HAL Id: hal-00844915
https://hal.archives-ouvertes.fr/hal-00844915
Submitted on 16 Jul 2013
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
analysis
Michel Kieffer, Luc Jaulin, Eric Walter, Dominique Meizel
To cite this version:
Michel Kieffer, Luc Jaulin, Eric Walter, Dominique Meizel. Robust autonomous robot localization using interval analysis. Reliable Computing Journal, 2000, 3 (6), pp.337-361. �hal-00844915�
Robust Autonomous Robot Loalization
Using Interval Analysis
MICHELKIEFFER 1
,LUCJAULIN 1;2
,
ERICWALTER 1
*
ANDDOMINIQUEMEIZEL 3
1
LaboratoiredesSignauxetSystemes,CNRS-Supele
Plateau deMoulon,91192 Gif-sur-Yvette,Frane
fkieer,jaulin,walterglss.supele.fr
2
onleavefromLaboratoired'IngenieriedesSystemesAutomatises, Universite d'Angers,
2bdLavoisier,49045Angers,Frane
jaulinbabinet.univ-angers.fr
3
HEUDIASYC,CNRS,Universite deTehnologiedeCompiegne,
B.P.20529,60205Compiegne,Frane
meizelhds.univ-ompiegne.fr
Editor:
Abstrat. Thispaper dealswiththe determination of the position and orientation of a mo-
bilerobot fromdistane measurementsprovided bya belt of onboardultrasoni sensors. The
environmentisassumedtobetwo-dimensional,and a mapofitslandmarksisavailable tothe
robot. Inthis ontext, lassial loalization methodshavethree main limitations.First, eah
datapointprovidedbyasensormustbeassoiatedwithagivenlandmark.Thisdata-assoiation
stepturnsoutto beextremely omplexandtime-onsuming,and itsresultsanusuallynotbe
guaranteed.Theseondlimitationisthatthesemethodsarebasedonlinearization,whihmakes
them inherently loal. The thirdlimitationistheir lakof robustness tooutliers due, e.g., to
sensormalfuntionsoroutdatedmaps.Byontrast,themethodproposedhere,basedoninterval
analysis,bypassesthedata-assoiationstep,handlestheproblemasnonlinearandinaglobalway
andis(extraordinarily)robusttooutliers.
Keywords: IntervalAnalysis- Identiation -StateEstimation- Outliers- BoundedErrors-
Robotis.
1. Introdution
Robots are artiulatedmehanialsystems employed for tasks that may be dull,
repetitiveandhazardousormayrequireskills orstrength beyondthoseof human
beings. They rst appeared as manipulating robots with their base rigidly xed,
performing simple and well dened elementary tasks in a ontrolled workspae.
Sine then, muh of the researhin robotis hasbeendevoted to inreasingtheir
autonomy,e.g.,by addingsensors,mobility anddeisionapability. Mobile robots
maytakevariousformsdependingonthetaskandenvironment.Tobeautonomous,
theymustbeabletoestimate theirpresentstatefromavailablepriorinformation
andmeasurements.
Theproblemtobeonsideredhereistheautonomousloalizationofarobotsuh
asthat desribed by Figure 1from distane measurements provided by a belt of
*
onboardexteroeptivesensors. Here,ultrasonisensorsareused,whihareknown
to beheap but impreise. Other typesof sensors providing rangedata ould be
onsidered, with the samemethodology. Theenvironmentis assumed to be two-
dimensional(althoughathree-dimensionalextensionposesnoprobleminpriniple),
and amapof itslandmarks isavailableto therobot. Nospeial beaonsneed to
beintroduedin theenvironmenttofailitateloalization.
Figure1. Robuter mobilerobotbyRobosoft.
Inthis ontext, lassialloalization methods [4℄, [2℄, [17℄, [6℄, [18℄, [21℄and [5℄
havethree main limitations. First, eah data pointprovided by an exteroeptive
sensormustbeassoiatedwithagivenlandmark.Thisdata-assoiationstepturns
outtobeextremelyomplexandtime-onsuming,anditsresultsanusuallynotbe
guaranteed.Theseondlimitationisthatthesemethodsarebasedonlinearization,
whihmakestheminherentlyloal. Thethirdlimitationistheirlakofrobustness
to outliers due, e.g., to sensor malfuntions or outdated maps. By ontrast, the
method proposedhere, whih is basedon bounded-error set estimation (see, e.g.,
[27℄, [22℄, [23℄and [20℄, and thereferenes therein),bypassesthe data-assoiation
step, handles the problem asnonlinearand in aglobal way(see also [19℄) and is
(extraordinarily)robusttooutliers.
Thispaperisorganizedasfollows.Theproblemisstatedin mathematialterms
in Setion 2. Setion 3desribes theelementarytests that will beused to loate
therobot. Extensiontointervalsand ombinationofthese testsareonsidered in
Setion4.Setion5desribesthealgorithmemployedtoharaterizethesetofall
values of theloalization parametersthat satisfy the tests hosen. Theresulting
methodology isillustrated on three tests asesin Setion 6, before drawingsome
2. Formulationof the problem
Computationwillinvolvetwoframes,namelytheworldframeWandaframeR,of
origin=(x
;y
)in W, tiedtothe robot. Theangle betweenRandW, denoted
by,orrespondstotheheadingangleoftherobot(seeFigure2). Pointsandtheir
oordinates will be denoted by lower-ase letters in W and by tilded lower-ase
lettersin R. Thus, apointm~ withoordinates(~x;y)~ in R willbedenoted by m
inW,with
m=
x
y
+
os sin
sin os
~ x
~ y
: (1)
Three parametersare to be estimated, namely the oordinates x
and y
of the
originofRinW andtheheadingangleoftherobot. Theyformtheonguration
vetor p=( x
;y
;) T
(Figure 2). Given some(possiblyverylarge)initial searh
box [p
0
℄ in onguration spae, robot loalization anbe formulated asthe task
ofharaterizingthesetS=fp2[p
0
℄ jt( p)holdstrueg, wheret(p)isasuitable
test or ombination of tests expressing that the robot ongurationis onsistent
withthemeasurementsandpriorinformation.
W
R
xc yc
µ c
Figure2. Congurationoftherobot.
2.1. Measurements
The robot of Figure 1 is equipped with abelt of n
s
onboard Polaroid ultrasoni
sensors (sonars). The position of the ith sensor in the robot frame R is ~s
i
=
(~x
i
;y~
i
): This sensor emits in aone haraterizedby its vertex~s
i
, orientation
~
i
and half-aperture ~ (Figure 3). As ~ is frame independent, ~ = . This one
willbedenotedbyC(~s
i
;
~
i
;~
i
). Thesensormeasuresthetimelagbetweenemission
and reeption of the wave reeted or refrated by some landmark. This time
lag is then onverted into a distane d
i
to some obstale. To take measurement
inaurayinto aount, eah data pointd
i
is assoiated with the interval [d
i
℄=
[d
i
(1
i );d
i ( 1+
i
) ℄, where
i
is the known relativemeasurementauray of
sensori. Thus,[d
i
℄isassumedtoontaintheatualdistanetothelosestreeting
landmarkintereptingatleastpartoftheithemissionone.
R
aj
bj-1
bj aj+1
bj+1
aj-1
µi
:
°i
:
:yi
:
di
® di i
si
xi :
Figure3. Emissionone.
2.2. Priorinformation
Two types of prior information will be onsidered.The rst one is amap M =
f[a
j
;b
j
℄jj =1;:::;n
w
gof theenvironment,assumedto onsistofn
w
orientedseg-
ments whih desribe the landmarks (walls, pillars, et.). By onvention, when
goingfroma
j tob
j
,thereetingfaeofthesegmentisontheleft. Thehalf-plane
ajbj
situatedonthereetingsideofthesegment[a
j
;b
j
℄ isthereforeharater-
izedby
a
j b
j
= n
m2R 2
det
!
a
j b
j
;
!
a
j m
0 o
: (2)
Theseond typeofinformation(optional)istheknowledgeofasetdesribedby
3. Loalization tests
This setion enumerates various elementary tests that will be used to build the
globaltestt(p) employedtodeneS.
3.1. Data-assoiation test
Toestimate the robot onguration from range data provided by ultrasoni sen-
sors, it is of interest to build a test that heks whether agiven onguration is
onsistentwith thesedata, given theirimpreision. Forthis purpose, information
available in therobot frameRwillbetranslatedin theworldframeW. Consider
any ultrasoni sensor ofthe robot, with emissionone C
~s;
~
;~
(in this setion,
theindiesiandjwillbeomittedtosimplifypresentation). Foranygivenongu-
rationp=(x
;y
;) T
,CanbeequivalentlydesribedinWbyitsvertexs(p)and
bytwounit vetors
!
u
1
p;
~
;~
and
!
u
2
p;
~
;~
orrespondingtoits edges,given
by
!
u
1
= 0
os
+
~
~
sin
+
~
~
1
A
;
!
u
2
= 0
os
+
~
+~
sin
+
~
+~
1
A
: (3)
So one may write C =C(s;
!
u
1
;
!
u
2
) (omitting the dependeny in p;
~
and ~). By
onvention,
!
u
1 and
!
u
2
have been indexed so that
!
u
2
is obtained from
!
u
1 by a
ounterlokwiserotationof2~. Sine~isalwayslessthan =2;theonditionfor
anym2R 2
to belongto theemissiononeis
m2C(s;
!
u
1
;
!
u
2
),(det(
!
u
1
;
!
sm)0)^(det(
!
u
2
;
!
sm)0): (4)
Thealgorithmfortestingagivenongurationisbasedonthenotionofremote-
nessofasegmentfrom asensor,whih willnowbedened. Considerrstasingle
isolatedsegment[a;b℄ .Its remotenessfrom thesensors,assoiatedwiththeone
C( s;
!
u
1
;
!
u
2
),isdened as
r(s;
!
u
1
;
!
u
2
;a;b) =1ifs2=
ab
orif [a;b℄\C=;;
= min
m2[a;b℄\C k
!
smk otherwise.
(5)
The remotenessfuntion (5) is evaluated asfollows.Equation (2) is used rst to
hekwhether s 2
ab
: If this is so, minimization of k
!
smk over [a;b℄\C is at-
tempted.Thisrequirestakingdierentsituationsintoaount.Lethbetheorthog-
onalprojetionofsontotheline (a;b) . Ifh2[a;b℄\C, thenr(s;
!
u
1
;
!
u
2
;a;b)=
!
sh
. Tohekwhetherh2[a;b℄\C,withoutatuallyomputingit,onemayuse
thefollowingrelation:
h2[a;b℄\C , D
!
ab ;
!
sa E
0
^ D
!
ab;
!
sb E
0
^ D
!
ab ;
!
u E
0
^ D
!
ab;
!
u E
0
: (6)
If h 2= [a;b℄\C, the minimum distane is either innite (if [a;b℄\C = ;) or
obtained for oneof the extremities of the segment [a;b℄ \ C. Let h
1 and h
2 be
the intersetions of the line (a;b) with the lines (s;
!
u
1
) and (s;
!
u
2
): The set of
possibleends of [a;b℄\C is thus K =fa;b;h
1
;h
2
g: Therefore,if h2=[a;b℄\C;
thenr(s;
!
u
1
;
!
u
2
;a;b)iseitherinniteorequaltok
!
smk,forsomeminK . Forthe
exampleofFigure4,r(s;
!
u
1
;
!
u
2
;a;b)=
!
sb
:
b a
s
h h
h2 1
u2 u
1
Figure4. Remotenessofanisolatedsegment[a;b℄fromthesensors.
Atestofwhether anyelementofKbelongsto[a;b℄\C iseasilyderivedfrom(4).
Forv2fa;bg
v2C()(det(
!
u
1
;
!
sv)0)^(det(
!
u
2
;
!
sv)0): (7)
Byonstrution,h
i
2C\(a;b) ;onethus hasonlytohekwhether h
i
belongsto
[a;b℄;whihisequivalenttoprovingthath
i 2C
s;
!
sa
k
!
sak
;
!
sb
k
!
sbk
:Thus,fori=1;2,
h
i
2[a;b℄\C()( det(
!
sa;
!
u
i
)0)^
det
!
sb ;
!
u
i
0
: (8)
Finally,ifneitherhnoranyelementofK belongsto [a;b℄\C;then[a;b℄\C=;;
andtheremotenessisinnite.
AppendixApresentsafuntion,basedonthesetests,evaluatingr(s;
!
u
1
;
!
u
2
;a;b)
foranisolatedsegment[a;b℄ .
Remark. This versionof remoteness doesnot takeinto aount thefat that if
the inideneangle of theemitted waveis greater thana given angle(depending
on the nature of the landmark), no wave will return to the sensor. This ould
easily be taken are of by modifying the denition of remoteness so as to take
theinideneangleinto aount. Anotherphenomenonnotonsideredismultiple
reetion taking plae, for instane, in onave orners. Aountingfor multiple
notworthwhile. Aswill beseeninSetion4.3, amuhsimplerrouteistoonsider
suhmeasurementsasoutliers. }
In the normal situation where n
w
segments are present, the fat that a given
segment maynot be deteted,beauseit liesin the shadowof anotheroneloser
to the sensor, must be taken into aount. Let r
ij
(p) be the remoteness of the
jthsegment,takenasisolated,from theith sensor iftheongurationisp. This
remotenessis givenby
r
ij
(p)=r
s
i ( p);
!
u
1i
p;
~
i
;~
i
;
!
u
2i
p;
~
i
;~
i
;a
j
;b
j
: (9)
Theremotenessofthemapfrom theith sensoriftheongurationispisthen
r
i
( p)= min
j=1;:::;nw r
ij
(p): (10)
Themeasurementprovidedbytheithsensormaybeexplainedbyasegmentlying
ataproperdistane ifthefollowingtestis satised:
Test dat
i
( p): dat
i
( p)holds trueifandonlyifr
i
(p)2[d
i
℄.
3.2. Inroomtest
Assume that the map partitions the world into two sets, the interior, whih the
robotshould belong to,
P
int
= 8
<
: m2R
2
nw
X
j=1 arg
!
ma
j
;
!
mb
j
=2 9
=
;
; (11)
andtheexterior
P
ext
= 8
<
: m2R
2
nw
X
j=1 arg
!
ma
j
;
!
mb
j
=0 9
=
;
; (12)
wherearg
!
ma
j
;
!
mb
j
;theanglebetween
!
ma
j and
!
mb
j
,isonstrainedtobelong
to ℄ ;℄. The fat that is exludedimplies that theboundarybetweenP
int
andP
ext
belongstotheinterior. Figure 5illustrates asituation wherepartofthe
roomisforbiddenbysuitablyorientedinternalpolygons. Foreahsegment[a
j
;b
j
℄,
thearrowindiatesthediretionfroma
j tob
j
:Reallthatthereetingfaeison
theleftwhengoingfroma
j tob
j :
Ifm~ isanypointoftherobotwithoordinates(~x;y)~ inR,thenitsoordinatesm
inWevaluatedaordingto(1)dependontherobotongurationp=(x
;y
;) T
and the following test will make it possible to eliminate some ongurations for
whihitwouldnotbein P .
-12 12 -12
12
aj
bj
Figure5. Partitionoftheworld.Theinteriorisinwhite.
Test in room( m):
8
<
:
inroom( m)=1 if nw
P
j=1 arg
!
ma
j
;
!
mb
j
=2;
inroom( m)=0 otherwise.
Whenmistheprojetionofpontothexyplane,in room( m)willberewritten
asin room(p).
AsshowninSetion6,thistestwillontributetoeliminatingongurationsmore
eÆientlythanthedata-assoiationtestalone,onpurelygeometrialgroundsand
in the absene of any measurements. However, it is reliableonly when the map
and thepartition itindues are reliable.Evenwhen thisis notthease, this test
remains of interest, as it forms the basisfor the leg in test presented below and
stillappliable.
Remark. FititiousnonreetingsegmentsmaybeneededtodeneP
int andP
ext .
Theymaybetransparent(opendoorsandwindows),orabsorbing. Thereetivity
ofeahof thesesegmentsould betakeninto aountwith amoreelaborate de-
nitionofremoteness.With thedenition adoptedhere,suhsegmentsmayleadto
outliers,see Setion4.3. }
3.3. Legintest
Considerarobotongurationp=(x
;y
;) T
,theith robotsensors
i
,withoor-
dinates (~x ;y~) in R, and its assoiated interval measurement [d℄. Let bethe
point at adistane equalto thelowerbound d
i of[d
i
℄ from s
i
in thediretion of
emission
~
i
. Theoordinatesof
i
inW satisfy
i
=
x
y
+
os sin
sin os
0
~ x
i +os
~
i
d
i
~ y
i +sin
~
i
d
i 1
A
: (13)
Assuming, as for inroom, that theworldis partitionned into P
int and P
ext , one
andene
Test leg in
i
(p): leg in
i
(p)=in room(s
i
(p))_in room(
i ).
Thefollowingresultexplainswhythistestanbeusedinonjuntionwithdat
i to
eliminateongurations.
Proposition1 leg in
i
(p)=0 )dat
i
(p)=0. }
Proof: leg in
i
( p)=0impliesthats
i isinP
int and
i inP
ext
(seeFigure6).Then
there exists j suh that [s
i
;
i
℄\[a
j
;b
j
℄ 6= ;and s
i
2
ajbj
. Let m
ij
=[s
i
;
i
℄\
[a
j
;b
j
℄. Theithoneintersets[a
j
;b
j
℄atleastatm
ij
. Sotheremotenessof[a
j
;b
j
℄
from s
i
is less thanor equalto k
!
s
i m
ij
k. As m
ij 2 [s
i
;
i [; k
!
s
i m
ij k< k
!
s
i
;
i k=
d
i
, and the remoteness of [a
j
;b
j
℄ from s
i
is therefore inompatible with [d
i
℄; so
dat
i
(p)=0.
Thetestlegin
i
(p)thusprovidesaneessaryonditionforptobeonsistentwith
theith measurement. Asthis onditionisnotsuÆient,leg in
i
(p)mayholdtrue
even when dat
i
(p) holds false.It will only be useful to eliminatesome unfeasible
ongurationsmorequikly.
4. Interval tests
Thetestspresentedinthepreedingsetionforpointongurations,shouldnowbe
extendedto intervalongurations. The notionof Boolean intervals will be used
totakethepossibleambiguityoftest resultsinto aount. Itwillthenbepossible
to give interval ounterpartsof the loalization tests, whih will be assoiated to
inreasetheireÆieny.
4.1. Booleanintervals andinlusion tests
A Boolean intervalis anelementof IB = f0;[0;1℄;1g, where 0stands forfalse, 1
for true and [0;1℄ for indeterminate. It is a onvenient objet for implementing
three-valuedlogi.
Table1speiestheAND(^)andOR(_)operationsbetweentwoBooleanintervals.
As Boolean intervalsare sets, standardset operatorssuh as[and \also apply.
They should not be onfused with the logial operators _ and ^. For instane,