Neural Predictive Control for Mobile Robot Navigation in a Partially Structured Static Environment
Abstract
This paper presents a way of implementing a Model Based Predictive Controller (MBPC) for mobile robot navigation when unexpected static obstacles are present in the robot environment. The method uses a non-linear model of mobile robot kinematics and thus allows an accurate prediction of the future trajectories. An ultrasonic ranging system has been used for obstacle detection. A Multilayer Perceptron is used to implement the MBPC, allowing real-time and also eliminating the need for data sensor high level processing. The perceptron has been trained to reproduce the MBPC behaviour in a supervised manner. Experimented results obtained when applying the neural network controller to a mobile robot are given in the paper.
Full text
8125
8126
8127
8128
NSF
(x)
= L 9i(X) +
19i(X)1
i=1
is ze o inside
he
con ex egion
and
inc eases linea ly
ou
o i , as
he
dis ance
o
he
on ie is augmen ed.
N S F is
he
numbe
o segmen s
o
he
obs acle on ie .
The
ollowing po en ial unc ion
is
used:
1
Pe x(x) =
[8
+
(x)]-1
=
--N-S-F------
8 + L (9i(X) +
19i(X)I)
i=1
whe e 8 is a small cons an
ha
limi s
he
alue
o
Pc x(x) inside
he
con ex egion. Pe x(x) eaches i s max-
imum alue
8-
1 inside
he
egion occupied by
he
obs a-
cle,
and
dec eases wi h
he
dis ance be ween he obo
and
he
obs acle. A g aphic example o his unc ion is
shown in igu e
5,
whe e wo ec angula shape s a ic
obs acles a e p esen in
he
p oximi y o
he
obo .
Fig.
5.
Con ex egions po en ial unc ion.
Po en ial
unc ion
o
a conca e polygon. A conca e e-
gion will be desc ibed by a se o inequali ies
ex)
:::::
0, E
Lm,
x E
IR
n
The
po en ial unc ion used in his case
is
1
Peel'
( x) = 8 +
()
gee"
X
whe e 8
is
a small cons an
and
9cc (X)
is
he
minimum
o
he
dis ances be ween
he
obo posi ion and e e y
s aigh
line
ha
de ines
he
obs acle on ie . This unc-
ion has
he
same cha ac e is ics as Peex(x).
An example
o
cos unc ion J, o a one s ep p edic ion
ho izon (which is
he
only one
ha
can be g aphically
ep esen ed),
is
shown in igu e 6.
In
his igu e
he
alue
o J o di e en le
and
igh wheel eloci ies is ep-
esen ed.
The
exis ence o a ec angula obs acle can
be no iced. Fu he mo e,
he
in luence o
he
non linea
p edic ion model can be obse ed.
Fig.
6.
Obje i e unc ion J.
3.
THE
NEURAL
NETWORK
APPROACH
As was men ioned be o e,
he
minimiza ion o
he
cos
unc ion J has
o
be ca ied
ou
by a nume ical op i-
miza ion me hod which equi es
oo
much compu a ion
o
be used in eal ime. A Neu al Ne wo k solu ion is
p oposed, which gua an ees eal ime o
he
obo con-
ol. Neu al Ne wo k app oaches o obo guidance has
been p oposed by o he esea che s (Pome lau, 1990),
(Meng and Kak, 1993).
The
modules o
he
con ol scheme used in his wo k
(see igu e
7)
a e:
Fig.
7.
P edic i e neu al ne wo k scheme o mobile
obo na iga ion.
•
A i icial
neu al
ne wo k
con olle .
The
ANN
a chi ec u e chosen he e
is
a Mul ilaye Pe cep on,
wi h one hidden laye (see igu e 8).
The
inpu laye consis s o wel e neu ons (see
igu e 8).
The
i s wo inpu s co espond
o
he
p e ious linea
and
angula eloci ies o
he
obo .
The
nex h ee inpu s a e associa ed
o
he
pa ame iza-
ion o
he
desi ed
ajec o y
o e
he
p edic ion
ho izon. In o de
o
educe
he
numbe
o
inpu s,
he
pa ame e s gi en
o
he
ne wo k a e
he
dis-
ance d om
he
obo guide poin
o
he
pa h,
he
angle 8 be ween
he
obo heading
and
he
pa h
o i-
en a ion and
an
a e age o
he
in e se o
he
cu a-
8129
8130