1998 Second InternationalConferem on Knowledge-Based Intelligent Electronic Systems, 21-23 April 1998, Adelaide, Australia. Editors, L.C.

Jain and R.KJain

Control of an mnidirectional Mobile Robot

Keigo Watanabe

Department of Mechanical Engineering

Faculty of Science and Engineering
Saga University
l-Honjomachi, Saga 840-8502, Japan
wat anabeeme. saga-u.ac.j p

Keywords: Holonomic System, Nonholonomic System, Control of Mobile Robot , Om-

nidirectional Mobile Robot

Abstract focused on the study of omnidirectional mobile

Holonomic or nonholonomia omnidirectional robots. The omnidirectional mobile robot has
mobile robots are known to be constructed by been studied by using a variety of mechanisms. In
applying a specialized wheel or mobile mecha- particular, a holonomic vehicle has a full omnidi-
nism. In particular, some representative control rectionality with simultaneous and independently
approaches are described for a holonomic and controlled rotational and translational motion ca-
omnidirectional mobile robot with three lateral pabilities.
orthogonal-wheel assemblies. The uses of resolved In this paper, we first review several specialized
acceleration control method, PID method, fuzzy wheel or mobile mechanisms for constructing an
model method, and stochastic fuzzy servo method omnidirectional mobile robot. Then, we develop
a dynamic model for an omnidirectional mobile
are highlighted.
robot, in which it is assumed that the platform
1. Introduction is based on three lateral orthogonal-wheel assem-
blies. Then, in order to develop a control system
Recently, a wheeled mechanism is now widely with high performance for such a mobile robot, we
used as a mobile mechanism. Two independent discuss on resolved acceleration control method,
driving wheels mechanism, or front-wheel steering PID method, fuzzy model method, and stochastic
and rear-wheel driving mechanism is well-known fuzzy servo method.
as a representative one, whose mechanism is of-
ten used for a conventional mobile robot. Note, 2. Review of Specialized Wheel or
however, that such a mechanism has a nonholo- Mobile Mechanisms
nomic constraint, so that it prevents the robot
from realizing a high mobility. It is required that The holonomic or nonholonomic omnidirec-
a suitable desired trajectory satisfying the above tional mobile robots has been studied by using
constraint must be designed to control a nonholo- a variety of mechanisms. In other wards, several
nomic mobile mechanism. In addition, there ex- omnidirectional platforms have been known to be
ists a problem that a feedback controller can not realized by developing a specialized wheel or mo-
easily designed to control the nonholonomic robot bile mechanism. From this point of view, such
accurately. specialized mechanisms suitable for construct-
On the other hand, an industrial mobile robot ing an omnidirectional mobile robot are reviewed
has been required to have some abilities that here.
it can freely travels in a confused factory and 2.1 Steered wheel mechanism
achieve an accurate positioning in a work sta-
tion. For this purpose, an active research is now Nakano et al. [1],[2]has already developed a
noholonomic-type omnidirectional mobile robot

0-7803-4316-6/98/$10.00 01998 EEE

51
 1998 Second Intemational Conference on Knowledge-Based Intelligent Elecmnic Systems,21-23 April 1998, Adelaide. Australia. Editors, LC.Jain and R.K.Jain

by using a driving wheel with steering, which 2.5 Orthogonal wheel mechanism
needs a preparatory motion for arranging the Pin and Killough [7] developed a holonomic and
wheel directions in a translational or rotational omnidirectional mobile robot with an orthogonal
motion. Nevertheless, this mobile robot has some wheel concept that provides normal traction in a
merits that the mechanism is simple, a specialized given direction while being free-wheeling in the
tire is not used, and there is no coupling between other perpendicular direction. Two possibly as-
the actuator for driving the wheel and the actu- sembly configurations of the wheels, labeled lon-
ator for steering [2]. gitudinal and lateral assemblies were discussed
and a prototype platform was also constructed
2.2 Universal wheel mechanism by using the former assembly, in which its cor-
responding control system was implemented in
In the development of holonomic and omni- a teleoperated mode or in a user-provided tra-
directional mobile robot, a universal wheel (or jectory mode. A holonomic and omnidirectional
Swedish wheel) concept [3] is well known for us. mobile robot was also constructed in Tang et al.
The wheel mechanism, which involves a large [8] by applying three lateral assemblies; however
wheel with many small rollers mounted on its its construction is slightly more complicated than
rim, can easily realize a holonomic and omnidirec- the longitudinal one, because it needs a gear or
tional motion, if such wheels more than three are transmission belt to synchronize two rounded-tire
mounted on a platform and allocated at different wheels.
directions. It is known that this type of wheels
suffers from the successive shocks caused when 2.6 Crawler mechanism
individual rollers make contact with the ground. West and Asada [9] and Nishikawa et al. [lo]
studied a holonomic and omnidirectional mobile
2.3 Omni-alpha mechanism robot called the Omnitrack by utilizing two rows
of full spheres as rolling units, in which the ball
Omni-alpha mechanism was developed to im- wheels or spherical tires are arranged in two con-
prove the remedy of the universal wheel mecha- veyer belts which produce forward and rotational
nism. The original omni-alpha wheel is a combi- motion of the platform. Two controlled rods in
nation of two plates, one of which has three bar- each track contact the top of the balls and, by
reled free-rollers. The three free-rollers of each rotating around an axis parallel to the track, pro-
plate are allocated with 120 [deg] so that the ex- vide sideways motion of the platform. Although
ternal form is consistent with the curvature of the mechanism provides holonomic and omnidi-
the wheel. Combining of two plates with 60 [deg] rectional motion, the rotational degree of freedom
shift, the outer envelope of six free rollers con- of the platform is extremely difficult to control be-
structs the external form of the wheel. Thus, cause, like all tracked or skid-steer vehicles, signif-
this wheel can make successive contact with the icant sideways slippage of part of the track must
ground, keeping a constant distance from a con- occur during turns. Note also that it may lead to
tact point. Asama et al. [4] developed an om- significant odometer and position tracking prob-
nidirectional mobile robot with three degree-of- lems, because some of the ball wheels can tem-
freedom decoupling drive mechanism and with a porarily lose contact with the ground on uneven
modified omni-alpha wheel in which one plate of terrain.
the wheel consists with four barreled free-rollers. In order to improve this shortcoming, Hirose
and Amano [ll]developed a special-type crawler
2.4 Ball wheel mechanism called the Vuton crawler, whose crawler has some
cylindrical free rollers. As a result, a holonomic
A nonholonomic mobile robot with two ball and omnidirectional mobile robot with four such
wheels was developed by Machida et al. [5]. This crawlers has no sideways slippage and can bear
nonholonomic vehicle has been shown to have an a relatively high payload. Hirano et al. [12] also
omnidirectionality with controlled rotational and developed a composite crawler with free rollers,
translational motion capabilities by using a par- so that a holonomic and omnidirectional mobile
ticular control system based on the concept of robot with four such crawlers can run on offload.
a chained form. However, note that this type However, note that this vehicle can not bear a
of robots can not achieve any three degree-of- high payload, because it has relatively narrower
freedom velocities at any time, because of its non- contact space with the ground, compared to the
holonomic property. On the other hand, West vehicle with Vuton crawler.
and Asada studied a holonomic omnidirectional
mobile robot using three ball wheels to achieve 2.7 Offset steered wheel mechanism
a full mobility and improved dead reckoning per- As suggested in above, the steered wheel mech-
formance [6]. anism of Nakano et al. [l],[2] requires a prepara-

52
1998 Second International Conference on Knowledge-BasedIntelligent Hecmnic Systems, 21-Z April 1998, Adelaide, Australia. Editors, L.C. Jain and R.KJain

z
tory motion for arranging the wheel directions
in a translational or rotational motion, because &o"edd
z

the steering axis is taken to be sideways (or per-

pendicular ) to the rolling direction of the wheel,
keeping a constant offset. In this steered wheel
with sideways offset, when rotating the steering Unconstrained
axis, the steering axis itself rotates with a circu-
lar motion around the contact point of the wheel,
due to the reaction of the driving torque. Then,
3
the translational velocity generated around the
steering axis is always parallel to the wheel, and
therefore the moving velocity of the steering axis
is consequently parallel to the wheel when the
wheel axis and the steering axis are simultane- Figure 1:The lateral orthogonal-wheel assembly
ously driven. Thus, the steered wheel with side-
ways offset can achieve the only one degree-of-
freedom motion.
On the contrary, a steered wheel mechanism
with forward offset was proposed by Wada et al.
[13], in which the steering axis is taken t o be for-
ward to the rolling direction of the wheel, keeping
a constant offset. In this mechanism, the trans-
lational velocity generated around the steering
axis is always perpendicular to the wheel, and
the translational velocity due to the wheel driv-
ing directs the direction of the wheel movement,
so that they always give the orthogonal velocities
to the steering axis. Thus, by controlling their ve-
locities, the moving velocity of the steering axis
Figure 2:The appearance of an omnidirectional mo-
bile robot
can be generated at any direction and with any
largeness.
cated with less than 90 [deg] offset.
3. OmnidirectionalMobile Robot In the following, it was assumed that a plat-
with orthogonal Wheels form is based on three assemblies allocated at an
In this section, it is assumed that the omnidi- equal distance from the center of gravity (c.g.) for
rectional mobile robot consists of the orthogonal- the robot, in which one assembly consists of two
wheel assembly mechanism proposed by Pin and wheels. The actual experimental mobile robot is
Killough [7]. appeared in Fig. 2.
The basic lateral orthogonal-wheel assembly is
illustrated in Fig. 1. The major components are 4. Derivation of Dynamic Model for
two spheres of equal diameter which have been the Omnidirectional Mobile Robot
sliced to resemble wide, i.e., rounded-tire wheels. Let the mobile robot be rigid moving on the
The axle of each wheel is perpendicular to the work space. It is assumed that the absolute coor-
sliced surfaces and is mounted using ball bearing dinate system 0, - XwYw is fixed on the plane
so that the wheel is freewheeling around its axle. and the moving coordinate system Om - XmYm
Through a bracket which is holding the extrem- is fixed on the c.g. for the mobile robot as shown
ities of the wheel axle, each wheel can be driven in Fig. 3.
to roll on its portion of spherical surface, rotat- When defining the position vector of the c.g.
ing around an axis 2,perpendicular to the wheel for the mobile robot such as S, = [z, p W l T , we
axle. When these axes 21 and 2 2 are maintained have
parallel and at a constant distance from each
M S , = F, (1)
other, and when the wheel rotations around these
axes are synchronized, contact with the ground where F , = [F, FYITis the force vector in the ab-
can be assured by at least one wheel, while al- solute coordinate system applied to the center of
lowing enough space for the brackets holding the gravity for the mobile robot and M is a symmet-
wheel axles to clear the ground. These assem- ric positive-definite matrix as M = diag(M, M )
blies can be constructed by using more than two with the mass M .
wheels, in which each wheel axle should be allo- Let 4 denote the angle between Xw-and X,-

53
1998 Second Intemational Conference on Knowledge-Based Intelligent Electronic Systems, 21-23 April 1998, Adelaide, Australia. Editors, LC.Jain and R.K.Jain

In addition, the driving system property [15] for

each assembly is assumed to be given by

Iw&i + cwi = kui - rDi (i = 1,2,3) (12)

where L is the distance between any assembly and

the c.g. of the robot; c is the viscous friction fac-
tor for the wheel; Di is the driving force for each
assembly; r is the radius of the wheel; Iw is the
moment of inertia of the wheel around the driv-
ing shaft; wi is the rotational rate of the wheel;
k is the driving gain factor; and ui is the driving
input torque.

0,
I Assembly 3

)XW
On the other hand,. the geometrical relation-
ships among variables 4, x,, y,, and wi, i.e., the
inverse kinematics are given by

Figure 3:Model of an omnidirectional mobile robot

coordinates, i.e., the rotational angle of the mov-

ing coordinate system with respect to the abso-
lute coordinate system. When introducing the
coordinate transformation matrix from the abso-
lute coordinate system to the moving coordinate Therefore, using equations (6)-( 15) gives
system such as
2
, = U I & + ~!,y,$ - bl(u1 +U:! - 2~3)(16)
- ukxm$ + f i b l ( u 1 - 742) (17)
"R - [ sin4 -sin41
cos4 (2) ji,
4
=
+ +
= U 3 4 b 2 ( u 1 + U 2 u3) (18)
it follows that
where
Sw = "&Sm (3)
a1 = -3C/(3Iw + 2Mr2)
F w = w&fm (4)
U; = 2Mr2/(31, +2Mr2)
where sm = [zm ymJT and f, = [fz fyIT are U 3 = -3cL2/(31,L2 +Id2)
the position vector of the c.g. and the force vec-
tor applied to the c.g. in the moving coordinate bl = kr/(31w + 2Mr2)
system, respectively. b2 = I C ~ L / ( ~ I , +
L ~I,,r2)
Therefore, transforming equation (1) to the
moving coordinate system gives It is easy to find that combining equations (2),
(3), (16) and (17) yields the appropriate dy-
M("R2W&Sm + Sm) = f m (5) namic equations in the absolute coordinate sys-
tem for the robot. Thus, defining the state vari-
Then, the dynamic properties for the mobile able for the robot as x = [z, y, 4 5 , yw
robot can be described as [14] 4]* , the manipulated variable as U = [ul 112 u3IT,
and the output variable as y = [&, iw$1' yields
M(zm - i m $ ) = f z (6) the following state equation [16]:
+
M(jim i n $ ) = f y (7)
1 ~ 4= M~ (8)
i = A ( x ) z+ B(z)u
y=cx
where I,, is the moment of inertia for the robot,
MI is the moment around the c.g. for the robot, where
and fzl f y l MI are given by
- 0 0 0 1 0 0
0 0 0 0 1 0
0 0 0 0 0 1
A(z) =
0 0 0 U1 -U24 0
0 0 0 U24 a1 0
- 0 0 0 0 0 U3

54
 1998 Second International Conference on Knowledge-Based Intelligent Electronic Systems, 21-UApril 1998.Adelaide, Australia. Editors, L.C.Jain and R.KJain

0 0 0
0 0 0
0 0 0
B(x)=
blPl b l P 2 2blCOS4
blP3 blp4 2blsin4
b2 b2 62

1
0 0 0 1 0 0
C= 0 0 0 0 1 0 where x; ,y; and $* are given.by adding PI-servo
0 0 1 0 0 0 or PD-servo to x w d , &,d and 4 d such that
= 1 - a; = 3IW/(3IW+ 2Mr2)
U2

~1 = -&sin 4 - cos4, p2 = &sin4 - cos4 x; = xwd + Kzpe;. + Kzi

I' ekdt (27)

~3 = &cos 4 - sin 4, ~4 = -&cos 4 - sin 4

Let the translational velocity of the robot in
the absolute coordinate system be V = (k: +
gG = ywd
$* = $d
+ Kype,j + Kyi
+ K$,kp + K+pe+
I' eydt (28)

(29)
y i ) $ and the azimuth of the robot in the absolute
+
coordinate system be 4 = 6' 4. Here, 6' denotes Here, K., is the proportional gain, K.i is the in-
tegral gain, K., is the derivative gain, and each
the angle between X,-coordinate and f,, i.e.,
the azimuth of the robot in the moving coordinate error is defined by
system. Then, it is found that
e;. = x w d - xw (30)
xw = v c o s + (21) ey = Ywd - Yw (31)
yw = VsinII, (22) e4 =4 d -4 (32)
4 = arctan - YW
(23) where xwd,iw, and 4 d denote the references for
XW
where note that a counterclockwise rotation de- x w , y w and 4, respectively. It should be noted
notes the positive direction for the rotational mo- that the solutions u i , i = 1,2,3 exit for all t , be-
tion of the robot. cause the input distribution factors b l and b2 are
It should be noted that Xw-and Y,-directional always non-zero constants.
motions in (19) are coupled, because the equation For a control result using this resolved acceler-
of motions is derived in the absolute coordinate ation control system, see [17].
system. However, since the rotational angle of
6. PID Control Method
the robot can be assured as 4 = - 8 even if the+
azimuth of the robot is changed arbitrarily, the The state and output equations (19) and (20)
mobile robot can realize a translational motion are described in the absolute coordinate system.
without changing the pose. It should be noted, however, that the control in-
put U is the quantity in the moving coordinate
5. Resolved Acceleration Control system. Therefore, the control input in the abso-
Since the dynamic equation is finally consid- lute coordinate system must be transformed into
ered in the absolute coordinate system, solving the control input for each assembly, i.e., signal
equation (19) with respect to ui, i = 1,2,3 gives expressed in the moving coordinate system, if the
the following resolved acceleration control: control input is designed in the absolute coordi-
nate system. Here, defining the control input in
U1 = Pl(.;: - aliw + a2diw) the moving coordinate as mu and one in the ab-
6bi solute coordinate as w u ,then the transformation
p3(Yw
"* - UlYw - a24.w) for each input can be derived from the kinematics
+ 6bi as follows
"U = J T W a (33)
(24) where J is the Jacobian matrix defined by

In the following, it is enough to only consider the

design of the control system in the absolute coor-
(25) dinate.

55
 1998 Second International Conference on Knowledge-Based Intelligent Electronic Systems, 21-U April 1998, Adelaide, Australia. Editors, L.C. lain and R.K.Jain

6.1 An adaptive PI controller for yw, where K,jc,ayl and a,j2 are some posi-
In this section, we show an adaptive PI con- tive constants. This control strategy is a kind
troller. Let uj.,u,j and u4 be the control in- of PI type controller with adaptive gain tuning,
puts using "i,"$ and 4, respectively and denote where P and I gains are adjusted automatically
W 21 = [uz,u,j,u4]. according to the error signals. By using this con-
In the conventional research, it is known 1181 trol strategy, when the error signal becomes large
that PD control is effective to control the rota- because of the plant parameter changing, the PI
tional angle, i.e., gains are tuned adaptively and look forward to
converging the error signal. The output response
u4 = K+peb K+ueb + (34) can be improved by setting the greater and
K,jc,but it requires a large initial input. On the
and P control is effective for controlling each ve- other hands, large aj.1, a i l in the P gain and
locit y, 0 5 2 , "$2 in the I gain may contribute to the im-
uj. = Kkpej., u,j = K,jpejr provement of rise time and offset removing with-
But the concrete method does not exist for the out causing a large initial input.
selection of the 'P' gains, KhP and Kjlp. Thus, a For the rotational angle control, we also use the
trial and error procedure is required until a de- conventional PD control as shown in (34).
sired performance will be obtained. Moreover, if
the system parameters are changed due to mount- 7. Fuzzy Model Method
ing a weight on the robot or changing of the phys- For the fuzzy servo, a rough construction pro-
ical parameters, then the design parameters must cedure of fuzzy models is as follows:
be selected again. 1. Substitute a representative angle 4 into
On the other hands, a simple output feedback Eq.(19) and construct a linear model.
strategy 2. Discretize the linear model.
3. Construct the augmented system.
u(t)= K(t)e(t), k(t)= e 2 ( t ) 4. Design an optimal regulator for the aug-
is known [19],where U is the scalar input, e is the mented system and determine the feedback
scalar error, and K is a scalar time-varying gain. gain matrix by solving the associated Ric-
This control strategy is a kind of P type con- cati equation using any numerical approach.
troller with an adaptive gain tuning, where P gain Here, we prepare four kind of models within the
is adjusted automatically according to the error range of -27r 5 4 5 27r. The value of 4 at each
signal. This control strategy is very attractive, model rule is given by
but a steady state error may remain, if we only
Model rule 1 :4 = 0 or 27r or - 27r
use the P-type controller. In a constant output
feedback problem, this can be easily overcome by Model rule 2 :4 = 7r/2 or - 3 ~ / 2
introducing an integral term to the control law. Model rule 3 :4 = 7r or - 7r
In the paper of [20], the following adaptive PI Model rule 4 :q5 = 37r/2 or - n/2
controller for iw was adopted,
7.1 Linear fuzzy models
Since the representative angle is assumed to be
constant in each model rule, the system matrix of
linear model is same for all of model rules such as
with
+
Kj.l(t) = K i p @ ) akiIX'j.i(t)
Kj.z(t)= &52Kj.i(t)
~ ~ j .= t )( t ) , &i(t) = e:(t)
~ e( %
where KkC,a i l , and aj.2 are some positive con-
stants. Similarly, The distribution matrix B,i is provided by

r 1
+ I' K,j2(7)e,j(~)dr
1
(36)

56
1998Second Intemational Conference on Knowledge-Based Intelligent Electronic System, 21-23 April 1998, Adelaide, Australia. Editors, L.C. Jain and R.KJain

- -
r 1 73 -73 0
71 71 -72
74 74 74
r4 = 77 -77 0
75 75 -76
- 78 78 78 -

where the subscript i of Bci denotes the number

of model rule.
7.2 Discrete-time models 7.3 Construction of augmented system
The linear model in the discrete-time is given If the type-2 servo system is considered here,
by using the polynomial equation with the time tran-
sition operator z described by

f ( z - 1 ) = (1- 2 - ) - 1 - 22-1 + 2-2 (39)

where it was assumed that @ and I'i were calcu- yields

lated by using the physical parameter conditions f(.-l)Yd(k) =0 (40)
given by for the reference output vector y,(k). Here, 2-l
denotes the one-step delay operator, and the ref-
I,, = 11.25 Fgm2], M = 9.4 Fg] erence error vector e ( k ) is defined by
L = 0.178 [m], k = 0.448, c = 0.1889 [kgm2/s]
1, = 0.02108 [kgm2], T = 0.0245 [m] e ( k ) = Y d k ) - Y(k) (41)
Operating f ( z - l ) to the discrete-time model (37)
and using the sampling width of 50 [ms]. The
and the above reference error vector, and defining
results are as follows:
v ( k ) = [ f ( z - l ) z ( k ) e ( k - 1) e(k - 2)JT as the
0.0416 0 0 state variable vector for the augmented system,
13x3 O 0.0416 O we have

@=
0 0 0.0475
0.684 0 0
03x3 O 0.684 0 where
0 0 0.903

-71
-73
72
0
- @e= [ -c @

03x3
06x3
2I3x3
13x3
06x3
-13x3
03x3 1
74 74

[
-75 76
-77 0 rei = 03x3

78 78 - 0"3]

- 7.4 Design of optimal regulator

-73 73 0 After selecting the suitable weighting matrices
-71 -71 72 Q i and R i for each model rule, solve the posi-
r2= "14 74 "14 tive definite matrix Pi to the associated algebraic
-77 77 0 Riccati equation given by
-75 -75 76
78 78 78 - Pi = Qi +@:Pi@ e - Tpirei(R,
- +r: pire)- r$pi@e (43)
71 71 72
-73 73 0 It should be noted, however, that the augmented
system does not assure the controllability. That
r3 = 74 74 74
is, it is difficult to obtain a unique stabilizable
75 75 -76
-77 77 0 solution Pi > 0. But, since there exist uncontrol-
78 78 78 lable modes on the unit circle [21], it is possible

57
 1998 Second International Conference on Knowledge-Based Intelligent ElectronicSystem, 21-23 April 1998, Adelaide, Australia. Editors, LC.Jain and R.KJain

to obtain Pi 2 0 as a strong solution [22] by ap- the manipulated quantity can be consequently ex-
plying an iterative approach. The strong solution pressed [23] by
Pi is obtained as a steady-state solution of the 4
following iterative relation:
u(k)= ui(k)
Pili + 11 = Qi + @TPib]cPe- cPTPiIj]rei(Ri i=l
r 4 i r 4 1
+r: h]re)-'r: Ij]ae (44)
with initial condition of p i [ O ] = 0. Anyway, if Pi r 4 1
is calculated, then the optimal feedback law can
be realized as (53)

f(z-')ui(k)= -Giv(k) (45) The corresponding simulation results are shown

in [24].
Here,
8. Stochastic Fuzzy Servo Method
Gj = [Ri + l$PJei]-' I':Pi@,
It is assumed that the controlled object is com-
= [Fi Ki Li] (46) pletely known, linear, and dynamic system de-
scribed by the following model:
7.5 Construction of fuzzy servo system
Since the actual manipulated input to the robot z(lc+ 1) = oz(lc)+ ru(q + w(k) (54)
is ui(k), it is necessary to arrange Eq.(45). There-
fore, introducing an intermediate variable ((k) y(k) = C Z ( k ) w ( k ) + (55)
that satisfies where z ( k ) E Rn is the state variable vector,
u(k)E RP is the control input vector, w(k) E Rn
is the system noise vector, v ( k ) E Rn is the
measurement noise vector, and y(k) E R" is
gives the following state equation: the measurement vector. Here, it is assumed
that { w (k), (k)} are independent each other for
k = 0,1, ... and zero-mean white Gaussian noises
with covariance matrices Q and R. The initial

+[ ] e(k) state vector z(0) is assumed to be subject to the

-
13x3 normal distribution with its mean and covari-
03x3
ance PO,i.e., z(0) N ( Z 0 ,Po), and be indepen-
dent of { w ( k ) , v ( k ) }In
. addition, let the data
Dividing the both sides of Eq.(45) by f(2-l) and A
using Eq.(47) in it, we have set of y(k) be described by Yk = {y(O), ...,y(k)}
and similarly the data set of u(k)be denoted by
Ui(k)= -&z(k)-Ki<(k-l)-L&(k-2) (49) A
Uk = {u(o),...,u(k)}.
where it is assumed that €(-2) = €(-1) = 0. 8.1 Multiple-model optimal servo system
The membership function is of a Gaussian type Under the assumption that in equations (54)
with the center value w,, the reciprocal value of and (55) the distribution matrix r and the ini-
standard deviation Wd,and the input variable x, tial state z(0) are uncertain, consider a multiple
whose confidence p is given by model adaptive control which treats with the can-
didate models as the hypotheses. Let the hypoth-
p = exp {ln(0.5)(z- W,)~W;} (50) esis Hi be described by
Since there exist two or three multiple labels in Hi : ~ ( +k1) = h ( k ) + I'+(k) + w(k) (56)
one model rule, the confidence of i-th model rule
in the antecedent part is assumed to be given by +
y(k) = C z ( k ) w(k),i = 1,...,M (57)
where M is the number of candidate models
pi = m y { p i j )

where j denotes the number of fuzzy labels in the

(51)
-
and the initial state is assumed to be subject to
z(0) N ( z ; , P ; ) .
The multiple model adaptive control in the
i-th model rule. When defining the normalized steady-state solution satisfying the above hy-
confidence given by potheses can be obtained [22] by
A M
u(k)=- F&j(k)P(Hilk)
i=l

58
1998 Second International Conference on Knowledge-Based IntelligentElectronic System, 21-23 April 1998, Adelaide, Australia. Editors, L.C. Jain and R.K.Jain

M
fact, as can be seen from equation (69), Si are
- Hizi(k - l)~(HiIk) (58)
all equal for i = 1, ...,M ) , as in the conventional
i= 1
stochastic fuzzy control [25]. Then, the multiple
From the Kalman filter theory, the elemental model adaptive control in steady-state is reduced
n to
state estimate &(k) = E[~(k)lHi,Yk-l,Uk-l]
can be obtained as follows: M M
U ( k )= - C F i & i ( k ) P ( i l k-
) CHiri(k)p(ilk)
i= 1 i=l
(70)
where the a posteriori probability p ( i l k ) for the
ith membership function (or ith control rule) can
be written by

Here, p(ilk-1) denotes the a priori probability for

the ith membership function, and the confidence
of data y ( k ) with respect to the label Ai is given
by

In this case, note that the membership function

moves on the support set with a fixed width, be-
cause {Gi(k)} can change in time. In addition,
if the a priori probability is not assigned to the
membership function (or control rule) as in the
case of a conventional fuzzy control, then the a
posteriori probability can be expressed by

where Several simulation examples are given in [26].

9. Conclusions
In this paper, we have reviewed several special-
Moreover, the a posteriori probability for the hy- ized wheel or mobile mechanisms for construct-
pothesis Hi,p ( H i l k ) , is calculated by ing mainly a holonomic and omnidirectional mo-
bile robot. Then, we concentrated our atten-
tion to a particular omnidirectional mobile robot,
whose platform was composed of the three lateral
orthogonal-wheel assemblies. After introducing
Pi (Y(k)) the dynamic model for such a mobile robot, we
1
=(~T)-”~Is~ I-lexp[-Sllv(k) - ~ ~ ( k ) l l ~ discussed
~~] on the four control approaches to de-
velop a feedback control system.
(68)
in which pi(y(k)) denotes the probability density
References
function of the data y(k) under the hypothesis Nakano, E. and Koyachi, N., “An Advanced
Hi,and Mechanism of the Omni-Directional Vehicle
Si = CPCT i- R (ODV) and Its Application to the Working
(69) Wheelchair for the Disabled,” Proc. of ’83
8.2 Stochastic fuzzy servo system Int. Conf. Advanced Robotics, 1983, pp. 277-
For the multiple model optimal servo system 284.
Nakano, E., Mori, Y., and Takahashi, T.,
using some linear dynamic models, let the hy- “Study of the Mechanism and Control of
pothesis of Hi,i = 1,...,M be regarded as the Omni-Directional Vehicle,” Procs. of the
fuzzy label Ai,i = 1,..., M for one vector data 12th Annual Conference of RSJ, Fukuoka,
y(k). Assume that the distributions of y(k) un- Japan, Nov. 1994, vol. 1, pp. 369-379 (in
der the hypothesis Hi are all equal forcedly (in Japanese).

59
1998 Second International Conference on Knowledge-Based Intelligent Elecmnic Systems,21-23 April 1998, Adelaide, Australia. Editors, LC. Jain and R.K.Jain

Muir, P.F. and Neuman, C.P., “Kinematic (SICE), vol. 27, no. 1, pp. 70-76, 1991 (in
Modeling of Wheeled Mobile Robots,” J. Japanese).
Robotic Systems, vol. 4, pp. 281-340, 1987. Saito, M. and Tsumura, T., “Collision
Asama, H., Sato, M., Bongoni, L., Kaetsu, Avoidance Among Multiple Mobile Robots -
H., Matsumoto, A., and Endo, I., (‘Develop- A Local Approach Based on Non-linear Pro-
ment of an Omni-Directional Mobile Robot gramming -,” Trans. of the Institute of Sys-
with 3 DOF Decoupling Drive Mechanism,” tems, Control and Information Engineers,
vol. 3, no. 8, pp. 252-260, 1990 (in Japanese .
Procs. of 1995 IEEE Int. Conf. on Robotics
and Automation, Nagoya, Japan, May 1995, Watanabe, K., Shiraishi, Y., Tang,
Fukuda, T., and Tzafestas, S.G., “Con-
2,
vol. 2, pp. 1925-1930.
Machida, K., Luo, Z.-H., and Funaki, M., struction of the Dynamic Model for an
“Prototyping and Control Experiments of a Omnidirectional Mobile Robot ,” Procs. of
Cartesian Nonholonomic Robot,” J. of the CESA ’96 IMA CS Multiconference, Sympo-
Robotics Society of Japan, vol. 15, no. 2, pp. sium on Robotics and Cybernetics, Lille,
243-248, 1997 (in Japanese). France, July 1996, pp. 643-647.
West, M. and Asada, H., “Design and Con- Watanabe, K., Shiraishi, Y., and Tang, J.,
trol of Ball Wheel Omnidirectional Vehi- “A velocity type of computed torque control
c l e ~ ,Procs.
~ ~ of 1995 IEEE Int. Conf. on system for an omnidirectional mobile robot ,”
Robotics and Automation, Nagoya, Japan, Proc. of the 15th SICE Kyushu Bmnch An-
May 1995, vol. 2, pp. 1931-1938. nual Conference, Saga, Japan, Nov. 1996,
Pin, F.G. and Killough, S.M., “A New pp. 185-188 (in Japanese).
Family of Omnidirectional and Holonomic Watanabe, K. and Shiraishi, Y., “A PID
Wheeled Platforms for Mobile Robots,” control system for an omnidirectional mo-
IEEE Tmns. on Robotics and Automation, bile robot,” Procs. of the 35th SICE An-
vol. 10, no. 4, pp. 480-489, 1994. nual Conference Domestic Session Papers,
Tnag, J., Watanabe, K., and Shiraishi, Tottori, Japan, July 1996, pp. 167-177 (in
Y., “Design of Traveling Experiment of an Japanese).
Ilchmann, A., Non-identijier-based hzgh-gain
Omnidirectional Holonomic Mobile Robot,”
Procs. of 1996 IEEE/RSJ Int. Conf. on In- adaptive control, Lecture Notes in Control
telligent Robots and Systems (IROS961, Os- and Information Science, no. 189. Berlin,
Springer-Verlag, 1993.
aka, Japan, Nov. 1996, vol. 1, pp. 66-73. Sato, K. and Watanabe, K., “Robust PI
West, M. and Asada, H., “Design of a Holo- Control System for an Omnidirectional Mo-
nomic Omnidirectional Vehicle,” Procs. of bile Robot with Three Orthogonal-Wheel
1992 IEEE Int. Conf. on Robotics and Au- Assemblies,” Proc. of JSME Kyushu Bmnch
tomation, Nice, France, May 1992, vol. 1,pp. Memorial Conference, Fukuoka, Japan, Oct.
97-103. 1997, pp. 132-134 (in Japanese).
Nishikawa, A., West, M., and Asada, H., Watanabe, K., Tang, J., and Han, F., “A
“Development of a Holonomic Omnidirec- Rotational Control for an Omnidirectional
tional Vehicle and an Accurate Guidance Mobile Robot Using a Fuzzy Modeling Ap-
Method of the Vehicles,” J. of the Robotics proach,” Procs. of the 14th Annual Confer-
Society of Japan, vol. 13, no. 2, pp. 249-256, ence of RSJ, Niigata, Japan, Nov. 1996, no.
1995 (in Japanese). 1, pp. 53-54 111 Japanese).
Hirose, S. and Amano, S., “The VUTON: Watanabe, J i., Adaptive Estimation and
High Payload, High Efficiency Holonomic Control, Hertfordshire, UK, Prentice-Hall,
Omni-Directional Vehicle,” Proc. of JSME 199’1:
--V-.
Annual Conference on Robotics and Mecha- [23] K. Tanaka, Advanced Fuzzy Control, Tokyo,
tronics, Sapporo, Japan, July 1993, pp. 350- Kyoritsu-syuppan, pp. 95-192, 1994 (in
355 (in Japanese). Japanese).
Hirano, T., Chen, P., and Toyota, T., “A [24] Watanabe, K., Izumi, K., Tang, J., and Han,
Fundamental Study on Omni-Directional- F., “Trajectory Control of an Omnidirec-
Vehicle (1) - Omni-Directional Running tional Mobile Robot Using a Fuzzy Servo
Mechanism for Offroad -,” Proc. of the 14th System,” Proc. of the 2nd Znt. Workshop
SICE Kyushu Branch Annual Conference, on Advanced Mechatronics (IWAM’971, Un-
Kitakyushu, Japan, Nov. 1995, pp. 273-274 zenn, Japan, Dec. 1997, pp. 74-79.
(in Japanese). 1251 Watanabe, K., “Stochastic Fuzzy Control
Wada, M. and Mori, S., “Development of (Theoretical Derivation) ,” JSME Int. Jour-
a Holonomic and Omnidirectional Mobile nal, vol. 40, no. 2(C), June, pp. 224-230,
Robot,” J. of the Robotics Society of Japan, 1997.
vol. 15, no. 8, pp. 1139-1146, 1997 (in [26] Watanabe, K., Izumi, K., and Han F.,
Japanese). “Stochastic Fuzzy Servo Control Using Mul-
Iwatsuki, M. Nakano, K., and Ohuchi, T., tiple Linear Dynamic Models, Procs. of
“Target Point Tracking Control of Robot Ve- 2nd Int. Conf. on Knowledge-based Intelli-
hicle by Fuzzy Reasoning,” Tmns. of the So- gent Electronic Systems, Adelade, Australia,
ciety of Instrument and Control Engineers April 1998, to be published.

60