Glass-Box Au oma ed D i ing: Insigh s and Fu u e T ends
Mau o Bellone1 a, Rai o Sell2 b and Ral -Ma in Soe1 c
1FinEs Cen e o Sma Ci ies, Tallinn Uni e si y o Technology, Es onia
2Depa men o Mechanical and Indus ial Enginee ing, Tallinn Uni e si y o Technology, Es onia
Keywo ds: Glass-Box Models, Au oma ed D i ing, Pe o mance s In e p e abili y T ade-o .
Abs ac : Au oma ed d i ing has ad anced signi ican ly h ough he use o black-box AI models, pa icula ly in pe cep-
ion asks. Howe e , as hese models ha e g own, conce ns o e he loss o explainabili y and in e p e abili y
ha e eme ged, p omp ing a demand o c ea ing ’glass-box’ models. Glass-box models in au oma ed d i ing
aim o design AI sys ems ha a e anspa en , in e p e able, and explainable. While such models a e essen ial
o unde s anding how machines ope a e, achie ing pe ec anspa ency in complex sys ems like au onomous
d i ing may no be en i ely p ac icable no easible. This pape explo es a gumen s on bo h sides, sugges ing
a shi o he ocus owa ds balancing in e p e abili y and pe o mance a he han conside ing hem as con-
lic ing concep s.
1 INTRODUCTION
Fully au oma ed d i ing on oads has been a long-
sough goal, wi h li le signi ican p og ess o many
yea s (S an on and Young, 1998). The ield has only
ecen ly ad anced no ably, d i en by imp o emen s in
ha dwa e compu a ional capabili ies and da a-d i en
models wi h he p omise o end- o-end au oma ed
d i ing in he nea u u e (Yu se e e al., 2020).
Howe e , his p og ess comes a he cos o losing a
clea connec ion o he undamen als o p ocess con-
ol (Omeiza e al., 2021). Au oma ed d i ing on
oads exempli ies he con ol o a complex sys em
whe e da a-d i en models o e ad an ages, as c ea -
ing de ailed analy ical models o e e y componen is
nea ly in easible.
Classical op imal con ol heo y aims o design
con ol p ocesses ha achie e op imal acking o a
desi ed e e ence signal, o e ing elegan analy ical
solu ions o many p oblems. The heo y begins wi h
linea , scala sys ems and ex ends o mul i a ia e,
non-linea p ocesses, signi ican ly inc easing compu-
a ional complexi y and he sys em’s le el o abs ac-
ion, pa icula ly due o he challenges o sol ing
highe -dimensional and non-linea models.
F om his pe spec i e, a model ha analy ically
cap u es eali y a i s undamen al le el would be
he ideal ounda ion o building de e minis ic, e o -
ah ps://o cid.o g/0000-0003-3692-0688
bh ps://o cid.o g/0000-0003-1409-0206
ch ps://o cid.o g/0000-0002-6782-1677
p oo con ol sys ems. By de ini ion, such a sys em
would quali y as a glass-box model, o e ing com-
ple e in e p e abili y and explainabili y. On he o he
hand, many undamen al physical p ocesses emain
poo ly unde s ood, and while he human ins inc is o
seek clea explana ions, echnological ad ancemen
o en elies on app oxima e desc ip ions o manage
complex con ol sys ems in an unce ain eal wo ld.
A deepe unde s anding leads o imp o ed unc ion-
ali y, highligh ing he impo ance o in e p e abil-
i y. While black-box models may enable con ol sys-
ems o unc ion e ec i ely, only ull analy ical in e -
p e abili y can unlock hei ue po en ial.
Following he example o au oma ed d i ing, a
ull analy ical and compu a ional desc ip ion o dy-
namic beha io —such as ae odynamics, i e ic ion,
and engine esponse o d i e commands—is com-
plex and no en i ely p ac ical ye achie able. S ee -
ing obo s capable o d i ing ehicles in s uc u ed
en i onmen s, such as es acks, ha e exis ed since
he 1980s (Weisse e al., 1999). The eal challenge,
howe e , lies in eal-wo ld in e ac ions, whe e model-
ing he unp edic able beha io o o he oad use s be-
comes inc easingly di icul . In such scena ios, use s
expec a obo ic d i e o ac de e minis ically and
make sa e y-c i ical decisions wi hin ac ions o a
second.
Following his line o esea ch, his wo k dis-
cusses he ollowing esea ch ques ions while ocus-
ing on he p oblem o au oma ed d i ing:
RQ1.Wha a e he p ac ical mo i a ions o build-
ing glass-box analy ical models?
880
Bellone, M., Sell, R. and Soe, R.-M.
Glass-box Au oma ed D i ing: Insigh s and Fu u e T ends.
DOI: 10.5220/0013384300003890
Pape published unde CC license (CC BY-NC-ND 4.0)
In P oceedings o he 17 h In e na ional Con e ence on Agen s and A i icial In elligence (ICAART 2025) - Volume 1, pages 880-885
ISBN: 978-989-758-737-5; ISSN: 2184-433X
P oceedings Copy igh ©2025 by SCITEPRESS – Science and Technology Publica ions, Lda.
RQ2.Wha a e he challenges in achie ing ull ex-
plainabili y and in e p e abili y in he ield o
au oma ed d i ing?
RQ3.Wha iable s a egies exis o achie ing
anspa ency in au oma ed d i ing sys ems?
This pape is s uc u ed as ollows: Sec ion 2 p o-
ides he mo i a ion o applying glass-box models
and p esen s ou pe spec i e on RQ1. Sec ion 3 ex-
plo es he challenges o de eloping such models in
he ield o au oma ed d i ing, add essing RQ2. Fi-
nally, ou p oposed s a egy o RQ3, which ocuses
on modula izing models and balancing in e p e abil-
i y wi h pe o mance, is discussed in Sec ion 4.
2 MOTIVATION
Le ’s assume we wan o design a sys em capable o
aking he bes ac ion in e e y si ua ion, ollowing a
da a-d i en app oach. Fo simplici y, we can en i-
sion i as a pe ec d i e : one ha always selec s he
op imal ou e om A o B while ensu ing ene gy e -
iciency, minimal a el ime, sa e y, and passenge
com o .
The p ocess o building such a sys em would in-
ol e eco ding pai s o in o ma ion poin s and op i-
mal ac ions. Each in o ma ion poin can be consid-
e ed as an abs ac ep esen a ion o e e y hing he
obo ic d i e pe cei es, including he in e nal ehi-
cle s a us and any ex e nal sou ces o in o ma ion
(e.g., d i ing scena io, a ic condi ions, o he oad
use s, e c.). Fo simplici y Fig. 1 p o ides an ab-
s ac ep esen a ion o hese in o ma ion poin s on an
ac ion-in o ma ion Ca esian map, whe e i ep esen s
ou in o ma ion con en and d ep esen s ou bes
eco ded decision in such a scena io. By conside -
ing all eco ded in o ma ion, ep esen ed as g ay do s,
one can cons uc a unc ion—using any analy ical
i ing me hod— hus ob aining an ac ion-in o ma ion
mapping unc ion.
The goal o building such a unc ion is o use he
ac ion-in o ma ion model as a decision-making ool
o es ima e he bes ac ion as new in o ma ion be-
comes a ailable. I we assume ha he blue line in
Fig. 1 ep esen s ou bes model, when new in o ma-
ion inew a i es, he model’s guess should be no o he
han ˆ
dnew as he in e sec ion poin be ween he in o -
ma ion and he model. Howe e , he op imal decision
in ha si ua ion may be some hing sligh ly di e en ,
such as d∗
new, esul ing in a gap be ween he model’s
guess and he ue op imal decision.
Such an gap in he op imali y migh esul om:
1. Measu emen e o s in bo h he in o ma ion used
o build he model and he new da a poin .
2. App oxima ion e o s due o complex in e ac ions
be ween a iables no ully cap u ed by he model.
3. T uly unknown ci cums ances, highligh ing a di -
e ence be ween he cu en in o ma ion poin and
p e ious knowledge o he sys em.
4. A change in he ope a ional domain, whe e he
model was no ained o calib a ed o new con-
di ions, leading o disc epancies in he decision-
making p ocess.
S icking o he example o au oma ed d i ing, he
gap in he op imal decision can lead o a wide ange
o consequences, om mino delays in a el ime o
passenge discom o o mo e se e e ou comes like
c ashes. How can one iden i y he cause o a miscal-
cula ion in he decision-making p ocess i he sys em
is a ully enclosed black box?
F om his pe spec i e, all poin s men ioned ea -
lie —measu emen e o s, app oxima ion e o s, uly
unknown ci cums ances, and changes in he ope a-
ional domain—a e equally alid sou ces o e o wi h
e y li le possible ac ion o add ess and co ec such
e o s o ensu e he sys em unc ions p ope ly ac oss
di e en scena ios. The cu en black-box app oach
o sol ing his p oblem ends o add in o ma ion o he
sys em inde ini ely, leading o memo iza ion a he
han meaning ul knowledge abs ac ion wi h minimal
alida ion and e i ica ion possibili ies (Pikne e al.,
2024).
Assuming ha ou ac ion-in o ma ion model is an
analy ical mul i a ia e unc ion unde pinning ou p o-
cess, a gap in decision-making can be in es iga ed,
s udied, and po en ially debugged i , and only i , he
sys em can be ully explained and in e p e ed.
O en, in e p e abili y and explainabili y a e seen
as a w i en o m o ex p o iding an explana o y
Figu e 1: Ac ion - in o ma ion model depic ion. G ay do s
cons i u e in o ma ion poin s used o gene a e he i ing
unc ion in blue. The op imali y gap a ise om he di -
e ence be ween he op imal decision d∗
new and he model
guess ˆ
dnew.
Glass-box Au oma ed D i ing: Insigh s and Fu u e T ends
881
s a emen abou d i ing decisions and speci ic ac ions
(Omeiza e al., 2022). Howe e , his app oach should
no be con la ed wi h ue in e p e abili y o accoun -
abili y. The gene a ed ex , o en p oduced by black-
box LLMs (la ge language models), me ely desc ibes
he si ua ion wi hou o e ing use ul debugging in o -
ma ion. As LLMs ha e shown, i is always possible
o gene a e ‘human-unde s andable’ explana ions ha
can be e ibly misleading and ail o e lec he sys-
em’s ac ual beha io a i s undamen al le el. Thus,
his ype o e bal explana ion o en alls mo e in o
he ealm o psychological pe spec i e a he han a
genuine in e p e a ion o sys em beha io .
3 CHALLENGES
Wi hou any doub , ull black-box models ha e
demons a ed compelling capabili ies in con olling
complex sys ems such as au onomous ehicles (Chen
e al., 2024). This achie emen canno be neglec ed,
as i ep esen s a signi ican oppo uni y o add ess
long- e m p ac ical p oblems. Howe e , glass-box
models o e he oppo uni y o debug and e ine hese
sys ems by p o iding ull con ol and anspa ency.
This allows o a deepe unde s anding and imp o e-
men o each sys em’s beha io (Kuznie so e al.,
2024).
The i s signi ican challenge in building a glass-
box con ol sys em o au onomous d i ing lies in
managing he complexi y o he en i onmen . As p e-
iously men ioned, he challenge does no s em om
d i ing along a p ede ined pa h a a p ecise eloci y
bu a he om in e ac ing e ec i ely wi h he su -
ounding en i onmen . Au onomous d i ing equi es
p ocessing as amoun s o eal- ime da a om sen-
so s such as came as, LiDAR, ada , and GPS. The
decision-making p ocess en ails in ica e in e ac ions
be ween pe cep ion, p edic ion, and con ol sys ems,
which a e o en modeled using complex deep lea ning
o neu al ne wo ks—me hods ha a e inhe en ly di -
icul o in e p e . Figu e 2 illus a es he con ol low
om low-le el o high-le el d i ing models. The con-
cep is ha any s a e-o - he-a con olle , om PID
(Emi le e al., 2014) o Lyapuno -based con olle s
(Alcala e al., 2018) (Ka a yllis e al., 2022), can
e ec i ely and p ecisely manage ehicle speed and
s ee ing angle, p o ided a simpli ied physical model
o he ehicle is known. The assump ions equi ed
o hese con olle s o unc ion a e o en un ealis-
ic and ine ec i e a p edic ing dynamic en i onmen-
al changes. The highe he complexi y o he e-
hicle model, he be e and mo e e ec i e he con-
olle . This capabili y is su icien o d i ing in
s uc u ed en i onmen s. The sys em’s in e nal com-
ponen s, highligh ed wi h a dashed line in Fig. 2, can
be ully explainable and in e p e able PIDs. Li e a-
u e o e s a obus ounda ion o analy ical analyses
and o mal solu ions o he p oblem o ollowing e -
e ence signals, enabling de e minis ic con ol in hese
scena ios (Fleming and Rishel, 2012).
As he esea ch communi y app oached his p ob-
lem, i quickly became e iden ha add essing i e-
qui ed mul iple le els o abs ac ion. A he e y low
le el, classical au oma ic con olle s, such as PID o
model p edic i e con olle s, pe o m hei asks e -
ec i ely. Howe e , ackling ehicle pe cep ion and
ou ing wi h he same le el o de ail as low-le el con-
ol, ep esen ed in he ou e cycle o Fig. 2, is com-
pu a ionally and p ac ically in easible.
The ou ing p oblem i sel can be subdi ided in o
wo dis inc componen s: he gene a ion o kinema -
ically o dynamically easible ajec o ies o sho -
ange ehicle con ol, and waypoin gene a ion o
high-le el ehicle ou ing. The o me equi es sol -
ing di e en ial equa ions, which a e compu a ionally
demanding and o en limi ed o localized egions.
The la e ypically employs simple , non-physically
complian algo i hms, such as Dijks a’s algo i hm
o apidly-explo ing andom ees (RRT) (LaValle,
2006). A emp ing o sol e di e en ial equa ions o
e e y poin along a long pa h is no only compu a-
ionally p ohibi i e bu also ine ec i e, as i assumes
a s a ic en i onmen , he eby educing he sys em’s
abili y o adap o dynamic changes. Con e sely, ig-
no ing di e en ial equa ions al oge he can esul in
un easible pa hs ha iola e he ehicle’s physical
cons ain s.
The ou ing p oblem exempli ies he necessi y
o di e en le els o abs ac ion o add ess com-
plex challenges in au onomous d i ing. Each le el
e ains i s own in e p e abili y: long- ange planning
may sac i ice physical p ecision bu p o ides compu-
a ional e iciency, while sho - ange ajec o y gen-
e a ion main ains a de ailed physical in e p e a ion.
This balance allows he sys em o adap o a dynamic
en i onmen while ensu ing easibili y a a local le el.
On a di e en le el, he pe cep ion p oblem,
which includes asks such as objec de ec ion and seg-
men a ion, is p edominan ly add essed using black-
box models (Huang e al., 2022). On one hand, hese
models excel in p o iding de ec ion capabili ies ha
can e en su pass human pe o mance in ce ain sce-
na ios, such as low-ligh condi ions o when in eg a -
ing da a om mul iple senso sou ces. An illus a i e
example is shown in he wo images in Fig. 2 which
a e ex ac ed om he IseAu o da ase (Gu e al.,
2023) and include he addi ion o a black-box seg-
IAI 2025 - Special Session on In e p e able A i icial In elligence Th ough Glass-Box Models
882
Figu e 2: Le els o con ol abs ac ions ea u ing low-le el
p ocess con ol and high-le el in elligen unc ions. Visual
images om IseAu o da ase (Gu e al., 2023).
men a ion esul (Gu e al., 2024).
On he o he hand, hey lack an unde s anding o
he seman ic meaning o objec s wi hin a scene, ea -
ing a ee and a pe son me ely as objec s wi hou ec-
ognizing hei dis inc oles o con ex ual impo ance.
Con ex ual impo ance is inhe en ly challenging o
cap u e, as e en human obse e s may assign a y-
ing in e p e a ions and signi icance o simila objec s
in a scene. An in e p e able and explainable model
should no aim o p oduce e bose desc ip ions o a
scene bu a he o classi y objec s accu a ely while
accoun ing o hei high-le el oles and impo ance.
Fo ins ance, in such a amewo k, e o s o laws in
de ec ing ehicles o pedes ians, compa ed o non-
c i ical objec s like ocks o oliage, could be iden i-
ied, debugged, and add essed in a a ge ed manne .
3.1 Sa e y-C i ical Na u e
While i may seem ha au onomous ehicles mus p i-
o i ize spli -second decision-making o e immedia e
explana ions, i is p ecisely in hese high-s akes sce-
na ios ha he e o o build simple, classical con ol,
and ule-based sys ems becomes essen ial. T ans-
pa ency is c i ical o diagnos ics and egula o y com-
pliance, and i is o en mischa ac e ized as comp o-
mising he ehicle’s abili y o espond e ec i ely in
eme gencies. On he con a y, such anspa ency en-
su es ha misunde s andings o misin e p e a ions a e
minimized, pa icula ly in si ua ions whe e hey can-
no be ole a ed. In (Ab ech e al., 2024), he au-
ho s clea ly emphasize he sa e y conce ns ha deep
lea ning poses o au oma ed d i ing, co e ing aspec s
such as ope a ional domain de ini ion and limi a ions,
as well as he me hods used o da a p epa a ion and
algo i hm de elopmen . Mo eo e , egula ions a e
placing inc easing emphasis on he impo ance o AI
explainabili y in sa e y-c i ical indus ies like ans-
po a ion. Glass-box models a e equi ed o comply
wi h s ingen indus y s anda ds, such as ISO 26262
(ISO, 2011) ( unc ional sa e y) and ISO/PAS 21448
(Sa e y o he In ended Func ionali y - SOTIF), o en-
su e eliabili y, accoun abili y, and sa e y (Ki o skii
and Go elo , 2019). Addi ionally, amewo ks such
as he Eu opean Union’s E hics Guidelines o T us -
wo hy AI and he U.S. NIST’s ini ia i es on AI ex-
plainabili y ad oca e o mo e anspa en and ac-
coun able AI sys ems. The ecen ly adop ed AI Ac
(Regula ion (EU) 2024/1689 laying down ha monized
ules on a i icial in elligence) se s clea equi emen s
and obliga ions o AI de elope s, emphasizing ex-
plainabili y and accoun abili y in AI-based sys ems,
including au oma ed d i ing unc ionali ies.
4 STRATEGIES AND FUTURE
TREND
Se e al s a egies exis o achie e anspa ency, ac-
coun abili y, and explainabili y in au oma ed d i ing,
many o which can be applied o o he complex sys-
ems ha bene i om black-box models. While a
ully glass-box model o au oma ed d i ing may be
imp ac ical in i s pu es o m—pa icula ly in com-
plex eal-wo ld scena ios—i emains an ambi ious
long- e m goal. In he in e im, hyb id models ep-
esen a p ac ical solu ion, le e aging he s eng hs
o bo h in e p e able and black-box app oaches. A
p omising s a egy is o combine in e p e able models
o high-le el decision-making wi h black-box mod-
els o pe cep ion asks, s iking a balance be ween
in e p e abili y and e icacy wi h he goal o educing
black-box models o he miminum le el. Fo exam-
ple, ule-based logic can be applied o lane-change
policies (Malayje di e al., 2022), p o iding clea
and explainable decision-making, while deep lea n-
ing models handle objec de ec ion, which o en bene-
i s om he da a-d i en adap abili y o black-box ap-
p oaches
The applica ion o simple , domain-speci ic mod-
els also o e s no able s eng hs in achie ing ans-
pa ency and eliabili y. In cons ained en i onmen s,
such as au onomous shu les o las -mile deli e y e-
hicles, ule-based sys ems o simple machine lea n-
ing models can e ec i ely balance explainabili y wi h
unc ionali y. This app oach aligns wi h he concep
o main aining a human-in- he-loop amewo k, al-
lowing human ope a o s o o e see decisions and in-
e ene when necessa y. This ensu es ha ac ions
aken by he sys em adhe e o e hical and sa e y con-
side a ions. Mo eo e , adop ing a modula sys em
design—b eaking he d i ing s ack in o smalle , in-
e p e able modules— u he aids in achie ing ans-
pa ency. Fo example:
Glass-box Au oma ed D i ing: Insigh s and Fu u e T ends
883
•Pe cep ion: Explains how objec s a e de ec ed and
classi ied.
•Planning: P o ides insigh s in o why a pa icula
ajec o y o decision was selec ed.
•Con ol: Demons a es how he ehicle execu es
he planned commands.
These modula explana ions help ensu e ha he
sys em emains in e p e able and debuggable, while
s ill bene i ing om ad ancemen s in AI and au oma-
ion. E en in such scena ios, designe s mus ely on
ex ensi e simula ion en i onmen s and o mal e i i-
ca ion me hods o unde s and and alida e sys em be-
ha io s unde di e se condi ions.
Au onomous sys ems ope a e in he physical
wo ld, a domain go e ned by he p inciples o physics
(e.g., Maxwell’s equa ions, New on’s laws). These
p inciples p o ide a obus ounda ion o de ining
a go e ning amewo k ha can enhance bo h ex-
plainabili y and in e p e abili y. A c i ical ques ion
eme ges:
How migh one le e age he undamen al p op-
e ies o physics o build a alida ion go e no
a ound AI-based au onomy sys ems?
This ques ion emains an open a enue o e-
sea ch, challenging he communi y o explo e inno-
a i e ways o in eg a e physical laws in o he ali-
da ion and explainabili y amewo ks o au onomous
sys ems (Pikne e al., 2024).
The applica ion o Pos -Hoc Explainabili y such
as saliency maps, SHAP (Shapley Addi i e Explana-
ions) (Lundbe g and Lee, 2017), (Lundbe g e al.,
2020), o LIME (Local In e p e able Model-Agnos ic
Explana ions) (Ribei o e al., 2016) o analyze and ex-
plain decisions made by black-box models cons i u e
also eme gen solu ions. One migh no e ha Pos -
hoc me hods do no equi e he model o be inhe -
en ly in e p e able bu a he a emp o in e p e and
explain he ou pu s o a AI model a e i has been
ained, which makes hem p ac ical o use wi h com-
plex models ha p io i ize pe o mance bu sac i ice
in e p e abili y.
5 CONCLUSION
A ully glass-box model o au oma ed d i ing is
likely imp ac ical in i s pu es o m, pa icula ly in
complex eal-wo ld scena ios as highly in e p e able
models (e.g., decision ees o ule-based sys ems)
may s uggle o cap u e he nuanced decision-making
equi ed in dynamic d i ing en i onmen s. How-
e e , hyb id app oaches ha blend in e p e abili y
wi h black-box models o e a p ac ical way o wa d
while keeping he aspi a ion o ull glass-box designs
ali e. As is o en he case in enginee ing, he op imal
solu ion lies in he middle g ound—making he c i -
ical componen s o he sys em anspa en enough o
ensu e sa e y, eliabili y, and accoun abili y, wi hou
comp omising pe o mance.
The d i e o pu sue glass-box models is deeply
oo ed in he undamen al cu iosi y o enginee s. The
ques ion, ’Why does a sys em wo k he way i does?’
o ms he e y basis o scien i ic explo a ion. Wi h-
ou such cu iosi y, humani y migh s ill accep a la
Ea h as u h, ne e challenging expe iences ha
seem coun e in ui i e o seeking al e na i e pe spec-
i es. The essence o p og ess lies in ques ioning and
e aming ou unde s anding o he wo ld and i s com-
plexi ies.
ACKNOWLEDGMENT
Pa o his esea ch has ecei ed unding om he Eu-
opean Union’s Ho izon 2020 Resea ch and Inno a-
ion P og amme unde g an ag eemen No. 856602
(Fines Twins) and om he Eu opean Union’s Ho i-
zon Eu ope Resea ch and Inno a ion P og amme un-
de g an ag eemen No. 101135988 (PLIADES:
AI-Enabled Da a Li ecycles Op imiza ion and Da a
Spaces In eg a ion o Inc eased E iciency and In e -
ope abili y)
REFERENCES
Ab ech , S., Hi sch, A., Raa a nia, S., and Woeh le, M.
(2024). Deep lea ning sa e y conce ns in au oma ed
d i ing pe cep ion. IEEE T ansac ions on In elligen
Vehicles, pages 1–12.
Alcala, E., Puig, V., Que edo, J., Escobe , T., and Comaso-
li as, R. (2018). Au onomous ehicle con ol using a
kinema ic lyapuno -based echnique wi h lq -lmi un-
ing. Con ol enginee ing p ac ice, 73:1–12.
Chen, L., Wu, P., Chi a, K., Jaege , B., Geige , A., and
Li, H. (2024). End- o-end au onomous d i ing: Chal-
lenges and on ie s. IEEE T ansac ions on Pa e n
Analysis and Machine In elligence, 46(12):10164–
10183.
Emi le , M. T., Uygan, ˙
I. M. C., Aksun G¨
u enc¸, B.,
and G¨
u enc¸, L. (2014). Robus pid s ee ing con-
ol in pa ame e space o highly au oma ed d i -
ing. In e na ional Jou nal o Vehicula Technology,
2014(1):259465.
Fleming, W. H. and Rishel, R. W. (2012). De e minis ic
IAI 2025 - Special Session on In e p e able A i icial In elligence Th ough Glass-Box Models
884
and s ochas ic op imal con ol, olume 1. Sp inge
Science & Business Media.
Gu, J., Bellone, M., Pi oˇ
nka, T., and Sell, R. (2024).
Cl : Came a-lida usion ans o me o seman ic
segmen a ion in au onomous d i ing. a Xi p ep in
a Xi :2404.17793.
Gu, J., Lind, A., Chhe i, T. R., Bellone, M., and Sell, R.
(2023). End- o-end mul imodal senso da ase col-
lec ion amewo k o au onomous ehicles. Senso s,
23(15):6783.
Huang, K., Shi, B., Li, X., Li, X., Huang, S., and
Li, Y. (2022). Mul i-modal senso usion o
au o d i ing pe cep ion: A su ey. a Xi p ep in
a Xi :2202.02703.
ISO (2011). Iso 26262 oad ehicles– unc ional sa e y. ISO
S anda d (2011).
Ka a yllis, I., Theodosis, D., and Papageo giou, M. (2022).
Lyapuno -based wo-dimensional c uise con ol o
au onomous ehicles on lane- ee oads. Au oma ica,
145:110517.
Ki o skii, O. and Go elo , V. (2019). D i e assis ance sys-
ems: analysis, es s and he sa e y case. iso 26262 and
iso pas 21448. In IOP Con e ence Se ies: Ma e ials
Science and Enginee ing, olume 534, page 012019.
IOP Publishing.
Kuznie so , A., Gye na , B., Wang, C., Pe e s, S., and
Alb ech , S. V. (2024). Explainable ai o sa e and
us wo hy au onomous d i ing: A sys ema ic e iew.
a Xi p ep in a Xi :2402.10086.
LaValle, S. M. (2006). Planning algo i hms. Camb idge
uni e si y p ess.
Lundbe g, S. M., E ion, G., Chen, H., DeG a e, A., P u kin,
J. M., Nai , B., Ka z, R., Himmel a b, J., Bansal, N.,
and Lee, S.-I. (2020). F om local explana ions o
global unde s anding wi h explainable ai o ees. Na-
u e machine in elligence, 2(1):56–67.
Lundbe g, S. M. and Lee, S.-I. (2017). A uni ied ap-
p oach o in e p e ing model p edic ions. In Guyon, I.,
Luxbu g, U. V., Bengio, S., Wallach, H., Fe gus, R.,
Vishwana han, S., and Ga ne , R., edi o s, Ad ances
in Neu al In o ma ion P ocessing Sys ems 30, pages
4765–4774. Cu an Associa es, Inc.
Malayje di, E., Sell, R., Malayje di, M., Udal, A., and Bel-
lone, M. (2022). P ac ical pa h planning echniques in
o e aking o au onomous shu les. Jou nal o Field
Robo ics, 39(4):410–425.
Omeiza, D., Web, H., Ji o ka, M., and Kunze, L. (2021).
Towa ds accoun abili y: P o iding in elligible expla-
na ions in au onomous d i ing. In 2021 IEEE In elli-
gen Vehicles Symposium (IV), pages 231–237. IEEE.
Omeiza, D., Webb, H., Ji o ka, M., and Kunze, L. (2022).
Explana ions in au onomous d i ing: A su ey. IEEE
T ansac ions on In elligen T anspo a ion Sys ems,
23(8):10142–10162.
Pikne , H., Malayje di, M., Bellone, M., Bayka a, B. C.,
and Sell, R. (2024). Au onomous d i ing alida ion
and e i ica ion using digi al wins. VEHITS, pages
204–211.
Ribei o, M. T., Singh, S., and Gues in, C. (2016). ”why
should I us you?”: Explaining he p edic ions o any
classi ie . In P oceedings o he 22nd ACM SIGKDD
In e na ional Con e ence on Knowledge Disco e y
and Da a Mining, San F ancisco, CA, USA, Augus
13-17, 2016, pages 1135–1144.
S an on, N. A. and Young, M. S. (1998). Vehicle au oma ion
and d i ing pe o mance. E gonomics, 41(7):1014–
1028.
Weisse , H., Schulenbe g, P., Gollinge , H., and Michle , T.
(1999). Au onomous d i ing on ehicle es acks:
o e iew, implemen a ion and ehicle diagnosis. In
P oceedings 199 IEEE/IEEJ/JSAI In e na ional Con-
e ence on In elligen T anspo a ion Sys ems (Ca .
No. 99TH8383), pages 62–67. IEEE.
Yu se e , E., Lambe , J., Ca ballo, A., and Takeda, K.
(2020). A su ey o au onomous d i ing: Common
p ac ices and eme ging echnologies. IEEE access,
8:58443–58469.
Glass-box Au oma ed D i ing: Insigh s and Fu u e T ends
885
