"SOZZY:" A Hormone-Driven Autonomous Vacuum Cleaner

Masaki Yamamoto

Matsushita Research Institute, Tokyo3-10-1, Higashimita, Tama-ku,

Kawasaki 214, Japantyama@mritmei.co.jp

andMIT Artificial Intelligence Laboratory

545 Technology Square,Cambridge, MA 02139, USA

masaki@ai.mit.edu

Abstract

Domestic robots are promising examplesof the application of robotics to personal life.There have been many approaches in this field,but no successful results exist. The problem isthat domestic environments are more difficult forrobots than other environments, such as factoryfloors or office floors. Consequently,conventional approaches using a model of humanintelligence to design robots have not beensuccessful. In this paper, we report on aprototyped domestic vacuum-cleaning robot thatis designed to be able to handle complexenvironments. The control software is composedof two layers, both of which are generallyinspired by behaviors of living creatures. Thefirst layer corresponds to a dynamicallyreconfigurable system of behaviors implementedin the subsumption architecture. The ability of therobot to support alternate configurations of itsbehaviors provides the robot with increasedrobustness. We have conveniently labeledparticular configurations as specific "emotions"according to the interpretation of observers of therobot’s behavior. The second layer simulates thehormone system. The hormone system ismodeled using state variables, increased ordecreased by stimuli from the environment. Thehormone condition selects the robot’s mostsuitable emotion, according to the changingenvironments. The robot hardware is built of off-the-shelf parts, such as an embedded CPU,inexpensive home-appliance sensors, and smallmotors. These parts keep the total building costto a minimum. The robot also has a vacuumcleaning function to demonstrate its capability toperform useful tasks. We tested the robot in ourlaboratory, and successfully videotaped its robust

behaviors. We also confirmed the hormonesystem to enhance the robot’s plasticity andlifelike quality.

1. Introduction

Commercial cleaning robots are gainingsome success in their field; for example [7]. Theyoperate in commercial halls or other large spacesusually taken care of by commercial janitorialservices. Their task is to navigate over horizontalsurfaces, without colliding with obstacles on thefloor, and to clean all open areas. In this case,the environment is quite predictable andunchanging, allowing the robot to build aconvenient representation of the work space.Furthermore, the user of the robot is not anindividual but a commercial enterprise, so areasonable capital cost is acceptable, and therobot can be equipped with sophisticatedmechanisms. Domestic environments are not aseasy for robots to manage as commercial places.

In the 1980s, a variety of primitive homerobots appeared on the market [3]. These robotshad insufficient capability of sensing, moving,and computing, so the performance of the robotswas disappointing. The popular approach takenby these robots is called the "sense-model-plan-action cycle." In this approach, a robot firstsenses the outer world; second, makes a CAD-type world model in its memory; third, makes anaction plan according to the world model; andfinally moves according to the action plan. Thisstrategy is based on human intelligence, but inthe real world, the robot’s poor sensingcapability cannot perceive the environment asprecisely as a human does and the robot’s

116

From: AAAI Technical Report FS-93-03. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved.

actuators are far less efficient and dexterouscompared to human hands and legs. In order tocompensate for these poor capabilities, designershave tried to use more complex sensors andactuators, or to use more computationallyexpensive algorithms. As a result, these robotsbased on the human intelligence model tend to beexpensive and big. However, the real world isstill hard to handle for these robots.

In contrast to the previous approach, thesubsumption architecture doesn’t rely on thehuman intelligence model [1]. It is, rather, basedon the intelligence of lower creatures, such asinsects or animals. The subsumption architectureis composed of prioritized layers, each of whichhas an independent control and can handle thesensor inputs and the actuator outputs. Each layeror cluster of layers realizes a "behavior" of therobot. There are several merits to programmingthe robot in the subsumption architecture. First,the robot is robust in the real environment,because the robot’s intelligence is distributed toeach layer which runs in parallel. Second,intelligence can be incrementally programmed,starting from conditioned reflexes to a highertask-oriented behavior, allowing the robot to gainits ability step by step. Third, as the sensorinformation fed to the layers is quickly processedand is reflected on the actuator output, the robot’sresponse speed to the environment is faster.

A weak point of the subsumptionarchitecture is that each layer worksindependently without consideration of the totalstrategic plan. So the robot is sometimes trappedin a dead-lock situation, repeating the samereaction many times. In order to solve thisproblem, there should be a higher module overthe subsumption architecture module to decide along-term vision of what the robot should do.

As a result, to build an creature-typerobot, we adopted the simulated hormone systemas a higher module. The hormone system issimulated with several state variables that areinfluenced by stimuli from inside and outside ofthe robot. Each state variable also has a functionto stabilize its own values [9]. The hormonesystem reflects long-term environmental changeson the subsumption architecture module bychanging the priority between layers, andactivating or deactivating layers. This makes therobot behave differently and adapt to thechanging situation. Moreover, this mechanismgives the robot several characteristics which

appear and disappear according to its hormonalcondition.

According to a classic psychologicalstudy, observers who view a simple display ofmoving figures endow the figures with "human"qualities of intention and personality throughperceiving their causal interrelationships andmutual relationships [5]. When the robot interactswith the world in several ways that changeaccording to the situation, people will feel therobot has a "human" quality, such as a mood oran emotion. This can make the robot appear morefriendly --- an important feature for a domesticrobot. From this point of view, we utilize thehormone system not only to increase robot’sflexibility but also to endow the robot withchanging emotions according to the situation.

Our robot programmed in thesubsumption architecture and the hormonesystem also has several merits in its physicalappearance. As the control program does notrequire either a large computational power orsophisticated sensors and actuator, the robot canbe small, light, quick, and cheap. The smallrobot navigates well in the domesticenvironment, as houses are usually builtaccording to the size of a human. The fight robotis also safe to its environment as any damage thatit causes to a piece of furniture or a human incase it should fail is small, in proportion to itsfight mass. The small and quick robot does notoccupy a great deal of space and can avoid ahuman quickly. On the other hand, a big andslow robot can be an obstacle to a person livingthere. Finally the robot made of off-the-sheffparts can be reasonably inexpensive, and, as aresult, possible to purchase.

The prototyped robot based on the abovementioned concepts is named "SOZZY," and isround-shaped, ten inches in diameter, andincludes all necessary parts in its body. The robothas many varieties of sensors, such as bumpsensors, proximity sensors, a beacon detector,pyro human sensors, and a dust sensor. Itslocomotion actuators are two geared-motor-driven wheels. It is also equipped with avacuuming blower. The subsumption architecturemodule realizes the behaviors, such as obstacleavoidance behavior, human-interacting behavior,homing behavior, and foraging behavior. Thehormone system controls the four emotions ofthe robot: joy, desperation, fatigue, sadness.Each emotion appears in the robot according tothe situation.

117

The robot was tested in the realenvironment of the laboratory. The room wasscattered with random objects, such as chairs,empty boxes, and garbage cans. First, the robotsuccessfully demonstrated its robust navigationcapability, starting from its home station,navigating around the room and returning to thehome station. Second, the dynamic change of thehormone system was recorded in the real run ofthe robot. The change corresponds well with therobot’s condition and situation, showing that thehormone system increases the robot’s adaptivity.Finally, the robot demonstrated its real-taskcapability, the vacuum cleaning on the floor,which is programmed as a foraging behavior.The following sections present the hardware ofSOZZY (Section 2), the subsumption architecturemodule and the hormone system (Section 3), andexperimental results (Section 4).

2. Robot Hardware

~ Blower fan

I11 Polycarbonate

CdS light sensor ’’ [ ’ , dome

Pyro ~ I ! [ ~ /receiversenso~ I i I~tP(I \

IR proximit3~ I [ [ I I ~ sensor

sensor ~Bump -~..ik~-]l Dust[tank ILX..

eoso JIII /t ~ \l -"".----71N°zzle

//" ’ IPowere~ wheel ~ Forward

Rear caster caster

Figure 1. Schematic view of the cleaning robotSOZZY.

A schematic view of the robot is shownin Figure 1. General specification of the robot isshown in Table 1. The robot includes allnecessary parts, such as CPU, actuators, sensorsand energy source in its body. The upper half ofthe body is made of a polycarbonate plastic domewhich consists of a vacuum chamber. Thevacuum blower on top of the robot works as avacuum source. Just in the middle of the robot isthe dust filter which stops the dust in the dust

chamber. From the point of view of maximizingthe efficiency of the dust-vacuuming ability,microprocessors and other electronics are storedin the vacuum chamber. This design allows thedust filter to be wider and the vacuum loss to beminimal.

Total weightWidthHeightMaximum speedLocomotionVacuumWorkspaceB artery duration

7pounds10 inches in diameter11 inches2 feet per second2 geared motor differential drives30W DC motor blowerScattered bare floorAbout 30 miniutes

Table 1. General Specifications for SOZZY.

The second design principle is to use off-the-shelf parts. This results in low building costand short development period. The CPU boardswe used are IS Robotics Plsystem which arebased on 68HCll (Motorola). These boardsoffer an environment to program the robot in abehavior based way and interface to off-the-shelfsensors. Vacuum cleaning usually requires a lotof power and is not so suitable for a smallcleaning robot. But in order to avoid theexcessive engineering challenge to develop a newcleaning mechanism, we used an off-the-shelfvacuum blower and a cleaning filter fromPanasonic vacuum cleaners. A vacuum blowerhas a maximum power of 30W, but is usuallyused in much smaller power in the robot. Two1500mAh NiCd batteries for radio controlmodels are used for the robot, as IS Robotics’boards are designed for 7.2V NiCd batterypacks.

Each of the sensors is relatively small,cheap and low-power. So the robot is equippedwith as many sensors as possible. The sensorsare listed in Table 2. In the future stage of thedevelopment, redundant or unuseful sensors canbe omitted. Pyro sensors are sensitive to 8-10micrometer wavelength infrared light, which isemitted from human body, and they can tell thepresence of a human. The beacon receiver isshown in detail in Figure 2. This sensor utilizesan IR LED and several decoding ICs typicallyused in the tone decoders of touch-tonetelephones. By the rotation of the mirror, thephotodetector may scan around 360 degrees tofind the maximum direction of the beacon signal.The direction of the mirror is also monitored by apotentiometer. This sensor gives the robot aminimum sense of the space its navigates, and

118

enables the robot to return back to its home (i.e.,the recharging station). The detail of the dustsensor is shown in Figure 3. To endow therobot with as acute a dust sensing ability aspossible, this sensor utilizes the same techniquesas a dust counter in a semiconductermanufacturing facility. The principle is that dustparticles which pass through the laser diode beamshine instantaneously, and the shining is countedby the dust counter. This sensor also can adjustits sensitivity and detects a wide range of dustlevel from several thousandths of an inch sizeddust particles to paper confetti which were oftenused in the experiments.

# Sensor name Function

2 Motor encoder to measure wheel velocity8 IR proximity sensor to detect obstacles8 Bump sensor to detect collisions4 Pyro sensor to detect the presence of a human2 Mechanical switch to detect collisions of the nozzle2 Motor current monitor to detect motor stalling2 Battery monitor to detect battery level1 Beacon receiver to determine homing direction1 Dust sensor to detect dust particles

Table 2. Sensors used in SOZZY.

From both programming and constructionstandpoints, the differential drive can be the leastcomplicated locomotion system [6]. The robot isaiming for insect-like robust and quick movementrather than accurate navigation capability. So thissimple locomotion mechanism is employed.

IR LEDs

Beacon detecting sensor Beacon station(Installed in the robot) (Fixed somewhere in the room)

Figure 2. Beacon sensors to endow SOZZY witha homing capability

Photo Transistor~

Optical Filter.~..~ , J

tight Trap[ ,11~_ @~ [Laser Diode[

d Optics providingsheet beam

Dust Particles

Figure 3. Acute dust sensing using a laser diode.

3. Software Architecture

Each behavior is programmed in theBehavior Language [2]. For example, SOZZY’scleaning behavior is programmed as follows.SOZZY has a dust sensor with which it can tellthat it has "eaten its bait." So the cleaningbehavior is programmed as a foraging behaviorto try to follow the dust distribution to eat more.After trial and error, we have found that aswinging motion to the left and right seems to bemost effective for the robot. Figure 4 shows howthe foraging behavior is programmed in thecombination of AFSMs (Augmented Finite StateMachines). The oscillator AFSM generatestiming for the swing AFSM to control the basicwheel movements. The dust detect AFSM sendsmessages when the dust count exceeds a certainlevel. These messages first stop forwardmovement to stay there for a longer period, andthen, second, reset the oscillator AFSM so thatthe robot swinging center should be the currentlocation where dust is found, and then, third,trigger the monostable to lower its sensitivityresulting in ignoring small dust density for acertain period. The dust counter also triggersvacuuming AFSM to make vacuuming stronger.

In the subsumption architecture, severallayered behaviors compete over control of therobot. As each behavior has direct connection tosensors and actuators, a tight connection ofsensors and actuators are realized. This results inquick response of the robot and robust behavior.This layered structure also facilitates

119

programming and debugging. Each behavior canbe programmed and debugged separately, asmost of them can operate independently. As aresult, programming and debugging can beprogresmve and incremental. On the other hand,the weak point of the subsumption architecture isits fixed combination and prioritization ofbehaviors. This often results in a lack offlexibility or local minimum problems.

~ ger

’

[~"~1 I ’ ’ ~t~pl ,

vimw !elI stable I trigger blower motors

higher priority of Z over Y. There are severalgood points of this mechanism. First, it doesn’tincrease the amount of computation, as it onlyadds a network of inhibition and suppressionconnection wires. Second, debugging eachemotion is easy, as each emotion is completelyindependent and can be tested separately. Using amacro function of the Behavior Language, thenetwork of the connection wires can be generatedautomatically.

~3ehavior X[

IBehavior Y]

]Behavi°r Z I

Emotion A l

L2® (

~um°ti°n . ]ppressmg

ehavior

Emotion Benablinginput

~ --~Actuator

Figure 4. Foraging behavior (Vacuum cleaningbehavior).

Figure 5. Emotion switching mechanism usingconnection wire network.

In the process of programming therobots, we prepared several versions of theprogram which have a different combination andprioritization of behaviors. Basically each versionhas the same repertoire of behaviors, but theconnections between behaviors and actuators aredifferent. This difference of connections betweenbehaviors gives the robot a different tendency ofinteraction to the environment. From the point ofview of programming the robot in a creature-likeway, we call each combination of behaviors as anemotion of the robot.

In order for the robot to have severalemotions, which appear and disappear accordingto the situation, an emotion switching mechanismis implemented. The emotion switchingmechanism is realized in behavior language asshown in Figure 5. The connections of eachbehavior are prewired and form a network ofconnection wires. Their connection to actuatorsare usually inhibited by emotion suppressingbehavior. Once some inputs are given to emotionenabling input nodes A or B, this inhibition isremoved and the emotion appears on the robot,as shown in Figure 6. When messages are sent tonode A, behaviors X and Z can control theactuators with higher priority of Z over X. On theother hand, when messages are sent to node B,behaviors Y and Z can control the actuators with

]Behavior X]

[Behavior Y[

[Behavior ZI

i) When Emotion A isenabled.

Actuator

]Behavior X]

]Behavior YI

[Behavior Z~"-l

ii) When Emotion B isenabled.

[-------~" Actuator

Figure 6. One emotion is selected by messagessent to the emotion enabling input.

The code example of "joy" emotion islisted in List 1. The keyword :motor-behaviorsintroduces the priority and combination of motionrelated behaviors. In this case, the highestpriority is the escape-stall behavior whichprevents motors from overheating by stalling.The keyword :blower-behaviors introduces thepriority and combination of vacuuming relatedbehaviors. The keyword :connections introducesthe connection wires which should be enabledonly when this emotion is active. These wires areused to connect a behavior to a virtual sensorbehavior. The virtual sensor behavior is

120

responsible for processing sensory data intomore understandable data [4]. By changing thevirtual sensor behaviors connected, the behaviorcan effectively change its reaction withoutmodifying its own function. As a result, in "joy"emotion, the robot concentrates in eating a baitand exploring to other places as long as it canfind a beacon signal. Once it loses the beaconsignal, it tries to go back as soon as possiblealong the way it has come. Based on this "joy"emotion, other three emotions are programmed inslightly different ways. In "desperation"emotion, the robot behaves a little roughlywithout using the obstacle avoidance. In"fatigue" emotion, the robot tries to go home asfast as possible without trying to do anythingelse. In "Sadness" exnotion, the robot gives upfurther navigation and vacuuming, and when itdetects a human, it asks for help by edging up tothe person.

(defemotion Joy

:motor-behaviors(exploresweephomingavoiddead-reckoning

bumpescape-stall)

:blower-behaviors(suck-weaksuck-strong)

:connections((connect (beacon-recover pbl)

(dead-reckoning pb))

(connect (beacon-recover go-dirl)(homing go-dir))))

List 1. Code example of "joy" emotion.

In order to appropriately switch theemotion of the robot, there should be a highermodule over the emotion switching mechanismand behaviors. From the analogy of livingthings, the emotion should be controlled by someinternal state, such as hormones. Action selectionusing hormone-like state variables have beenresearched[8]. The emergent behaviors bycontrolled quantity with disturbers and stabilizerare examined[9]. In both cases, the appropriatebehavior is selected according to the condition ofits internal state variables. These processes aredistributed over behaviors and are robust anddynamic.

Eat-Dust?

Tim e~mfl~~r

Stalled? Enclosed?

~ilizer

Beacon-lost?

@ilizer

Figure 7. Simulated hormone systemcorresponding the four emotions.

In our robot, we utilized this techniquefor selectively activating one of four emotions.Figure 7 shows the hormones corresponding tothe four emotions: joy, desperation, fatigue, andsadness. All hormone values have a disturber,such as "Eat-Dust?." If these condition aresatisfied, the hormone value increases ordecreases slowly. When the robot vacuums thedust, "joy" hormone increases (Eat-dust?disturber). When the battery or preprogrammedoperation time is expiring, "joy" hormonedecreases and "fatigue" hormone increases(Time-up? and Battery-empty? disturbers). Whenthe robot stalls or it detects too many obstaclesaround it, "desperation" hormone increases(Stalled? and Enclosed? disturbers). Finally,"sadness" hormone is increased, when the robotloses the beacon signal (Beacon-lost? disturber).Among these four hormones, the greatesthormone value realizes its emotion on the robot.There are also stabilizers to try to keep thehormone value to an initial level and avoidsaturation. This mechanism prevents anyhormone from always being dominant in certaincircumstances resulting in the loss of robotflexibility. The hormone system usually changesrelatively slowly on the order of several secondsor several tens of seconds, getting influence fromexternal or internal conditions. As a result, theemotion switching doesn’t happen frequently. Soa lower module takes care of fast response andthe hormone system tries to increase the robot’sability to change itself according to the situation.

4. Experimental Results

SOZZY’s navigation capability was testedin a real laboratory environment and videotaped.

121

In each experiment, the robot is programmed inadvance to return to its home (the beacon station)after five minutes. Figures 8 to 11 show how therobot behaved in the laboratory. The movementof the robot was recorded manually usingvideotape. The reason why the robot can go backto the beacon station is that it is programmed notto go out of the view of the beacon station. Evenwhen the robot happens to lose the beaconsignal, it tries to backup its way to the pointwhere it can find the signal.

The hormone condition of the robot wasalso recorded in the real run, as shown in Figure12. In order to record the conditions of thequickly moving robot, we used a radio modem.To obtain a space for the radio modem, we had toremove the vacuum blower and disable thevacuum cleaning ability. As a result, joy-hormone doesn’t change by dust detection. In thefigure, during the first 40 seconds, joy-hormone is dominant. And in the next 20seconds, as we intentionally shadow the beaconstation from the robot, sadness-hormoneincreases. From around 90 seconds,desperation-hormone becomes dominant, as therobot headed into a messy area of cables andsmall equipment. After 120 seconds, as it tries togo back to the beacon station, fatigue-hormonegets high and joy-hormone gets low. The fatigue-hormone and joy-hormone are oscillating in thisperiod because the stabilizers are trying toprevent saturation and are resetting the hormonevalues to their initial levels. Usually, a stabilizergradually forces an unsaturated hormone toconverge to its original level, as seen in thesadness-hormone graph after 60 seconds.

20 ......................................... t ....................................... ~ .................................Fatigue fI s~ |

| Desperation[C..’Y...!~z ........... l

10

0 6O 120

Time (second)

Figure 12. Hormone variation in the real run ofSOZZY.

As a by-product of hormone-drivenemotion switching, the robot seems to be more

friendly and more lively. For example, during theexperiments, we tried to enclose SOZZY withpieces of wood. First, the robot wanderedaround, but finally it began to try to push a pieceof wood (like a small child to get angry). Next,we moved a piece of furniture and intentionallyobstructed the robot’s view to the beacon station.First, the robot tried to back up by itself to thepoint where it could see the beacon again. Butfinally knowing that it could not find the beaconsignal any more, it changed its attitude toapproach to a human for help (in sadnessemotion).

e position

¯ indicates the dust count on the spot

Figure 13. Dust following using the dust sensor.

To experiment with dust is not easy, asdust distribution in every experiment is different.In order to facilitate the experiment, we haveused paper confetti as simulated dust. Figure 13shows how the robot behaves in an island shapeddust distribution. The experiment was done onthe real robot, and the data of dust count, rightwheel velocity, and left wheel velocity areretrieved by a tether and shown onto thecomputer display. White circles indicate thenozzle position of the robot, and black dots showthe dust count on the spot. This figure showswell how the robot follows the dust distributionby a swinging motion. Once the dust isvacuumed, the distribution of dust changes. Thismakes it difficult for the robot to follow the dustdistribution perfectly. But together with therobust navigation ability, SOZZY should be ableto eat up dust.

Programs written in the behaviorlanguage are compact and require minimalcalculation. At this stage, SOZZY’s behaviorprogram occupies 17kbytes of memory, and stillreqmres only 5% of the calculation power of the

122

68HC11 CPU running at 8MHz. The prototypealso contains two slave CPUs for motor servocontrol and sensor processing. Thisconfiguration of three CPUs makes programdevelopment easy. Judging from thecompactness of the behavior program, it seemspossible to control all of SOZZY’s function withone 68HC11 CPU.

There are several points to improve on therobot. First, the wheeling motor’s power turnedout to be weak compared to the total weight ofthe robot. This makes a delay of motion inresponse to a command. Such a delay in therobot makes programming troublesome. In orderto guarantee the quick movement of the robot, theratio of the motor power and the robot’s weightshould be improved. The vacuuming mechanism,which includes the vacuum blower, vacuumchamber, dust chamber and nozzle, seems to beinefficient. A lot of improvement will be requiredin this field. Sensor configurations should alsobe optimized. In SOZZY’s case, as it uses totallyoff-the-shelf parts, there was a limitation in thesensor configuration. From our experience,bump sensors seem to be most important sensorin the unstructured environment. So in order tomaximize its capability, we think the robotshould be as light as possible (because a lightrobot won’t give damage to others even if itbumps), and the bump sensor should cover allthe surface of the robot (as the bump to a pieceof furniture can happen at any height)

5. Conclusion

A prototype of a hormone-drivenautonomous domestic robot has been developed.This robot is programmed in a behavior basedmanner, and can behave quickly and robustly inan unstructured environment. This robot has fourdifferent configurations of behaviors, which wecall four emotions of the robot, as each of themmakes the robot behave in a characteristic way.The robot also has a simulated hormone systemwhich controls emotion switching. This emotionswitching driven by the hormone system not onlysuccessfully increases the flexibility of the robotwithout increasing CPU computation, but alsomakes the robot more friendly and more lively.Although from the point of view of efficiency oraccuracy, the robot still cannot execute a perfectjob at this stage, these elements of the robotshould add some value to its existence. Using

the features, such as robustness, quickness, andfriendliness, we would like to realize a small,affordable, and enjoyable domestic robot in thefuture.

6. Acknowledgments

I would like to express my great thanks toProf. Rod Brooks, who allowed me to buildSOZZY, IS Robotics Co., who not only suppliedthe robot’s electronics, but also offered technicalassistance, the members of Mobot lab, andMichael Caine, Erik Vaaler and Ron Wiken, whohelped me a lot in building the robot. I also thankmy colleagues in Matsushita Research InstituteTokyo, Inc. who named the robot S OZZY as acombination of "Soji" (means cleaning inJapanese) and "Fuzzy" (Fuzzy control is popular keyword in Japanese home appliances),wishing the robot also to be popular. And finallyI would like to express my thanks to GaryBorchardt and Lynne Parker for smartsuggestions and reviews of my English.

7. References

[1] Brooks, R. (1986) A Robust Layered ControlSystem for a Mobile Robot. IEEE Journal ofRobotics and Automation. Volume RA-2,Number 1.

[2] Brooks, R. (1990) The Behavior LanguageUser’s Guide. Massachusetts Institute ofTechnology Artificial Intelligence Laboratory,A.I. Memo 1227, April, 1990.

[3] Engelberger, J. (1989) Robotics in service.The MIT press.

[4] Ferrell, C. (1993) Robust Agent Control an Autonomous Robot with Many Sensors andActuators. Masters Thesis, MassachusettsInstitute of Technology Artificial IntelligenceLaboratory.

[5] Heider, F. & Simmel, M. (1944) experimental study of apparent behavior.American Journal of Psychology, 57, 243-259.

[6] Jones, J. & Flynn, A. (1993) Mobile Robots.A K Peters, Ltd.

123

[7] Kobayashi, T., et. al. (1991) Systems,Control and Information, Volume 35, Number 8,1991 (in Japanese)

[8] Maes, P. (1991) Bottom-Up Mechanism forBehavior Selection. Proceedings of the firstinternational conference on simulation of adaptivebehavior, The MIT press.

[9] Steels, L. (1993) Building Agents out Autonomous Behavior Systems. The ’artificiallife’ route to ’artificial intelligence’, LawrenceErlbaum, Ass. New Haven.

124