126
In e na ional Jou nal o Ad ance and Applied Resea ch
www.ijaa .co.in
ISSN – 2347-7075
Impac Fac o – 8.141
Pee Re iewed
Bi-Mon hly
Vol. 6 No. 38
Sep embe - Oc obe - 2025
Enhanced G ound T a e sabili y Es ima ion o Quad uped Robo s using
Imp o ed CNN A chi ec u es and Expanded Heigh map Da ase
Nee a Bonde
Dep o Compu e Science.
D . D.Y. Pa il Science and Compu e Science College, Aku di, Pune.
Co esponding Au ho –Nee a Bonde
DOI - 10.5281/zenodo.17312926
Abs ac :
T a e sabili y es ima ion is a key p e equisi e o sa e and e icien na iga ion o quad uped
obo s in uns uc u ed en i onmen s. While p e ious esea ch demons a ed he use o con olu ional
neu al ne wo ks (CNNs) o classi ying e ain heigh map pa ches in o a e sable and non- a e sable
ca ego ies, limi a ions in da ase size and shallow ne wo k a chi ec u e es ic ed model gene aliza ion.
In his pape , we ex end p io wo k by (i) expanding he simula ed da ase om 12 o 60 di e se
heigh maps and (ii) imp o ing he CNN a chi ec u e by in oducing deepe con olu ional laye s, ba ch
no maliza ion, d opou egula iza ion, and he Adam op imize . These modi ica ions inc eased
classi ica ion accu acy om 82% o 96%, signi ican ly enhancing obus ness ac oss di e se e ain
condi ions. The p oposed model is in eg a ed in o a a e sabili y-awa e pa h planning amewo k,
enabling quad uped obo s o selec sa e and smoo he ajec o ies in complex e ains. Unlike
handc a ed geome ic ea u es, he CNN-based me hod lea ns o ex ac ele an spa ial pa e ns
au oma ically. The inc eased da ase di e si y no only imp o es classi ica ion accu acy bu also
ensu es obus ness ac oss di e en e ain mo phologies, including hills, slopes, ocky su aces, and
une en pa ches. This wo k hus b idges he gap be ween limi ed simula ion-based a e sabili y s udies
and eal-wo ld deploymen challenges.
Keywo ds: CNN, quad uped obo s, a e sabili y es ima ion, deep lea ning, pa h planning, Gazebo
simula ion
In oduc ion:
Na iga ion in uns uc u ed e ains
equi es eliable a e sabili y es ima ion o
p e en obo s om becoming immobilized in
ough, obs acle- ich, o i egula egions.
Na u al en i onmen s such as ocky hillsides,
collapsed s uc u es, o c a e -like dep essions
o en p esen unp edic able condi ions whe e
small e o s in e ain assessmen can esul in
signi ican ins abili y. Ea lie me hods based
on slope h esholds, e ain oughness indices,
o geome ic heu is ics o en ail in hese
en i onmen s, as hey canno cap u e he
complexi y and a iabili y o na u al e ain
ea u es (Bal a e al., 2013; Sil e e al., 2010).
To o e come hese limi a ions, ecen
esea ch has le e aged con olu ional neu al
ne wo ks (CNNs) applied o heigh maps,
whe e local image pa ches a e labeled as
a e sable o non- a e sable based on obo
simula ion ou comes (Cha ez-Ga cia e al.,
2017; Gius i e al., 2016). CNNs p o ide he
ad an age o lea ning hie a chical spa ial
ea u es di ec ly om da a, educing eliance
on manually enginee ed ea u es. Howe e ,
p io app oaches aced wo majo challenges:
(i) da ase limi a ions—mos s udies used only
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Nee a Bonde
127
12 e ains, which p o ided insu icien
a iabili y o obus gene aliza ion (Dhande
& Ohol, 2022), and (ii) a chi ec u al
limi a ions—shallow CNNs wi h only a small
numbe o il e s pe con olu ional laye (e.g.,
5) lacked he ep esen a ional powe equi ed
o dis inguish sub le e ain di e ences. These
limi a ions led o subop imal accu acy o
a ound 82% and poo adap abili y o unseen
e ains.
In his pape , we add ess hese challenges wi h
wo main con ibu ions:
Da ase Expansion. We c ea ed 60
di e se cus om e ains in simula ion, esul ing
in a signi ican ly la ge and iche da ase ha
cap u es a wide ange o e ain mo phologies
(Fankhause e al., 2018).
Imp o ed CNN A chi ec u e. We
designed a deepe CNN model wi h inc easing
il e sizes (32–64–128) ac oss laye s,
combined wi h ba ch no maliza ion, d opou
egula iza ion, and he Adam op imize
(Kingma & Ba, 2015). This a chi ec u e
achie ed a classi ica ion accu acy o 96%.
Reliable a e sabili y es ima ion is
no only aluable o simula ion s udies bu is
also c i ical o eal-wo ld obo ic applica ions
such as plane a y explo a ion o e s,
au onomous mili a y sys ems, and sea ch-and-
escue ope a ions in haza dous en i onmen s
(Cunningham e al., 2013; Gladisch e al.,
2025). Quad uped obo s, compa ed o
wheeled o acked obo s, p o ide supe io
mobili y in une en and i egula e ains, bu
hei highe deg ees o eedom make hem
mo e ulne able o ins abili y when
a e sabili y p edic ions a e inaccu a e
(Wellhausen e al., 2020).
Rela ed Wo k:
Ea lie a e sabili y es ima ion
me hods p ima ily elied on slope h esholds,
oughness indices, and geome ic analysis o
classi y e ain (Bal a e al., 2013; Zhou e al.,
2022). While compu a ionally simple, hey
o en s uggled in na u al en i onmen s wi h
i egula ocks, s eep inclines, o mixed
obs acle dis ibu ions. These handc a ed
me hods we e sensi i e o noise, en i onmen -
speci ic assump ions, and senso inaccu acies,
limi ing scalabili y.
Mo e ecen CNN-based me hods
demons a ed ha lea ning disc imina i e
ea u es di ec ly om heigh maps can
signi ican ly ou pe o m geome ic heu is ics
(Cha ez-Ga cia e al., 2017; Shen e al., 2025).
Howe e , da ase ichness and ne wo k dep h
emain c i ical o pe o mance (Bansod e al.,
2025). Ou wo k ex ends hese s udies by
enla ging he da ase om 12 e ains (Dhande
& Ohol, 2022) o 60 and edesigning he CNN
wi h deepe laye s, ba ch no maliza ion, and
d opou , he eby imp o ing gene aliza ion
(Cos a e al., 2025).In addi ion, ein o cemen
lea ning (RL) app oaches ha e been applied o
e ain-awa e locomo ion (Qu eshi & Ayaz,
2014), bu RL equi es ex ensi e simula ion
and aces a ― eali y gap‖ when ans e ed o
physical obo s (Zhang e al., 2024). In
con as , ou supe ised lea ning app oach
p o ides e icien , in e p e able, and
ans e able e ain classi ica ion sui able o
in eg a ion in o highe -le el planning
(Fankhause e al., 2016).
App oach:
A.
Da ase Gene a ion:
We simula ed an Anymal quad uped
obo in Gazebo and c ea ed 60 cus om
e ains (heigh maps), (simila o Fankhause
e al., 2018).compa ed o only 12 in p io
wo k. The obo was commanded wi h
cons an eloci y ac oss hese e ains. I he
obo a e sed he p ede ined h eshold
dis ance, he co esponding pa ches we e
labeled a e sable; o he wise, non-
a e sable. This gene a ed app oxima ely
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Nee a Bonde
128
350,000 labeled pa ches a 20 Hz sampling
a e. The e ains included syn he ic hills,
s ai - like s uc u es, andom noise-based
su aces, and obs acle- ich maps. By
inc easing he a ie y o e ains, he da ase
co e s a wide spec um o di icul y le els,
ensu ing ha he model lea ns ea u es
beyond simple slopes. Addi ionally, each
pa ch was no malized and esized be o e
being ed in o he CNN, ensu ing consis ency
in aining.
Fig.1. Cus om Heigh Maps
The igu e below shows he simula ion
o he Anymal quad uped obo ac oss
di e en e ains. These simula ions illus a e
he da ase c ea ion p ocess, whe e e ain
pa ches we e labeled as a e sable o non-
a e sable based on he obo ’s in e ac ion
wi h he en i onmen .
Fig. 2. Te ain 1
Fig. 3. Te ain 2
B.
Baseline CNN
The baseline model consis ed o h ee
con olu ional laye s wi h 5 il e s each (3×3
ke nels), ollowed by max pooling and wo
dense laye s (128 neu ons and 2 ou pu
neu ons). The model was ained using
Adadel a op imize and achie ed 82 (Dhande
& Ohol 2022)
C.
Imp o ed CNN (P oposed)
Ou p oposed CNN in oduces he
ollowing imp o emen s:
•
Con olu ional il e s inc eased o 32,
64, and 128 in successi e laye s.
•
Ba ch no maliza ion applied a e each
con olu ional block.
•
MaxPooling laye s o spa ial down
sampling.
•
Fully connec ed laye wi h 128
neu ons, ollowed by D opou (0.5) o
p e en o e i ing.
•
Final dense laye wi h 2 neu ons and
So Max ac i a ion.
•
Adam op imize used ins ead o
Adadel a. (Kingma & Ba, 2015)
This modi ica ion signi ican ly boos ed
pe o mance, achie ing a classi ica ion
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Nee a Bonde
129
accu acy o 96%, which ep esen s a
subs an ial imp o emen o e he baseline
design. The p og essi e inc ease in
con olu ional il e s (32 → 64 → 128) allows
he ne wo k o i s cap u e ine-g ained
local s uc u es such as edges, slopes, o small
i egula i ies, and hen p og essi ely abs ac
highe -le el e ain cha ac e is ics like
obs acles, hills, o une en egions. This
hie a chical ea u e ex ac ion is essen ial o
complex e ain unde s anding. Ba ch
no maliza ion u he s abilizes he lea ning
p ocess by educing in e nal co a ia e shi
and main aining consis en ac i a ion
dis ibu ions ac oss laye s, he eby enabling
as e and mo e eliable aining. In addi ion,
he inclusion o d opou egula iza ion
p e en s excessi e co-adap a ion o neu ons,
o cing he ne wo k o lea n mo e obus
ep esen a ions ha gene alize be e o
unseen e ains. Finally, he adop ion o he
Adam op imize , ins ead o adi ional
me hods such as Adadel a o s ochas ic
g adien descen , p o ides adap i e lea ning
a es ha dynamically adjus o each
pa ame e , signi ican ly imp o ing
con e gence speed while a oiding local
minima. Toge he , hese a chi ec u al and
op imiza ion e inemen s c ea e a model ha
is no only mo e accu a e bu also mo e s able,
e icien , and sui able o deploymen in
sa e y-c i ical obo ic applica ions.
Fig 4. Accu acy G aph
D.
In eg a ion wi h DWA
As in he p e ious wo k, he classi ie
ou pu is in eg a ed in o he cos unc ion o
he Dynamic Window App oach (DWA).
Wi h he imp o ed classi ie con idence, he
algo i hm selec s smoo he and mo e eliable
ajec o ies, u he educing he likelihood o
he obo en e ing non- a e sable egions.
(Dhande & Ohol, 2022),
Resul :
The imp o ed CNN demons a ed a
14% accu acy gain o e he baseline model.
Da ase expansion wi h 60 e ains u he
enhanced obus ness and gene aliza ion,
while Adam op imize educed aining ime
compa ed o Adadel a. In addi ion,
quali a i e esul s om isual inspec ion o
classi ica ion maps con i med ha he
imp o ed CNN was able o be e dis inguish
sub le e ain a ia ions ha he baseline
model o en misclassi ied. The educed
aining ime highligh s he e iciency o he
Adam op imize , which dynamically adjus s
lea ning a es. Fu he mo e, he con usion
ma ix indica ed ha alse posi i es
(classi ying non- a e sable pa ches as
a e sable) we e educed signi ican ly,
which is c i ical o eal-wo ld obo ic
deploymen whe e sa e y ma gins a e
essen ial. (Wellhausen e al., 2020).
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Nee a Bonde
130
Conclusion and Fu u e Wo k:
This s udy p esen ed an enhanced
CNN-based a e sabili y es ima ion
amewo k o quad uped obo s. By
expanding he da ase om 12 o 60 e ains
and adop ing a deepe CNN a chi ec u e wi h
ba ch no maliza ion, d opou , and he Adam
op imize , he p oposed me hod imp o ed
accu acy om 82% o 96%. These esul s
unde line he impo ance o bo h da ase
di e si y and deepe ne wo ks o eliable
e ain classi ica ion. The educ ion in alse
posi i es and sho e aining imes u he
alida e he model’s p ac icali y in obo ic
applica ions whe e bo h sa e y and e iciency
a e c i ical.
Beyond hese di ec accu acy
imp o emen s, his wo k con ibu es mo e
b oadly o he long- e m ision o enabling
quad uped obo s o ope a e au onomously in
highly a iable and unce ain e ains.
Reliable a e sabili y es ima ion is a
co ne s one capabili y o such obo s, as i
di ec ly a ec s mission success in domains
such as plane a y explo a ion, unde g ound
mining, ag icul u al au oma ion, and disas e
esponse. The abili y o co ec ly iden i y sa e
pa hs no only ensu es ope a ional con inui y
bu also educes main enance cos s and he
isk o mission-c i ical ailu es.
Fu u e esea ch can ex end his wo k in
se e al p omising di ec ions:
1)
Real- ime gene a ion o heigh maps
om onboa d senso s wi hou elying
solely on p e-simula ed en i onmen s.
2)
Deploymen and alida ion on eal
quad uped ha d- wa e.
3)
Inco po a ing mul i-modal senso
usion (LiDAR + came a) o u he
obus ness.
4)
Le e aging la ge-scale image da ase s
o p e aining, ollowed by ine- uning
on e ain da a, may signi ican ly
educe da a equi emen s while
imp o ing gene aliza ion. Semi-
supe ised echniques could u he
help in scena ios wi h limi ed labeled
e ain da a.
5)
Me ging a e sabili y es ima ion wi h
ein o cemen lea ning locomo ion
s a egies would enable obo s no jus
o classi y e ain bu also o adap hei
gai s dynamically based on e ain
di icul y.
6)
De eloping in e p e able CNN models
ha p o ide explana ions o hei
p edic ions will inc ease us -
wo hiness, especially in sa e y-c i ical
deploymen s.
Ul ima ely, his esea ch en isions he
deploymen o in elligen quad uped obo s
capable o eliable, adap i e, and sa e
na iga ion in uns uc u ed eal-wo ld
en i onmen s. Such sys ems could play a
ans o ma i e ole in socie y, assis ing in
disas e eco e y ope a ions, enabling
sus ainable ag icul u al p ac ices, conduc ing
haza dous inspec ions in mines, and
ad ancing plane a y coloniza ion missions.
Wi h con inued imp o emen s, he p oposed
amewo k se es as a s epping s one owa d
b idging he gap be ween con olled
simula ions and he demanding equi emen s
o ully au onomous quad uped na iga ion in
Model
Da ase Size
Fil e s
Op imize
Accu acy
Baseline CNN
12 heigh maps
5-5-5
Adadel a
82%
P oposed CNN (This Wo k)
60 heigh maps
32-64-128
Adam
96%
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Nee a Bonde
131
he eal wo ld.
Re e ences:
1. Jianming Zhang, Ruiqiang Zhang,
Zilong Dong, Yuxin Wang, Yu Song,
Guangming XieTRIP: Te ain
T a e sabili y Mapping wi h Risk-
Awa e P edic ion o Enhanced Online
Quad upedal Robo Na - iga ion. a Xi
p ep in , No 2024.
2. Jens Gladisch, Luca Ba olomei, Ma c
Ho¨p , Thomas S oyano , Achim
3. J. Lilien hal Online Adap i e
T a e sabili y Es ima ion h ough In-
e ac ion o Uns uc u ed, Densely
Vege a ed En i onmen s. a Xi
p ep in , Feb 2025.
4. Yash Bansod, Abhishek Sha ma,
Subodh Mish a Towa ds Ze o-Sho
Te ain T a e sabili y Es ima ion:
Challenges and Oppo uni ies. a Xi
p ep in , Aug 2025.
5. Filippo Cos a, Lo enzo Saba ini,
Rugge o Ca li On Te ain
T a e sabili y Analysis in Uns uc u ed
En i onmen s: Recen Ad- ances in
Fo es Applica ions. Jou nal o
In elligen Robo ic Sys ems, Sp inge ,
Ma 2025.
6. S. Dhande and S. S. Ohol, ‖G ound
T a e sabili y Es ima ion o
Quad uped Robo wi h Modi ied DWA
Algo i hm,‖ 2022 In e na ional
Con e ence on Signal and In o ma ion
P ocessing (IConSIP), Pune, India,
2022, pp. 1-5, doi:
10.1109/ICoNSIP49665.2022.1000749
6.
7. Ho nung, W. Bu ga d, K. M. Wu m, C.
S achniss, M. Wu m, ―Oc omap: An
e icien p obabilis ic 3d mapping
amewo k based on oc ees,‖ in
Au onomous obo s, 2013, ol. 34, no.
3, pp. 180–200.k and A. Pen land,
Jou nal o Cogni i e Neu oscience,
ol.3, No.1, 1991
8. P. Fankhause , M. Bloesch, and M.
Hu e , ―P obabilis ic e ain mapping
o mobile obo s wi h unce ain
localiza ion,‖ p esen ed a he IEEE
Robo ics and Au oma ion Le e s, ol.3,
2018, pp. 3019- 3026.
9. P. Fankhause , C. Da io Bellicoso, M.
Bjelonic, M. Bjelonic, T. Miki, and M.
Hu e , ―Robus ough- e ain
locomo ion wi h a quad upedal obo ,‖
p esen ed a he IEEE In e na ional
Con e ence on Robo ics and
Au oma ion (ICRA),2018, pp. 57600–
57677.
10. R. Buchanan, M. Bjelonic, T.
Bandyopadhyay, L. Wellhausen, M.
Hu - e , and N. Ko - ege, ―Pe cep i e
whole-body planning o mul ilegged
obo s in con ined spaces,‖ p esen ed in
Jou nal o Field Robo ics, 2020, pp.
68-84.
11. M.We melinge , M.Hu e , P.K usi,
R.Siegwa , P.Fankhause , ¨
R.Die helm, ―Na iga ion Planning o
Legged Robo s in Challenging Te ain‖
p esen ed in IEEE/RSJ In e na ional
Con e ence on In elligen Robo s and
Sys ems (IROS),Oc obe 2016, pp.
1184-1189.
12. D. Sil e , J. And ew Bagnell, A.
S en z,‖ Lea ning om demons a ion
o au onomous na iga ion in complex
uns uc u ed e ain‖ p esen ed in The
In e na ional Jou nal o Robo ics
Resea ch 29(12), Oc obe 2010, pp.
1565–1592.
13. H. Bal a, Gee De Cubbe , D. Do o ei
, Y. Baudoin , H. Sahli, ‖Te ain
a e sabili y analysis o o - oad
obo s using ime-o - ligh 3d sensing‖
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
Nee a Bonde
132
in 7 h IARP In e na ional Wo kshop on
Robo ics o Risky En i onmen -
Ex eme Robo ics , Oc obe 2013. pp.
1223-1231.
14. C. Cunningham, U. Wong, K. Pe e son
and W. Whi ake p esen ed in
P oceedings o 9 h In e na ional
Con e ence on Field and Se ice
Robo ics (FSR ’13), Decembe 2013,
pp. 61 – 74.
15. R. Oma Cha ez-Ga cia, J. Guzzi, L.
Ma ia Gamba della, ―Image
Classi ica ion o G ound T a e sabili y
Es ima ion in Robo ics‖, p esen ed in
In e na ional Con e ence on Ad anced
Concep s o In elligen Vision
Sys ems,No 2017, pp. 225-236.
16. R. Oma Cha ez-Ga cia, J. Guzzi, L.
Ma ia Gamba della, ―Image
Classi ica ion o G ound T a e sabili y
Es ima ion in Robo ics‖, p esen ed in
In e na ional Con e ence on Ad anced
Concep s o In elligen Vision
Sys ems,No 2017, pp. 225-236.
17. L. Zhou, J. Wang, S. Lin, Z. Chen,‖
Te ain T a e sabili y Mapping Based
on LiDAR and Came a Fusion‖,
p esen ed in 8 h In e na ional
Con e ence on Au oma ion, Robo ics
and Applica ions, Feb 2022, pp. 217-
222.
18. H. Qu eshi and Y. Ayaz, ―Po en ial
ield based ein o cemen lea ning o
mo ion planning in mobile obo s,‖
IEEE In e na ional Con e ence on
Mecha onics and Au oma ion (ICMA),
2014, pp. 519- 524.
19. Gius i, J. Guzzi, D. C. Ci esan, e al., ―A
machine lea ning app oach o isual
pe cep ion o o es ails o mobile
obo s,‖ IEEE Robo ics and
Au oma ion Le e s, ol. 1, no. 2, pp.
661–667, 2016.
20. J. Wellhausen, R. Ran l, and M.
Hu e , ―Sa e obo na iga ion ia
mul i-modal anomaly de ec ion,‖ IEEE
Robo ics and Au oma ion Le e s, ol.
5, no. 2, pp. 1326–1333, 2020.
21. Y. Gao, M. E e e , B. Huang, and J.
How, ―In en ion-Ne : In eg a ing
planning and deep lea ning o goal-
di ec ed au onomous na iga ion,‖
Con e ence on Robo ics: Science and
Sys ems (RSS), 2017.
