From: FLAIRS-01 Proceedings. Copyright © 2001, AAAI (www.aaai.org). All rights reserved. Exploring UnknownStructured Environments Jonathan F. Diaz and Alexander Stoytchev and Ronald C. Arkin Mobile Robot Laboratory College of Computing GeorgiaInstitute of Technology Atlanta, Georgia 30332-0280U.S.A. Abstract Thispaperpresentsperceptualalgorithmsfor corridornavigation, doordetectionandentry that canbe usedfor exploration of an unknown floor of a building.Thealgorithmsare fairly robustto sensornoise, donot use odometric data, and can be implementedon a low end PC. Theprimary sensor usedis a laser rangescanner. Introduction Manypractical robot applications require navigation in structured but unknownenvironments. Search and rescue missions, surveillance and monitoringtasks, urban warfare scenarios, are all examplesof domainswhere autonomous robot applications would be highly desirable. Deploying robots in such scenarios is expected to reduce the risk of losing humanlives (Blitch 1999). Onescenario, for example, wouldbe a biohazard detection mission, wherea robot or a teamof robots wouldbe deployed to enter and navigate an unknown building to search and report the existence of any hazardous materials. Although the building layout maybe unknown,it is safe to assumethat the building will consist of several floors, where each floor will haveone or morecorridors, and each corridor will have several roomswhoseentrance doors are located on the corridor. This paper presents several perceptual algorithmsthat exploit available structure to allow mobilerobots to navigate through corridors, detect doors, and enter rooms. In combination the algorithms can be used for exploration of an unknownfloor of a building in the absence of any map. The algorithmsare fairly robust to sensor noise, do not rely on odometry,and are computationally efficient. The primarysensor used is a laser range scanner. The algorithms presented here have been demonstratedsuccessfully at the DARPA Tactical Mobile Robotics (TMR)Demonstration Rockville, Marylandin September2000. Related Work Lu and Milios (Lu &Milios 1997a; 1997b) have previously described algorithms for robot pose estimation with laser Copyright©2001,American Associationfor Artificial Intelligence(www.aaai.org). All rights reserved. data. Theapproachthat they take is registration and matching. Their technique relies on constraint-based search to align a current laser frame with a mapbuilt from previous laser scans. The alignment of range scans is computationally expensiveand can be unnecessarywhenwivigating in a structured environment. Otherresearchers (Foxet al. 1999)haveused laser range scanners for mapbuilding and robot localization. Their approachrelies heavily on the assumptionthat the features of the environmentused for localization are not changingover time. Therefore,it is not robust to changesin the environment(e.g., openingand closing doors) whichmaydisorient the robot causingit to decreaseits localization certainty. Consecutivelocalizations can take time which maynot alwaysbe available in a warfare scenario. The office delivery mobilerobot Xavier (Simmonset al. 1997) was capable of entering roomsin order to deliver or pick up printouts that users requested over a web-basedinterface. Xavier used a neural networkand cameraimagesto positionitself in front of a door in order to enter a room.The wholeprocess was discrete rather than continuous. The approachpresented in this paper for door entry uses a behaviorsimilar to the dockingmotorschemadescribed in (Arkin & Murphy1990). This schemaallows the robot enter the roomrapidly while traveling on a smoothsemi-arc trajectory (see Figure 2). Therobot never stops to makesure it is onthe right track. Mission Specification The MissionLabmission specification system (Endoet al. 2000)was used as a test-bed for the biohazardsurvey experiments (see ExperimentsSection for details). MissionLab allows a robot mission to be specified graphically in terms of a finite state acceptor (FSA). Figure I showsthe FSA used for the biohazard detection mission. Wepresent perceptual algorithms that support the ProceedAlongHallway and GoThroughDoorbehavioral assemblages. An assemblage is a collection of low-level behaviors aggregatedin supportof a task. The behavioral assemblage ProceedAlongHallway is composed of the motor schemas StaylnHallway, MoveDownHallway,and AvoidObstacles. StaylnHallway directs the robot in a direction orthogonalto the centerline of the hallway.It requires as input the widthof the hallwayandthe DECISIONANALYSIS 145 Figure 1: Finite State Acceptordescribing a biohazard detection missionat a highlevel. angle betweenthe centerline of the hallway and the current robot heading. MoveDownHallway movesthe robot parallel to the centedineof the hallway. Theangle to the centerline alone is sufficient input for this motorschema. AvoidObstacles repulses the robot fromobstacles. Eachof the three schemasproduces a vector as its output. The vectors are adjusted accordingto gain values suppliedby the user. The behavioral assemblage GoThroughDoor consists of EnterDoorwayand AvoidObstacles schemas. EnterDoorway uses a docking schema(Arkin & Murphy1990) to compute the instantaneousforce on the robot relative to the door. The schemarequires three points to computethe vector response: two end points of the door to enter and a point on the far side of the door that the robot should movetowards (Figure 2). ========================,: Iii; I J ~l ~//,’///.% "-:.:-....’,:+X,:,’,’, :....-..,,<,,,,,,,,,,:,:, d,ll //,~% ,*.%%#/,%%" f lll lld~d’[~.l¢¢,ll’¢~’¢’, ::::::::::::::::::::: bq,,’,;,/,:.:.-:-:..-.-:-: ,:/.:/.:.: 2~-:-:-t~a,wtenemM -...,,, ;’;,,,,,,,, -%-,.P’PttPf ¯ tl. .:.:.:.:, , ,,,, :,, ,,,,,,,,,,,. ;-.7.:.7.:,l,bi,X,b.,:,. Corridor Detection Webegin with the assumption that hallway parameters (widthand angle withrespect to the robot) can in fact be determinedfroma single laser reading. This is a fairly strong but generally true assumption.Twofactors contribute to this result. First, hallwayshavedistinct features; long, straight, parallel walls. Second,the laser range scanneris extremely accurate~ allowing proper perceptual algorithms to detect these features. 2 200 used in our experiments has a The SICK LMS field-of-view of 180 degrees and returns 361 distance readings (two per degree). These distances are converted software to Cartesian coordinates to form a laser frame F =< pl, P2, ..., Pa61>, whereeach Pl is a point in 2D. A local coordinate systemis imposedon the laser frame such thatpxst is alwayson the x-axis, Pl on the -y axis, and Pa~ on the +y axis. All calculations that follow are relative to the origin of this coordinatesystem.Anglescalculated relative to the sensor are calculatedwithrespect to the x-axis. Since corridors consist of long, straight walls, it is expectedthat manyof the laser rays will intersect these walls, yieldinga framewitha large numberof relevant data points. It only needsto be determinedwhichpoints fall along these walls. Figure 3 showsa laser frametaken in a corridor with two open doors. m .. w leo i0 I ~ ’~’"""’"" -... ° . .tee im Figure 3: ASingle frameof laser scannerdata, positionedat the origin. Figure 2: The docking motorschemaused to enter rooms. To determine which points fall along the two parallel walls a voting schemeis adopted. Eachpoint Pi is allowed to vote for the dominantorientation of the corridor. The points cast their vote by observingthe local orientation, determined by a small neighborhoodof points around them, and declaring it the dominantorientation. The orientation with the mostvotes is declaredthe orientation of the hallway.This is done by fitting a line L to the local neighbor- Perceptual Algorithms Weare motivated by the approach of (Gibson1979) and use affordance-basedpercepts to interpret sensor data. For example,a door for the robot is defined as an affordance for passage;a corridor is an affordancefor navigation. ~Maximum error of +/-3 cmper 80 meters. 2SICK is an industrial sensorengineeringcompany. TheLaser Measurement System(LMS)calculates the distance to an object usingthe timeof flight of pulsedlight. Arotatingmirrordeflects the pulsedlight beamto manypoints in a semi-circle. TheLMS 200has a serial port interface, a peakpowerdemand of 35 Watts, andcosts approximately five thousand dollars. -’:-:.:":,’7.’:. Z,’,;.;’:-b’ //...:.,.7,;,¢,;,;,;,’,,;,:, 146 FLAIRS-2001 hoodof points about pi. A line, L, in 2Dcan be described by twoparametersp and 4, wherep is the perpendiculardistance to the line L and $ is the angle fromthe positive x-axis to that perpendicularof L (Figure 4). ¯ y po ¯ rotated by 180 degrees. In this way,angles correspondingto parallel lines will nowbe overlapping.Figure 5 showsthe histogramfor the laser framedisplayedin Figure3. Step 3. Thehallwayangle is selected. A triangle filter is used to smooththe histogramand the dominantangle in the histogramis selected. This is the angle betweenthe perpendicular to the left-hand wall and the laser scanner. The angle betweenthe perpendicular to the right-hand wall and the sensor is 180degrees larger. This latter angle, labeled H~,is the one used in subsequentcalculations. Thebin size of the histogramwasselected to be twodegrees and the triangle filter usedwas[0.1, 0.2, 0.4, 0.2, 0.1]. Figure4: Bestfit line for a set of points. Aleast-squares approachis usedto fit a line to the neighborhoodof points. A closed form solution exists and can be calculated efficiently (Lu &Milios 1997a). The size the local neighborhoodwas empirically determinedto be 7. Smaller values for the neighborhoodcan cause the lines to be overly sensitive to small changesin contour and noise, while larger values incur morecomputationwith little gain in performance. / Figure 6: Hallwaytransformation parameters. Step 4. Thewinningpoints are divided into two sets. Thewidth of the hallwaywill be estimatedfrom those points that voted for the winninghallwayangle. Thesepoints can be divided into twosets correspondingto the left and right wall: L and R. The set L consists of points pi for which ~i E [0,180]. The set R consists of points Pl for which ~i E[-180,0]. Figure 5: Histogramfor the dominanthallwayangle for Figure 3. Eachbin correspondsto a twodegree interval. Step1. A line is fit througheach datapoint. To be more precise, for eachdata point Pl in a laser frameF, a best fit line Li is foundfor the set of data points {pj E F : i - 3 < j<_i+3}. Step 2. A histogramis calculatedover all ~i. After all (Pl, ~i) pairs have beencalculated, a histogramover ~i computed.For a typical hallway and a laser scanner position within the hallway, the histogramwill have two large humpscorrespondingto the two opposingwalls. Exploiting the fact that the walls are parallel, or near parallel, the two humpscan be mergedif all ~i outside the range [0,180] are Step 5. Thedistance to each wall is calculated using a histogrammethod.A histogram is calculated over pi for each of the two sets. Triangle filtering is again used. The most common value is selected from each histogram. Let dt be the distance to the left-hand wall calculated fromthe set L, and let dr be the distanceto the right-handwall calculated fromthe set R. Step6. Thewidthof the hallwayis calculated. Thenthe widthof the hallway, Hto, is the sumof the two distances: H~, = dt + dr. Step 7. Thelocation of the robot within the hallwayis calculated. The distance betweenthe laser scanner and the hallwaycenterline, can also be calculated: S~ = dl - H,,,/2. The sign of this variable determineswhether the robot is to the right (positive) or to the left (negative)of the centerline. There is nowsufficient informationto navigate along the hallway.Let the laser scannerbe positioned at the front center of the robot and oriented in the samedirection. Then the angle of the hallwaylocal to the robot is -He- 90. The widthof the hallwayis Hi, and the robot is located S=units awayfromthe hallway centerline. DECISIONANALYSIS 147 t. +oo ~o ~l°°-im¯ llml~llm) soo imo "too Experiments The experiments described in this section were conducted during the DARPA TMRDemoin Rockville, Maryland in September2000. The site for our demoportion was inside a firefighters’ training facility, also known as the Burn Building (Figure 9). The conditions inside the building were quite different from the pristine environmentsfound in manyrobotics labs. Mostwalls (both corridor and room walls) werenot precisely straight due to the routine incinerate/extinguish exercises that have beengoing on for more than a decadein that building. Light was very limited and the air wasquite dusty. Figure 7: Alignedlaser scan data. Door Detection Thetechniquefor doorwaydetection relies on the definition of a door as an affordancefor passage. It looks for openings in the hallway that are within someacceptable width. In the experiments,fifty centimeters was used for the minimum doorway width and two meters for the maximum.These numbers were used to accommodateour RWIurban robot (Figure 8). A larger or smaller robot mayhavevery different acceptable ranges of doorwaywidths. Step 1. All points in the frameare transformed so that the hallwaycenterline coincideswiththe x-axis. This step allows for efficient detection of openingsin the twowalls, since the walls will nowbe horizontal. This transformation requires a translation of (cos(H~)¯ S~, sin(H~)* S~=) lowed by a rotation of -H~ - 90, where again H~ is the angle betweenthe laser scanner and the perpendicularto the right-hand wall. The wall positions are now(0, -H~/2) and (0, H,~/2)and the newlaser scannerposition is (0,-S~:). Figure 7 displays the laser frameafter the translation and rotation have been performed. Step 2. Openingsin the wall are detected using a box filter. To detect openingsa box filter of size 50 cmx 50 cmis passed along each wall at 5 cmoffsets. For each filter position the numberof points that fall within the boxis counted. Consecutiveregions with zero points are mergedto forma potential door opening.Thewidth of that openingis checkedagainst the acceptabledoor widthsspecified earlier. Step 3. False positives are eliminated. This methodcan generate false positives, however,due to obstacles located betweenthe laser and the wall. To eliminate this problem the laser frameis checkedfor points that are closer to the laser scannerthan the expectedposition of the wall and fall in the sameviewing angle as the opening. If such points are found the opening is rejected. The coordinates of the two endpoints of the opening(whichcan be represented as a line segment)are tracked continuously, transformedinto coordinates local to the robot, and fed to the EnterDoorway motor schemaduring doorwayentry. 148 FLAIRS-2OOl Figure 8: The RWIUrbanRobot equipped with a SICKlaser scanner and a Sonypan/tilt/zoom camera. Nevertheless, the RWIUrbanrobot (Figure 8) was up the test. Its missionwas to start fromone end of the building, navigate downthe corridor, find a roomthat is labeled as one containingbiohazardmaterials, enter that room,find the hazardousmaterial, and navigatetowardsit. If a biohazard is found,the robot sendsa picture of the materialbackto the humanoperator. The roomin question was labeled with a biohazard sign (Figure 8). A bucket labeled ’biohazard’ wasused for the biohazardmaterial. To detect the sign and the bucket the robot used simple color tracking. A Newton Labs Cognachrome board performing real time color blobbingwasused to do the tracking. Thesize and position of the blob wasused for detection. This simple techniqueserves as a place holder for a moresophisticated sensor that wouldbe able to detect a biohazarddirectly. TomlMissionsSuccessesFaituresiMeanRunTimc I STDl 32 29 3 62.3 seconds 5.6 seconds Table1: Summary of the experimentalresults. Thirty-two missions were attempted during a two day period before the demoday. Of those twenty-nine were successful - 91%success rate. Twomission failures were causedby unreliable detection of the biohazard due to the limited light in the building, whichcausedthe computervision codeto fail. Onefailure resulted froman unsuccessful Figure 9: The BumBuilding in Rockville, MD:site of the DARPATMR Demo. attempt to enter the door. TableI summarizes the results of the experiments. The corridor and door detection algorithmsperformedexceptionallywell since the laser rangescanneris not affected by the amountof light present in the environment.In fact, the samemission(without the biohazard sign detection) was run successfully in completedarkness. References Arkin, R., and Murphy,R. 1990. Autonomousnavigation in an manufacturing environment. IEEETransactions on Robotics and Automation 6(4):445-454. Blitch, J. 1999. Tactical mobilerobots for complexurban environments. In Mobile Robots XIV, 116-128. Endo, Y.; MacKenzie,D.; Stoychev, A.; Halliburton, W.; Ali, K.; Balch, T.; Cameron,J.; and Chen, Z. 2000. MissionLab: User Manualfor MissionLab. Gerogia Institute of Technology,4.0 edition. Fox, D.; Burgard, W.; Dellaert, E; and Thrun, S. 1999. Montecarlo localization: Efficient position estimation for mobile robots. In AAA/,343-349. Gibson, J. 1979. The Ecological Approachto Visual Perception. Boston, MA:HoughtonMifflin. Lu, F., and Milios, E. 1997a. Globally consistent range scan alignment for environment mapping. Autonomous Robots 4:333-349. Lu, F., and Milios, E. 1997b.Robotpose estimation in unknownenvironments by matching 2d range scans. Journal of Intelligent and Robotic Systems 18:249-275. Simmons,R.; Goodwin,R.; Haigh, K.; Koenig, S.; and O’Sulli van, J. 1997. A layered architecture for office delivery robots. In International Conferenceon Autonomous Agents, 245-252. Summary This paper describedperceptual algorithmsfor corridor navigation, door detection and entry that can be used for exploration of an unknownfloor of a building. The algorithms use the inherentstructure in buildingdesignto extract a few parametersneededfor correct navigation. To navigate down a corridor, for example,the robot needs to knowonly the width of the corridor and the angle betweenthe centerline of the corridor and its current heading.To enter a door, the robot needs to knowonly the two endpoints of the doorway.Thealgorithmspresentedin this paper use a laser range scanner to detect these parameters. Experimentalresults presented aboveshowthat this techniqueis quite reliable. In fact, the samemotorschemasand perceptual algorithms have been used without modification to control a Nomad 200 robot that has a completelydifferent drive mechanism than the Urbanrobot. Acknowledgements This research has been supported by DARPA under the Tactical MobileRobotics (TMR)program. Additional funding for the Georgia Tech MobileRobotLaboratory is provided by C.S. Draper Laboratories and Honda R&D,Co. The authors wouldlike to thank WilliamHalliburton, Yoichiro Endo, MichaelCramer, and TomCollins for their help with the implementationof the systemand for their hard workat the DARPA TMRDemoin Rockville, MD. DECISION ANALYSIS149