From: AAAI Technical Report FS-96-05. Compilation copyright © 1996, AAAI (www.aaai.org). All rights reserved. TheTAO Project: Intelligent wheelchairs for the handicapped Takashi Gomi Applied AI Systems, lnc. (Add) 340 A4arch Road, Suite 600 Kanata, Ontario, CANADAK2K 2E4 E-mail: gomi@dpplied-,4Lcom HomePage: ht~o://fox.nstn.ca/-aai Abstract AnR&D project to build a series of intelligent autonomouswheelchairs is discussed. A standardized autonomy management systemthat can be installed on commerciallyavailable well-engineeredpowerchairs has beendeveloped and tested. Abehavior-basedapproachwasused to establish sufficient on-boardautonomy at minimalcost and material usage,whileachievinghighefficiency, maximum safety, Iransparencyin appearance,and extendability.So far, the add-on systemhas beeninstalled and tried on twopowerwheelchairmodels.Initial results are highly encouraging. 1. Introduction In recent years, with the conceptof applyingrobots to service tasks [Gomi,92] and with the accelerated rate of aging of the populationbeing reported in manypost-industrial countries, demandfor more robotic assistive systems for people with physical ailments or loss of mental control is expected to increase. This is a seeminglymajorapplication area of service robots in the near future. For the past five years, we have been developing a range of autonomousmobile robots and their software using the behavior-based approach [Brooks, 86] [-Maes, 92]. In my experience the behavior-based approach [Brooks86, 91a][Steels, 93] [Pfeifer and Scheier, 96] [Maes, 92] allows developers to generate robot motions whichare more appropriate for use in assistive technology than traditional Cartesian intelligent robotic approaches [Gomi, 96a]. In Cartesian robotics, on which most conventional intelligent robotics approaches are based, planning for the generation of motion sequence and calculation of kinematics and dynamics for each planned motion occupy the center of both theoretical interest and practice. By adopting a behavior-based approach, I felt, wheelchairs which can operate daily in complexreal-world environments with increased performance in efficiency, safety, and flexibility, and greatly reduced computational requirements can be built at less cost. In addition, improvementsin the robustness and graceful degradation characteristics wereexpected. In the summerof 1995, an autonomymanagementsystem for a commercially available Canadian-madepower wheelchair was successfully designed and implemented. The system looks after both longitudinal (forward and backward) and angular (left and right) movements of the chair as well limited vocal interactions with the user. The results were exhibited in August1995at the Intelligent WheelchairEvent 28 organized by David Miller at the International Joint Conference on Artificial Intelligence (IJCAI’95) held MontrealDespite a very short period (a little over a month), the chair performedremarkablywell at the exhibition. Encouragedbythe initial success, I implemented a three year plan to develop a fully autonomouspowerwheelchair for use by people with various types and degrees of handicap. The intelligent wheelchair project, now called TAOProject, intends to establish a methodologyto design, implement,and test an effective add-onautonomymanagement systemfor use in conjunction with most commoncommercially available powerwheelchairs. In order to demonstratethe principle, the project will build, during its life, an autonomymanagement system for four well-established electric wheelchair models currently available on the market throughout North America and Japan. In late 1995, a sister R&D companywas established in Japan exclusively for the development of intelligent robotic technologiesfor the disabled. Withthe initiative of this new R&Dgroup, the development of TAO-2 autonomous wheelchairbegan in the spring of 1996. Based on the experience gained in the past year, methods used and some issues related to the application of the behavior-basedapproach to realize an intelligent wheelchair and possibly other assistive technologies are discussed. 2. Desirable characteristics of robots for the handicapped 2.1 Background Since around 1992, AAIbegan a number of exchanges with people with various handicaps and Individuals whoassist them. This was proceeded by a few years of on-going Interactions with the handicapped community through marketing,installing, servicing, and training individuals on a speech-to-text voice interface system for computers. This device provedto be effective for people with several types of handicap, particularly for individuals whohad lost arm/hand usage. Since late 1995, voluntary work has been attempted by membersof AAIat two institutions for the mobility handicappedin Japan: a senior citizen’s hospice for severe physical/mental problems,and an institution for people with severe physical handicaps. A considerable amountof time practising physical assistive workhas been carried out by membersof the R&Dteam, including the designer involved in the conceptual design of the robots, engineers and a technician responsible for the construction of the robots, and the project manager and administrators of the robotics projects. In early 1995, an individual with a severe physical disability (a quadriplegie) joined AAIas a regular data entry/bookkeepingclerk and as a future tester of autonomous wheelchairs. Basedon these exposures, as well as earlier volunteer work by myself and someof mycolleagues, a preferable approach to robotics for service tasks [Gomi,96b]and a tentative list of deskablecharacteristics for future robots built for the purpose of interacting directly with severely handicappedor fully disabled individuals has been compiled. Some of the desirable characteristics are discussed below. 2.2 Softness and flexibility Establishment of rapport between the handicapped person and the earegiver is essential for a care to be successful. So muchso, there will be a great deal of anxiety in the mindof those treated by future robotized arms, support boards, and wheels.Theneedfor set’messrealized at the physical interface of the end effectors of such a robot and the humanbody surface or limbs does not stop at simple paddingof otherwise solid effector surfaces, or use of softer materials, or passiveor active compliance of effectors. The softness must also be architectural in that the entire physical supportstructure must be able to alter, reconfigure, and evencompletelyrestructure momentto momentreactions and responses to accommodate, whenevernecessary, changesinnot only the physical but also the perceivedpsychologicalsituation of the user. Theflexibility of the systemas a whole,as well as that of the end effectors, must essentially comefrom this "structural softness". The flexibility must be foundedon the opennessof the design of motionsthe systemcan generate so that it does not rely on fixed modesof operation or rigid scenarios defined a priori. Humans in general and in most circumstances behave without a prepared set of motion patterns, and since we are dealing with such an existence, a man-madesystem itself must not act with a fixed set of motionswhichare algorilhmieally describable. This places the appropriateness of most existing systemcontrol methodsin doubt as a tool to seriously deal with manytypes of physieaUy handicapped people. Learning has often been hailed as a schemewith which a 29 system can be mademore adaptable. Wewouldalso have to question this relishable notion as a candidate that would sufficiently increase adaptability of systems such as service robots dealing directly with humans. Learning schemes, particularly those so far studied to a great extent and depth in symbolic AI community, have failed to make significant contributions to robotic systems operating in highly dynamic application areas. In general, learning research has focusedon methodsto improvethe chosen performanceindex of systems but variables involved in the schemeare most often not groundedthrough sensors or actuators. 2.3 Fail safe and robustness A robot arm holding a fragile humanbody cannot drop the person whena bugis hit for the first time. Theconceptof fail safe implies readiness of a systemagainst possible failure. In traditional system engineering disciplines, such as Fault Tolerant ComputerSystems(FTCS)research and practice, this typically translates into the preparation of additional capabilities in the form of a standbyin computerhardwareand software. The concepts of hot-standby and cold-standby are commonly employedin system design. Since it is impossible to prepare for every possa’ole failure, the provision of readiness should exist, however, more in the form of capabilities spread across the system in atomic form and meshedfine grain with the competencestructure which also functions in the normalexecutionof tasks. This is analogous to the wayreadiness to failure is implementedin life forms found in nature. If a small animal or an insect temporarily loses the use of a limb, it tries to adjust to the situation by immediatelyenlisting the use of other limbs or even other portions of the body. Theadditional capability readied in this formwouldbe quickly organized and mobilized the moment a fault is detected. 2.4 Graceful degradation A cousin to the conceptof fail safe, graceful degradationis more important in systems that physically interface with humans~hanin systemsthat deal with materials and artifacts. A conlrol system designed as a monolith or componentswith relatively larger granularity would have less chance of realizing the concept fully. Whenone loses a limb, the resulting transition is not smooth,causing great suffering to the individual However,as we lose a large numberof brain cells every day that we knowwon’t reproduce, we do not deteriorate or lose capabilities as drastic as loosing a limb. Systemscomposedof fmer grain active units seem to offer moredesirable results. 2.5 Evolvability Anotherreason for the failure of learning in symbolic AI wouldbe the relatively short time the methodshavetypically tried to achieve the"resulf’. In fact, weprobably do not know what desirable results are as muchas we think we do. Both shortcomings,this and the lack of grounding, are due mostly to the very nature of being symbolicrather than pragmatic. In evolution, changesoccur along a muchlonger time scale. In situated andembodiedsystems, such as life formsin nature and well-built autonomousrobots, a search through a very high dimensional space of the real world for adaptation demands"experiments" on a vast numberof combinations of dimensional parameters, if such dimensionalization or parameterization makesense at all. Evolutionary Robotics (ER) is an emergingfield of science and technology[Harvey, 92], wherephysical or virtual robots’ autonomy structures are evolved to achieve collective trans-generational learning. ER seems to be a schemethat could well be applied to robots operating to tend and care for humansbecause of the open nature of humanautonomyand ER’sbasic principle that can provide long term learning. Here, the concept of learning should probably be replaced by a more comprehensive conceptof evolution, whichimplies perpetual adaptation of an autonomoussystem to a constantly changing operational environment rather than optimization of one or more performanceindices of such a system. 2.6 The development plan The development of autonomous wheelchairs at AAI is carried out in the following four phases. Someof the phases overlap in their execution. (1) The basic safety phase, (2) The mobility phase, (3) The humaninterface phase, and (4) The exploration phase. Currently, we are in the secondphase of the project which beganon April 1, 1996.Prior to the start of the project on July 20, 1995, a study was conducted to identify various requirementsby potential users of the autonomouswheelchair both in Canadaand Japan through interactions with people with various types of handicap. Causes of the handicaps we cameacross included gradual mobility loss by aging, recent sudden loss of body control due to brain damage, and prolonged motion limitations and bodily contortion due to stroke suffered at ayoungage. Theproject continues to enjoy cooperation from institutions for the handicapped and individuals with disabilities. The TAO project is scheduledto end in the summer of 1998. For a description of the developmentplan, please refer to [Gomiand Ide, 1996]. 3. Implementationof the first prototype, TAO-1 A regular battery poweredwheelchair (a motorized chair) produced and marketed in Canada (FORTRESS Model 760V) was used as the base of the first implementation of the concept. A set of sensors, a computerized autonomy managementunit, and necessary harnesses were built and added to TAO-1(Figure 1) through the summerof 1995. 3O 3.1. Planned functions of the chair The selection of functions to be implementedon TAO-1was somewhat influencedby the rules for exhibition set out for the IJCAI’95robotics contest. However,the later demonstrations of our prototype and observations madeat an institution for the aged confirmedthe guideline was in fact appropriate. Of the following functions which we nowfollow, only the fLrst two were attempted at our IJCAI’95entry. However,all free of themare currently pursued. (a) Basic collision avoidance This is achieved by behaviors which monitor and respond to inputs from on-board ted camerasor those which respond to active infrared (IR) sensors. Whenthe chair encounters obstacle, it fkst reduces its speed, and then dependingon the situation it faces, stops or turns awayfrom the obstacle to avoid hitting it. Theobstacle can be inanimate(e.g., a column in a halhvay, a light pole on the sidewalk, a desk, a standing human)or animate (a passerby, a suddenlyopeneddoor in its path, an approaching wheelchair). Encountering a moving obstacle,the chair first hies to steer aroundit. If it cannot,it stops and backs off if the speed of the advancingobstacle is slow enough(e.g., 20 centimeters per second). Otherwise, stays put until the obstacle passes away. Thus, if the chair encountersanother wheelchair, both chairs can pass each other smoothlyas long as there is enoughspace in the passage for two chairs. A fast paced humanusually does not affect the chair’s progress and at most causes the chair to temporarily slow downor steer away. (b) Passage through a narrowcorridor Whensurroundedby walls on each side of the path, as in a hallway, the chair travels autonomouslyfrom one end to the other in parallel to the walls. (c) Entry through a narrow doorway The chair automatically reduces its speed and cautiously passes through a narrow doorway which mayleave only a fewcentimeters of space on each side of the chair. Sometypes of ailment such as Parkinson’s disease or polio often deprive a humanof the ability to adjust the joystick of a power wheelchairthrough such a tight passage. (d) Maneuverin a fight comer Similarly, whenthe chair is surroundedby obstacles (e.g., walls, doors, humans), it is often dil~cult to handle the situation manually. The autonomouschair should try to fmd a break in the surroundings and escape the confinementby itself unless instructed otherwiseby the user. (e) Landmark-basednavigation Twoccd color cameras on-board the chair are used for functions explainedin (a), (b), and (c) above. Theyconstantly detect the depth and size of free space aheadof the chair. The cameras are also used to identify landmarks in the environment so that the chair can travel from its present location to a given destination by tracing them. An on-board topological mapis used to describe the systemof landmarks. 3.2 Hardware Structure As a standard powered wheelchair, model 760Vhas two differentially driven wheels and two free front casters. Although they are designed to rotate freely around their vertical and horizontal axis, these casters typically give fluctuations in delicate maneuvers due to mechanical hysteresis that exists in thembecause of design constraints (the rotating vertical shaft of the supportstructure of the caster cannot be at the horizontal center of the caster). This sometimescauses the chair to wiggle particularly whenits orientation needsto be adjusted f’mely. Suchline adjustments are necessary typically whena wheelchair tries to enter a narrow opening such as a doorway. Theentire mechanicaland electrical structure, the electronics, and the control circuitry of the original powerwheelchair were used without modification. The prototype autonomy managementsystem still allows the chair to operate as a standard manuallycontrolled electric wheelchair using the joystick. The joystick can be used anytime to seamlessly override the control whenever the user wishes even in autonomy mode. Physical additions to the chair were also kept to a minimum. AI components added to the chair were made visually as transparent as possible. Twoprocessor boxes, one for visionbased behavior generation and the other for non-vision behaviorgeneration are tacked neatly underthe chair’s seat, hiddencompletelyby the wheelchair’soriginal plastic cover. Sensorsare hiddenunderthe footrests, inside the battery case, and on other supporting structures. Onlytwoccd camerasare a little morevisible: they are attached to the front end of two armrests for a good line of sight. A small keypad and miniaturetelevision set are installed temporarilyoverthe left armrest to enter instructions and for monitoring. The non-vision behavior generator is based on a Motorola 6833232-bit micro controller. A multi-tasking, real-time operating systemwas developedand installed as the software framewolk.This combinationgave the systemthe capability to receive real-time signals from a large numberof sensors and to send drive outputs to the two motorswhichgovernthe wheels. The chair currently has several bumpsensors and 12 active infrared (IR) sensors whichdetect obstacles in close vicinity (less than 1 meter) of the chair. Signals from the camerasare processed by a vision-based behavior generation unit based on a DSPboard developed by a group at MIT. Vision processingis discussed in Section 5.5 below. 3.3 Software structure Theover-all behaviorstructure of TAO-l is shownin Figure 3. Smaller behaviors are lumpedup to save space on the diagram~Softwarefor the vision systemis also built according to behavior-based principles. The major difference between this and conventionalimageprocessingis that it consists of behaviors, each of whichgenerates actualbehavior output to the motors.It can presentlydetect depthand size of free space, 31 Figure 1 Autonomous wheelchair TAO-1.Cover is removed to showautonomyunit. vanishing point, indoor landmarks, and simple motionsup to 10 meters ahead in its path. Indoor landmarksare a segment of ordinary office scenery that naturally comesin view of the cameras. Nospecial markings are placed in the environment for navigation. There are also a large numberof behaviors invoked by IRs and bumperswhich collectively generate finer interactions with the environmentVision-based and non-vision behaviors jointly allow the chair to proceed cautiously but efficiently through complexoffice spaces. Note that there is no main programto coordinate behaviors. ~tly, the autonomy program occupies about 35 KBytes for all ofthe vision related processingand 32 KBytesfor other behavior generation and miscellaneous computation. Of the 35 KBytes for vision related processing, only about 10 KBytesare directly related to behaviorgeneration.Therest are involved invarious forms of signal preprocessing: generation ofdepthmap,calculation of the size of free space, estimation of the vanishingpoint, and detection of specific obstaeles in the immediatefront of the chair. Of the remaining 25 KBytes, approximately 20 KBytes are used in the neuralnetwolksystemfor detecting landmarksand referencing atopological map. The current implementationof the landmarksystem consumesonly 256 Bytes per landmark, although this figure may change in the future as more sophisticated landmarkdescription might becomenecessary. The current systemhas space for up to 64 landmarksbut this can also be adjusted in the future version. Of the 32 KBytesof non-vision processing (i.e., processing of inputs from IR’s, bumpsensors, voice I/o, etc.), again no morethan several KBytesare spent for generating behaviors. Altogether, there are some 150 behaviors in the current version of TAO-1.A considerable amountof code has been wlitten to deal with trivial periphery,suchas keypadinterface, voice I/o, and LCDdisplay. The comparableinefficiency of ceding is because these non-behavioral processing had to be described in more conventional algorithms. for a public demonstration in the town of Kosaka, Akita Prefecture. 4. The second prototype, TAO-2 Encouragedby the success of TAO-1,in late 1995 a sister company of AAI (AAI Japan, Inc.) was established in northern Japan. AAIJapan is dedicated to the developmentof advanced intelligent robotics to aid people with various handicaps. In May1996, AAIJapan purchased anew power wheelchair (Suzuki MC-13P),which is a model widely used in Japan. MC-13P has a form of powersteering in which the twofront casters alter their orientation in synchronywith the drive wheelswhena turn is indicated by joystick. The servo controller also halts the inside turn wheelof the two drive wheels while the chair is makinga tight turn. This is a significant departure from the way the FORTRESS model makesa turn. The latter simply turns the two differentially driven mainwheelsin opposite directions, allowingthe chair to turn in place. The intent of providing a powersteering feature on the Suzukichair is obviously for ease of use, and the user is freed from the wiggly caster problemdescn’bed above. However,this prevented the chair from makingturns in a tight turn circle. The feature wasfelt undesirablefor an autonomouschair. Immediatelyfollowing the purchase of the Suzukichair, the development team began building an autonomymanagement system for TAO-2;a new prototype autonomouschair based on MC-13P.The over-all computer hardware and software structures as well as sensors are almost identical to those for TAO-1, except for a few changes listed below to accommodate the above mentioned and other minor differences in characteristics. (1) The behaviors responsible for turning TAO-2needed their parametersadjusted. (2) The locations of touch sensors madeup of lhin piano wires needed to be moved forward in order to compensate for a larger turn circle. (3) The back bumperwas not activated since it was hardly used. The difference in turning characteristics reduced the chance of the Suzuki chair performing frequent switch backs. (4) Two prominent side bumpers were added to protect bystanders when the chair makes a turn in their directions. This wasnecessitated by the lack of structure on which to mountsensors. TAO-2is shownin Figure 2. It was fitted with the autonomy management system at AAIin Canadain the span of a week. After twodays of testing, it wasshippedbackto Japanin time 32 5. Evaluation of the prototypes 5.1 Demonstrations WhenTAO-1was demonstrated at IJCAI’95 in Montreal on the 22rid of August, itwas the 33rd day of the developmentof the first prototype. Everythingfrom the motherboard,vision system, sensor arrangementsand their harnessing, operating system (based on an earlier prototype), a large number behaviors (some 60 by that time) were all developed and Figure 2 TAO-2autonomouswheelchair tested in that period. The chair could performfunctions (a) and (b) in Section 3.1 well and functions (c) and moderatelywell although they were not initially targeted. Function (e) wasnot yet implemented.In all, it performed well as other chairs at the exhibition mostof whichtook much longer time to develop. All five functions are now implemented on TAO-1and are undergoing continuous improvement. TAO-2was demonstrated on June 4, 1996 at a gymnasiumof a local school in Kosaka. The chair ran smoothlythroughout the 1 hour demonstrationpersistently avoiding by-bystanders, other obstacles and the walls. Unsolicited, a severely handicappedspectator whocould not even reach the joystick volunteeredto test ride the chair. The chair performedto her satisfaction and excitementas it went through the gymnasium amonga large numberof spectators. The success of the twoprototypes suggests that myintention to build a standardized add-onautonomyunit is a valid one. The concept has at least been provenin two powerwheelchair types whichcomefrom drastdcaI~ different backgrounds.The divergence in design philosophy and practical variances in ~] Speaker Voice handler /~ Joystick r con~’o]ed) vlslonsystem Cameras ~a Vanlshrng po~t I Motors I @’°"’, Depthmap r~s- Frontovoid L LCD j/ Cameral Keypad ~Itor I "- IR~ B4 ,~~Rll 1 12 B3 B’left~ ’~" --IR 10 -IR 3 B2 B-right IR 9 ¢ B5 Camera2 Sensor arrangementfor TAO-1 Figure3 TAO-1 behavior structure(Notall behaviors arcshown) implementation,some fairly significant, of a base power wheelchaircan be absorbedby relatively minorhardwareand software alterations madeon the standardizedadd-onunit. TAO-2also showedthat the installation, testing, and adjuslmentof a separatelybuilt autonomy unit canbe madein 33 a very short period of time. In both TAO-1and TAO-2,no cooperationfromthemanufactureswas sought. In eaoh ease, charaotedsties of the joystick werestudied anda seamless interface wasdesignedaroundit. 5.2 Development time The extremelyshort development time required for the initial prototype for both TAO-1and TAO-2can largely be attributed to the behavior-based approach. To achieve the demonstratedlevel of mobility and flexibility wouldnormally have required another several months to a few years in conventional AI-based mobile robotics. In behavior-based robotics, the operationalcharacteristics of the sensorsneednot be as precisely uniformas in conventional mobilerobotics. For example, emission strength and angular coverage of the emitter, and the sensitivity and shapeof the reception coneof the receptor of on-boardIR sensors need not be homogeneous across all sensors, allowingthe use of inexpensivesensors and simplertesting. All sensors, including the ccd cameras,neednot be installed at precise translational and angular coordinates. Theyalso do not need calibration. They were placed on the chair in a relatively ad hoc mannerat First, and continually moved aroundfor better results as the development wenton. In fact, the camerasand someof the sensors are attached to the chair by velcrodetachabletape, so that their location and orientation can be adjusted easily. Suchloose treatmentof sensors is not common in conventional robotics where robot’s motions are derivedafter high-precisionmeasuremeuts of the relationships betweenits extremities and environment.The large tolerance for signal fluctuation is due also to flexibility of processing and greater adaptability inherent in Subsumption Architecture [Brooks,86]. With the absence of "sensor fusion", sensor inputs are directly linked to motor output only with a simple signal transformation and amplification (e.g., from sensor output voltage to motordrive current). The developer only needs to adjust the appropriateness of the definition and performance of the sensor-action pair or behavior in terms of its output without a detailed and precise analysis of input signal characteristics and elaborate planning and computation of output signal generation. Readers not familiar with the theoreticalbasis of behavior-basedAI are encouragedto read [Brooks,91b]. Thesetheories are fuUyput into practice in our development. 5.3 Software structure During development,sensor-actuator pairs or behaviors are simply "stacked up". Theyare addedto the system one by one without muchconsideration for the design of the over-all software structure. Our operating system provided an adequate frameworkfor the incremental developmentprocess allowing for shorter developmenttime. Thus, software developmentwent totally incrementally side by side with finer adjustment of the sensors. Only general functions neededassigned to each sensor-actuator pair type first. For example,depth map- motorpairs are excellent for dealing with obstacles that suddenlyappear in the path of the chair a few meters away. But the same sensor-actuator pair type is not at all effective for the management of the situation in whichthe chair has actually madephysical contact with an obstacle. Sometimes,competition or contradiction occurs betweentwo or morebehaviors. Suchcontradicting definitions of behavior areinmost cases easily observable and corrected quickly. An example of more complex contradiction occurs when two IR collision-detection sensors placedon left and right front sides of the chair detect an approaching doorway in quick succession~ Since the doorwayis normally quite narrow, the .... : offnebul[d~ m ::~:::::: :~.................... ~:: ....~iiiiiiii~ii~ ::iiiiiiii~iiiii~iiiiiif ~iiiiii:J ============================== .... ® ::::i::iii::iiiiiiiiiiiiiii::i::i~ "%~iiiiiiiiiiiiiiiiiiii ........ ~:::::: cow .... ~iiiiii~i ...... ~#i::i::i::i~::~:: o9 %iiiii _ ============================================================== ....~~ Figure 4 Theoffice space whichcontains the test loop 34 reflection of infrared signals received by these sensors is usually strong enoughto cause the chair’s immediate evasive action. As both sensors react alternatingly, the chair can get into an oscillatory motion, commonly known as "Braitenburg’soscillation" after [Braitenburg, 84]. In this specific situation, other frontally-mounted IR sensors take in "just go ahead" signals that invoke behaviors which can breakthe tie. 200 0 200 0 t~ a: 2O0 0 200 0 IR4 ".51--~j1 ~jAl~l’~’--°---~e ~----~j~l~ I R3 flv _IR2 200 0 200 IR1 I 0 (11 (21 I (3) I (4) (5) (6) Figure5a Outputof active infrared OR)sensors runningfast, medium,and slow speed, respectively. In another 5.4 Priority scheme modeof obstacle detection using an IR, changesin value are The priority arrangement is shownin top fight comerof monitored for several sensor cycles. If the change is Figure 4, whereseveral output lines to the motorsare joined sufficiently large, detection of an obstacle is reported. TheIR by O nodes or suppression nodes. Input from the left of the sensors take invalues at 64Hzand several consecutive values node is suppressed and replaced by one comingin vertically are compared.Onceinvoked, a behavior corresponding to a wheneverit is active. Inputs fromthe joystick take the highest specific IR sensor generates a predeterminedreactive motion, priority in deciding whichaction the chair should take. The altering the speedand orientation of the chair. electronically and mechanically seamless interface between the joystick controller and the autonomymanagement system allows the chair to nmas a standard powerwheelchairsimply 5.5 Vision processing by operatingthe joystick. Thesecondhighest priority is given Inputs from 2 cc, d cameras are alternatively processed to behaviorswhichtake in signals fromleft and fight bmnpers through a single flame grabber into two primary vision planes and somekey frontal IR sensors. Behaviors are bundled up of 256 x 128 pixels each at about 8 frame sets per second. in Figure 4 with implied logical relationships amonginput Imagesin these primary vision buffers are averaged downto lines to simplify the diagram. After several groups of 64 x 32 pixel secondary vision plane by combiningthe left behaviorsthat mostlydependtheir invocation on signals from and right vision inputs after dividing each primaryplane into IR sensors, behaviorsinvokedby signals from the voice input left, center, and fight. All vision processingdescribed below systemare arrangednext, followedby vision-driven behaviors occur using imagedata in this secondaryvisual plane. as the lowest priority behavior groups. They are, in Figure 5b plots three depth values (Y1, Yoand Yr) in terms descendingorder of priority, depth map,vanishingpoint, and of the number of pixels in the secondary visual plane determinedaccording to HorswiU’shabitat constraint vision free area. Figure 5a showsIR signals from a test run in which TAO-1 processing [Horswill, 92]. In the absence of active bumper wentaroundthe test loop in our office floor shownin Figure and IRinvokedbehaviors, the parameterset directly dictates 3 (shaded area). Note that signals from only 6 of the 12 the orientation and speed of the wheels. sensors are plotted here. The x axis in Figures 5a # pixies [3 o-5 [] 5-10 ¯ 10-15 [] 15-20 [] 20-25 [3 25-30 through 5d shows the passage of time and its length corresponds to the time required to complete the loop from the workshopwhere the chair began its run and back there counterclockwise. Note that checkpoints (1) through (6) shown in Figure 3 are also marked on the diagrams. Whenthere is no reflection from an obstacle, output of an IR is kept at 255. 111 121 (31 (41 (51 (6) Depending on the strength of the reflected signal, Figure 5b Depthmapparametersfrom the vision subprocess a receptor mayreport lower values, 0 being the lowest. Whenthe value becomes less than a threshold, the sensor wouldhave "detected an obstacle." The Outputfromthe vanishingpoint detector of the vision system thresholdis set as a functionof speedof the chair, andin this is shown in Figure 5c. The detector attempts to find a specific test set at 210, 180, and 150, for whenthe chair is vanishingpoint in the secondaryvisual plane and outputs its IIIllllNliilli llt ll IlJl 35 domestiotranaportation needs. In April 1996, TAO-1was brought to a local shoppingmall in Ottawa to freely roam around for an hour or so. TAO-1 skilfully skirted all internal structures of the mall such as esoalators, flower pots, benches,signs, and showcases,as well as allemoun shoppers. TAO-1 andits rider visited stores as if lOO ~C~O 00~ ~ O O~ ¢~> he was window shopping or X 9o Bestfit lines sloped sameway just strolling the mall. Virtually Invalid Vanishing vanishing point set to appropriate point- use Area to 8o all fellow shoppers failed to edge turn 70 notice that it was not driven "6 Q. manually. This made me feel 6O Mid point @ t:D e°_ that with proper engineeringto e5O 0 make the chair further to C 0 4O 0 t~ o ~>o dependable, it can already > 30 O O serve as an autonomouschair O o~o t20 O in limited application areas, O lO O O such as strolling or window O O 0 ~ zt~D 0 ~> IX. shopping for the severely (3} (4} (5) t6~ (1) (2) handioapped. After the modest Figure 5c Output of vanishing point detector demonstration of TAO-2in Japan, which was reported in local television newsand several looal newspapers,we have Figure 5d depicts output from the area deteetor. The number receivedinquiries for the chair’s availability. Needlessto say, of pixels representingfree space in the left, center and right it wouldbe at least a few moreyears before even a modestly visual fields are calculated by the detector, and steering and autonomouschair can be released for use by the handicapped speed of the chair are determinedby the size of available at large and put into daily use only with affordable amountof space as in depth mapprocessing. The behaviors associated support. Maintenancewouldbe another issue if we proceed, with area detection are invokedonly whenall other behaviors not to mention various public liability issues that, are not invoked. unfortunately but undoubtedly, foUow. I am not at all As the project proceeds the vision systemwill be enhanced optimlstio about the efforts required to establish an to detect moreobjects and events such as outdoor landmarks, infrastructure for physical and moral support that indoor landmarks that change in time, more complex and encompasses all these and other yet to be found issues. dynamicobstaoles, and traffic signals in the path. In the fall Nevertheless, I can foresee that we will be able to answer,in of 1996, a popular Americanwheelchairwill be fitted with the the near future, someof the sincere wishes that already come autonomy managementsystem to become TAO-3. from people whowouldbe most benefitted by the technology. Gettinginto technioal issues, the list of things yet to be done is still quite long. Landmark # pixels ¯ 400-600 [] 0-200 [] 200-400 detection, for example, requires a lot more left work. Although we have su~ed in navigating the chair to go through a series of landmarks arbitrarily chosen in the chair’s present operational environment,this is still a far cry frombeing able to state that it can run freely in any environment traversable by a 141 151 11) 12~ 131 wheelchair by detecting landmarks. Apartfrom these and other shortcomings,I feel Figure 5d Output of the area detector the technologyas it is, is alreadyuseful in real world applications by individuals with certain types of 6 Lessonslearned so far fromthe chair project handicap. Persons with bodily contortions such as those who Althoughthe experienceis stiU very limited, I can state that suffered polio in earlier life, or individuals with involuntary there is a strong expectation amongthe population for the hand/annmovements such as patients of Parkinson’s disease, developmentof an autonomouswheelchair for assisting and nowcould travelthrough confined and narrow spaces such as eventually fully taking care of the handicapped person’s corridors and doorwayswithout assistance. Other interface x axis value whenit finds one. Thevalue 0 correspondsto the left mostangle in the visual plane, and 63 to the right most. Whenit fails to comeup with a vanishing point, value 99 is output. Thecombinedhorizontal viewingangle of the left and the right camerasis approximately100 degrees. ~: ~ 36 mechanismssuch as neck control and voice recognizer would also makethe inlroduction of the autonomouschair easier. Less handicappedusers can use the chair as a manualpower wheelchair whenever desired, while autonomymanagement can assist in mundanesituations and emergencies. Everybodywe have interfaced with so far, from a passer-by at a shoppingcenter whereTAO-1was tested, fellow robotics researchers, several handicappedpeople and caregivers who heard about the project and cameto see and even volunteered to try an early prototype,willing investors, to journalists all gave us positive feedback. They agree in principle that mobility should be provided as much as and as soon as possible to those whootherwise are not capable of going to places by themselves. Althoughthe developmentis still far from complete, TAO-1and 2 have so far been covered by several TV programs and a few dozen newspaper and magazine articles in Europe, Japan, USA, and Canada, indicating the keen level of interest the public has on this subject. I wishto report our next interim results at IJCAI’97(Kyoto, Japan) and other occasions aroundthat time so that results can be shared amongresearchers and the concept of the autonomous wheelchair which is manually overridable becomesfamiliar to the populationat large. Needsof the users and capabilities that could be providedby the technologyare matchedand better understood through frequent disclosures through technical papers, presentations, and serious media coverage. 7. Conclusions Two prototype autonomous wheelchairs based on commerciallyavailable motorizedwheelchairshave been built using behavior-based AI. The initial prototyping went very rapidly and the size of the software is significantly smaller than control programsfor similar vehicles operating in real world environment implementedusing conventional AI and robotics methodologies.Oneof the chairs is nowcapable of travelling to its indoor destinations using landmark-based navigation~The performanceof the prototypes indicates there is a cautiouspossibility todayto build a functionalintelligent wheelchairthat is practical and helpful to peoplewith certain types and degrees of handicap. Acknowledgements KoichiIde is responsible for the detailed design and most of the implementation of both TAO-1and TAO-2. Reuben Martin and Richard Edwardsassisted Ide in implementation and testing of both chairs. AnnGfifl~th proofread earlier manuscriptsof this paper and aided the author in editing. References [Braitenberg, 84] V. Braitenberg, "Vehicles: ExperimentsIn Synthetic Psychology", MITPress, Cambridge,MA.1984. 37 [Brooks, 86] 1L A. Brooks, "A Robust Layered Control System For a Mobile Robot", 1EEEdournalof Robotics and Automation, RA-2(1986). [Brooks, 91a] 1L A. Brooks, "NewApproachesto Robotics", Science 253, 1991, pp 1227-1232. [Brooks, 91b] 1LA.Brooks, "Intelligence WithoutReason".In Proc. IJCAI-91, Sydney, Australia, August1991. [Gomi, 96a] Takashi Gomi, "Non-Cartesian Robotics", Robotics and AutonomousSystems 18, 1996, Elsevier, pp 169-184. [Gomi, 96b] T. Gomi, "Aspects of Non-Cartesian Robotics" in proc. Artificial Life and Robotics (AROB’96),Japan, February 1996. [Gomi, 92] T. Gomi, "Subsumption Robots and the Application of Intelligent Robotsto the Service Industry", AppliedAI Systems,Inc. Internal Report, 1992. [Gomi and Ide, 96] T. Gomi, T., and K. Ide, "The Developmentof an Intelligent Wheelchair". In proc. IVS’96, September1996, Tokyo, Japan. [Gomi et al, 94] T. Gomi, K. Ide, H. Matsuo, "The Development of a Fully Autonomous Ground Vehicle (FAGV), Intelligent Vehicles Symposium(1V’94), Pads, France, October 1994. [Harvey., 92] I. Harvey, "Evolutionary Robotics and SAGA: the case for hill crawling and tournament selection", In Artificial Life 3 Proceedings, C. Langton, editor, Santa Fe Institute Studies in the Sciences of Complexity,Proc. Vol XVI, pp 299-326. [HorswiU,92] I. Horswill, "A simple, cheap, and robust visual navigation system", Simulation of Adaptive Behavior (SAB’92), Honolulu, Hawaii, 1992. [Maes, 92] Pattie Maes, "Behavior-based Artificial Intelligence", SecondInternational Conferenceon Simulation of Adaptive Behavior (SAB’92), 1992. [Pfeifer and Scheier, 96] 1L Pfeifer, C. Scheier, Introduction to "NewArtificial Intelligence", Ins3itut fttr Informatikder Universit/it ZtLrich,July 1996. [Steels, 93] Luc Steels, "Building agents out of autonomous behavior systems", The Biology and Technologyof Intelligent Autonomous Agents, NATOAdvanced Study Institute, Trento, 1993, Lecture Notes.