Dual-Link Da a Resilien Edge- o-cloud Communica ion F amewo k
o Ag icul u al Robo s
Iman Es andiya
Łukasiewicz Resea ch Ne wo k
Poznan Ins i u e o Technology
Poznan, Poland
e-mail:[email p o ec ed]wicz.go .pl
Kamil Młodzikowski
Łukasiewicz Resea ch Ne wo k
Poznan Ins i u e o Technology
Poznan, Poland
e-mail:kamil.mlodziko[email p o ec ed]wicz.go .pl
Abs ac —Reliable and high- h oughpu communica ion be-
ween ield obo s and cloud se ices emain a key challenge
in p ecision ag icul u e, whe e emo e u al a eas o en lack
consis en high-bandwid h connec i i y. In his wo k, we in o-
duce a new dual-link edge- o-cloud da a ans e amewo k
ha combines long- ange Low-Powe Wide A ea Ne wo king
(LPWAN) o essen ial con ol and moni o ing wi h IEEE 802.11
Wi-Fi ha ca ies bulk da a o e a Zenoh p o ocol. In addi ion,
a da a ou e dynamically swi ches he obo be ween ’T ans e
Mode’, in which senso s eams and image y da a a e being
o wa ded ia Wi-Fi, and ’S o age Mode’, in which da a a e
locally eco ded in Robo ic Ope a ing Sys em (ROS) 2 bags o
p e en loss when connec i i y deg ades. To p eemp i ely de ec
Wi-Fi link ailu es and issue ou ing ins uc ions o he da a
ou e , an onboa d anomaly de ec ion node moni o s hea bea
iming using a machine lea ning-based algo i hm, namely he
XGBoos model. Field ials demons a e ha (1) Wi-Fi ans e s
main ain sub-100 ms la ency wi hin 240 m o he ga eway, (2)
Long Rang (LoRa) communica ion pe sis s eliably beyond 350
m wi h ≈0.1 s la ency, (3) he ou e achie es an a e age o
0.8 s o e lap when en e ing S o age Mode, and (4) he anomaly
de ec o success ully lags link deg ada ion ahead o an ou age.
Ou amewo k scales o mul i- obo deploymen s ia ROS 2
namespaces and Zenoh mul icas , laying he g oundwo k o
esilien swa m ope a ions in u al en i onmen s.
Keywo ds-Au onomous Ag icul u al Robo ; Anomaly De ec ion;
IoT-cloud con inuum; LoRa; Zenoh.
I. INTRODUCTION
P ecision a ming and au onomous machine y a e wo con-
cep s ha a e becoming inc easingly p e alen in mode n
ag icul u e, o simpli y key aspec s o ag icul u al wo k by
ans e ing physically demanding asks o machines and max-
imizing c op yields, he e o e conse ing esou ces [1]. Tasks
such as weed and pes con ol o p ecise plan e iliza ion
a e among hose pe o med by au onomous machines in
ag icul u e, such as unmanned au onomous obo s. Fo he
de ec ion o en i onmen al and/o soil pa ame e s, he obo s
a e equipped wi h In e ne o Things (IoT) senso s ha can
de ec objec s on si e, gene a ing a subs an ial amoun o da a.
The analysis, u iliza ion and s o age o his da a equi es he
a ailabili y o ex ensi e compu ing esou ces, which can be
p o ided h ough cloud compu ing [2]. Howe e , a challenge
a ises in he ans e o da a om he obo o he cloud,
as an in e ne connec ion in he ields is o en un eliable o
una ailable [1]. Gi en he expansi e and spa sely popula ed
a eas ypically u ilized o ag icul u al pu poses, he e is
a clea necessi y o a communica ion solu ion capable o
ope a ing o e conside able dis ances while simul aneously
ansmi ing subs an ial quan i ies o da a. The sole use o
LPWAN echnologies a e no a iable op ion due o he high
da a olumes in ol ed. While LPWAN enables da a o be send
o e he necessa y dis ances, hei da a a es and payload sizes
a e inadequa e o ansmi ing mo e han a ew kiloby es pe
day [3]. E en in he licensed domain o LPWAN solu ions, he
h oughpu would be insu icien . Con e sely, applica ion laye
p o ocols ope a ing o e Wi-Fi do no achie e he equi ed
dis ance, ye can accommoda e he necessa y da a olume [4].
To deli e he necessa y da a while main aining a cons an
connec ion o he edge/cloud, we p opose he in eg a ion o
bo h solu ions in an ag icul u al use case ha inco po a es an
au onomous obo in o he edge/cloud con inuum.
In his pape , we p opose a no el da a ans e me hod ha
employs unlicensed spec um physical laye LoRa o ansmi
con ol messages as well as minimal i al messages o an
ag icul u al obo , he eby p o iding in o ma ion ega ding he
obo s a us and i s loca ion a a sel -p o ided ga eway and
enabling an eme gency shu down o he sys em i necessa y.
Addi ionally, he ecen ly, om he eclipse ounda ion and
Ze ascale de eloped Zenoh p o ocol is u ilized o da a ex-
change be ween he h ee pa icipan s o he da a exchange,
namely he obo , he ga eway and he cloud.
Zenoh is a publishe /subsc ibe /que y p o ocol designed o
ope a e in he mic ocon olle o cloud con inuum, suppo ing
pee - o-pee , ou ed and b oke ed communica ion ia WiFi [5].
To de ec packe loss du ing da a ansmission in an ag i-
cul u al se ing, whe e he dis ance be ween he obo and he
ga eway is a he high, o minimize he dis ance a eled by
he obo , we employ an anomaly de ec ion mechanism in he
obo o assess whe he da a ansmission wo ks p ope ly o
i i is be e o s a eco ding backup da a. The pe o mance
o he p oposed sys em is e alua ed h ough ield expe imen s,
which demons a es he e icacy o he da a exchange be ween
he obo and he ga eway.
The es o he pape is o ganized as ollows. Sec ion II
p esen s he ela ed wo k. Sec ion ?? shows he sys em a -
chi ec u e. Sec ion VI desc ibes he expe imen me hodology.
Sec ion VII discusses he expe imen esul s. Finally, Sec ion
VIII concludes he pape .
27Copy igh (c) IARIA, 2025. ISBN: 978-1-68558-286-9
Cou esy o IARIA Boa d and IARIA P ess. O iginal sou ce: ThinkMind Digi al Lib a y h ps://www. hinkmind.o g
IoTAI 2025 : The Second In e na ional Con e ence on IoT-AI
II. RELATED WORK
The au ho s o [6] used open-sou ce so wa e o build a
LoRa ne wo k connec ing se e al senso nodes o a ga eway
node o be used in an ag icul u al scena io. Communica ion
om he ga eway o he se e is done using Message Queu-
ing Teleme y T anspo (MQTT) o e Long Te m E olu ion
(LTE). Ou sys em di e en ia es om hei wo k by u ilising
Zenoh o e Wi-Fi o communica ion be ween he senso node,
ga eway and se e , while ha ing he senso node connec ed
o an au onomously mo ing obo . In [7], a salable hyb id ne -
wo k o moni o ing an ag icul u al en i onmen is p oposed.
This wo k elies on LoRa o ansmi all he ga he ed senso
da a and aims o co e a size o land ha makes i necessa y
o include LoRa elay nodes o each he ga eway om whe e
i uploads he da a o he cloud using Wi-Fi. In con as
o his wo k, ou p oposed sys em elies on Wi-Fi o da a
ansmission, u ilizing LoRa only o minimal communica ion
o he obo o de ec i s posi ion and o send eme gency
commands. Using LoRa as a con ol link has been done by
he au ho s o [8] as well. In hei case, he con ol link is
es ablished o an Unmanned Ae ial Vehicle (UAV) o inc ease
i s ope a ional ange. Expe imen al esul s we e ob ained om
simula ions only. In compa ison o his wo k, he use o LoRa
is limi ed o he ansmission o minimal con ol messages,
a he han he encapsula ion o o he p o ocol messages wi hin
he LoRa payload.
In hei s udy, he au ho s o [9] examine he po en ial
o Zenoh in he e ogeneous ne wo ks. They demons a e ha
Zenoh can ac as a middlewa e o pee s in di e en ne -
wo ks, enabling communica ion using a pub/sub app oach in
eal- ime. We u ilize Zenoh o in e communica ion be ween
de ices ope a ing on dispa a e sys ems, including ROS and
Linux. Zenoh has been used as he backbone o a cloud-
o-edge communica ion F amewo k, in oduced in [10]. The
p oposed amewo k aims o c ea e a domain o dis ibu ed
compu ing o IoT scena ios, le e aging decen alized pub/sub
communica ion using Zenoh, ligh weigh i ualiza ion and
o ches a ion o he sys em and i s componen s. I is ou
objec i e o le e age he capabili ies o Zenoh o ex end o he
IoT nodes. Ou in en ion is no o limi ou scope o he com-
munica ion be ween he edge and cloud compu ing sys ems.
The au ho s in [11] compa ed he pe o mance o h ee Wi-Fi
s anda ds, IEEE 802.11ax, 802.11ac and 802.11, in ou doo
IoT scena ios. T ansmission h oughpu was e alua ed in he
ange o 2 o 125 m. Ou app oach is simila , bu we u ilized
Zenoh o e WiFi and analysed packe loss and delay while
inc easing and dec easing ansmission dis ances. In [12] he
au ho s ha e p oposed an algo i hm ha p edic s he quali y
o WiFi and Blue oo h Low Ene gy (BLE) communica ion
wi h accu acies o 94 % and 92 %. They a e using Recei ed
Signal S eng h (RSS) as he assessmen me ic o he quali y
o he connec ion. The basis o hei p edic ion is a suppo
ec o eg ession model using a adical base unc ion. Ou
sys em di e s om his by using a linea eg ession model in
o de o ind ou lie ansmission beha iou o ind an ideal
spo o s a ing/s opping he ansmission o packe s. The
anomaly de ec ion in WiFi signals is being done by [13] as
well. Thei esea ch ocuses on he de elopmen o a Radio
F equency (RF) inge p in ing sys em o de ices used in a
WiFi da ase , in o de no only o de ec abno mal ansmi e
bu also o lea n om hei beha iou and ejec hem in he
u u e. Ou app oach u ilizes he de ec ion o anomalies in he
WiFi connec ion o iden i y any issues wi h he ansmission
o WiFi signals o a ga eway. This is necessa y since WiFi can
be dis u bed a any ime be o e he obo c osses he dis ance
h eshold. By de ec ing anomalies in da a ans e we ensu e
as li le da a is los as possible.
III. AGRICULTURAL ROBOT
The ag icul u al obo ic pla o m, Ag oRob, is designed
o au onomously na iga e ields o p ecision a ming asks
such as c op sp aying and weeding. Equipped wi h ad anced
senso s, communica ion modules, and a modula so wa e
a chi ec u e, i ensu es eliable localiza ion, e icien ope a ion,
and seamless da a exchange wi h he cloud-based sys ems.
This sec ion de ails he au onomous unc ionali ies, ha dwa e
se up, so wa e a chi ec u e, and communica ion p o ocols
employed by he obo .
1) Robo Au onomous Func ionali y: The ag icul u al pla -
o m (Ag oRob) au onomously na iga es ields o p ecise
c op sp aying and weeding. I achie es accu a e localiza ion
by using da a om dual Global Na iga ion Sa elli e Sys ems
(GNSS) modules, an Ine ial Measu emen Uni (IMU), and
wheel odome y. This enables i o ollow c op lines and sha e
i s posi ion wi h a cloud-based sys em. A Deep Neu al Ne -
wo k (DNN) model p ocesses came a images o de ec c ops
and weeds, o p ecise sp aying. Compu e ision minimizes
chemical use, educing e ilize and he bicide consump ion
while imp o ing e iciency and sus ainabili y.
2) Robo Ha dwa e: The obo ea u es an onboa d com-
pu e managing con ol and communica ion. I includes lo-
caliza ion senso s, came as, a Wi-Fi ou e , and a LoRa
ansmi e o cloud da a ans e . Communica ion occu s ia
a Con olle A ea Ne wo k (CAN) o USB adap e , handling
s a us upda es and con ol commands. Figu e 1 illus a es he
ha dwa e se up.
3) Robo So wa e: The obo ’s so wa e is de eloped us-
ing ROS 2 [14], u ilises pee - o-pee communica ion ia he
Da a Dis ibu ion Se ice (DDS). Modules, such as mission
handling, localiza ion, and na iga ion, communica e ia UDP
o TCP, based on Quali y o Se ice (QoS) se ings. Da a is
ansmi ed h ough a publish-subsc ibe model o b oadcas
communica ion o se ices o di ec in e ac ions. Ce ain
opics enable cloud communica ion o con ol and moni o ing,
as shown in Figu e 2.
4) Messages: LoRa communica ion in ol es sending s ing
da a. Messages o he ga eway con ain h ee comma-sepa a ed
alues, o alling up o 29by es:
•ID: Unique packe iden i ie .
•Coo dina es: La i ude and longi ude.
•Time: UTC imes amp.
28Copy igh (c) IARIA, 2025. ISBN: 978-1-68558-286-9
Cou esy o IARIA Boa d and IARIA P ess. O iginal sou ce: ThinkMind Digi al Lib a y h ps://www. hinkmind.o g
IoTAI 2025 : The Second In e na ional Con e ence on IoT-AI
Figu e 1. Ha dwa e con igu a ion o he ag icul u al obo ic pla o m.
The ga eway sends wo boolean con ol commands unde
0.5Hz:
•Da a Con ol Command: Chooses Wi-Fi ansmission
o local logging.
•Eme gency S op Command: Immedia ely hal s he
obo .
Wi-Fi communica ion ansmi s ope a ional da a, including
senso eadings, mission s a us, and ield analysis, essen ial
o obo pe o mance and ag icul u al insigh s.
GNSS IMU wheels
Localiza ion
Na iga ion
Plan
de ec ion
Cams Mission handle
Sp ay ool
LoRa clien
Zenoh b idge
ROS bag
ROS Topics
C i ical da a
Full da a
LoRa
WiFi
Figu e 2. Da a low a chi ec u e o he Ag oRob pla o m.
IV. DATA TRANSFER AND STORAGE
As shown in Figu e 3, ou p oposed da a ans e me hod
assumes ha bo h he obo and he ga eway can communica e
using LoRa and Wi-Fi, wi h bo h sys ems ha ing independen
GNSS localiza ion a ailable ia a u-blox ZED-F9P GNSS
module onboa d. The obo is equipped wi h an indus ial
Wi-Fi ou e , he NR600 om Na iga eWo x, along wi h a
Hel ec WiFi LoRa 32 V3 module, while he ga eway ea u es a
Nigh hawk® AXE3000 Wi-Fi USB adap e and a Hel ec WiFi
LoRa 32 V3 module. In addi ion, he ga eway is also equipped
wi h in e ne connec i i y h ough a 5G/LTE modem. Bo h
he obo and he ga eway a e pe o ming localiza ion using
GNSS. Robo geoloca ion is sen h ough LoRa o he ga eway.
Based on his in o ma ion and i s own posi ion, he ga eway
WIFI access poin
GPS an enna
LoRa ansmi e
Ag oRob Communica ion Ha dwa e Ga eway Communica ion ha dwa e
GPS an enna
WIFI adap o
5G/LTE
LoRa ansmi e
Figu e 3. The communica ion ha dwa e con igu a ion o he Ag oRob LoRa-
based ne wo k.
calcula es he dis ance o he obo . This da a oge he wi h
hea bea delay calcula ed based on UTM ime is hen used
by Da a Rou e so wa e o swi ch be ween wo s a es:
•T ans e Mode: While he obo is in an e icien ange
and s eng h o Wi-Fi connec ion wi h he ga eway, i
ope a es in T ans e Mode. Selec ed da a ( ep esen ed by
ROS2 opics) is being sen o e Wi-Fi o he ga eway
using a Zenoh b idge.
•S o age Mode: Con e sely, when he e is a isk o losing
he Wi-Fi connec ion be ween he obo and he ga eway,
he obo is swi ched o S o age Mode. In his mode,
he Zenoh b idge is u ned o o minimise he isk o
da a in e cep ion. The da a ha no mally in T ans e
Mode would be sen o he ga eway is ins ead being
eco ded using ROS2 bags and s o ed locally o u u e
synch oniza ion. The p oposed communica ion low is
p esen ed in Figu e 4.
Ag icul u al obo ( a edge)
Da a
ou e
Ga eway
Geoloca ion
Con ol command
Robo
Geoloca ion
Cloud
Compu ing
Robo S a e
Robo minimal
S a e and command
Robo S a e
/command
Clien Clien
Rou e Local s o age
Ga eway
Figu e 4. The communica ion a chi ec u e o he Ag oRob sys em.
A. Da a ou e
E e y ime he ga eway ecei es a geoloca ion o he obo
(la i ude ϕand longi ude λ) h ough LoRa, i calcula es he
dis ance be ween i sel and he obo using he ollowing
equa ions:
c=1
qcos2(ϕ) + (1 − )2·sin2(ϕ)
(1)
s= (1 − )2·c(2)
29Copy igh (c) IARIA, 2025. ISBN: 978-1-68558-286-9
Cou esy o IARIA Boa d and IARIA P ess. O iginal sou ce: ThinkMind Digi al Lib a y h ps://www. hinkmind.o g
IoTAI 2025 : The Second In e na ional Con e ence on IoT-AI
x= (R·c+h)·cos(ϕ)·cos(λ)
y= (R·c+h)·cos(ϕ)·sin(λ)
z= (R·s+h)·sin(ϕ)
Gi en wo poin s: ga eway = (xG, yG, zG)and obo =
(xR, yR, zR), he Euclidean dis ance dGR is:
dGR =p(xG−xR)2+ (yG−yR)2+ (zG−zR)2(3)
Whe e:
R= 6356752.3142 (Ea h adius)
= 1/298.257223563 (Ea h la ening ac o )
ϕis he la i ude
λis he longi ude
The calcula ed dis ance, oge he wi h he alue o delay
calcula ed based on he imes amp o messages ecei ed om
he ga eway is con inuously ed o he anomaly de ec ion so -
wa e. The so wa e analyzes his da a in eal- ime o de ec any
abno mali ies in communica ion beha iou . Upon iden i ying
an anomaly, i issues commands o he da a ou e on he
obo . The da a ou e hen swi ches be ween S o age Mode
and T ans e Mode as needed, ensu ing an o e lap be ween
da a eco ding and ansmission o p e en any po en ial da a
loss.
In cases o S o age Mode da a loss is mi iga ed h ough local
s o age on he edge de ice. Howe e , since he ope a ional
da a o he ag icul u al obo p ima ily consis s o nume ical
alues and image da a ha a e p ocessed locally, he olume
o s o ed in o ma ion emains ela i ely low and does no
necessi a e la ge-scale s o age solu ions. I should be no ed,
howe e , ha he local s o age o da a is inhe en ly cons ained
by he physical s o age capaci y o he edge de ice. Despi e
his limi a ion, e aining ope a ional da a is essen ial o he
obo ’s con inued unc ionali y— o example, o main ain a
eco d o he loca ion and s a us o indi idual c op ins ances,
which is c i ical o planning and execu ing u u e ope a ions
on he same ield plo s.
V. ANOMALY DETECTION
Ou anomaly de ec ion sys em employs a machine lea n-
ing app oach o iden i y abno mal beha iou in Wi-Fi da a
ansmission. The sys em moni o s obo ’s hea bea iming
da a and he dis ance be ween he obo and he ga eway, o
de ec po en ial ailu es o mal unc ions. I implemen s a wo-
s age p ocess: i s , a model aining phase using XGBModel
wi h KMeansSco e o lea n no mal ope a ional pa e ns om
his o ical da a; second, a de ec ion phase, whe e he Anoma-
lyDe ec ionNode con inuously analyzes incoming da a agains
hese lea ned pa e ns in eal ime.
The de ec ion mechanism combines IQRDe ec o and
Th esholdDe ec o me hodologies o iden i y s a is ical ou -
lie s, publishing ale s when anomalies exceed a con igu able
pe cen age h eshold. Ope a ing independen ly on ROS2, he
sys em samples da a a egula in e als (con igu able, se o 2
seconds in he cu en implemen a ion) and main ains a sliding
window o obse a ions o balance de ec ion sensi i i y wi h
compu a ional e iciency.
Fo ep oducibili y, he XGBModel was ained on 3285
hea bea in e als. The model uses lags=64. Anomaly sco es
a e p oduced by a KMeansSco e wi h k= 20 clus e s and
a 32-sample window (componen _wise=False).
By con inuously analyzing hea bea iming and lagging
ou lie s in eal- ime, he anomaly de ec o enables he obo
o swi ch p eemp i ely be ween T ans e and S o age modes,
beginning local da a logging be o e Wi-Fi b eaks down and
eenabling Zenoh he ins an link quali y eco e s, hus elim-
ina ing da a gaps and nega i e o e laps. Mo eo e , when
sus ained anomalies indica e wo sening channel condi ions,
he sys em can dynamically h o le non-c i ical s eams (e.g.,
educe image esolu ion) o p ese e essen ial eleme y, while
simul aneously elaying “link deg ading” ale s back o he
ope a o o e LoRa. I anomaly a es c oss a c i ical h eshold,
he de ec o can e en igge an immedia e eme gency-s op
command, ensu ing bo h da a in eg i y and ope a ional sa e y
wi hou human in e en ion.
Ou implemen a ion inco po a es adap i e sensi i i y ad-
jus men s based on en i onmen al condi ions and ope a ional
con ex . Du ing pe iods o known ne wo k conges ion o when
he obo a e ses a eas wi h documen ed Wi-Fi in e e ence,
he sys em au oma ically adjus s de ec ion h esholds o e-
duce alse posi i es while main aining igilance o genuine
anomalies.
This app oach allows o ea ly wa ning o de eloping issues
be o e hey cause c i ical ailu es, making i possible o ac
acco dingly o p e en da a loss as much as possible. The
modula design o he sys em also enables easy in eg a ion
o addi ional de ec ion algo i hms as hey become a ailable,
ensu ing u u e ex ensibili y.
VI. EXPERIMENTS METHODOLOGY
This sec ion desc ibes he me hodology used o e alua e ou
app oach’s pe o mance and accu acy. The expe imen s es
he hypo hesis unde a ious condi ions o ensu e comp ehen-
si e and eal-wo ld- ep esen a i e esul s.
A. Expe imen Se up
In ou expe imen s, we mimic eal-wo ld applica ions. The
ga eway is s a iona y while he obo mo es owa d and away
om i as shown in Figu e 5.
To ensu e accu a e ime synch oniza ion o one-way com-
munica ion measu emen s, messages a e imes amped using
GNSS-based UTC ime [15], as bo h he obo and he ga eway
a e equipped wi h GNSS ecei e s. The ga eway also unc ions
as an Real-Time Kinema ic (RTK) base s a ion, p o iding
localiza ion co ec ions o he obo o imp o ed accu acy.
GPS signals se e as a common ime e e ence, enabling
imes amp compa isons o calcula e one-way communica ion
delays, especially when swi ching be ween LoRa and Wi-Fi.
30Copy igh (c) IARIA, 2025. ISBN: 978-1-68558-286-9
Cou esy o IARIA Boa d and IARIA P ess. O iginal sou ce: ThinkMind Digi al Lib a y h ps://www. hinkmind.o g
IoTAI 2025 : The Second In e na ional Con e ence on IoT-AI
100 m
Ga eway
Robo Pa h
Figu e 5. A sa elli e iew map illus a ing he obo ’s pa h du ing he es
expe imen . The ga eway compu e was posi ioned a a ixed loca ion, while
he obo began i s mo emen nea he ga eway, a eled away, and e en ually
e u ned o i s ini ial posi ion. The ed line on he map ep esen s he pa h
ollowed by he obo .
This me hod a oids complexi ies in ound- ip measu emen s,
which can obscu e pa h delays in asymme ic ne wo ks.
Two se ies o connec i i y and da a ans e expe imen s
we e conduc ed:
1. The obo mo es away om he ga eway while ansmi -
ing da a ia bo h LoRa and Wi-Fi (using he Zenoh b idge).
As he dis ance inc eases, Wi-Fi e en ually goes ou o ange.
Du ing his p ocess, da a is logged, including imes amps o
messages gene a ed by he obo and ecei ed a he ga eway.
This in o ma ion is used o analyze ans e cha ac e is ics and
o gene a e aining da a o he anomaly de ec ion model.
2. The same p ocedu e is epea ed wi h he anomaly de ec-
ion and da a ou ing sys em enabled; he esul is illus a ed
in Figu e 8.
B. Expe imen Me ics
Key pe o mance me ics include:
•Communica ion La ency:
Wi-Fi: The ime aken o da a ans e o e Wi-Fi wi hin
ange. I is compu ed as:
∆ = cu − s amp (4)
τ= UT CG
− UT CR+ ∆ (5)
Whe e:
cu is he cu en ROS2 ime,
s amp is he UTC message imes amp,
UT CGis UTC ime on he ga eway,
UT CRis UTC ime on he obo ,
τis he delay.
LoRa: Measu ed simila ly, wi h UTC ime added o LoRa
messages.
•Packe Loss:
LoRa: Reliabili y o posi ion da a sen om he obo .
Packe loss is calcula ed by acking message ID gaps.
•Ne wo k Co e age:
Wi-Fi: Maximum eliable connec ion dis ance.
LoRa: Maximum dis ance o eliable command ecep-
ion.
Range-based Swi ching: E ec i eness o ansi ioning
be ween Wi-Fi and LoRa.
•Da a O e lap:
As desc ibed in Sec ion IV-A, swi ching be ween T ans e
Mode and S o age Mode mus ensu e da a o e lap. The
a ge o e lap is 1s, hough ac o s like Zenoh b idge
s and-up ime may in luence i .
Zenoh o Bag: Time be ween he i s message s o ed in
osbag2 on he obo and he las ecei ed a he ga eway.
A posi i e alue indica es o e lap, while a nega i e alue
means da a loss.
Bag o Zenoh: Time be ween he i s message ecei ed
a he ga eway and he i s s o ed in osbag2. A posi i e
alue means o e lap; a nega i e alue indica es loss.
These e alua ion me ics ensu e a balanced assessmen o
he me hod’s pe o mance. Each con igu a ion unde wen mul-
iple uns o ensu e consis ency and accoun o a iance. The
inal esul s a e epo ed as a e ages wi h s anda d de ia ions,
whe e applicable.
VII. EXPERIMENTS RESULTS
A se ies o ield es expe imen s ha e aken place in ol ing
he obo mo ing away om he ga eway while measu ing he
de ined me ics o Wi-Fi and LoRa a he ga eway.
A. WiFi and LoRa delay
Figu e 6 illus a es he delay expe ienced by bo h Zenoh
and LoRa communica ion o e ime and dis ance om he
ga eway. In his expe imen , he dis ance be ween he obo and
he ga eway is g adually inc eased. As he dis ance be ween
he wo de ices inc eases, he a e age delay o he Zenoh
messages also inc eases un il app oxima ely 240 me e s, a
which poin connec i i y o he ga eway is los . The dis ance
is hen ex ended o 350 me e s, which has no impac on he
delay o he LoRa messages.
In o de o egula e he ans e o da a ia Wi-Fi and
o acili a e local logging, a h eshold o 50 me e s was
implemen ed o Wi-Fi ansmissions in he cou se o he
ollowing expe imen s. The esul s o his can be obse ed
in Figu e 7a. In his expe imen , he dis ance be ween he
obo and he ga eway ini ially inc eases and subsequen ly
dec eases. Upon eaching he h eshold o 50 me e s, he Wi-Fi
ansmissions a e e mina ed, while he LoRa con ol messages
con inue o be exchanged. Figu e 7b illus a es he numbe o
Wi-Fi packe s ecei ed and LoRa packe s los . The impac o
he h eshold can be obse ed he e, as Wi-Fi packe s a e only
ansmi ed when he dis ance is less han 50 m and he delay
he e o e emains below 0.1s. Howe e , a a dis ance o 60 m,
loss o LoRa packe s occu s. The packe loss has no in luence
31Copy igh (c) IARIA, 2025. ISBN: 978-1-68558-286-9
Cou esy o IARIA Boa d and IARIA P ess. O iginal sou ce: ThinkMind Digi al Lib a y h ps://www. hinkmind.o g
IoTAI 2025 : The Second In e na ional Con e ence on IoT-AI
Figu e 6. WiFi and LoRa packe delays as he obo mo es away om he
ga eway. As he obo ecedes om he ga eway, packe ansmission delays
inc ease; he Wi-Fi link ails beyond app oxima ely 300 m, whe eas he LoRa
channel con inues o deli e low-bandwid h da a wi h an almos cons an
la ency.
on he delay o he subsequen LoRa packe s as his alue
luc ua es a ound 0.1 s o LoRa packe s.
In o de o egula e he ans e o da a ia Wi-Fi and
o acili a e local logging, a h eshold o 50 me e s was
implemen ed o Wi-Fi ansmissions in he cou se o he
ollowing expe imen s. The esul s o his can be obse ed
in Figu e 7a. In his expe imen , he dis ance be ween he
obo and he ga eway ini ially inc eases and subsequen ly
dec eases. Upon eaching he h eshold o 50 me e s, he Wi-Fi
ansmissions a e e mina ed, while he LoRa con ol messages
con inue o be exchanged. Figu e 7b illus a es he numbe o
Wi-Fi packe s ecei ed and LoRa packe s los . The impac o
he h eshold can be obse ed he e, as Wi-Fi packe s a e only
ansmi ed when he dis ance is less han 50 m and he delay
he e o e emains below 0.1s. Howe e , a a dis ance o 60 m,
loss o LoRa packe s occu s. The packe loss has no in luence
on he delay o he subsequen LoRa packe s as his alue
luc ua es a ound 0.1 s o LoRa packe s.
Time o e lap be ween da a sen o e Wi-Fi and s o ed in
osbag2 was measu ed in mul iple expe imen s. The esul s
a e p esen ed in Table I. The da a shows ha he a e age ime
o da a o e lap o swi ching om T ans e Mode o S o age
Mode is 0.8seconds. This means ha o an a e age o 0.8
seconds da a is s o ed bo h locally a he obo and sen o e
Wi-Fi (using Zenoh) o he ga eway. The e o e, he p ocess o
s a ing bag eco ding akes an a e age o 0.2seconds (as he
desi ed o e lap was se o 1second).
In he second case, whe e he sys em swi ches om S o age
Mode o T ans e Mode, he a e age o e lap is −1.0667
seconds. The nega i e alue indica es ha he e was a gap
be ween he da a s o ed locally on he obo and he da a sen
using he Zenoh b idge. The esul is illus a ed in Figu e 8.
Such a esul indica es ha a much highe o e lap is needed
when swi ching om S o age Mode o T ans e Mode. The
mos p obable cause o his beha iou is he s and-up ime
o he Zenoh b idge, as he p ocess is s opped each ime he
sys em swi ches o S o age Mode.
(a) WiFi and LoRa packe delays
(b) WiFi Recei ed packe s and LoRa packe los
Figu e 7. (a) WiFi and LoRa packe delays, and (b) WiFi ecei ed packe s
and LoRa packe loss as he obo mo es away om and ge s close o
he ga eway, swi ching be ween T ans e and S o age Modes a a 50-me e
dis ance h eshold.
0 100 200 300 400
Time [s]
0
25
50
75
100
125
150
175
Dis ance [m]
Dis ance
Anomaly de ec ed
Wi-Fi Packe s Recei ed
Figu e 8. Anomaly de ec ion sys em ecognizes issues wi h Wi-Fi da a
ans e , pa icula ly as he dis ance be ween he obo and he ga eway
inc eases and signal quali y begins o deg ade.
TABLE I. AVERAGE DATA OVERLAP TIME (IN SECONDS)FOR TRANSFER
TO STORAGE AND STORAGE TO TRANSFER MODE SWITCHES,WITH
STANDARD DEVIATION AND DIFFERENCE FROM DESIRED 1SECOND.
T ans e →S o age [s] S o age→T ans e [s]
A g. 0.8000 -1.0667
S d. 5.19×10−91.0263
Di . 0.2000 2.0667
B. Anomaly de ec ion model pe o mance
The pe o mance o he anomaly de ec ion model was e al-
ua ed using s anda d classi ica ion me ics: p ecision, ecall,
and F1-sco e. These me ics we e calcula ed by compa ing
32Copy igh (c) IARIA, 2025. ISBN: 978-1-68558-286-9
Cou esy o IARIA Boa d and IARIA P ess. O iginal sou ce: ThinkMind Digi al Lib a y h ps://www. hinkmind.o g
IoTAI 2025 : The Second In e na ional Con e ence on IoT-AI
he p edic ed labels (p ed_labels) agains he g ound u h
labels (g _labels). The calcula ions we e pe o med using
he p ecision_sco e, ecall_sco e, and 1_sco e
unc ions, wi h he ze o_di ision pa ame e se o 0
o handle any po en ial di ision by ze o issues g ace ully.
The esul s o hese e alua ions a e summa ized in Table II,
p o iding a iew o he model’s abili y o co ec ly iden i y
anomalies in Wi-Fi communica ion while minimizing alse
posi i es and alse nega i es.
TABLE II. PERFORMANCE METRICS OF THE ANOMALY DETECTION
MODEL: PRECISION, RECALL,AND F1-SCORE.
P ecision Recall F1-sco e
Sco e 0.951 0.966 0.958
VIII. CONCLUSION AND FUTURE WORK
This s udy p esen s a no el da a ans e me hod ha
in eg a es LoRa communica ion wi h Wi-Fi o enhance he
ope a ional capabili ies o au onomous ag icul u al obo s.
U ilizing LoRa o essen ial con ol messages and minimal
s a us upda es acili a es eliable communica ion in u al a eas,
whe e connec i i y is equen ly limi ed. The esul s indica e
ha employing LoRa o suppo Wi-Fi communica ion can
signi ican ly imp o e he unc ionali y o obo s ope a ing in
emo e egions.
In his wo k, we demons a ed he iabili y o ou ame-
wo k using a single obo ic pla o m while inhe en ly e ain-
ing he capabili y o suppo mul iple obo s and ins ances.
Ou a chi ec u e le e ages he Robo Ope a ing Sys em’s
namespace and opic emapping ea u es, allowing each obo
o publish and subsc ibe o uniquely p e ixed opics (e.g.,
/ obo _<ID>/cmd_ el), he eby isola ing and managing
concu en Wi-Fi message s eams. A single ins ance o he
Zenoh b idge a he ga eway is su icien o inges and p ocess
hese pa allel communica ions. Once ecei ed, messages a e
a chi ed in he cloud along wi h hei o igina ing edge-de ice
iden i ie s. Con e sely, command messages can be a ge ed o
indi idual obo s by publishing o he app op ia e namespaced
opic.
Mo eo e , ou amewo k accommoda es LoRa communi-
ca ions: mul icas downlink enables he simul aneous deli e y
o iden ical packe s o a g oup o obo s ia a single ga eway
module. In p inciple, one ga eway can o ches a e he bidi ec-
ional da a low o an en i e obo ic swa m, seamlessly linking
edge de ices wi h he cloud.
While he amewo k al eady suppo s mul i- obo and
mul i-communica ion mechanisms, empi ical alida ion wi hin
a ue swa m se ing emains o be conduc ed. Fu u e wo k
will ocus on: (1) deploying and s ess- es ing he sys em
wi h a he e ogeneous lee o obo s; (2) e alua ing ne wo k
pe o mance and la ency in high-densi y LoRa mul icas sce-
na ios; (3) ex ending he cloud-s o age schema o inco po a e
ad anced me ada a and secu e access con ols; and (4) op i-
mizing he sys em’s beha iou du ing communica ion mode
swi ching, speci ically add essing he s and-up ime o he
Zenoh b idge. In he cu en implemen a ion, he Zenoh b idge
p ocess is e mina ed when swi ching om T ans e Mode
o S o age Mode and es a ed when swi ching back. This
eini ializa ion in oduces addi ional la ency due o he Zenoh
b idge’s s and-up ime, a ec ing he con inui y o da a ans e .
To mi iga e his, we p opose de eloping an imp o ed swi ch-
ing mechanism ha a oids e-execu ing he Zenoh p ocess.
This wo k con ibu es o he de elopmen o use-case sce-
na ios o he alida ion o he IoT Cloud Ope a ing Sys em
(ICOS), a me a ope a ing sys em unde de elopmen wi hin
he Eu opean Union’s Ho izon p og am. A no able limi a ion
o he p oposed amewo k is ha , when da a is s o ed locally
in he absence o Wi-Fi, he logged da a mus be ans e ed o
he cloud manually. I is an icipa ed ha ICOS will ul ima ely
manage he da a ansmission unc ionali ies associa ed wi h
his use case, encompassing da a ans e and s o age be ween
edge de ices and he cloud.
ACKNOWLEDGMENT
This p ojec has ecei ed unding om he Eu opean
Union’s HORIZON esea ch and inno a ion p og am unde
g an ag eemen No 101070177.
REFERENCES
[1] O. Kopishynska, Y. U kin, O. Galych, M. Ma enych, and I.
Sliusa , “Main aspec s o he c ea ion o managing in o ma ion
sys em a he implemen a ion o p ecision a ming,” in 2020
IEEE 11 h In e na ional Con e ence on Dependable Sys ems,
Se ices and Technologies (DESSERT), 2020, pp. 404–410.
[2] S. S. Baghel e al., “Ai, io and cloud compu ing based sma
ag icul u e,” in 2022 5 h In e na ional Con e ence on Con em-
po a y Compu ing and In o ma ics (IC3I), 2022, pp. 1658–
1661.
[3] M. Iqbal, A. Y. M. Abdullah, and F. Shabnam, “An applica ion
based compa a i e s udy o lpwan echnologies o io en i-
onmen ,” in 2020 IEEE Region 10 Symposium (TENSYMP),
2020, pp. 1857–1860.
[4] B. Jackson, A. S. A. Nisha, S. Va alakshmi, N. Da win,
and M. Va un, “Enhancing wi i ange and h oughpu wi h
beam s ee ing an ennas,” in 2023 7 h In e na ional Con e ence
on Elec onics, Communica ion and Ae ospace Technology
(ICECA), 2023, pp. 1749–1754.
[5] A. Co sa o e al., “Zenoh: Uni ying communica ion, s o age
and compu a ion om he cloud o he mic ocon olle ,” in
2023 26 h Eu omic o Con e ence on Digi al Sys em Design
(DSD), 2023, pp. 422–428.
[6] X. Wang e al., “Spa c-lo a: A scalable, powe -e icien , a -
o dable, eliable, and cloud se ice-enabled lo a ne wo king
sys em o ag icul u e applica ions,” in 2024 IEEE In e na-
ional Symposium on Dynamic Spec um Access Ne wo ks
(DySPAN), 2024, pp. 151–156.
[7] S. Ding e al., “A scalable hyb id ne wo k o ag icul u e
en i onmen moni o ing,” in 2021 IEEE 9 h In e na ional
Con e ence on In o ma ion, Communica ion and Ne wo ks
(ICICN), 2021, pp. 566–570.
[8] Z. B. A slanbenze , B. Bay am, T. T. Sa ı, and G. Seçin i,
“U ilizing lo a o con ol link in so wa e-de ined ae ial
ne wo ks: Analysis and implemen a ion,” in 2023 IEEE 28 h
In e na ional Wo kshop on Compu e Aided Modeling and
Design o Communica ion Links and Ne wo ks (CAMAD),
2023, pp. 164–169.
33Copy igh (c) IARIA, 2025. ISBN: 978-1-68558-286-9
Cou esy o IARIA Boa d and IARIA P ess. O iginal sou ce: ThinkMind Digi al Lib a y h ps://www. hinkmind.o g
IoTAI 2025 : The Second In e na ional Con e ence on IoT-AI
[9] C.-S. Shih, H.-J. Lin, Y. Yuan, Y.-H. Kuo, and W.-Y. Liang,
“Scalable and bounded- ime decisions on edge de ice ne wo k
using eclipse zenoh,” in 2022 IEEE 28 h In e na ional Con-
e ence on Embedded and Real-Time Compu ing Sys ems and
Applica ions (RTCSA), 2022, pp. 170–179.
[10] J. J. L. Escoba , F. Gil-Cas iñei a, and R. P. D. Redondo, “De-
cen alized se e less io da a low a chi ec u e o he cloud-
o-edge con inuum,” in 2023 26 h Con e ence on Inno a ion in
Clouds, In e ne and Ne wo ks and Wo kshops (ICIN), 2023,
pp. 42–49.
[11] I. ElKassabi and A. Abd abou, “An expe imen al compa a i e
pe o mance s udy o di e en wi i s anda ds o sma ci ies
ou doo en i onmen s,” in 2022 IEEE 13 h Annual Ubiqui ous
Compu ing, Elec onics & Mobile Communica ion Con e ence
(UEMCON), 2022, pp. 0450–0455.
[12] B. Sudha san, J. G. B eslin, and M. I. Ali, “Adap i e s a egy
o imp o e he quali y o communica ion o io edge de ices,”
in 2020 IEEE 6 h Wo ld Fo um on In e ne o Things (WF-
IoT), 2020, pp. 1–6.
[13] T. Zhao, S. Sa ka , Y. Tian, and D. Cab ic, “Anomaly ans-
mi e ecogni ion and acking,” in 2024 IEEE In e na ional
Symposium on Dynamic Spec um Access Ne wo ks (DySPAN),
2024, pp. 357–364.
[14] S. Macenski, T. Foo e, B. Ge key, C. Lalance e, and W.
Woodall, “Robo ope a ing sys em 2: Design, a chi ec u e, and
uses in he wild,” Science Robo ics, ol. 7, no. 66, eabm6074,
2022.
[15] L. D. Vi o, S. Rapuano, and L. Tomaciello, “One-way delay
measu emen : S a e o he a ,” IEEE T ansac ions on Ins u-
men a ion and Measu emen , ol. 57, no. 12, pp. 2742–2750,
2008.
34Copy igh (c) IARIA, 2025. ISBN: 978-1-68558-286-9
Cou esy o IARIA Boa d and IARIA P ess. O iginal sou ce: ThinkMind Digi al Lib a y h ps://www. hinkmind.o g
IoTAI 2025 : The Second In e na ional Con e ence on IoT-AI
