Intelligent Control And Navigation Of An Outdoor AGV

Ping Ping Khaw, W.S. Wijesoma, and, Earn Khwang TeohIntelligent Machines Research Lab

School ofElectrical and Electronic . eeringNanyang Technological University, Singapore 639798

e-mail: eswwijesoma@ntu.edu.sg

AbstractPiloting AGVs in an unstructured, dynamicand complex outdoor environment has becomean emerging area of research. In this paper wedescribe the development of a fuzzy na torto perform the local navigation and obstacleavoidance tasks for an outdoor AGV which isrealised from a golf cart. The hybridarchitecture incorporates a high level 'model­and-plan' and a low-level 'sense-and-execute'layer. Each of the complex local navigationaltasks that need to be carried out is analyzed interms of primitive behaviours and expressed asan aggregation of such behaviours. The fuzzybehaviours are synthesized based onappropriately fused sensory data received fromthe complementary sensor devices, especially,proximity sensors and a laser scanner. A fuzzysimulator developed using Matlab isintroduced which is ca able of monitoring thevehicle performance gh travelling profile,heading angles, change in steering commands,as well as the logic patterns of the proximitysensors. A method to associate these logicpatterns to linguistic variables is discussed.Lastly, enhancements to the simulator arepresented.

1 Introduction

Pigure 1: Outdoor AGV realised from a Jolf cart

240

The development of techniques for autonomousnavigation and control of vehicles has become animportant and active research topic in the face ofemerging markets for advanced AGV and mobilerobots. A key requirement for autonomous navigation inan unconstrained and uncertain environment is that thesystem must be capable of sensing the surroundings todetermine where the AGV is at present (localisation)and where it is moving. This is in order for the AGV tobe able to respond intelligently to a changing situationor environment.

As in the case with many autonomousnavigation systems, all the sensors are placed on boardthe vehicle. Sensory modules on - board the actualvehicle include optical encoders, magnetic sensors,DGPS, vision system, laser scanner and proximitysensors. The sensors provide complementaryinformation as regards the internal state of the vehicleand the current state of the environment. Figure 1 showsthe AGV, which is realised from a golf cart.

2 Navigational and Control Architecture

In the hybrid architecture illustrated in Figure 2 basedon [Tunstel and Jamshidi, 1997] and [Biewald, 1996],the AGVs functions are partitioned into a high level'model-and-plan' layer (global path planning ornavigation) and a low level 'sense-and-execute' layer(local navigation), with a behaviour-baseddecomposition of the latter layer. This architectureclosely mimics the highly successful navigationalbehaviour exhibited by humans in executingnavigational tasks in a complex, unstructured anduncertain environment without the aid of complete andaccurate spatial maps of the environment.

Figure 2 shows the integration and interactionof the various modules, comprising the low level­vehicle controller, sensors, the obstacle avoidance andlocal navigation system, localization system, and thepath planner for global navigation. Overall operation ofthe autonomous .navigation system can be described asfollows. Given a particular goal task for the AGV, theglobal path planner determines an optimum route or anavigational plan to execute the task. The output of the -­planner is a symbolic description of the path in terms ofplaces, paths and actions (primitive behaviours) andmay be in the form of context or applicability rules[Tunstel and Jamshidi, 1997]. Based on thisnavigational plan, and the current context of the AGVas determined by the localization subsystem, the

Navigation Block

Behaviors

External Sensory Data

FusionBlock

Limit themax v,t/J

esiredx,y, v,t/J Threshold

Logic __~---tUnit

Sensors

Controller Block

Dynamicof AGV

Figure 2: Hybrid Architecture of AGV Navigation and Control

behaviour based local navigation subsystem generatesthe necessary commands to the vehicle controller at aspecified sampling rate. The vehicle controller executesthese commands continuously to ensure that the vehicleaccomplishes the goal task.

The local navigation block is the sense-and­execute layer of the overall structure of the navigationsystem. This layer is implemented using a behaviouralapproach. That is, each of the complex localnavigational tasks that need to be carried out is anaIysed--~-

in terms of primitive behaviours and expressed as anaggregation ,of such behaviours. A fuzzy logic approachto behaviour synthesis and integration has beenadopted. Th~ fuzzy behavioural methodology providesa natural means of incorporating human navigationskills in terms of linguistic information. The fuzzybehaviours which are considered to be necessaryinclude, Curb Following, Obstacles Avoidance,Obstacles Contouring, Narrow Path Maneuver,Cornering, Route Following and Wandering. Each ofthe behaviours is synthesised based on appropriatelyfused sensory data received from the complementarysensor devices [Wijesoma et al., 1999].

3 System Setup And Sensor Modelling

The solid state proximity sensors used aresuitable for harsh outdoor environments. The downsideof using proximity sensors is that they can sense thepresence or absence of an obstacle to within a limitedprecision. That is, they provide a high logic outputwhen an obstacle is sensed and a low when theperimeter is clear. Thus, it is necessary to appropriatelyrelate the logic outputs with range information, in order!()~'!hanse tb~J.J.~~~!!YQ!Jb~PI"Q~tmity. S~I1SQ:rS.

-- ~ -- In the simulation, a total of 11 proximitysensors are used to detect the obstacles in thesurroundings of the vehicle. The challenge is, how touse the minimum amount of sensors to give asatisfactory performance in the local navigation andobstacle avoidance task. Among the three proximitysensors located on the left, the one closest to the frontcomers of· the vehicle is tuned to the longest(maximum) sensing range of 3m, followed by 2m and1m. The sensing range and mounting location of theproximity sensors are shown in Figure 3.

Front Senaora

Figure 3: Sensing range and mounting positions of theproximity sensors

In order to achieve a satisfactory performance in localnavigation and obstacle avoidance, two types of sensorsare utilised, namely solid state infra red proximitysensors and a laser scanner. The SICK laser scanner(LMS 220) [Sick Optic Electronic], equipped with asensing range of 50m, and, 180 degree sweeping angle,with 0.5 degree resolution is mounted on the front end.The blind spots of the laser scanner are properlycovered by three photoelectric proximity sensors of 3mcoverage (maximum) along each side of the vehicle. Inorder to track the road curb and able to provideemergency stop function, two pairs of sensors aremounted on the bumper, and linked directly to theemergency stop circuitry.

241

Sensing Range

3m

2m

1 m

Lett Senaora

Mounting Locations

0.5 m

1.2 m

1.8 m

Right Senlora

R Ig h tP ro x (.. )

RS

SteerAngle (9)

agv

64 rules

(mamdani)

Figure 5: Mamdani type Fuzzy Inference System

System agv: 3 inputs, 1 outputs, 64 rules

In the local navigational module, the set offuzzy rules can be regarded as the "brain" commandingthe steering direction of the vehicle to the vehiclecon ere The inputs to the Mamdani type FuzzyInference System (FIS) are signals received from theleft (LejtProx), front (FrontProx) and right (RightProx)proximity sensors of the vehicle. The output of the FISwill be the steering angle (SteerAngle) conUnanded tothe vehicle to move to a safer or favorable direction inorder to accomplish the goal task. '

Several proximity sensors mounted along thefront and the left of the vehicle. are used to .implementthe Curb Following (Left) behavior. The linguisticvalues associated with the array of left proximitysensors are Near, Regular, Medium, and, Far. In orderto increase the steering performance, there are ninelinguistic terms to describe the FIS output, SteerAngle,over the universe of discourse ranging from -30 to 30degree. Figure 6 shows the membership functions andtheir distribution.

Figure 4: Art example iUus.ng the logic patterns andlinguistic variables represented by proximity sensors at

sampling instance t

To make the sensor inputs fuzzy, the followingalgorithm has been inc~rporated. It has, in oile way oranother, imposed a certain degree of uncertainty andcreated a more realistic simulation environment, whichmimics the real world. When the first two sensors alongthe left (right) side indicate high, it signifies that theobstacle lies within the sen . es of the second(2m) and the third (1m) sensors. e actual d ce,however, is not known. From Figure 3, it can beobserved that the worst case is such that the obstacle isvery close to 1m (the last sensor), and the safestsituation is when the object is just less than 2m apartfrom the AGV. Realizing the uncertainties in therelative distance between the vehicle and the obstacle, itis logical to generate a random number between 0 and1, and add it to the 1m sensing distance of the thirdsensor. In other words, the error tolerance in rangemeasurement is within 1m. In order to improve theperformance, the sensing range, however, can bereduced. Adding a few more proximity sensors isanother alternative.

Figure 4 illustrates the logic patterns and thelinguistic variables represented by proximity sensors atsampling instance t. Assume 1 represents ObstacleDetected and 0 indicates Clear, then the logic pattern 1­1-0 is associated with the linguistic value ObstacleRegular (the relative distance between the obstacle andthe AGV is regular). By the same token, 1-1-1corresponds to Obstacle Near.

Left Front Right

Logie Logic

IT-Logie

Proximity

~tl-1

~ttSenior

(3m range)

o IProximity 'm ~hL tilSensor

(2m range)

o I o I'lim.

Proximity

~Lt. ~U- 'ldlSensor

o tim.

(1m ringe)

t t t

Logic110 100 111Pattern

linguistic Obstacle Obstacle ObstacleTerms Regul., "edlum Ne.,

4 Fuzzy Inference System

Fuzzy inference system serves as a means oftransforming the linguistic control strategy based onexpert knowledge into an automatic control strategy.Fuzzy control appears to be very useful for handlingproblems that are too complex for analysis usingconventional quantitative techniques or when theavailable sources of information provide qualitative,approximate, or uncertain data. The navigational controlof an outdoor AGV falls into this class of problem.

242

Figure 6: Membership functions of SteerAngle

It is observed that the distribution of themembership functions over the 60 degree range is notuniform. The spacing is tighter at smaller angles. Bydoing so, it is able to steer at a finer scale when thevehicle is close to the curbs and obstacles. It can beexpected that the steering motion will be a smootherone, as abrupt steer is often not desired.

100

150

illustrated the concept of localization as well as globalpath planning. The commands listed in the global pathplanner are: Left Curb Following ~ Turning Right atJunction ~ Left Curb Following. Initially, the AGV istrying to look for the road curb to follow in order toaccomplish the Keep Left goal task. Along the journey,the vehicle is able to avoid the obstacles nicely, asObstacle Contouring is blended into each primitivebehaviors, such as the Left Curb Following and TurningRight at Junction.

Secondly, the localization modules, whichcomprise of the vision system and DGPS, reports thatthe vehicle has reached the desired junction and it isready to make a .turn. In the simulation, however, apseudo-localization module is utilized. Currently, theregions on the map are defined by stating thecoordinates of the comers of the block. The centroid ofeach region is then determined and the symmetrical Xand Y distance to the border of the region is obtained.At each sampling instance, the program will examinethe current location of the vehicle by monitoring thewhereabouts of the centroid of the rectangular boxwhich signifies the AGV. Hence, the Tum Right atJunction behavior will be fired only when the pseudo­localization module indicates the vehicle has reachedthe desired junction. Hence, the possibility of executingthe incorrect goal task at the wrong place is thereforeminimized.

The sensing range provided by the proximitysensors is fairly limited, thus the vehicle has no priorknowledge on the location and width of the subsequentpath that it is going to tum into. Hence, it can beobserved that the vehicle is attempting to make a verycareful turn when it approaches the junction, byreducing the speed to half of the original and taking asmaller turning radius. The turning profile is determinedby assuming the next path to turn into has a scale ratio,SR, of the current junction. In the simulation resultshown here, SR has been set to 1 The width of thejunction can be obtained from the measure~ent

provided by the proximity sensors on the left and nghthand side of the vehicle. As shown in .Figure"8, upon,reaching the next region, the vehicle will try to locatethe curb to follow and avoiding the sharp comergracefully before continuing the Left Curb Followingbehavior again.

50

QUt

St5t SiYUati:rt

f;1 Showtrace

Q,ooaeWap .

ISchool d EEE CIIpark 3Nl.f'ItJerdn.n

1200 ::oJEriargef9.re

IFig A :.N;V nrnobcn ::oJ

20

Figure 7: AGV simulator

NAN YAN GTE CHN0 LOG ICA L UN IVER SIT YINdigent WKhinea R....ch UIb

AGV Simulator

10

3) ,.------,

Figure 7 shows the AGV simulator developed inMATLAB / SIMULINK environment. The simulatorserves as a test bed to evaluate the effectiveness of thefuzzy rules in the PIS before they are to be implementedon the golf cart controller. The development flow chartof. the fuzzy simulator follows closely the hybridarchitecture of navigation and control of the AGV, asillustrated in Figure 2. Upon the completion of eachsimulation, four figures will be shown for performanceevaluation, namely AGV in motion, Change in headingangles, Change in steering commands and Logicpattern ofproximity sensors.

5 Simulation Results

.100 100 2(])

rig C:. Steemg angies

10 15 2D 25 30

Figure 8: Travelling profile of the AGV when perf~rming

Turning Right at Junction and Left Curb FollOWIng

The travelling profile of the AGV performinga right turn at the junction is shown in Figure 8, whichis the enlarged figure taken from Fig A locates at theupper left comer of the simulator. This figure has

180 180 2IXl

Figure 9: Change in heading angles when performing TurningRight at Junction and Left Curb Following

243

25

6 Conclusions

The logic patterns shown in Figure 11 will beassociated to the linguistic values as defined in themembership functions of the linguistic variableLeftProx, FrontProx and RightProx, respectively. Thefirst proximity sensor on the left of the vehicle is atlogic high for most of the time. This indicates that thevehicle is following the I a distance less thanMedium distance. The third sensor in the front alwaysreports no obstacle detected. It demonstrates theeffectiveness of the designed fuzzy rule base. Underthese rules vehicle will never have a heading distance ofless than 1m with respect to the curb.

Besides better accuracy in range finding, th~

laser scanner is capable of providing a longer sensingrange. Due to the fact that these two kinds of sensorsare mounted at different elevations, a two layer planarsensing approach is adopted. Using this method, it isable to define regions, which are common to both kindsof sensors, or only visible to either one. A goodexample is, when the AGV is approaching vehiclesparked in the lots. The proximity sensors are only ableto sense the wheels (or tires) which are at a lowerelevation from the ground. As for the laser scanner itcan correctly locate the range ofthe'bumper or the bodyof the car, which is at a higher elevation with respect tothe wheels. The use of laser scanner has certainly drawnthe gap between the simulated environment and the realworld closer by reducing the ambiguities in obstacledetection.

A hybrid navigational and control strategy for outdoorAGVshas been proposed. The hybrid architecture

:comprises "of a global path planner and a localnavigation module. The latter one is implemented using.a· fuzzy.' :behavioral approach. The fuzzy behaviours.which.:have.been.incorporated include, Curb Following,.Obstacles Avoidance, Obstacles Contouring, NarrowPath Manoeuvre, Cornering, Route Following andWandering.

A simulation framework based on MATLAB ISIMULINK environment gives a convenient method toverify the strategy for its reliability, and effectiveness.Also the simulation environment provides for thetesting of the applicability of different sensortechnology. An effective methodology of usingproximity sensors (as opposed to ultrasonic sensors) asa means of range sensing for realizing fuzzy behaviorsnecessary for certain navigational tasks is proposed andits effectiveness is demonstrated through simulations.To overcome limitationsof.the proximity sensors, viz.,the limited sensing range and the precision, and also tobe able to synthesize other behaviors necessary foroutdoor navigation a laser range sensing system isproposed. It is discussed how the laser range scannercan complement and augment the reliability andnavigability of the outdoor AGV. The authors proposeto implement the fuzzy behavior based navigational ..strategy on the actual AGV to ascertain its practicalefficacy.

50 100 150 200

50 100 150 200

'1 V

0.5

0.5

50 100 150 200

50 100 150 200TIM

I

I0.5

0.5

50 100 150 200

50 100 150 200

I'II II

I

.1 I/. I

I

II I

LeftlR FrontIR RightlR1

i~

0.5 0.5 0.5

I L.....- i50 100 150 200 50 100 150 200 50 100 150 200

Figure 10: Change in steering commands when performingTurning Right at Junction and Left Curb Following (Fig 8)

_10 ..........-"-------1_--"- ---100_.....-.........._00lo0o--..1

o 20 40 10 10 100 120 140 110 110 200

10

15

Change of St_ring Ang'- VS Time

Figure 9 shows the change in vehicle headingangles during Turning Right at Junction and Left CurbFollowing. A positive value indicates a right steer andnegative signifies a left tum. These data will be fed tothe fuzzy vehicle controller. As graceful steer is moreprefe pt turn, these steering angles are ableto provide some insights, which are quite useful intuning the d" . tion of the mem "functions. Ofcourse, this will be restricted by the sensing range of thesensors. For instance, for tracking road curbs, thevehicle has to perform a turn within a radius of 3m, dueto the fact that the maximum sensing range of theproximity sensors used is 3m

Figure 10 shows the change in steeringcommands, with respect to time. The relative directionof the vehicle is taken into account. Hence, a 0 steeringcommand indicates that the vehicle is maintaining theprevious heading angle, instead of steering to the so­called home position. It can be observed that the outputsof the FIS have been limited to ±30 degree. In addition,higher steering rate is required when the AGV isapproaching sharp corners.

Figure 11: Logic pattern of proximity sensors collected duringTurning Right at Junction and Left Curb Following (Fig 8)

0.5

Logjc0.5

244

7 References

[Tunstel and Jamshidi, 1997] E.W.Tunstel andM.Jamshidi" Intelligent Control and Evolution ofMobile Robot Behavior, Application of Fuzzy LogicTowards High Machine Intelligence QuotientSystems, pages 1-24, Prentice Hall, 1997

[Biewald, 1996] R.BiewaId, A neural networkcontroller for the navigation and obstacle avoidanceof a mobile robot, Neural Network for RoboticControl, Theory and Application, pages 162-191,Ellis Horwood Ltd, 1996

[Wijesoma et al., 1999] W.S. Wijesoma, Ping PingKhaw, and Earn Khwang Teoh, Sensor ModellingAnd Simulation For Navigation Of An OutdoorAGV, accepted by lASTED International Conferenceon Modelling and Simulation, Philadelphia,Pennsylvania, USA, May 5-8, 1999

[Wijesoma et al., 1999] W.S. Wijesoma, Ping PingKhaw, and Eam Khwang Teoh, Control AndNavigation Of An Outdoor AGV Using FuzzyReasoning, submitted to IEEE Conference OnIntelligent Transportation Systems, Tokyo, Japan,October 5-8, 1999

[Saffiotti et aI., 1993] Alessandro Saffiotti, EnriqueH.Ruspini, Kurt Konolige. Blending reactivity andgoal-directedness in a fuzzy controller, Proc. of 2nd

IEEE conference on Fuzzy System, pages 134-13, SanFrancisco, California, March 1993

[Li et aI., 1996] Wei Li, Chen ZuShun, et al., FuzzyLogic Based:,Behavior Fusion for Navigation of anIntelligent Mobile Robot;'.,? Journal of ComputerScienve antiTechnology, 1fl(4):385-394, 1996

[Sick Optic Electronic] Technical Description of LMS210/220/290 Laser Measw:ement System (OutdoorVersion), SICK Optic Electronic

[Ollera et al.,A1997] A.OI~~~;o A.Garcia-Cerezo et.al,Fuzzy Tracking Methods for Mobile Robots,Application of Fuzzy Logic Towards High MachineIntelligence Quotient Systems, pages 347-364,Prentice Hall, 1997

[Ramirez-Serrano and Boumedin.e, 1996] AlejandroRamirez-Serrano and Marc Boumedine, Real timeNavigation in Unknown Environments Using FuzzyLogic and Ultrasonic Sensing, Proceedings of theIEEE International Symposium on IntelligentControl, pages 26-30, Dearborn, MI, September 1996

[AoId et al., 1994] Takesbi Aoki, Moriyasu Matsuno,et.al, Motion Planning for Multiple ObstaclesAvoidance of Autonomous Mobile Robot UsingHierarchical Fuzzy Rules, Proceedings of the IEEEInternational Conference on Multisensor Fusion andIntegration for Intelligent Systems, pages 265-271,Las Vegas, NV, 1994

[Xu and Tso, 1996] W.L Xu and S.K Tso, Real-timeSelf-reaction of a Mobile Robot in Unstructured

. Environments using Fuzzy Reasoning, EngineeringApplication of Artificial Intelligent, 9(5):475-485,1996

245