Do Virtual Sheep Smell Emotion?

Do Virtual Sheep Smell Emotion ? Carlos Delgado, Ruth Aylett Centre for Virtual Environments The University of Salford Business House, University Road Salford, M5 4WT, United Kingdom Tel: +44 161 295 2918 Fax: +44 161 295 2925 C.Delgadopgr.salford.a .uk R.S.Aylettsalford.a .uk Abstra t In this paper we des ribe an emotional- behavioural ar hite ture. layer than the The emotional behaviour engine system, is and it a higher an alter behaviour patterns, the engines is designed to simulate Emotionally-Intelligent Agents in a Virtual Envi- ronment, where ea h agent senses its own emotions, and other reature emotions through a virtual smell sensor; senses obsta les and other moving reatures in the environment and rea ts to them. The ar hite ture onsists of an emotion engine, behaviour synthesis system, a motor layer and a library of sensors. KEYWORDS: autonomous agents, multiple agents, emotion, virtual environment, behavioural arhite ture, virtual sensors, virtual smell. 1 Introdu tion This paper dis usses work-in-progress in the use of emotion as a form of intera tion between (pseudo)embodied agents in a virtual environment (VE). It attempts to integrate emotions at the pseudophysiologi al level into an existing behavioural ar hite ture whi h will be brie y des ribed. It then seeks me hanisms for the transmission of emotion between agents, and for the per eived emotion of one agent to in uen e the emotion and then e the behaviour of another. One bran h of this work on nes itself to the behavioural level, taking sheep as exemplar agents, and it this we will dis uss here. The other bran h onsiders how a planning system might be integrated into the ar hite ture and takes the idea of the Holode k, a soap-opera type drama played out by virtual pseudohuman agents, as its test-bed. While emotional intera tion between agents naturally requires an internal emotional ar hite ture, this ar hite ture has to be linked both to transmission and re eption me hanisms. Thus the fo us is somewhat di erent from mu h existing work whi h either onentrates on the internal agent ar hite ture or onsiders the agent-human user relationship. As dis ussed below, transmission and re eption are asymmetri : transmission is the sole step in ommuni ating emotion to another agent, but re eption may also involve an intera tion between the sensed emotion and the existing emotional state of the agent before feeding into its behaviour. Note that we do not in lude intentionality in this system - emotional intera tion is being modelled as essentially involuntary. 2 Behavioural ar hite ture - previous work Previous work had taken a behavioural ar hite ture developed for multiple ooperating robots - the Behavioural Synthesis Ar hite ture or BSA [4℄ - and reapplied it to agents in a virtual environment (VE) in the Virtual Teletubbies proje t [2℄. We here give an overview of this ar hite ture. The BSA in orporated three stru tures at in reasing levels of abstra tion: behaviour patterns, behaviour pa kets, and behaviour s ripts. 2.1 Behaviour patterns At the most primitive level, a behaviour pattern, (bp) as illustrated in gure 1, was de ned as a pair of fun tional mappings, one from in oming sensory stimulus to outgoing desired motor response, and the other from sensory stimulus to utility, a mapping de ning the importan e of the motor response for the given level of stimulus. An agent possesses a repertoire of behaviour patterns, with ea h a tive pattern at any given time proposing its desired motor response a ording to its urrent sensory input. These responses were weighted by their utility values and synthesised together to produ e an emergent response, whi h was the a tual behaviour of the agent. Thus se ond-tose ond variation in emergent behaviour was dealt with via weighted synthesis on a ontinuous basis, unlike the time-sli ed Brooksian ar hite ture. 1 ooperant a tivities e.g. two agents arrying an obje t together, two agents hugging ea h other. A Universe level ontains bps spe i to a parti ular task, e.g. navigating to the initial lo ation of an obje t to be relo ated (su h as the lo ation of a sli e of bread for example), then subsequent navigation to the desired goal lo ation (the toaster in this ase). These strategy levels be ame an implementational feature as dis ussed below when the ar hite ture was moved from robots to virtual agents. 2.2 Utility 0 Sensor 1 1 Utilitor at time t Sensory stimulus at time t 1 0 Motion 1 0 Sensor 1 Behaviour Pattern Figure 1: Behaviour pattern example Consider the situation where the sensory stimulus relates to an agent's forward fa ing distan e-toobsta le measuring sensor and the asso iated motion response relates to the forward translate velo ity for that agent. From Figure 1 it an be seen that as the agent gets nearer to the obje t then its forward translate velo ity will be redu ed to zero. At the same time, the asso iated utility for this motion response in reases. A similar pair of fun tions for rotation produ es an in reasing rotation away from the obsta le, also with in reasing utility, as its distan e from it deminishes. Thus as the agent gets nearer to an obje t in its path, it be omes more important for the agent to slow down and turn away from it. In the BSA four behaviour levels (often known as strategy levels) were identi ed, originally for purely on eptual onvenien e [3℄. A Self level ontains those bps on erned with the maximisation and replenishment of internal resour es, e.g. making sure an agent does not go hungry, or not walking up hills when energy is low An Environment level ontains bps asso iated with a tivities involving obje ts within the agent's environment, e.g. ollision avoidan e, ollision dete tion, playing with a toy. A Spe ies level ontains those bps asso iated with Behaviour pa kets If all the bps in an agent's repertoire were a tive at the same time then the overall emergent behaviour of the agent might be of little value. For example patterns designed to produ e obsta le avoidan e as des ribed above are not useful if you want an agent to sit down on a hair or hug another one of its spe ies. The bp designer must always bear in mind that the low-level ar hite ture is sensor-driven, and not task or even sub-task dependent. What is needed in this ase is an automati me hanism for dea tivating the 'obsta le avoidan e' bps when the 'sit' bps or 'hugging' bps are a tive. Asso iated therefore with every bp within an agent is an 'a tive ag', whi h enables or disables it. Thus obsta le avoidan e bps for example an be turned o and on when required. A bp is 'dea tivated' in the BSA by for ing the respe tive utility to zero. The a tion e e tively produ es a bp of zero importan e and hen e one whi h does not ontribute to the overall emergent behaviour of the agent. This me hanism is applied by grouping together bps in goal-a hieving sets known as behaviour pa kets. A behaviour pa ket is a small data stru ture whi h inludes a sensory pre- ondition for a tivating the bps it referen es, and a sensory post- ondition whi h ontrols dea tivation of the named bps. Behaviour pa kets show some similarity with AI produ tion rules [7℄, though they work at the sub-symboli level and are driven by in oming sensor data rather than by an inferen ing system. They support behavioural sequen ing for agents performing at a task (universe) behaviour level. Thus a sensory pre- ondition of 'being near the hair' ould be used to move from a beahviour pa ket in whi h obsta le avoidan e bps were a tive to one in whi h they are not. Thus behaviour pa kets provide a me hanism for ontextually sensitive behaviour swit hing, whi h is seen as a more exible me hanism than the nit-state ma hine de nition of inhibition and exitation between behaviours of the subsumption ar hite ture. 2.3 Behaviour S ript: high-level sequen ing and agent drives A behaviour s ript is simply a set of behaviour pa kets assembled for the a hievement of a parti ular task, using the sensory rep-and post- onditions of Figure 2. The original approa h was to generate behaviour s ripts on the y using a re e tive agent in orporating a symboli AI planner, and then send the individual s ripts to behavioural-based agents. This hybrid approa h was taken with the o-operative robots in MACTA [1℄ and is appropriate where the domain is predominantly task-based. However, while the lower levels of the ar hite ture were moved from robots to virtual agents with no hange, virtual agents are less likely to live in a task-oriented environment. It was at this point that the issue of in orporating emotion into the ar hite ture ame to the fore sin e if behaviour is not task-dire ted then drives or emotions are required instead. Broadly, if an agent is not being told what to do, then it does 'what it likes'. Thus a sequen ing engine linking behaviour to internal motivations or drives was developed as seen in Figure 3. This was applied in the Virtual Teletubbies proje t [2℄. A set of drives were developed for the virtual agents - hunger, ex itability, happiness, uriosity, and sleep - omparable to the homeostati variables found in work of Bru e Blumberg [5℄. These drives play the same role of ontextually-driven behaviour swit hing that had previously been played by an AI planner. The framework developed ontains four queues, one for ea h of the on eptual ategories self, spe ies, environment and universe dis ussed earlier. The entries in this queue onsist of groups ontaining one or more behaviour pa kets, e e tively sub-s ripts known as behaviour s riptlets, ea h with an atta hed priority. The priority is generated automati ally and is typi ally related to a predetermined threshold level of a drive, so the more hungry an agent be omes, the greater the priority. The s riptlet with the highest priority is then sele ted for pa ket exe ution. Thus if one adds fear to the set of emotions/drives, a ertain level of fear may trigger pani behaviour in sheep whi h up until then have been quietly eating grass. The behaviour-sequen ing engine although always pro essing stimuli might not be exe uting a s riptlet and therefore has a default s ript at the environment level. The default s ript exe utes a single pa ket ontaining bps that e e tively lets the low-level module handle wandering in the environment while avoiding obsta les. In the ase of sheep this might be modi ed to grazing behaviour. The default s ript is hanged when another sensory pre ondition from another set of pa kets is met. This is typi ally at another strategy level so if a agent sensed the presen e of another agent this ould trigger the behaviour at the spe ies level. It should be lear that to sequen e behaviour, it is enough to model emotion/drives as a meter with a threshold above whi h swit hing o urs, though this is of ourse not at all biologi ally plausible. The level of the drive is itself determined by either per eption and a tuation in most ases: uriousity rises when a new obje t is seen, hunger rises more qui kly if an agent is more a tive. 3 Emotion as a behaviour pattern or pattern modi er The use of emotion/drives to swit h behavioural mode is however only one way of dealing with the relationship between emotion and behaviour. The ar hite ture also supports a mu h lower level relationship at the level of the synthesis me hanism. Two approa hes are possible here. One is to onsider a drive or emotion as a behaviour pattern in its own right, driven by a parti ular sensor input. Its output is an a tuator response whi h is synthesised into that of the other bps, thus having an e e t upon the global emergent behaviour. Thus if an unfamiliar smell is sensed on the prevailing wind, a grazing sheep might slowly graze in the opposite dire tion to it so as to move away from it, with its fear behaviour . Or alternatively it might graze loser to other sheep, produ ing a more ompa t o k. Alternatively, one ould view it as an extra input into the stimulus-utility fun tion omponent of a behaviour pattern. By a e ting the utility omponent of more than one pattern, an emotion/drive in e e t ouples utilities together, reintrodu ing the idea of inhibition/ex itation between patterns but in a exible fun tionally de ned manner. This would not only allow new behaviour to be added, but would allow existing behaviour to be suppressed via a attening of its utility fun tion. It would also allow for a emotional behaviour to gradually be ome more dominant for example a grazing sheep be oming more and more jumpy in the fa e of a threatening stimulus. These approa hes are not mutually ex lusive, sin e a freestanding emotional behaviour ould be di erent from other behaviour patterns pre isely through its ability to ross- ouple with utilities. Analysis of real sheep behaviour is needed to onstrain the design of this emotional system. A sheep subje t to an alarming sensory stimulus may be ome 'restless' - dgeting while it grazes, 3.1 Motor Layer High Level Forward Kinematic Animation Behaviour Packet Behaviour Pattern Emotion Patterns Behaviour Script Behaviour System Drives Virtual Nose Other Sensors World Virtual Sensors Moods Emotion Engine Environment Conditions Simulator (wind, temp) Figure 2: Ar hite ture Blo k diagram twit hing its ears. As its level of fear rises, it may break o grazing to look around, before resuming; at a higher level still it may stop grazing and enter a state of alert, before nally pani king and running. With several drives/emotions a ting on urrently, possibly with some modelled as a temporal y le, it ought to be possible to model the hara teristi s put forward by Neary [12℄. He notes that sheep will move more sluggishly on hot days and during the middle part of the day. This is simulated in our ar hite ture by a e ting the moods, as illustrated in gure 2. The mood an a t as a lter for the a tivation of emotions patterns. Thus if the sheep is in a bad mood it ould get angier fast, or if the sheep is sluggish it ould rea t (the sensing sensitivity ould be a e ted) slowly and move awkwardly. The sheep are reatures of habit and they usually graze, drink and are more a tive in the morning and during the evening. Thus, if you are running (i.e. herding them) in the middle part of the day, the sheep may o k tighter, move slower, and spend more time looking at and ghting a dog. This suggests that areful analysis is needed in order to establish the most appropriate set of drives/emotions - Neary also points out that for sheep to be sheep, there needs to be a minimum number of ve in the group. Less than ve and their behaviour is not as predi table, suggesting a drive/emotion one might des ribe as 'group se urity' a ting as a behavioural modi er. Communi ating emotion Be ause agents exist in a VE and not in the real world, in prin iple the transmission of emotion between agents ould just be arried out by ' heating', that is by allowing agents to read ea h other's internal state dire tly. We hoose not to do this however, sin e we see advantages in reusability and in mat hing realworld behaviour (that is in real sheep for example) by trying to model emotional intera tion in a slightly more prin ipled way. In the real-world however, emotional transmission may well be multi-modal, with ertain modes su h as the per eption of motion (or in the general ase 'body language') being parti ularly diÆult to model. Thus we have limited ourselves for now to a single mode, and the one we have hosen is s ent, to be per eived by a virtual nose sensor. In our emotional ar hite ture we use a virtual nose sensor, be ause the nose has been linked with emotional responses and intelligen e. Goleman [10℄ states "The most an ient root of our emotional life is the sense of smell, or, more pre isely, in the olfa tory lobe, the ells that take in and analyse smell. Every living entity, be it nutritious, poisonous, sexual partner, predator or prey, has a distin tive mole ular signature that an be arried in the wind." Neary [12℄points out that sheep, parti ularly range sheep, will usually move more readily into the wind than with the wind, allowing them to utilise their sense of smell. Figure 3: Mammalian olfa tory system (from Gardner 1999) There is ongoing resear h in the eld of ele troni noses, Gardner [9℄ but we are not aware of the use of virtual noses in a virtual environment, so important issues arise like modelling mole ules and wind ow, and how the wind moves them in the environment. The stru ture of a parti ular mole ule is important in determining odour. This requires sensors whi h are responsive to the shapes or stru tural features of the organi mole ules being modelled. In real animals hemore eptors (extero eptors and intero etors) are used to identify hemi al substan es and dete t their on entration. Smell exists even among very primitive forms of life . In our ar hite ture we intend to model the extero eptors whi h dete t the presen e of hemi als in the external environment. In vertebrates the olfa tory re eptors are primary sensor neurons with dendrites that extend as ilia into a mu ous layer, as illustrated in Figure 3. These axons from the olfa tory nuerons synapse with the dentrites of the mitral and tufted ells in the olfa tory bulb. At the level of the glomeruli the quality of the odorous stimulus is en oded in the form of a tivated glomeruli, Axel [15℄suggests that\The brain is essentially saying something like, I'm seeing a tivity in positions 1, 15, and 54 of the olfa tory bulb, whi h orrespond to odorant re eptors (glomeruli)1, 15, and 54, so that must be jasmine, ". Most odours onsist of mixtures of odorant mole ules. Therefore, other odours would be identi ed by di erent ombinations. In the ele troni noses industry an empiri al approa h is used, making use of available sensor types and attempting to modify sensor designs to meet the requirement of the ele troni nose [9℄. To sense smell in the virtual environment, the mole ules must be distributed in the environment; this is modelled by setting the density of ea h of the mole ules within the environment represented as a grid. To simplify the omputation the urrent grid is 2D, but we plan to use a voxel-based grid in the future. The smell sensor and the virtual mole ules used to represent the signature of di erent smells, is an important feature of our ar hite ture, be ause it an be used to ommuni ate emotions between agents through the environment. For example if a sheep pani s it ould exude a distin tive odour, or mole ular signature, to the environment using a density fun tion, and through time the mole ules would disperse depending on several arti ial environment fa tors like wind, rain, season and time of day. Other sheep will sense the pani smell and they will pani themselves and also exude the distin tive odour for pani so that the odour will propagate through the o k in few simulation steps. Primary odour Camphora eous Musky Floral Pepperminty Ethereal Pungent Putrid Familiar substan e Moth repellant Angeli a root oil Roses Mint andy Dry- leaning uid Vinegar Bad egg Table 1: Primary odours with more familiar examples (Adapted from M Farland) To simulate volatile parti les, i.e. s ent whi h an be dete ted with the virtual ofa tory system, we use parti le animation theory, in whi h ea h parti le represents a on entration of mole ules with a di erent primary odour as in Table 1 or other hemi al signal, like pani king sheep. The parti les behaviour depend on the emitter propeties: 1. Emitting dire tions 2. Initial position. 3. Initial velo ity. 4. Initial size. 5. Initial transparen y. 6. Shape and olour 7. Lifetime The number or parti les at a parti ular time an be derived from the following formula. N (t) = M (t) + I (t) Where M(t) is the mean number of parti les perturbed by the intensity of the s ent I(t) in the ase of sheep pani . The position of the parti les will be in uen ed by weather ondition, in parti ular wind and time of day. Pi (t) = Pi (t 1) + W (t) Where Pi (t) is the Position of the parti le at time t and W(t) is the weather ve tor at time t. Mole ule properties are en oded in shape and olor of ea h parti le, we are going to use John Amoore's lassi ation, illustrated in Table 1, whi h is one of the most widely a epted[11℄. Re ent s ienti studies [15℄ show that many mammals use a separate set of sensory re eptor ells in their nose to re eive so ial and sexual information from members of their own spe ies, and there is growing suspi ion that we (humans) do, too. We will en ode this virtual \s ents" as a di erent olour to trigger emotions, su h a pani in sheep. The rea tion time of the re eptor is given by: trea tion = tmin + 1 kS n 4 The ognitive dimension In sheep it seems that the re eption of emotion is rather straightforward - a pani king sheep auses others around it to pani . In humans though, it seems that emotion re eption is often mediated by a base emotional stan e towards an individual as well as by the general sensitivity of re eption that forms one of the omponents of empathy. Again, for sheep, emotion an be modelled as a purely behavioural phenomenum, while it lear that in a human it also enters into ognitive fun tions. It is not intended to arry out a omprehensive disussion in this paper of how a ognitive apability in the form of an AI planner an be added into the behavioural ar hite ture dis ussed above in su h a way as to take input from the emotional system. However the extra omplexity required in the system an be illustrated from the following s enario onsidered in the Holode k ontext. In this s enario, one agent, representing a daughter, enters rying, into a room holding an agent representing her mother. The mother's rea tion is to hug the daughter, and say "What's wrong?". The daughter says "I've split up with Dave', where Dave is her boyfriend. After a ertain amount of hugging, the mother says "Sit down and let me make you a uppa" (this refers to a up of tea, a ulturally- onditioned omforting re ex in the UK). While it is true that the daughter's grief is in reallife dete ted by sound and vision, there seems to be no reason why it ould not be modelled as a smell just like pani in sheep. However the out ome is not that the mother re eives the emotion of grief and bursts into tears too. Her base emotional state towards the daughter might be des ribed as 'prote tive', and the intera tion of per eived grief with this base state produ es a ' omforting' behaviour. There seems to be no reason why this annot be modelled at the behavioural level. The idea of a modi ed emotional input triggering a ' omfort' behaviour seems at least as plausible as a ognitive a ount whi h would require a symboli level ategorisation of the in oming emotion ('if a person ries then they are unhappy') followed by some reasoning about an appropriate response. 5 Virtual Sheep To simulate the \emotional" virtual sheep in a virtual environment we are using the ar hite ture des ribed in Delgado [8℄, with the di eren e we have substituted Performer for Maverik, be ause the former o ers better performan e and reliability for appli ations developed for the CAVE see gure 4. Figure 4: Users in the CAVE 6 Con lusion In this brief paper we have tried to explain the behavioural ar hite ture being used for virtual agents and to look at how drives/emotions an be integrated into it in a more sophisti ated manner than the existing use of a meter representation. We believe that by modelling transmitted emotion as a smell and using a virtual nose for re eption, we are employing a biologi ally plausible me hanism. Work will ontinue both to produ e hopefully a urate olle tive sheep behaviour as well as to investigate the more varied intera tion between emotion re eived and indu ed in more human-like agents. 7 A knowledgments The authors thank Carlos Delgado Flores, AgroIndustrias CEJA, Universidad Bonaterra and The University of Salford for their support. Referen es [1℄ Aylett, R.S. Communi ating goals to behavioural , Pro eedings of the agents - A hybrid approa h Third World Congress on Expert Systems, Seoul, Feb 1996 [14℄ Pi ard, R., A e tive Computing, The MIT Press, 1997 [2℄ Aylett, R.S; Horrobin, A; O'Hare, J.J; Osman, A.& Polyak, M. Virtual teletubbies: reapplying a robot ar hite ture to virtual agents Pro eedings, 3rd International Conferen e on Autonomous Agents . [15℄ Pines, M., Seeing, Hearing and Smelling the [3℄ Barnes, D.P. A behaviour synthesis ar hite ture for ooperant mobile robots. Advan ed Roboti s and Intelligent Ma hines, eds. J.O.Gray and D.G.Caldwell, IEE Control Engineering Series 51.295-314, 1996 [4℄ Barnes, D.P., Ghanea-Her o k, R.A, Aylett, R.S. and Coddington, A.M. Many hands make light work? An investigation into behaviourally ontrolled ooperant autonomous mobile robots, in: Pro eedings of the rst International Conferen e on Autonomous Agents, 413-20, Johnson, L. eds., Marina del-Rey, Calif.: ACM Press [5℄ Blumberg, B.. Old Tri ks, New Dogs: Ethology and Intera tive Creatures. Ph.D. diss. Massa husetts Institute of Te hnology [6℄ Ca~namero, D. Modeling Motivations and Emotions as a Basis for Intelligent Behaviour, Pro eeding of the First International Conferen e on Autonomous Agents, 1997. [7℄ Davis, R., and King, J.An overview of Produ tion Systems, in Ma hine Intelligen e, E.W. Elo k and D. Mi hie (Eds.), Wiley and Sons, 300332, 1977 [8℄ Delgado, C. and Aylett, R. From Robots to Crea- tures: Extending a Robot Behaviour Based Ar hite ture to Create Virtual Creatures. In ISRA 2000, Monterrey Mexi o, 2000. J. and Bartlett, P. Ele troni Noses:Prin iples and Pra ti e.Oxford Univer- [9℄ Gardner, sity Press. United Kingdom. 1999 [10℄ Goleman, D. Emotional Intelligen e. Bloosmbury.United Kingdom. 1996 [11℄ M Farland, D., Animal Behaviour, Addison Wesley Longman, 1999 [12℄ Neary, M.Sheep Sense. http://www.workingborder- ollie. om/arti le3.html [13℄ Ortony, A., Clore, G, Collins, A. The ognitive Stru ture of Emotions, Cambdrigde University Press, 1988 World: New ndings help S ientists make sense of our senses, Report from the Howard Hughes Medi al Institute, 1997 [16℄ Reilly, S. and Bates, J., Building Emotional Agents, Te hni al Report CMU-CS-92-143, S hool of Computer S ien e, Carnegie Mellon University, May 1992. [17℄ Velasquez, J., When Robots Weep: Emotional Memories and De ision-Making