot a prory, and then an approprate expresson can be selected based on the actual value of a TAME parameter. We have mplemented the latter case on a humanod robot Nao. 4.3 Stmul Interpreter The raw sensor data from the robot themselves are useless unless some context s provded for them. A confguraton fle for Stmul Interpreter gves such contextual nformaton. For each TAME varable that s drectly affected by envronmental condtons or stmul (all but trats), the fle specfes whether each type of ncomng sensor data s relevant to the calculaton of that TAME varable. As emotons are nvoked n response to specfc stmul, certan object propertes would be used for stmulus strength calculaton. These propertes may correspond to preprocessed ncomng sensor data, such as dstance, sze, approach angle and acceleraton, or color of an object; they can also nclude more abstract propertes, such as frendlness or dsapproval of a person. A confguraton fle would specfy whch of these are relevant for generatng a partcular emoton, as well as weghts descrbng a relatve mportance of each. For example, for fear, object sze and speed of approach may play a larger role, whereas an nteractng person s personal attrbutes may be more mportant n case of joy. For moods, ncomng external and nternal sensor data would nclude battery level, nternal and external temperature, brghtness and nose level, and other potental nfluences. For example, postve mood s more susceptble to energy consumpton, and negatve to lghtng condtons, and these dfferences are reflected through assgnng approprate weght for each n the confguraton fle. Fnally, for atttudes, an object dentfer s used (such as an AR marker), whch encode specfc object propertes: color, sze, shape, category and materal. If a type of sensor data s relevant, a scalng factor s provded for normalzaton purposes, to translate t to an approprate strength snce each type can have a value n dfferent ranges. The confguraton fle also allows flexblty n specfyng whether multple sensor types should have a combned effect as an average (smoothng) or the stmulus wth the maxmum strength should have a domnatng effect n generatng the value of each TAME varable (wnner-take-all). 14

15 5 Implementaton The TAME Module was ncorporated nto MssonLab, a robotc software toolset that allows an operator to easly create and confgure a mult-robot msson usng a graphcal user nterface [36, 37] 1. In order to demonstrate the effectveness of our affectve system, t has been ntally tested usng Aldebaran Robotcs Nao humanod platform (Fgure 10). Fgure 10: Aldebaran Robotcs Nao humanod robot (source Aldebaran Robotcs) 5.1 MssonLab Overvew In MssonLab, an operator uses the Confguraton Edtor (cfgedt) to specfy a msson usng a graphcal representaton called an FSA, or fnte state acceptor [25]. In FSA representaton, a msson s composed of a combnaton of varous actons (behavors) to perform, and perceptual trggers act as condtons for movng from one acton to the next. The resultng msson s translated nto C++ code and compled to make Robot Executable. Then, t can be deployed on a wde varety of smulated and real robot platforms, and the operator can montor the executon of the msson n real-tme usng mlab GUI dsplay. HServer [37] s a control nterface to a varety of robotc hardware, and t s separate from Robot Executable to enable more flexble coordnaton wth dfferent robotc platforms. 5.2 Integraton wth MssonLab and Nao Robot Fgure 11 presents a graphcal vew of the ntegraton. Here, HServer acts as a brdge to the Nao robot to communcate between Robot Executable (whch contans the actual control code for the robot s current msson) and the TAME Module. In HServer, an nterface for the Nao robot has been created usng Nao s API for hardware control. When Robot Executable s n a certan behavoral state wthn a gven msson, the generated motor commands are transmtted to HServer, whch controls the Nao robot at the hardware level. HServer also contnuously receves perceptual data from the robot. Upon recevng the data, HServer sends them to both Robot Executable and the TAME Module. Robot Executable needs the sensor data for performng certan behavors and for determnng when to transton from one state to the next n the msson. When sendng the sensor data to the TAME Module, HServer organzes relevant data for the TAME module n accordance wth the confguraton fle for Stmul Interpreter, and sends each type of sensor data wth a unque ID. The TAME Module nterprets each datum n context usng ts Stmul Interpreter and then the updated values of ts TAME varables are calculated accordngly. Robot Executable possesses a smple database of the TAME varables, and ther values are updated at 3 hertz (to ease computatonal burden) by the TAME Module. These varables nfluence the robot s behavors by changng approprate behavoral parameters or selectng from a predefned set of expressve affectve behavors. 1 MssonLab s freely avalable for research and development and can be found at 15

16 Fgure 11: Archtectural vew of the TAME Module ntegrated wth MssonLab and Nao humanod robot. 5.3 Nonverbal Affectve Behavor Recognton Survey All components have been mplemented on an Aldebaran Nao robot. Based on an extensve lterature revew [38-46], we desgned expressons of Extraverson and Introverson, Postve and Negatve Mood, and Emotons of Fear and Joy. Fgure 12 (Left) and (Rght) provdes examples of statc poses of Joy and Fear, respectvely. To test the recognton of these affectve behavors, we conducted an onlne survey, n whch 26 partcpants were asked to watch a number of short vdeos of Nao producng the aforementoned affectve expressons [47]. The followng measures were used n ths survey: a shortened verson of PANAS (mood) questonnare [35] to assess Negatve and Postve Mood of the robot (1 clp wth the robot dsplayng Negatve Mood and 1 - Postve Mood); Extraverson subset of Mn-Markers Bg-Fve personalty questonnare [34] to assess to what extent the robot n the correspondng two clps was perceved as extraverted or ntroverted; and a multple-choce queston askng the partcpants to select one of sx emotons (Anger, Joy, Interest, Fear, Dsgust and Sadness) or suggest another one f not present n the choce, for the clps wth the robot exhbtng Joy and Fear. To the best of our knowledge, there has been no explct research that addresses nonverbal behavor for humanods across multple affectve constructs. Fgure 12: (Left) Statc pose for Joy. (Rght) Statc pose for Fear. On the Extraverson scale from 1 to 9, the Introverted Nao scored 3.6, and Extraverted 7.1 (almost twce as Extraverted); ths result was statstcally sgnfcant (p<0.001), see Table 2 for Mean and Standard Devaton. In terms of mood expressons, the robot dsplayng postve mood was rated low on Negatve and hgh on Postve Affect; the robot dsplayng negatve mood was rated medum on Negatve and low-medum on Postve Affect. For the postve robot mood, Postve Affect score was sgnfcantly hgher than that for the negatve robot mood (21 vs out of 30, p<0.001), and vce versa, ts Negatve Affect score was sgnfcantly lower than that of negatve robot mood (8.6 vs out of 30, p<0.001); see Table 2 for Mean and STD. Fnally, the recognton rates for emotons of joy and fear were hgh 85% and 81%, respectvely; these rates are comparable to those obtaned n judgments of joy and fear portrayals by human actors n move clps (facal features obscured), whch were 87% and 91%, respectvely [48]. Gven ths successful encodng of a number of affectve behavors, we are currently desgnng and conductng a set of human-robot nteracton studes to test the effect of the system on physcally present users. In the nterm, a number of vdeos demonstratng the results to date (ncludng the vdeos used for the survey) are avalable at: 16