mm-NOLOC: mmWave-based Localization for Mobile Networks without 3GPP Location Service

mm-NOLOC: mmWa e-based Localiza ion o Mobile Ne wo ks
wi hou 3GPP Loca ion Se ice
Phuc Dinh
†
, Yu ei Feng
†
, Edua do Baena
†
, Yunmeng Han
§
, Weiming Qi
¶
, Zihan Xu
‡
, Moinak
Ghoshal†, Pau Closas†, Dimi ios Kou sonikolas†, Joe g Widme
†
No heas e n
Uni e si y,
‡
Vi ginia
Tech,
§
Uni e si y
o
Cincinna i,
¶
Michigan
S a e
Uni e si y,
USA,
IMDEA Ne wo ks, Spain
Abs ac
Accu a e localiza ion in dense u ban a eas emains a signi ican
challenge due o he limi a ions o Global Na iga ion Sa elli e Sys-
ems (GNSS) in en i onmen s wi h obs acles and e lec ions, such as
u ban canyons. While he mos ecen 3GPP s anda ds o e sophis-
ica ed ne wo k-cen ic posi ioning echniques, hei widesp ead
deploymen will ake ime and is hinde ed by high in as uc u e
cos s and complexi y. In his wo k, we p esen mm-NOLOC, a UE-
cen ic localiza ion sys em, designed as a p ac ical allback when
GNSS ails o deli e high accu acy, ha le e ages he g owing
deploymen o 5G mmWa e in as uc u e in dense u ban a eas.
Unlike adi ional app oaches, mm-NOLOC ope a es independen ly
o 3GPP loca ion suppo and u ilizes only s anda dized con ol-
plane in o ma ion collec ed solely on he UE side – Synch oniza ion
Signal Block (SSB) Indices ha a e mapped o 5G mmWa e beam
di ec ions – o ob ain obus posi ion es ima ions. To add ess he
unce ain y in oduced by u ban mul ipa h, mm-NOLOC models
he SSB- o-angle ela ionship as a disc e e and mul imodal dis ibu-
ion, based on empi ical measu emen s in ope a ional 5G mmWa e
ne wo ks, and uses a pa icle il e o e ine posi ion es ima es by
in eg a ing p obabilis ic obse a ions wi h UE-side mo ion dynam-
ics. We alida e mm-NOLOC h ough expe imen s o e comme cial
5G mmWa e deploymen s, as well as ace-based simula ions. Ou
esul s show ha mm-NOLOC achie es a median localiza ion e o
below 3 m and a 95 h pe cen ile e o below 10 m, o e ing a p ac i-
cal allback localiza ion solu ion in u ban canyon scena ios o 5G
ne wo ks wi hou ne wo k loca ion suppo .
CCS
Concep s
•
Ne wo ks → Loca ion based se ices.
Keywo ds
5G,
mmWa e,
localiza ion,
beam
managemen
ACM Re e ence Fo ma :
Phuc Dinh
†
, Yu ei Feng
†
, Edua do Baena
†
, Yunmeng Han
§
, Weiming Qi
¶
,
Zihan Xu
‡
, Moinak Ghoshal
†
, Pau Closas
†
, Dimi ios Kou sonikolas
†
, Jo-
e g Widme . 2025. mm-NOLOC: mmWa e-based Localiza ion o Mo-
bile Ne wo ks wi hou 3GPP Loca ion Se ice. In The Twen y-six h In-
e na ional Symposium on Theo y, Algo i hmic Founda ions, and P o ocol
Design o Mobile Ne wo ks and Mobile Compu ing (MobiHoc ’25), Oc o-
be 27–30, 2025, Hous on, TX, USA. ACM, New Yo k, NY, USA, 10 pages.
h ps://doi.o g/10.1145/3704413.3764451
1 In oduc ion
Localiza ion se ices a e in eg al o mode n applica ions such as
na iga ion, asse acking, eme gency esponse, and au onomous
sys ems. Among hese, Global Na iga ion Sa elli e Sys ems (GNSS)
such as he Global Posi ioning Sys em (GPS), emain he mos
widely used solu ion due o hei global co e age and ease o use.
Howe e , GNSS se ice de e io a es o becomes una ailable in
ce ain en i onmen s, such as indoo s o in u ban canyons, whe e
sa elli e signals a e blocked o e lec ed. In such scena ios, echnolo-
gies such as WiFi and cellula posi ioning o e iable al e na i es.
While WiFi localiza ion can be e ec i e indoo s, i s accu acy di-
minishes in ou doo en i onmen s wi h spa se deploymen o lack
o WiFi access poin s. In con as , cellula posi ioning, s anda dized
by 3GPP, p o ides b oad ou doo co e age using mobile ne wo k
in as uc u e, making i a po en ial solu ion o u ban canyon use
cases.
To p o ide obus cellula posi ioning, 3GPP speci ies a ange o
echniques ailo ed o di e se scena ios. Me hods such as Enhanced
Cell ID (E-CID) and Obse ed Time Di e ence o A i al (OTDOA)
p o ide b oad co e age bu s ill lack he p ecision equi ed in dense
u ban a eas due o signal e lec ions and limi ed g anula i y [4].
Ad anced echniques, including Downlink Angle o Depa u e (DL-
AoD), Uplink Angle o A i al (UL-AoA), and Channel S a e In o -
ma ion (CSI)-based posi ioning, can signi ican ly imp o e accu acy,
bu equi e ad anced ea u es, such as high-accu acy synch oniza-
ion, dedica ed e e ence signals, and ad anced an enna con igu-
a ions, which add conside able deploymen cos s and complexi y.
These complexi ies pose signi ican challenges o deploymen a
scale, slowing down hei widesp ead adop ion in eal-wo ld com-
me cial ne wo ks. To he bes o ou knowledge, no p oduc ion 5G
ne wo ks cu en ly suppo he ad anced loca ion ea u es speci ied
in ecen 3GPP eleases.
While ope a o s con inue o wo k owa d e en ually enabling
high-accu acy ne wo k-cen ic solu ions, he cu en eali y neces-
si a es an al e na i e allback app oach o u ban en i onmen s,
whe e GPS pe o mance is o en comp omised. In his pape , we ill
his p ac ical gap by in oducing mm-NOLOC (
mm
Wa e-based Lo-
caliza ion o Mobile Ne wo ks wi h
NO
3GPP
LOC
a ion Se ice), a
UE-cen ic localiza ion solu ion designed o ope a e wi hou mobile
ne wo k loca ion suppo . mm-NOLOC capi alizes on he ac ha
ope a o s o en deploy 5G mmWa e in as uc u e in u ban a eas
o mee high da a a ic demands, making i uniquely sui ed o
add ess he localiza ion challenges in such en i onmen s.
MobiHoc ’25, Oc obe 27–30, 2025, Hous on, TX, USA
Phuc Dinh e al.
As a p ac ical UE-cen ic solu ion, mm-NOLOC adhe es o he
cons ain o using only s anda dized con ol-plane in o ma ion
cu en ly a ailable on he UE side, as de ined by 3GPP. This con-
s ain excludes he use o ad anced ne wo k-speci ic in o ma ion,
such as CSI, AoA, and AoD, which a e imp ac ical on comme cial
de ices. Signal s eng h me ics (RSRP, RSRQ) a e eadily a ail-
able on comme cial-o - he-shel (COTS) UEs bu o en un eliable
o high-accu acy localiza ion. The Timing Ad ance (TA) [3] ield,
which is used o synch onize he ansmission iming o a UE wi h
he 5G base s a ion (gNB) o compensa e o he p opaga ion delay
o he adio signals, could be le e aged o calcula e he dis ance
be ween he UE and gNB, bu i s coa se esolu ion (39 m o 100
MHz o bandwid h) makes i unsui able, especially o dense u ban
cellula deploymen s.
Ins ead, mm-NOLOC le e ages Synch oniza ion Signal Block
(SSB) Indices, 3GPP s anda dized signaling messages commonly
used o iden i y mmWa e beams. Le e aging s anda dized con ol
messages ensu es b oad compa ibili y ac oss de ices and simpli ies
in eg a ion o applica ion de elope s, hus accele a ing adop ion.
5G mmWa e gNBs ansmi di e en SSBs pe iodically and di ec-
ionally, ypically e e y 5–160 ms. Each SSB Index can se e as
an indica o o he ansmission di ec ion o he gNB, and hence
implici ly p o ides in o ma ion abou he angula ela ionship be-
ween he UE and he gNB. Di e en SSBs om neighbo ing gNBs
p o ide mul iple angula ela ionships, which can be combined wi h
knowledge o gNB posi ions in he se ice a ea and he o ien a ions
o hei an enna panels o localize he UE.
None heless, le e aging SSB indices o build a high-p ecision
localiza ion sys em aces a se o p ac ical challenges. Fi s , mm-
NOLOC elies on cons uc ing an SSB- o-angle mapping unc ion.
Fo such an app oach o be scalable, his unc ion should gene al-
ize e ec i ely ac oss a la ge numbe o gNBs wi hou equi ing
exhaus i e measu emen s a e e y gNB. Second, unlike p e ious
he a iabili y and mul imodali y obse ed in eal deploymen s. Fo
he hi d challenge, mm-NOLOC in eg a es he lea ned mapping
unc ion in o a pa icle il e ha uses hese di ec ional p io s o
sampling and esampling, achie ing accu a e localiza ion wi hou
elying on he s ic geome ic cons ain s o mul i-angula ion.
We e alua e mm-NOLOC ia eal-wo ld expe imen s in wo
ci ies and ex ensi e ace-based simula ions. Ou esul s demon-
s a e ha mm-NOLOC achie es high localiza ion accu acy, compa-
able o o e en be e han GPS, wi h a median e o below 3 m and
a 95 h pe cen ile below 10 m. O e all, mm-NOLOC o e s a depend-
able allback solu ion o loca ion-based se ices in u ban canyon
en i onmen s, whe e GPS accu acy is o en comp omised, by only
le e aging ea u es o oday’s 5G deploymen s and in o ma ion
eadily a ailable in COTS UEs. Ou da ase is publicly a ailable [1].
2
mm-NOLOC
Design
In his sec ion, we p esen he design o mm-NOLOC. We begin
by jus i ying wo cen al design decisions: (i) using SSB indices
as he p ima y inpu o localiza ion, a he han elying on aw
signal s eng h me ics; and (ii) adop ing a pa icle il e o posi ion
es ima ion, in place o mo e adi ional app oaches like he Kalman
il e , which is commonly used in simila con ex s. We hen desc ibe
he
co e
componen s
and
wo k low
o
he
mm-NOLOC
a chi ec u e.
2.1 Limi a ions o Signal S eng h
One common app oach in wi eless localiza ion is o use signal
s eng h, such as he Re e ence Signal Recei ed Powe (RSRP), o
es ima e he UE’s dis ance o nea by gNBs [8]. This app oach elies
on he assump ion ha signal s eng h deg ades p edic ably wi h
dis ance, allowing a sys em o app oxima e a UE’s posi ion using
ila e a ion, i.e., by es ima ing dis ances o mul iple gNBs and
sol ing o he in e sec ion poin .
0.8
heo e ical wo ks ha assume a di ec and de e minis ic ela ion-
ship be ween SSB measu emen s (beams) and angula posi ions,
his ela ionship in p ac ice is disc e e and mul imodal (due o i -
egula beam pa e ns gene a ed by phased a ays on comme cial
gNBs) and con ains some deg ee o andomness (due o blockage
and mul ipa h e ec s), hinde ing di ec applica ion o many a-
di ional localiza ion echniques. Thi d, in p ac ical 5G mmWa e
deploymen s, whe e gNBs a e o en ins alled on a ic ligh s o
−60
−70
−80
−90
−100
−110
Dis ance o gNB (m)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
Angle (deg ees)
lamp pos s, he UE may equen ly be wi hin he co e age ange
o only a ew gNBs, o , in he wo s case, jus a single gNB, due o
he inhe en ly limi ed ange o mmWa e signals o building block-
age. This limi ed co e age signi ican ly hinde s he accu acy o
Figu e 1: RSRP s. Dis ance
o wo gNBs in Bos on and
Cha lo e.
Figu e 2: Empi ical exam-
ples o mul i-modali y in an-
gle dis ibu ions.
adi ional mul i-angula ion echniques, which ely on inpu om
mul iple gNBs a he same ime o p ecise localiza ion.
To add ess he i s challenge, we ca ied ou a measu emen
campaign in wo US ci ies, co e ing 17 gNBs deployed by wo majo
mobile ope a o s. Ou analysis e eals a s ong simila i y in he SSB-
o-angle mappings among gNBs om he same ope a o wi hin a
ci y. This allows he cons uc ion o a unique mapping unc ion o
a gi en ci y by using only a small numbe o measu emen s a one
(o a mos a ew) gNB(s) o calib a ion, enabling ci y-le el deploy-
men wi h minimal e o . Fo he second challenge, mm-NOLOC
models he SSB index– o-angle ela ionship p obabilis ically, using
empi ical dis ibu ions de i ed om ou measu emen s o cap u e
To e alua e he eliabili y o RSRP as a localiza ion me ic, we
conduc a con olled walk es in on o wo a bi a y mmWa e
gNBs, one in Bos on and one in Cha lo e. In bo h cases, he UE
mo es s eadily owa d he se ing gNB – om app oxima ely 90
m down o 10 m – while eco ding RSRP and GPS coo dina es.
Acco ding o he heo e ical ee-space pa h loss (FSPL) model,
RSRP is expec ed o inc ease supe linea ly as he UE app oaches
he gNB. Howe e , as shown in Fig. 1, he obse ed co ela ion
be ween RSRP and dis ance is weak and inconsis en ac oss he wo
es loca ions. This disc epancy a ises p ima ily because he pa h
loss model assumes a ixed ansmi powe , whe eas, in p ac ice,
gNBs
employ
adap i e
beam
swi ching
and
powe
con ol.
No ably,
Bos on
Cha lo e
T imodal SSB Index
Bimodal SSB Index
RSRP (dBm)
P obabili y
mm-NOLOC: mmWa e-based Localiza ion o Mobile Ne wo ks wi hou 3GPP Loca ion Se ice
MobiHoc ’25, Oc obe 27–30, 2025, Hous on, TX, USA
he e is also a clea di e ence in he o e all RSRP ange be ween
he wo ci ies, e lec ing a ia ions in an enna gain and ansmi
powe among he deployed gNBs. These obse a ions indica e ha
RSRP is a poo p edic o o dis ance and, consequen ly, o loca ion
es ima ion in p ac ical mmWa e sys ems.
2.2 Handling empi ical SSB- o-Angle Mapping
Gi en he un eliabili y o RSRP o accu a e localiza ion in eal-
wo ld mmWa e ne wo ks, mm-NOLOC shi s ocus om aw signal
s eng h o sec o -le el in o ma ion p o ided by SSB indices. In u-
i i ely, SSB indices, ansmi ed pe iodically by he gNB, p o ide
di ec ional hin s abou he UE’s loca ion o e ing a ough es ima e
o he di ec ion om which he signal a i es. Fu he mo e, he
me ic p o ides mo e s able cues abou he UE’s spa ial o ien a ion
ela i e o he gNB.
Wi h
his
new
ocus,
ano he
s aigh o wa d
hypo hesis
is:
Since
SSBs ca y angula in o ma ion, we can simply in e he UE’s angle
based on he measu ed index and apply iangula ion om mul iple
and le e aged o imp o e localiza ion accu acy by combining hei
spa ial in o ma ion.
Tempo al in o ma ion using UE dynamics. While spa ial cues
om SSB indices o m he ounda ion o ou localiza ion sys em,
mm-NOLOC u he s eng hens i s accu acy and obus ness by in-
co po a ing empo al in o ma ion h ough mo ion modeling. This
empo al laye is especially c i ical in eal-wo ld scena ios, whe e
en i onmen al dynamics, such as sudden blockage, in e mi en is-
ibili y, and apid signal luc ua ions, can cause spa ial obse a ions
o become spa se o ambiguous.
By using spa ial and empo al in o ma ion, mm-NOLOC ans-
o ms un eliable, agmen ed SSB obse a ions in o a cohe en lo-
caliza ion ajec o y. This dynamic in e ence capabili y is especially
powe ul in he challenging and as -changing channel condi ions
o mmWa e links, enabling con inuous and accu a e acking wi h-
ou elying on ex e nal in as uc u e o in e -gNB coo dina ion.
Addi ionally, he UE-cen ic design a oids dependencies on ad-
anced 3GPP loca ion ea u es, making i compa ible wi h oday’s
gNBs o loca ion es ima ion. Howe e , his hypo hesis also b eaks
down in eal-wo ld scena ios. F om ou ex ensi e measu emen
comme cial
de ices.
I
P e ious
-
I
pa icles
campaign
(§3),
we
d aw
wo
undamen al
insigh s:
ssB Angle
I
I
"
,,
'T P edic ed
(1)
The ela ionship be ween SSB index and angula posi ion is Indices
P obabilis ic
dis ibu ion(
I
'
mapping
.._
'
I
Pa icle Dynamic pa icleC
nei he de e minis ic no one- o-one. A single SSB index may be
obse ed om mul iple di ec ions. This necessi a es modeling he
SSB-angle
ela ionship
p obabilis ically,
hinde ing
di ec
applicabil-
/
unc ion
,/
,
,,.
Mapping
unc ion
/
gene a o
model
I
I
IIAngula -
•domain
:pa icles Ca esian-domain
V
/
Pa icle
weigh
upda e
h
Weigh ed
i y
o
de e minis ic
geome ic
app oaches
such
as
iangula ion.
(2)
E en
a e
p obabilis ically
modeling
he
ela ionship
be ween
Cell
/Os
b
ina y
g
NB
s' I
loca ions
a
I
`³
3`
Domain
pa icles
'
7pa icles
Pa icle
esampling
SSB indices and angles, modeling SSB indices as inpu emains
non i ial due o he na u e o hei empi ical dis ibu ions. These
dis ibu ions a e disc e e and can be mul imodal, de ia ing sig-
Ope a o 's
V
,I
da abase
£
ı
ans o ma ion
._
n
I
g
NB
s'
l
on
.
en a
1
c
ons,
I
Resampled
,
7
pa icles
Loca ion
es ima ions
ni ican ly om he heo e ical Gaussian assump ions unde lying
commonly used es ima ion echniques such as he Kalman il e .
As a esul , common es ima ion app oaches canno be di ec ly ap-
plied. We show wo examples o angle dis ibu ion mul imodali y
o wo di e en SSB indices om ou da ase in Fig. 2. Physically,
his beha io is due o impe ec beam pa e ns, which a e com-
mon in comme cial mmWa e an enna a ays, whe e sidelobes may
poin o di ec ions di e en om he main lobe, p oducing mul iple
dominan angles.
These insigh s mo i a e aming he localiza ion p oblem as a
p obabilis ic, non-pa ame ic in e ence p oblem ha le e ages em-
pi ical angle dis ibu ions. Unlike heo e ical models (e.g., Gaussian)
ha may o e look c i ical ea u es like seconda y modes o asym-
me ies, empi ical likelihoods e ain he s uc u e o obse ed da a,
a oiding in o ma ion loss om o e simpli ied cu e i ing.
2.3 Sys em o e iew
A a high le el, mm-NOLOC combines
spa ial in o ma ion
de i ed
om SSB index measu emen s and empo al in o ma ion in e ed
om he UE’s mo ion h ough a pa icle il e o deli e obus and
accu a e loca ion es ima es.
Spa ial in o ma ion h ough SSB indices. As discussed ea lie ,
SSB index measu emen s p o ide di ec ional cues abou he UE’s
posi ion ela i e o a gNB’s posi ion. In mmWa e ne wo ks wi h su -
icien ly dense deploymen s, as is o en he case in u ban canyons,
SSBs
om
mul iple
nea by
gNBs
can
be
simul aneously
obse ed
P ep ocessing s age Fil e ing s age
Figu e 3: Sys em a chi ec u e o mm-NOLOC.
2.4 Sys em a chi ec u e
The sys em a chi ec u e o mm-NOLOC is shown in Fig. 3. We
o ganize he a chi ec u e in o ou s ages: Inpu , P ep ocessing
S age, Fil e ing S age, and Ou pu .
Inpu .
mm-NOLOC uses wo ypes o in o ma ion p o ided by he
5G gNBs o loca ion es ima ions.
•
Synch oniza ion Signal Block (SSB) Indices:
These indices a e
b oadcas pe iodically by su ounding gNBs in mmWa e ne wo ks.
They p o ide ela i e di ec ional cues abou he UE’s posi ion wi h
espec o he ansmi ing gNB.
•
Physical Cell IDs (PCIs):
This in o ma ion helps iden i y gNBs
in he ope a o ’s ne wo k. PCIs enable mm-NOLOC o e ie e he
co esponding gNBs’ loca ion, o ien a ion da a, and which mapping
unc ion o use om he ope a o ’s da abase.
Unlike me ics such as Angle o A i al (AoA), Angle o Depa -
u e (AoD), o Channel S a e In o ma ion (CSI), SSB indices and
PCIs a e undamen al pieces o con ol-plane in o ma ion de ined
by 3GPP and used in ou ine ne wo k ope a ions. Al hough hese
pa ame e s a e no exposed h ough public And oid o iOS APIs
by de aul , mul iple open-sou ce ools like MobileInsigh [20] ha e
demons a ed access o hem ia diagnos ic modes on common
phone chipse s (e.g., Qualcomm, Media ek, Samsung). Hence, ex-
posing his in o ma ion o highe laye s o he p o ocol s ack is a
Es ima ion
I
.._
I
I
I
I
I
I
MobiHoc ’25, Oc obe 27–30, 2025, Hous on, TX, USA
Phuc Dinh e al.
ma e o policy and in e ace exposu e, no ha dwa e o i mwa e
limi a ions, as i is al eady a ailable in modem logs.
P ep ocessing S age. This s age p epa es he inpu s equi ed o
he pa icle il e -based in e ence.
•
The ope a o ’s da abase p o ides h ee essen ial pieces o in-
o ma ion: he posi ions o he gNBs in he se ice a ea, he o ien-
a ions o hei an enna panels, and he co esponding mapping
unc ion used in his se ice a ea o gene a e angle dis ibu ions
o he il e ing s age.
•
The
p obabilis ic mapping unc ion
is cons uc ed empi ically
o iine and p o ided h ough he ope a o ’s da abase in a bina y
o ma . This unc ion ans o ms he SSB indices in o angle dis-
ibu ions, cap u ing he noisy and p obabilis ic na u e o u ban
en i onmen s.
We no e ha , since he p obabilis ic mapping unc ion is unique
o a la ge se ice a ea (e.g., a ci y), as will be shown in §3, and
gNB deploymen is gene ally ixed, he p ep ocessing s age can be
pe o med as a one- ime ope a ion upon en e ing he mm-NOLOC
se ice a ea. Hence, once deployed in a ci y, mm-NOLOC can ope -
a e wi hou u he on-si e measu emen s.
Fil e ing S age. This s age le e ages a pa icle il e amewo k
o deal wi h unce ain y in p obabilis ic SSB- o-angle inpu s o
posi ion es ima ion.
•
The pa icle gene a o c ea es angula -domain pa icles based
on he angle dis ibu ions ob ained om he p obabilis ic mapping
unc ion.
•
The domain ans o ma ion module con e s angula -domain
pa icles in o Ca esian-domain pa icles. This ans o ma ion uses
he e ie ed cell posi ions and o ien a ions o align di ec ional
in o ma ion in o a common posi ional e e ence ame. This module
ensu es seamless in eg a ion be ween he mapping unc ion and
he weigh upda e unc ionali y o he pa icle il e .
•
The dynamic model p opaga es pa icles based on he UE’s
mo ion model and sys em dynamics.
•
In he
pa icle weigh upda e
s ep, pa icles a e weigh ed based
on he likelihood o obse ing he ecei ed SSB indices a hei
hypo hesized posi ions.
•
The pa icle esampling module ocuses compu a ional e-
sou ces on pa icles wi h highe weigh s by esampling hem, mi i-
ga ing pa icle degene acy.
•
The
es ima ion
module p oduces he inal UE’s loca ion by com-
pu ing he weigh ed mean o he esampled pa icles.
Ou pu .
The ou pu p o ides es ima ions o UE’s loca ion, p oduced
by he il e ing s age.
In he ollowing, we p o ide a de ailed desc ip ion o he wo co e
componen s
o
mm-NOLOC–
he
p obabilis ic
mapping
unc ion
in
§3 and he pa icle il e in §4, including hei design and solu ions
o he iden i ied challenges.
3 SSB Index-Angle Mapping
Ob aining he SSB index-angle mapping aces wo majo challenges:
Challenge 1: P op ie a y Codebook In o ma ion. Al hough
he SSBs ecei ed a he UE a e di ec ly ela ed o he UE’s an-
gula posi ion, he speci ic mapping and implemen a ion de ails,
including he beam codebook used ha de ines he SSB pa e ns,
a e p op ie a y in o ma ion and no publicly disclosed. In ac , e en
o ope a o s and de elope s, he knowledge o he beam codebook
alone is no enough, as he SSB index-di ec ion mapping may a y
o a ce ain deg ee in di e en en i onmen s due o blockage, mul-
ipa h, UE o ien a ion, e c. Consequen ly, we adop a p obabilis ic
app oach ins ead o a de e minis ic one. We ob ain he mapping ia
a ield measu emen campaign including 17 di e en gNBs deployed
by wo majo ope a o s in wo ci ies.
Challenge 2: Reusabili y o Mapping Func ions. Ope a o s
may deploy di e en ypes o gNB ha dwa e, aising conce ns
abou he gene aliza ion o he cons uc ed mapping unc ion. I
he mapping was di e en o each gNB, cons uc ing mapping
unc ions would equi e signi ican e o , unde mining he scala-
bili y o ou app oach. Howe e , ou analysis in §3.2 demons a es
ha SSB- o-angle dis ibu ions wi hin a ci y show signi ican simi-
la i y, indica ing ha a ia ions ac oss gNBs s em p ima ily om
en i onmen -speci ic ac o s and no om di e ences in beam code-
books. Thus, we can use a unique, ci y-wide p obabilis ic mapping
unc ion, which can be de i ed using a small numbe o measu e-
men s a one (o a mos a ew) gNBs. In he es o his sec ion, we
desc ibe (§3.1) ou measu emen me hodology o ob aining he
SSB index-angle mapping o each ci y ha add esses
Challenge 1
.
We hen desc ibe (§3.2) ou analysis o he collec ed da ase , which
highligh s he eusabili y o he ob ained mappings om a small
se o gNBs by compa ing hem wi h he o e all dis ibu ions om
all su eyed gNBs, he e o e add essing Challenge 2.
3.1 SSB Index-Angle Measu emen Me hodology
3.1.1
Measu emen Loca ions and De ices.
We conduc ed measu e-
men s in wo U.S. ci ies – Bos on and Cha lo e – ha ea u e
comme cial 5G mmWa e deploymen s by Ve izon and AT&T. Fo
Ve izon in Cha lo e, we iden i ied a s ee wi h dense co e age,
ea u ing 7 consecu i e gNBs deployed on nea ly e e y block. Fo
Ve izon in Bos on, we iden i ied wo la ge zones wi h a o al o 8
mmWa e gNBs. On a e age, Ve izon’s mmWa e gNBs a e deployed
app oxima ely 100 m o less apa in bo h ci ies, enabling mo e con-
inuous mmWa e co e age. In con as , AT&T’s gNBs a e ypically
spaced a ound 1 km apa in bo h ci ies, esul ing in in e mi en
mmWa e se ice and equen allback o sub-6 GHz 5G.
Ou measu emen s we e pe o med using a Samsung Galaxy
S21 sma phone equipped wi h a Qualcomm 5G modem – he mos
widely adop ed mmWa e chipse in comme cial de ices in he U.S.
– ensu ing ha ou da a e lec s ypical use ha dwa e. Fo p ecise
localiza ion, we used an RTK-GPS module o ob ain cen ime e -
le el posi ioning accu acy. SSB and PCI in o ma ion was collec ed
using an Accu e XCAL-Solo [2], which connec s o he phone’s
USB-C po and in e aces wi h he diagnos ic subsys em o log
PHY-laye KPIs and signaling messages.
No e ha he use o specialized ha dwa e like XCAL-Solo is no a
necessi y o ex ac PCI and SSB in o ma ion. Many so wa e-based
solu ions ha e been de eloped o ex ac simila con ol-plane in-
o ma ion wi hou equi ing non-s anda d i mwa e modi ica ions,
e.g., [20]. Exposing his in o ma ion o app de elope s is inally a
ma e o policy, as we no iced in 2.4. Simila ly, we used an RTK
GPS solely o ob ain high-p ecision da a o ou e alua ion. In
p ac ice, ope a o s al eady know he exac loca ions and SSB con-
igu a ions o hei gNBs, and can ob ain beam-angle mappings
om he ha dwa e endo as pa o equipmen speci ica ions.
mm-NOLOC: mmWa e-based Localiza ion o Mobile Ne wo ks wi hou 3GPP Loca ion Se ice
MobiHoc ’25, Oc obe 27–30, 2025, Hous on, TX, USA
(a)
AT&T
Coo dina es
(b)
Ve izon Coo dina es
p o ides a good ade-o be ween measu emen o e head and
accu acy. Finally, al hough XCAL eco ds SSB indices e e y ew
hund ed ms, sho - e m blockages om ehicles o pedes ians a e
common in u ban se ings. Collec ing da a o e 5 s pe o ien a ion
mi iga es hese ansien e ec s, p oducing a mo e obus da ase .
3.2 Analysis o SSB Index Measu emen s
A e da a collec ion, we build he mapping om se ing SSB indices
o angles. Fo each angle, we ex ac he pe cen age o ime each
SSB Index is measu ed and hen analyze he SSB Index dis ibu ions
o
each
angle.
No e
ha ,
due
o
he
spa se
deploymen
o
AT&T
Figu e 4: Cons uc ed angula coo dina es.
3.1.2
Cons uc ing angula coo dina es. To build a mapping be-
ween SSB Indices and angles, we i s need o cons uc an angula
coo dina e sys em su ounding each gNB we expe imen wi h. Each
o hese coo dina e sys ems is cons uc ed using he ollowing s eps:
•
De ining he 0𝑜 angle: Since he beam codebook is no known,
and we only need o know ela i e angle measu emen s om gNBs
o loca e he UE, we can de ine he
0
𝑜
angle a bi a ily as a e e ence
o measu ing o he angles, as long as he de ini ion is consis en
among di e en gNBs. In ou cons uc ed coo dina e sys em, we
de ine he 0𝑜 as he di ec ion ha is pe pendicula o he an enna
panel. We ob ain he GPS coo dina es o he gNB and a andom
poin on he 0𝑜 di ec ion o cons uc he di ec ion ec o .
•
Building he angula coo dina e sys em: A e ob aining he
0𝑜 di ec ion ec o o a gNB, we use Google Ea h P o, which
suppo s
accu a e
angle
measu emen s,
o
manually
gene a e
he
coo dina es o ha gNB. Unlike simple plo ing ools, Google
Ea h P o ensu es ha he measu ed angles and poin s a e accu a e
eal-wo ld angles and measu emen poin s. We cons uc angles
in s eps o 10𝑜 su ounding each gNB. Then, we ob ain he GPS
coo dina es o he in e sec ions be ween an inne ci cle wi h adius
o 10 me e s and an ou e ci cles a adius 20 o me e s and used
hese in e sec ion as "expe imen poin s" o SSB measu emen s.
Fig. 4 shows examples o cons uc ed coo dina e sys ems o wo
gNBs om Ve izon and AT&T. The cen e o he ci cle is he gNB
posi ion, while he
0
𝑜
angles a e ma ked in ed. The ci cles in e sec
wi h he lines iden i ying he angles o c ea e expe imen poin s.
No e ha he angle no a ions used in coo dina e sys ems in Fig. 4
di e sligh ly be ween AT&T and Ve izon due o hei di e ing
app oaches o gNB iden i ica ion. Speci ically, 5G mmWa e gNBs
a e ypically equipped wi h h ee an enna panels. Ve izon assigns a
unique PCI o each panel, esul ing in h ee PCIs pe gNB, whe eas
AT&T uses a single PCI o iden i y an en i e 3-panel gNB.
3.1.3
Expe imen al P ocedu e. A e de ining he coo dina e sys-
em o each gNB, we collec SSB measu emen s a e e y expe i-
men poin using he ollowing p ocedu e: • S ep 1: Posi ion he
phone so ha i di ec ly aces he se ing an enna panel. • S ep
2: Run a sc ip ha sends a small amoun o a ic o 5 s. Du ing
his in e al, he sc ip logs he angle and de ice o ien a ion, while
XCAL-Solo eco ds he se ing SSB index. • S ep 3: Ro a e 45◦
clockwise and epea S ep 2 wi h he new o ien a ion. • S ep 4:
Con inue
S ep
3
un il
comple ing
a
ull
360
◦
sweep.
Since a UE in mo ion may ace any di ec ion and expe ience
pa ial o ull body blockage, he SSB index is no assumed cons an
ac oss
o ien a ions.
Cap u ing
SSB
indices
in
8
di ec ions
pe
poin
mmWa e gNBs and he loca ions o mos gNBs (o en in he middle
o he oad), we only managed o do expe imen s wi h 2 AT&T
gNBs wi h 11 di e en angles. Fo Ve izon, he deploymen is much
dense gi ing us mul iple op ions. Speci ically, we expe imen ed
wi h 8 gNBs in Bos on and 7 in Cha lo e. Due o space limi a ion,
we canno show he dis ibu ions o all 17 gNBs. Fig. 5 shows a
compa ison be ween wo ep esen a i e Ve izon gNBs in Cha lo e
wi h a Ve izon gNB and ano he AT&T gNB in Bos on. We make
wo obse a ions om ou analysis and Fig. 5.
Obse a ion 1: many- o-many ela ionship be ween measu ed SSB
Index and angle
. We obse e ha mul iple SSB indices can be associ-
a ed wi h he same angle. Fo example, in Bos on-Ve izon (Fig. 5a),
we o en see 9-10 SSB indices o a gi en angle. This can be a -
ibu ed o (i) he use o 3D beam o ming, due o which, he SSB
selec ion is a ec ed no only by he angles bu also he dis ance
o he gNB, and (ii) en i onmen al noise such as ansien o pe -
manen blockage, which causes he UE o ecei e SSBs om non-
line-o -sigh (nLOS) pa hs. Con e sely, he same SSB index can
co e mul iple angles, as illus a ed by epea ed colo s a di e en
measu ed angles in Figs. 5a-5d. This is unde s andable, since one
SSB index can co e a sec o la ge han
10
0
, especially in scena ios
wi h a low numbe o suppo ed beams. Ano he eason is he p es-
ence o i egula beam pa e ns ea u ing one main lobe poin ing a
one main angle and mul iple side lobes poin ing a di e en angles.
Obse a ion 2: SSB Index con igu a ions can a y ope a o -wise o
ci y-wise. In pa icula , Ve izon gNBs in Bos on ha e 48 ac i a ed
beams pe PCI, while AT&T in Bos on ha e 24 pe PCI (only a
subse o angles a e shown due o he limi ed numbe o easible
measu ed posi ions), and Ve izon gNBs in Cha lo e ha e 12 pe
PCI. They also ha e di e en dis ibu ions. Fo example, in Bos on-
Ve izon case, he mos equen ly obse ed beam a angle 0 is SSB
Index 25 (pink) while in Cha lo e-Ve izon, beam 11 (g een) is mos
equen ly obse ed a he same angle.
Obse a ion 3: SSB Index con igu a ions o he same ope a o in he
same a ea exhibi simila dis ibu ions. Speci ically, we obse e ha
SSB indices measu ed a he same angle o di e en gNBs wi hin
a gi en se ice a ea display simila pa e ns in hei dis ibu ion.
Simply pu , we o en obse e simila se s o o dominan indices pe
measu ed angle ac oss gNBs in he same ci y. An example is shown
in Figs. 5b and 5c. To o mally quan i y his obse a ion and com-
pa e indi idual gNBs’ dis ibu ions wi h he o e all dis ibu ions,
Fig. 6 p esen s hea maps o he a e age pai wise Jensen-Shannon
di e gence (JSD) o he SSB index-angle dis ibu ions o a ew
di e en gNBs as compa ed o he o e all dis ibu ions o all gNBs
in ou da ase . JSD quan i ies he dissimila i y be ween wo dis i-
bu ions, anging om 0 (iden ical dis ibu ions) o 1 (comple ely
110
100
140
5
70
180
190
200
210
40
AT&T gNB
30
20
10
240
250
340
330
260
320
270
280 290
300
310
4
60
50
-50
-40
-30
-20
Ve izon gNB
40
30
20
10
10
20
-20
-30
30 40 50 60 -50 -40
1
160
130
120 90 80
0
60
10 0 -10 -20
20 -30
30 -
40
0
-50
170
50
50
60
220
0
-10
0
230
350
0
-10

MobiHoc ’25, Oc obe 27–30, 2025, Hous on, TX, USA
Phuc Dinh e al.
1.0
0.5
0.0
Angle (deg ee)
(a)
Bos on-Ve izon
1.0
0.5
0.0
Angle (deg ee)
(b)
Cha lo e-Ve izon 1
1.0
0.5
0.0
Angle (deg ee)
(c)
Cha lo e-Ve izon 2
1.0
0.5
0.0
Angle (deg ee)
(d)
Bos on-AT&T
Figu e 5: Empi ical dis ibu ions o obse ed SSB indices o di e en angles ac oss gNBs in di e en ci ies and ope a o s. Each
colo ep esen s a dis inc SSB index de ec ed a he co esponding angle.
dissimila ). Ou esul s show ha , when compa ing di e en gNBs
agains each o he and he o e all dis ibu ion, mos JSD alues
a e close o 0.1, wi h he la ges di e gence (0.2) obse ed be ween
gNB 1 and gNB 2 o Ve izon in Bos on. This inding unde sco es
wo key poin s: (i) ope a o s gene ally deploy gNBs wi h he same
beam codebook ac oss a la ge se ice a ea, and (ii) sligh a ia ions
in SSB indices a simila angles a e p ima ily due o en i onmen al
ac o s a he han undamen al di e ences in codebook. In ui i ely,
i he beam codebooks we e en i ely di e en o each gNB, he SSB-
angle dis ibu ions would lack simila i y, esul ing in JSD alues
close o 1. Mo e impo an ly, his esul highligh s a p ac ical ad-
an age: he SSB index-angle mapping can gene alize o a la ge a ea
(e.g., a ci y) using measu emen s om only a small subse o gNBs.
This signi ican ly educes he measu emen e o equi ed, making
mm-NOLOC scalable and p ac ical o eal-wo ld deploymen .
4.1 Modeling Ra ionale
A i s co e, mm-NOLOC le e ages a nonpa ame ic, da a-d i en
likelihood mapping unc ion, quan i ying he likelihood o di e -
en angula posi ions o measu ed SSB Indices. Unlike adi ional
Gaussian-based modeling app oaches, ou modeling app oach cap-
u es he disc e e, spa se, and mul imodal na u e o mmWa e signal
p opaga ion, e lec ing eal-wo ld e ec s like sidelobes o blockage.
All
g1
g2
g3
All
g1 g2
All
g1
g2
g3
g3
All g1 g2
All
g1
g2
g3
All g1 g2
0.5
0.4
0.3
0.2
0.1
0.0
(a)
Bos on-Ve izon
(b)
Cha lo e-Ve izon
(c)
Bos on-AT&T
Figu e 6: Pai wise Jensen-Shannon Di e gences be ween di -
e en gNBs o he same Ci y/Ope a o .
We also no e ha , in p ac ice, ope a o s ob ain hei ha dwa e
om one (o a ew) endo (s) (e.g., Samsung, Nokia, and E icsson
in he US) and a e u he cons ained by 3GPP s anda ds ha
limi SSB con igu a ion di e si y. As a esul , only a small numbe
o SSB con igu a ions a e used na ionwide. This was con i med
expe imen ally by ou ecen s udy [10], in which we d o e in 11
US ci ies wi h mmWa e deploymen s, and ound only 2 dis inc
con igu a ions o each o he wo main 5G mmWa e ope a o s
(Ve izon and AT&T). Thus, ou obse a ion in his wo k ha a single
mapping unc ion gene alizes wi hin a ci y is ac ually conse a i e;
eal deploymen s sugges e en b oade euse.
4 Pa icle Fil e Model
While he use o pa icle il e ing o loca ion es ima ion is no new,
i s e ec i eness hinges on he ca e ul cons uc ion o measu emen
models. This sec ion p o ides he a ionale and in ui ion behind ou
modeling app oach be o e p esen ing he ma hema ical o mula ion.
In 4.1, we desc ibe in ui i ely how pa icle il e ing is pa icula ly
well-sui ed o ou se ing, gi en he unce ain y in oduced by
mmWa e-speci ic cha ac e is ics, such as signal blockage, NLoS
pa hs, and he inhe en ly mul imodal na u e o he obse ed SSB-
o-angle ela ionship in eal-wo ld deploymen s. The o maliza ion
o ou in ui ion is desc ibed in 4.2.
Figu e 7: A concep ual isualiza ion o loca ion es ima ions
based on SSB measu emen s wi h pa icle il e ing.
To p o ide a high-le el in ui ion, we p esen a concep ual i-
sualiza ion o ou sys em in Fig. 7, illus a ing how mm-NOLOC
uses a pa icle il e o ope a e unde a ealis ic mmWa e scena io.
Upon en e ing a mmWa e se ice egion, he UE may expe ience
pa ial blockage (e.g., be ween 𝑡 = 1 and 𝑡 = 3), du ing which i
ecei es SSBs only om a single gNB. Consequen ly, he pa icle
clus e , ep esen ing he se o hypo hesized UE posi ions gene a ed
using mm-NOLOC’s empi ical likelihood dis ibu ions, becomes
inc easingly unce ain, as indica ed by he g owing sp ead o he
pa icles. A 𝑡 = 4, when he UE exi s he blockage and ecei es
an SSB measu emen om a new gNB, his addi ional spa ial in-
o ma ion educes he unce ain y, as i aligns well wi h he UE’s
his o ical mo ion (mo ing om le o igh ), esul ing in a mo e
con iden posi ion es ima e.
Na u ally, mm-NOLOC’s measu emen model also inco po a es
SSB indices ecei ed om mul iple gNBs, as shown a 𝑡 = 5. This
u he inc eases obus ness in es ima ion and educes unce ain y.
In e ec , he sys em bene i s om a o ing-like mechanism, whe e
mul iple independen obse a ions (collec ed om di e en gNBs)
con ibu e o he o e all likelihood es ima e, mi iga ing he impac
o any single e oneous o ambiguous measu emen . This capabil-
i y is pa icula ly aluable in dense u ban en i onmen s, whe e
0.12 0.11
0
0.08
0.12 0 0.2 0.18
0.11 0.2 0 0.16
0.08
0.18 0.16 0
0.06
0.04
0
0.09
0.06 0 0.14
0.09
0.09
0.14 0 0.17
0.04
0.09 0.17 0
0.13 0.12
0
0
0.15
0.13
0.12
0.15
0
P obabili y
P obabili y
P obabili y
P obabili y
mm-NOLOC: mmWa e-based Localiza ion o Mobile Ne wo ks wi hou 3GPP Loca ion Se ice
MobiHoc ’25, Oc obe 27–30, 2025, Hous on, TX, USA
𝑘
.
.
𝑘
𝑘
𝑘
𝑘
o e lapping co e age om mul iple gNBs can be le e aged wi h-
ou equi ing in e -gNB coo dina ion o addi ional in as uc u e.
Mo eo e , his o ing-like mechanism e ec i ely handles cases
whe e he SSB index dis ibu ion is mul i-modal, i.e., when he e
a e mul iple dominan angles, by discoun ing unlikely angles ha
a e
inconsis en
wi h
he
di ec ionali y
o
o he
SSB
measu emen s.
4.2 Ma hema ical Fo mula ion
In his sec ion, we o malize he p oposed measu emen model
Fo simplici y, we call ec o s c
𝑘
and s
𝑘
he PCI ec o and he SSB
ec o , espec i ely.
Since he posi ion and o ien a ion o he cell deployed by he
se ice p o ide a e p o ided by he ope a o , le us deno e he
lookup unc ion as L(.). The ec o s o cell posi ions and hei
di ec ional ec o s can be ob ained as:
(P𝑘, D𝑘 ) = L(c𝑘 ) (6)
whe e P𝑘 = (p1,𝑘, p2,𝑘 , . . . , p𝑀,𝑘 ), (7)
D𝑘 = (d1,𝑘, d2,𝑘 , . . . , d𝑀,𝑘 ),
and how i i s in o he o e all pa icle il e amewo k oge he
wi h
p
( o
𝑚
=
1
,
2
, . . . , 𝑀
)
he
posi ion
o
he
𝑚
- h
se ing
cell,
wi h
ou
de i ed
empi ical
dis ibu ion
collec ed
by
ou
ex ensi e
𝑚,𝑘
and d ( o = 1 2 ) he di ec ional ec o o he cell’s
measu emen s.
𝑚,𝑘
𝑚 , , . . . , 𝑀
Gi en a se o 𝑁 pa icles {x𝑖 , 𝑤𝑖 }
𝑁
, whe e x
𝑖
is he s a e
an enna
o
he
𝑚
- h
cell.
ec o and
𝑖 𝑘 𝑘 𝑖=1 𝑘 Since we associa e he angle o he di ec ional ec o wi h i s
𝑤 is he weigh o he 𝑖- h pa icle, he il e ope a es
using he oll𝑘owing key s eps:
•
P edic ion:
P opaga e
each
pa icle
h ough
he
s a e
ansi ion
0𝑜 angle by con en ion, we de ine he ela i e angle o a UE om
a cell as he di e ence be ween he UE’s posi ion ec o and he
cell di ec ion
ec o .
The e o e,
he
angle
o
a
UE
a
he
ime s ep
𝑘
model wi h noise ec o 𝑖 : x𝑖 ∼ 𝑝 (x𝑘 |x𝑖
−1, 𝑖 )
om
PCI
𝑚
can
be
w i en
as
•
Upda e:
Adjus
weigh s
based
on
he measu emen
model:
𝑤
𝑖
∝
𝑝
(
z
𝑘
|
x
𝑖
)
𝜃
=
180
·
a an2
×
d ,
·
d
(8)
𝑘
𝑚,𝑘
•
𝑚,𝑘
𝑚,𝑘
𝑚,𝑘
𝑚,𝑘
Resampling: Resample pa icles o ocus on hose wi h highe
weigh s. whe e
𝑚,𝑘
= p
𝑘
–
p
𝑚,𝑘
is
he
posi ion
ec o
o
he
UE
ela i e
o
⊤
•
Es ima ion: Compu e he s a e es ima e as: xˆ𝑘 =
.
𝑤𝑖 x𝑖
PCI
𝑚
, wi h
p
𝑘
=
(
𝑥
𝑘
, 𝑦
𝑘
)
ep esen ing
he
posi ion
componen
o
Sys em Dynamic Model. 𝑖=1 𝑘 𝑘
he UE’s
s a e
ec o
x
𝑘
, while
𝑚,𝑘
×
d
𝑚,𝑘
is
he
c oss p oduc ,
and
We p esen he dynamic model o ou
sys em, which cap u es he e olu ion o he s a e o e ime. This
model is u ilized wi hin he pa icle il e amewo k o es ima e
he
s a e
o
he
sys em
om
noisy
measu emen s.
The
goal
o
he
𝑚,𝑘
·
d
𝑚,𝑘
is he do p oduc be ween he wo ec o s. The a an2
unc ion compu es he angle in adians be ween he posi i e x-axis
and
UE’s
posi ion,
ensu ing
he
co ec
quad an .
By
mul iplying
by 180 , he esul is con e ed o deg ees.
il e is o es ima e he posi ion
(
𝑥, 𝑦
)
o a Use Equipmen (UE) in
a
p ede ined
coo dina e
sys em.
We
deno e
he
s a e
ec o
o
he
Since he p obabili ies o obse ing
(𝑐
o all 𝑀 PCIs, we can w i e:
𝑚,𝑘 , 𝑠
𝑚,𝑘 ) a e independen
sys em
a
ime
𝑘
as
x = (𝑥 , 𝑦 , 𝑣
, 𝑣
)⊤ (1)
𝑀
𝑃 (z |x ) =
𝑃
(𝑐 , 𝑠 |x ) (9)
𝑘 𝑘 𝑘 𝑥,𝑘
𝑦,𝑘
𝑘 k
𝑚,𝑘
𝑚,𝑘 𝑘
whe e 𝑥𝑘 is he x-coo dina e o he UE a ime 𝑘, 𝑦𝑘 is he y-
coo dina e o he UE a ime 𝑘, 𝑣𝑥,𝑘 is he eloci y o he UE along
he x-axis a ime 𝑘, and 𝑣𝑦,𝑘 is he eloci y o he UE along he
y-axis a ime 𝑘.
Nex ,
we
de ine
he
s a e
ansi ion
model
as
x𝑘+1 = Fx𝑘 + 𝑘 (2)
whe e F is he s a e ansi ion ma ix, and p ocess noise 𝑘 ∼
𝑚=1
Fu he mo e since he s a e ec o x𝑘 ollows a de e minis ic ela-
ionship wi h i s ela i e angle 𝜃𝑚,𝑘 o a gi en 𝑚 as es ablished in
(8), (9) is equi alen o:
𝑀
𝑃
(
z
𝑘
|
x
k
)
=
𝑃
(
𝑐
𝑚,𝑘
, 𝑠
𝑚,𝑘
|
𝜃
𝑚
,
𝑘
)
(10)
𝑚=1
The e o e, o any s a e hypo hesis x𝑖 gene a ed by he pa icle
N (0, Q𝑘 ) is a join ze o-mean Gaussian dis ibu ion wi h co a i-
ance ma ix Q
𝑘
, ep esen ing p ocess noise; Δ
𝑡
ep esen s he ime
in e al be ween successi e s a e upda es, e lec ing he disc e e-
ime na u e o he sys em’s e olu ion.
Measu emen Model and Weigh Upda e. Nex , we discuss he
measu emen model o he il e . In o he wo ds, we p esen how
he pa icle weigh s a e upda ed using he measu emen s ecei ed
a he UE.
z𝑘 = (𝑐1,𝑘 , 𝑠1,𝑘 , 𝑐2,𝑘 , 𝑠2,𝑘 , . . . , 𝑐𝑀,𝑘 , 𝑠𝑀,𝑘 ) (3)
whe e
𝑐
𝑚,𝑘
ep esen s he PCI o he
𝑚
- h cell/gNB
1
and
𝑠
𝑚,𝑘
ep e-
sen s he SSB index om he 𝑚- h cell a ime 𝑘 o 𝑚 = 1, 2, . . . , 𝑀.
The measu emen ec o z𝑘 can be b oken down in o wo sub-
ec o s c𝑘 and s𝑘 as ollows:
c𝑘 = (𝑐1,𝑘 , 𝑐2,𝑘 , . . . , 𝑐𝑀,𝑘 ) (4)
s𝑘 = (𝑠1,𝑘 , 𝑠2,𝑘 , . . . , 𝑠𝑀,𝑘 ) (5)
1
Recall
ha
each
gNB
may
include
one
o
mo e
PCIs
depending
on
deploymen
ype.
il e , we can i s ans o m i o angula
𝑘
domain using (8), hen use
(10)
as
he
weigh
upda e
unc ion.
(10)
is
he
likelihood
o
obse ing
a gi en SSB index a a gi en angle, which can be ob ained using ou
collec ed measu emen da ase in §3. This concludes ou de i a ion.
The
o e all
algo i hm
is
shown
in
Algo i hm
1.
5 E alua ion
In his sec ion, we e alua e mm-NOLOC h ough eal-wo ld expe -
imen s and ace-d i en simula ions. Fo con ex , he GPS baseline
used in ou expe imen s is dual- equency GNSS wi h senso usion
enabled on mode n sma phone chipse s, ep esen ing a s a e-o -
he-a con igu a ion. We emphasize ha ou goal is no o show
ha mm-NOLOC ou pe o ms e e y p oposed localiza ion solu-
ion in he li e a u e, bu o demons a e a p ac ical solu ion ha
p o ides accu acy compa able o ha o GPS in u ban canyon en i-
onmen s, using only s anda dized con ol-plane in o ma ion cu -
en ly
a ailable
on
he
UE
side.
S a e-o - he-a
schemes
om
he
𝜋
𝜋
MobiHoc ’25, Oc obe 27–30, 2025, Hous on, TX, USA
Phuc Dinh e al.
G ound T u h
mm-NOLOC
GPS
G ound T u h
mm-NOLOC
GPS
Algo i hm 1 Pa icle Fil e Algo i hm o mm-NOLOC
80
60
40
20
0 0 20
40
60
80 100 120
ΔUTM X (m)
1.0
0.8
0.6
0.4
0.2
0.0
0
1
2
3
4
5
Euclidean Dis ance E o (m)
. (a) Type-1 ajec o y
80
60
40
20
0 0 20
40
60
80 100 120
ΔUTM X (m)
1.0
0.8
0.6
0.4
0.2
0.0
1 2 3 4 5 6 7 8
Euclidean Dis ance E o (m)
(b) Type-2 ajec o y
𝑘 1,𝑘
𝑀,𝑘
80
60
40
20
Figu e 8: Pe o mance o mm-NOLOC in Bos on.
80
60
40
20
12: Weigh Upda e
13:
Collec
measu emen
ec o
as
in
(3)
14:
Compu e he weigh o pa icle
𝑛
based on he measu e-
men ec o using (10)
15: end o
16: Weigh No maliza ion
17: Es ima ion
18: Compu e he s a e es ima ion o ime 𝑘 as he weigh ed
0 0 20 40 60 80
ΔUTM X (m)
1.0
0.8
0.6
0.4
0.2
0.0 0 1 2 3 4 5 6 7
Euclidean Dis ance E o (m)
(a) Type-1 ajec o y
0
0 20 40 60 80
ΔUTM X (m)
1.0
0.8
0.6
0.4
0.2
0.0 0 2 4 6 8 10 12 14
Euclidean Dis ance E o (m)
(b) Type-2 ajec o y
mean
pa icle
s a e
19: Resampling
Resampled new pa icles based on he upda ed weigh s, used
o nex -i e a ion p edic ion.
20: end o
li e a u e may indeed p o ide highe accu acy han mm-NOLOC.
Howe e , hey a e no cu en ly deployed in comme cial 5G ne -
wo ks, because o hei equi ed complexi y, making mm-NOLOC
he only easily deployable op ion.
5.1 E alua ion Me hodology
Despi e p omises o dense deploymen in u ban a eas, he majo i y
o comme cial 5G mmWa e deploymen s oday a e s ill in e mi en
and spa se, esul ing in co e age gaps and making i di icul o
achie e con inuous mmWa e se ice. This poses a g ea challenge
o eal-wo ld e alua ions. Du ing ou ex ensi e expe imen s, we
ound wo si es wi h dense enough deploymen o e alua ions o
Ve izon in Bos on and Cha lo e, while AT&T gNBs a e deployed
oo a apa . The e o e, we ocus on Ve izon deploymen s o ensu e
a ai and consis en e alua ion ac oss ci ies.
Es ablishing g ound u h. Since ou expe imen s a e conduc ed
in c owded u ban a eas wi h su ounding all buildings, GPS mea-
su emen s a e highly inaccu a e and canno be used o es ablish
g ound u h posi ion. Ins ead, we manually cons uc ed g ound
u h posi ion o e ime. Speci ically, we used Google Ea h o
i s iden i y he coo dina es o mul iple landma ks along ou in-
ended e alua ion ajec o ies. Du ing expe imen s, we mo ed a a
pedes ian speed and logged he imes amps upon a i al a each
landma k, while also logging PCI and SSB measu emen s in he
backg ound. Then, we econs uc ed he g ound u h using he
logged ime s amps and landma k coo dina es using in e pola ion
o ou e alua ion.
Choosing a coo dina e sys em. Ou doo posi ioning ypically
uses
he
Geog aphic
Coo dina e
Sys em
(GCS)
wi h
la i udes
and
Figu e 9: Pe o mance o mm-NOLOC in Cha lo e.
longi udes. Fo p ecise dis ance calcula ions and geome ic ope -
a ions,
we
ins ead
use
he
Uni e sal
T ans e se
Me ca o
(UTM),
a 2D Ca esian sys em. In ou e alua ions, we o se UTM alues
wi h an app op ia e e e ence poin o be e isualiza ion.
5.2 Real-wo ld E alua ion
In his sec ion, we conduc a eal-wo ld e alua ion by pe o ming
mobili y expe imen s in a eas wi h comme cially deployed Ve izon
mmWa e gNBs in Bos on and Cha lo e. We e alua e wo ypes o
ajec o ies. Fo Type-1 ajec o y, we mo e s aigh on he side-
walk o a s ee whe e mul iple gNBs a e deployed o a ound wo
minu es a pedes ian speed. Fo Type-2 ajec o y, we i s mo e
s aigh and hen make a U- u n and con inue walking o a o al
du a ion o a ound 4 minu es in he chosen mmWa e se ice a eas.
We also log he GPS coo dina es on he phone o compa ison wi h
mm-NOLOC. The pu pose o wo ypes o ajec o y is o see how
well he il e ing algo i hm pe o ms wi h sudden changes in he
eloci y ec o .
As shown in Figs. 8 and Fig. 9, mm-NOLOC achie es simila
median e o s in he wo ci ies – 3.2/4.5 m in Bos on and 2.5/4.5
m in Cha lo e o he wo ajec o y ypes. Addi ionally, he max-
imum e o is also kep e y low – below 6/9 m in Bos on and
below 7/14 m in Cha lo e. The accu acy is simila o o e en be e
han he GPS accu acy; in ac , in Cha lo e o Type-1 ajec o y,
mm-NOLOC ou pe o ms GPS o he whole ajec o y. We also
obse e ha Cha lo e has much highe maximum e o s compa ed
o Bos on, especially o he Type-2 ajec o y. This can be a ib-
u ed o Ve izon in Cha lo e using wide beams, leading o lowe
angle esolu ion and spa se pa icle weigh upda es.
We no e ha hese es a eas do no co espond o he mos se-
e e u ban canyon scena ios, and GNSS in ou measu emen s, while
deg aded, emains gene ally a ailable. In uly challenging en i on-
men s, such as hose documen ed in [12] wi h mean GNSS e o s
exceeding 10 m and maximum e o s o e 200 m, he ad an age
o mm-NOLOC would be e en mo e p onounced. We also s ess
mm-NOLOC
GPS
GPS
LOC
mm-NO
CDF
ΔUTM Y (m)
CDF
ΔUTM Y (m)
CDF
ΔUTM Y (m)
CDF
ΔUTM Y (m)
1:
2:
Inpu : 𝑁 , 𝑘 , Δ𝑡
Ini ializa ion:
3:
o 𝑛 = 1 o 𝑁 do
4:
5:
x𝑛 = FxGPS + 𝑘
0
−
1
𝑤𝑛 = 1/𝑃
6:
end o
7:
o 𝑘 = 1 o 𝐾 do
8:
o 𝑛 = 1 o 𝑁 do
9:
P edic ion
10:
P edic
new
s a e
o
pa icle
𝑛
using
(2)
11:
Compu e he angle ec o 𝜽𝑛 = [𝜃𝑛 , ...𝜃𝑝
] o 𝑀 asso-
cia ed
se ing
PCIs
using
(8)
mm-NOLOC: mmWa e-based Localiza ion o Mobile Ne wo ks wi hou 3GPP Loca ion Se ice
MobiHoc ’25, Oc obe 27–30, 2025, Hous on, TX, USA
1.0
0.8
0.6
0.4
0.2
0.0
0
5 10 15 20 25 30 35
1.0
0.8
0.6
0.4
0.2
0.0
0 5 10 15 20 25 30 35
To alida e his hypo hesis, we p esen a sca e plo in Fig. 11(d),
showing he maximum localiza ion e o obse ed in each ajec-
o y e sus he a e age dis ance a eled pe gNB, wi h highe
alues indica ing spa se deploymen s. We obse e a clea posi i e
co ela ion be ween gNB spacing and maximum e o , wi h a ew
excep ions. This highligh s he impo ance o dense gNB deploy-
E o s in Euclidean Dis ance (m)
(a)
Cumula i e pe o mance.
E o s in Euclidean Dis ance (m)
(b)
Impac o ajec o y leng h.
men s in educing ex eme e o s and ensu ing obus localiza ion
pe o mance.
Figu e 10: T ace-based simula ion o e 100 ajec o ies.
again ha he design goal o mm-NOLOC is no o eplace GPS bu
o complemen i , p o iding a obus allback when sa elli e-based
posi ioning becomes un eliable o una ailable. This complemen a y
ole is pa icula ly ele an o dense u ban deploymen s whe e
comme cial 5G mmWa e co e age al eady exis s.
5.3 T ace-based Simula ions
Ano he limi a ion o oday’s comme cial mmWa e deploymen s is
ha , in mos ci ies (including he wo ci ies whe e we conduc ed ou
expe imen s), mul iple gNBs a e ypically deployed on a s aigh
line along one majo s ee . As we u n away om his s ee , he
UE s ops ecei ing SSBs om mul iple gNBs o e en disconnec s
om mmWa e se ice due o blockage om all buildings. This
deploymen hinde ed e alua ions wi h long, u n-hea y ajec o ies
o s ess- es ou solu ion.
To o e come his limi a ion, we buil a simula o based on he col-
lec ed mobili y aces om eal-wo ld mmWa e gNB deploymen s
in mul iple loca ions in wo su eyed ci ies. Since ou measu emen
me hodology (§3) enables us o es ablish g ound u h posi ions
using high-accu acy GPS de ices, we can de i e he g ound u h
angles ela i e o hese gNBs o any gi en measu ed SSB index.
Le e aging hese eal-wo ld obse a ions, we c ea e mul iple "sim-
ula ed" gNBs wi h angle-SSB p o iles ha mimic he beha io o
ac ual ope a o -deployed gNBs unde simila condi ions (u ban en-
i onmen s). We place hese "simula ed" gNBs a po en ial deploy-
men si es o o m a g id-like deploymen opology. This simula o
allows us o e alua e mm-NOLOC on di e se syn he ic ajec o ies
ha include mo e complex mobili y pa e ns, such as equen u ns
and mul i-s ee a e sal, unde con olled bu ealis ic condi ions.
Fig. 10a shows he pe o mance o mm-NOLOC cumula i ely o
100 andomly gene a ed ajec o ies o a ying leng hs in ol ing
le -, igh -, and U- u ns. The esul s ag ee wi h hose in Figs. 8,
9,
showing
a
median
e o
o
3
m
and
a
95- h
pe cen ile
below
10
m. Fig. 10b plo s he esul s o he same 100 ajec o ies b oken
down by he ajec o y leng h. In e es ingly, we obse e ha he
ajec o y leng h (measu ed in e ms o he numbe o gNBs he
UE a e ses h ough) has li le impac on accu acy.
To u he analyze he high-e o cases in ou esul s, we ocus
on speci ic examples in Bos on, whe e we obse e wo s -case e -
o s exceeding 30 m. We p esen wo scena ios wi h high and low
maximum e o s in Fig. 11(a) and Fig. 11(b), espec i ely. Despi e
ha ing e y simila a el dis ances and numbe s o u ns, and sim-
ila median e o (Fig. 11(c)), Fig. 11(a) exhibi s a maximum e o
o 35 m, whe eas Fig. 11(b) exhibi s a maximum e o o only 10 m.
We conjec u e ha high-e o scena ios a e associa ed wi h spa se
gNB deploymen s, quan i ied by he a io o o al a el dis ance o
he o al numbe o unique gNBs obse ed.
6 Rela ed Wo k
Localiza ion in 4G and ea lie gene a ions. Ea ly GSM-based
posi ioning sys ems, e.g., [13, 14, 22], use signal s eng h measu e-
men s o Cell-ID sequences o localiza ion, e ec i ely handling
noisy en i onmen s by cap u ing empo al and spa ial dependen-
cies, wi hou speci ically add essing NLOS en i onmen s. Mo e
ecen wo ks a ge localiza ion in LTE ne wo ks and employ a
combina ion o echniques o deal wi h NLOS en i onmen s. How-
e e , hese sys ems s ill su e om poo localiza ion accu acy, e.g.,
150-220 m in [9]. [6] p oposes a single base s a ion localiza ion
sys em using ToA and AoA and demons a es accu acy o a ew m
in an LTE es bed, bu i pe o ms he e alua ion in indoo en i-
onmen s, which a e e y di e en om ou doo ones (e.g., sho e
UE-BS dis ances, absence o ansien blockage). None o hese
wo ks le e ages beam o ming o localiza ion.
Localiza ion in 5G and beyond. A ple ho a o ecen wo ks a -
ge localiza ion o 5G and beyond ne wo ks, including mmWa e
ne wo ks wi h beam o ming capabili ies – see [21], [28], [16] o a
comp ehensi e co e age o he opic. [15] p esen s ISLA, a sys em
enabling IoT de ices o localize using ambien 5G signals wi hou
coo dina ion wi h base s a ions. A numbe o schemes combine ML
echniques and/o comp essi e sensing wi h mmWa e signals o
enhance
accu acy,
especially
in
NLOS
en i onmen s
[7,
17–19,
27].
[29] discusses he p ac ical implemen a ion o hese echniques,
demons a ing he e ec i eness o using exis ing 5G in as uc u e
o cos -e ec i e localiza ion solu ions. These app oaches equi e
ad anced signal p ocessing and ha dwa e modi ica ions, and he
e alua ions o mos o hem ely on simula ions. In con as , we
p opose a p ac ical sys em o 5G mmWa e ne wo ks, which only
equi es access o SSB indices – a piece o in o ma ion eadily
a ailable a he lowe laye s o he p o ocol s ack in comme cial
o - he-shel sma phones, and demons a e i s easibili y and p ac-
icali y ia an ex ensi e measu emen campaign.
Indoo mmWa e-based localiza ion. A numbe o wo ks ha e
speci ically ocused on indoo mmWa e localiza ion le e aging
he beam o ming capabili ies o 802.11ad WiFi APs in he 60 GHz
band [5, 11, 23–26]. Mos o hese wo ks [11, 24–26] e alua e he
p oposed sys ems ia simula ions o ocus on heo e ical pe o -
mance me ics. [5] explo es he p ac ical implemen a ion o a lo-
caliza ion sys em using comme cial o - he-shel (COTS) 60 GHz
APs, add essing he limi a ions and pe o mance o consume -
g ade ha dwa e in indoo scena ios. [23] discusses a comp ehensi e
mmWa e localiza ion sys em designed o simul aneous commu-
nica ion and localiza ion on consume -g ade de ices, ocusing on
indoo en i onmen s and con olled es bed se ups. Simila o mos
wo ks p oposed o 5G and beyond mmWa e ne wo ks, hese ap-
p oaches equi e ad anced signal p ocessing echniques o ob ain
accu a e AoA es ima es. In con as , mm-NOLOC achie es high
Cha lo e
Bos on
e (1-10)
e (11-20)
e (21-30)
Bos on (1-10)
(11-20)
(21-30)
on
on
Bos
Bos
lo
lo
Cha
Cha
lo
Cha
CDF
CDF