Hiroyuki Kajiwara, Pierre Apkarian, and Pascal Gahinet

ain-scheduling control structures have proved useful in

many practical applications. As an example, most aircraft
control laws are based on the interpolation of individually de-
signed controllers or make use of some ad-hoc gain-switching
policy. Similarly, in robot control problems the controller dy-
namics are adjusted in real-time according to geometry and
inertias. However, in spite of numerous successful applications,
the construction of the overall control structure invariably calls
for the engineering insights of the designer and, more critically,
the resulting control laws do not provide any guarantees in the
face of rapid changes in the scheduled variables. These difficul-
ties have been the main motivation for the development of mod-
ern gain-scheduling control techniques, and have led to some
challenging research in the area of the analysis and synthesis of
LPV systems. Such systems are described in state-space form as
X =A(O)x+B(O)u,
y = C ( 0)x + D( O)U, (1)
where e:= e ( t ) is a time-varying parameter describing the range
of possible dynamics of the plant. Such systems are natural ex-
tensions of customary LTI systems.
Briefly speaking, the recently available LPV synthesis tech-
niques allow the construction of the global control law as a whole
entity for all admissible 0, that is, without requiring unnatural
separated design syntheses. They furthermore provide theoreti-
cal guarantees in terms of both stability and performance in the Fig. 1. Arm-driven inverted pendulum (ADIP).
presence of fast time-domain evolutions of the scheduled vari-
ables. Note also that these synthesis techniques reduce to solving e the so-called LFT (Linear Fractional Transformation) con-
a finite set of LMIs (Linear Matrix Inequality), which are easily trol techniques, which are well suited when the model
solved using currently available LMI codes. state-space coefficients are fractional functions of a vary-
In this work, we are considering the challenging application ing parameter (possibly a state), and
of an arm-driven inverted pendulum (ADIP) as depicted in Fig. polytopic techniques, which assume that the state-space co-
1. The ADIP was originally designed by Dr. Y. Nishi for training efficients evolve in a prescribed (bounded) polytope (as for
purposes at Kawasaki Heavy Industry in Japan. Here the pendu- instance a hypercube).
lum is the top link and is driven by the rotated arm (bottom link), The intention behind these different representations is to ac-
instead of a more classical cart. As the arm is rotated and gets count for the time-varying and nonlinear nature of the plant by
closer to the horizontal position, the horizontal motion of the arm formalizing a set of admissible trajectories instead of a single op-
tip becomes more limited and the inertias viewed from the arm erating point.
are modified. This naturally leads to the design of controllers that Our aim is to completely validate these techniques on the
adjust in real-time to the rotation of the arm. For this purpose, physical experiment. Therefore, in addition to the usual stability,
two kinds of LPV synthesis techniques are investigated: performance, and robustness requirements, we shall also be con-
cerned by the implementation constraints that inevitably show
up in real-world applications. The major implementation con-
straints are the following.
Hiroyuki Kajiwara (kajiwara@ces.kyutech.ac.jp)is with the Kyushu The high-frequency gain of the controllers must be compat-
Institute of Technology, Kawazu, Japan. Pierre Apkarian is with ible with the actuator bandwidth.
ONERA-CERZ Toulouse, France. Pascal Gahinet is with The Math- The controller dynamics must be consistent with the avail-
Works, Inc., Natick, MA. able sampling rate in this application (I1kHz.).

44 0272- 1708/991$10.0001999IEEE IEEE Control Systems

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on May 27,2024 at 10:00:14 UTC from IEEE Xplore. Restrictions apply.
 Note that the second of these constraints is especially difficult where
to handle, as it concerns the internal properties of the controller
and cannot be directly treated through the properties of the
closed-loop system. We shall see, however, that a suitable multi-
objective extension of the polytopic technique provides an effec-
tive means to overcome this difficulty. Another way to handle and
this implementation constraint would have been to tune the
weighting functions until adequate closed-loop plant and open-
loop controller specifications are met. However, this procedure
revealed itself to be intractable in this application and in most in-
stances led to high order controllers, as it requires complicated
weighting functions.
Another important issue is to evaluate the benefits of LPV
synthesis techniques in regard to classical robust control tech-
niques such as H _ andp syntheses. It turns out in this application 4
M,:= -m& , R = 2m2!,!,
that though the H _ and p controllers are capable of providing 3
some stability guarantees they are, as expected, outperformed by
LPV controllers at the performance level.
This article is organized as follows. The next section de-
scribes the modeling of the ADIP and introduces the problem
specifications. A brief review of the LPV synthesis techniques
used for the ADIP is given in the following section. The full
design procedure up to the nonlinear simulations and real ex-
periment results are then presented, followed by some con-
cluding remarks.
All LMI-related computations in the application were per-
formed using the LMI Control Toolbox [23], p controllers were
designed using the p-Analysis and Synthesis Toolbox [22], the
nonlinear simulations were obtained using MATLABI~IMULINK In this applicatiion,the first joint is actuated and the second joint
facilities and LPV controllers were implemented using the REAL- is free. From (2)-(9), we can derive the following equations.
TIMEWORKSHOP.
(MI -M,)cp, + Rcos(cp, -cp,)cp,
LPV Modeling of the ADIP +Rsin(cp, -cp,)$ + ( m , + 2m,)!,gsin(cp,)
In this section, the LPV synthesis model for the ADIP is de-
veloped and the design specification? are introduced. Consider -m,t,gsin(cp,)=%, , (10)
the two-link arm depicted in Fig. 2. It is well known [l], [ 2 ] that
the motion equation is described as
M ( q ) q + C ( q ,q ) + G ( q )= 7, (2)
The main control objective is to maintain the second arm in a ver-
tical position like an inverted pendulum using the rotation of the
(first) actuated arm. In the following, the first arm and the second
arm are called arm and pendulum, respectively. In the physical
experiment corresponding to Fig. 2 the arm is actuated by a mo-
tor driven by a velocity-control power amplifier. The physical
quantities are given as follows:
,
! = 0.13 m, !, = 0.15 m, m, = 0.05 kg,
m2 = 0.03 kg, g = 9.8 m l s 2 . (12)
As the velocity ip, of the first arm can follow the command in-
put voltage U to the amplifier because of the lightness of the sec-
ond arm, we cain assume that the dynamics from the input voltage
to the velocity ip, is almost equally given by
d .
-cp
1 .
=--cp,+-LL
ICu
dt I T , c
This means thait (10) can be simplified to (13) with reasonable ac-
curacy.
Fig. 2. For modeling of A D P On the other hand, (1 1) becomes

February 1999 45

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on May 27,2024 at 10:00:14 UTC from IEEE Xplore. Restrictions apply.
 4 we get
2& cos('pl - cpz I@,+ - 1z@, = g sin(9,) + 21 sin(cp, - cp,)$.
3
(14)
Then, defining
Gathering the equation = r (p with (13) and (19) the fol-
rx :=2t,sin(cp,), ry:=2&,cos(cp,), (15) y.1
lowing simple LPV model is obtained
and using (14), another description for the ADIP is as follows:
0 1 0 0
0 0 0 0
0 0 0 0
The pendulum has two kinds of equilibrium states: 0 0 0 - +
unstable equilibrium state: cp; = 0 (tip pointing upwards)
stable equilibrium state: cpi = n (tip pointing downwards).
P1
In this application, we will only consider the difficult situation
where the tip is pointing upwards. An immediate linearization of
(1 6) around 'pi = 0 then leads to
4
i--+ -1, cp, = ( g + I;)(P,.
3

Introducing the new variable z defined as

4
z:=r, + -&,cp2,
3 where both z and y, are assumed to be measured, ry is viewed as
an external time-varying parameter, and Y, is assumed to be zero
because its value is negligible in face of g around the vertical line
(cp, < 60"). Moreover, as we have already experimented, sched-
uling fydoes not bring any improvement in the design and our as-
sumption substantially simplifies the derivation hereafter.
In order to derive LPV models with bounds on the time-
varying parameters, it is assumed that the arm can rotate within
the angular range

This yields upper (<) and lower (7,) for the range of variation
-
of ry:

-
<I := [2e I cos(@,
ry E [ry, COS(O)].
(22)
From (22), ry is normalized as
I - ,

Fig. 3. Wide range stabilization for ADIE

Two different-though completely equivalent-LPV repre-
sentations can be used for the ADIP. This is described in the
sequel.

LPV model with LFT structure

In LFT representations, the parameter or state dependence of
the plant state-space matrices is reformulated as an outer loop
closed on a nominal plant, which involves the parameter in multi-
plicative form. Representing ry from (23) as
- -
r
r, + ry ry -ry
y 2 2 er

Fig. 4. Interconnection structure for ADIP. = t,(1+cos(cp,))+ &,(1

-cos(cp,))er, (24)

46 IEEE Control Systems

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on May 27,2024 at 10:00:14 UTC from IEEE Xplore. Restrictions apply.
 where 0, denotes the new normalized scheduling variables, the difficulty of this problem comes from the fact that the ADIP be-
following LPV model of LFT type is obtained: comes uncontrollable as the pendulum gets closer to the horizontal
position, hence the need for a gain-scheduled controller. Apart
0 1 0 0 - ,
from the maximization of the range where stabilization holds, we
must also proviide performance in terms of settling-timeand over-
shoot in response to reference signals. This will be detailed in a
later section. Also the controller should exhibit adequate roll-off
in the high-frequency range for noise attenuation.

LPV Control Techniques

This section provides a brief review of the design techniques
that will be used for the ADIP application. Two LPV design tech-
(25) niques will be investigated:
the LFT design technique,
~i the polytopic design technique.
Z
As indicated by their names, such techniques apply to LPV
Z plants with LFT and polytopic parameter-dependence, respec-
z , = [ O 0 0 11
tively. The first class of LPV plants can be described as

with

LPV Model with Polytopic Structure

In polytopic descriptions, the matrix coefficients are ex-
pressed as convex combinations of their extreme values. Hence,
similarly to the above, by remarking that

and introducing the notation

Fig. 5. Reference signal r to r,.

we obtain the following LPV polytopic model for the ADIP:

,
and it is easily verified that p and p2are polytopic coordinates,
thatis,p, >Oandp, >Oandp, +p2 =l.

Quick Look at the Wide-Range

Stabilization Problem
As already stated, the main control objectivefor the ADIP is to
stabilize the inverted pendulum using the rotations of the arm, as
depicted in Fig. 3 below, and simultaneouslyincrease (as much as
possible) the range where stabilizationis achieved. One important Fig. 6. LMI region for closed-loop system.

February 1999 47

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on May 27,2024 at 10:00:14 UTC from IEEE Xplore. Restrictions apply.
 x = Ax + B,w, + B,w + B,u,
z, = C,x + Doewe+ D,,w + D,,u,
z = C , x + D,,w, + D,,w + D,,u,
y = C,x + D,,w, + D,,w + D Z 2 u ,
w , = o ( t ) z, ,
(31)
where @(t)is a time-varying parameter matrix
and is usually assumed to have a block-diagonal
structure in the form

o ( t ) = diag(...,@ , ( t ,...,
) e,(t)I ,... ), (32)
and is normalized such that

o ( t ) ' o ( t ) 5 I , t 2 0. (33)
Blocks denoted 0, and 8, I are generally re-
ferred to as full and repeated-scalar blocks ac-
cording to the p analysis and synthesis literature
[ 3 ] ,[4]. Note that straightforward computations

[;]=I[;
lead to the state-space representation

A B, ;;j+[;;jw
$1 -DeeO(t))-l[CeDe1 0821

:1.1
hence the plant with inputs w and u and outputs z
,

(34)

and y has state-space data entries which are frac-

tional functions of the time-varying parameter
0 ( t ). Hereafter, we are using the following nota-
tion
0 U for the control signal

w for exogenous inputs

e z for controlled or performance variables
y for the measurement signal.
As an alternative to this description, we are
also considering polytopic systems which are de-
scribed by the state-space representation

x = A ( p ( t ) ) x+ B , ( p ( t ) ) w+ B*(p(t))u I

z = Ci(P(t))x+ Di,(P(t))w+ D,,(p(t))u,

Y = C,(P(t))x + D z i ( ~ ( t ) ) +
w D,,(p(t))u,
(35)
where generally, A(p), B,(p),...are affine func-
tions of the time-varying parameter p(t) evolving
in a polytopic set Po,i.e.,

{ p pv,1, t 2 0 ,
p ( t ) ~ P ~ : = c o,,,,..., (36)
where the notation CO{ .} stands for the convex
hull of the set { .}.
Clearly, the state-space data of the plant (35)
range over a matrix polytope and thus it trivially
Fig. 7. H , Control results (y = 2.5114, (plc = 45") holds that

48 IEEE Control Systems

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on May 27,2024 at 10:00:14 UTC from IEEE Xplore. Restrictions apply.
 i 1
B,(P(t)) B,(p(t))
A(P(t))
Ci(P(t)) Dii(P(t)) Dl,(p(t)) E

C,(P(t)) D,,(P(t)) D,,(p(t))

:=col[i: :::i], Bl I

D,,,
DZI, Q2i
i =1,2,...,r } ,

(37)
where

i‘~~q
czi
t;:;

D2I1 D22,
-4l0 5 10 15 20

I
:=
1
W , : ) B,(P,,,)B,(P,>,)
CAP”, 1 D,,(P”,1 Dl,(P”, 1
C,(PY, 1 D*,(P”,1 D,,(P”, 1
i =1, ...,r .
3

(38)
Time [sec]

For any of the LPV plants (31)-(33) or (35)-

(37), the LPV control problem (often referred to
as the gain-scheduling control problem) consists
in seeking an LPV controller
X K = A,(P)XK -tWP ) Y>
= C K ( ~ ) x+KD,(P)Y I
(39) -1.5
0 5 10 15 20
where p ( t ) = @(t) for the LFT-LPV plant (31)- Time [sec]
(33) and p(t) =p(t) for the polytopic LPV plant
(35)-(37) such that
the closed-loop system (3 1)-(33) and (39)
or the closed-loop system (35)-(37) and 100
(39) is internally stable,
0 the L2-induced gain of the operator con- 50
necting w to z is bounded by y,
for all parameter trajectoriesp(t)defined by either 0
(33) or (37).
It is now well-known that such problems can
be handled via a suitable generalization of the -50
Bounded Real Lemma and can be solved by
computing solutions to a set of LMIs. It is im- 100
portant to note that LMI problems are convex 5 10 15 20
and thus one can easily compute a solution to Time [sec]
these problems (whenever it exists) by using
very efficient codes of semi-definite program-
ming [14]-[17].

Controller Implementation Constraints

Using a more general approach, it is possible
to characterize multiple and multi-channel speci-
fications in the case of polytopic LPV systems
[lo], [19], [9]. Since it has proven useful in the
ADIP control problem, we briefly discuss the
LMIs characterizing L,-gain performance in
conjunction with LMI region constraints on the
poles of the closed-loop system. The reader is re-
ferred to [19] and [20]for a thorough discussion
on LMI regions and their use in robust control
theory and to the web reference [24] for a de- Fig. 8.1-1Control results (y = 0.7938 (pic = 45”).

February 1999 49

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on May 27,2024 at 10:00:14 UTC from IEEE Xplore. Restrictions apply.
 tailed description of the LMI characterizations involved. LMI re-
gions are self-conjugate regions of the complex plane which are
useful to specify transient constraints for the closed-loop dynam- where CO, is a design parameter. The family of test or reference
signals r for the arm rotation in the range (21) which are used
ics. Interesting examples of such regions include vertical and
hereafter can be described as follows.
horizontal strips, circles, conic sectors, etc., and their intersec-
tions thereof which again are LMI regions (see Fig. 6 for an illus-
tration). Moreover, any self-conjugate region of the complex
plane can be approximated to any desired accuracy with an LMI
region.
Another practical advantage of LMI region constraints is that
they can be exploited to indirectly enforce constraints on the con-
troller dynamics and thus on the controller bandwidth. This is of
primary importance in any realistic design. We stress here that ,
where q,,c,x:= 21 sin(cp,,), z = 0 5 and cp,,(5 Cp,) is a target value.
such an approach is only indirect and might require preliminary A simple drawing of such signals is depicted in Fig. 5. Then the
adjustments and trials to translate original constraints on the con- specifications for the control system are the following:
troller into constraints on the closed-loop dynamics. These ad-
justments are useful to avoid unnecessary restrictions on the (Sl) the closed-loop system is internally stable.
design problem. From a theoretical viewpoint, achieving con-
troller bandwidth specifications remains a delicate and unsolved (S2) the L,-induced gain of the operator connecting w = Y to
question, except for some classes of fixed-structure controllers
such as PID and the like. In the context of this paper, it is enough
to have in mind that LMI region constraints provide a sensible
though not direct approach to this problem. A more abstract jus-
z= [ is bounded by y, where CO,
is a design parameter.

(S3) specifications (Sl) and (S2) must hold on the largest

tification can be found in [ 191. range of (p, as far as possible.
This approach extends naturally to polytopic plants in (35)
but does not have yet any counterpart for LFT plants (31) (see (S4) the LPV controller must be implementable with a mini
[24]). Also important is the fact that pole constraints must be un- mum sampling interval 1 msec.
derstood in the time-invariant sense, that is, for “frozen” values
of the parameter in its range. Obviously, these concepts must be The specification (Sl) means that the LPV controller must
handled with care when one manipulates the time-varying plants stabilize the inverted pendulum in any vertical position in the
described previously. In the ADIP application, stability and per-
formance are time-variant specifications whereas LMI region
constraints are classical time-invariant pole constraints. It is
shown that the formulation is powerful enough to meet the im-
plementation constraints of our application.
When solutions to the LMI conditions have been found, the
state-space data of the controller in (39) can be computed for any
admissible value of the parameter @(t)or p(t), using well-known
schemes. The reader is referred to [181 and [21], and to [71-[91,
for a comprehensive discussion.

LPV Controller Synthesis for the ADIP

In this section, we first give a thorough descnption of the
specifications and control objectives together with the control
structure used for the ADIP. Next different robust and LPV syn-
theses are carried out and results are discussed with regard to the
specifications introduced earlier but also from the viewpoint of
the ease of implementation on the physical experiment. This last
phase has been found necessary since we have observed some
mismatches between computerized nonlinear simulations and
the “real” responses on the ADIP.

Problem Presentation and Specifications

For the designs, we shall be using the synthesis interconnec-
tion shown in Fig. 4. The serial connection of the ADIP and the
actuator (ACT) is described by (25)-(27) or (29)-(30) depending
on the particular LPV representation we are actually using. Note
that we have introduced some integral action in the forward
channel to ensure zero steady-state tracking error at an equilib-
rium position. The integral action is here described as Fig.9. LPV control with LFT structure.

50 IEEE Control Systems

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on May 27,2024 at 10:00:14 UTC from IEEE Xplore. Restrictions apply.
 range cpI E [-(PI, (PI]. Specification (S2) translates
performance tracking and high-frequency gain at-
tenuation objectives. The specification (S3) ex-
press that stabilization but also performance and
roll-off requirements must be achieved on the larg-
est range of possible dynamics of the ADIP. The
specification (S4)is directly dictated from physi-
cal hardware limitations.

Robust LTI Syntheses

Before utilizing the LPV synthesis techniques
introduced in the previous section, it is instructive
to investigate what can be achieved using custom-
ary robust control techniques such as H _ and p
syntheses. Note first that since the synthesis 0 5 10 15 20 0 5 10 15 20
problem depicted in Fig. 4 is completely singular Time [sec] Time [sec]
(D12= 0 and D,, = 0), any H _ synthesis steps were U [degkec] 10.1<1
performed using the LMI formulation in [19], t 1.5
which is not restricted by singularity problems.
Moreover, in order to satisfy the implementation 1
constraints (S4),which require reasonable control- 0.5
ler dynamics, we also have introduced LMI region
pole constraints on the closed-loop dynamics. This 0
requires using the refined H _ synthesis technique -0.5
in [ 191 (see also the prior section). The LMI region
under consideration is determined by the intersec- -1
tion of a half-plane, a conic sector and a disk, as
shown in Fig. 6. -1.5 1
0 5 10 15 20
Time [sec] Time [sec]
H _ synthesis based on a nominal model
Based on a nominal model (ADIP in vertical
position cpl = 0, i.e. (PI = 0), an H _ controller has
been computed leading to a performance level
y = 25774 where the a, = 0.02, o,= 05 and the 50 ......... ........
LMI region constraint is determined by aTeg = 05,
rreg= 50,0,, = 45". The poles of the H _ controller
are given as 0

(-67,166, -49.834f19.109j, -39.3169, -0,1319, 0).

-50 ......... .........
Note that this is in stark contrast with the result ob-
tained without pole constraints which yielded the
controller poles
100
0 5 10 15 20
I -4
0 5 10 15
I
20
Time [sec] Time [sec]
(-3.3110 x105, -5.8209x104, -3.97759x104,
U [degkec] 10,1<1
- 2.0037 x104, - 22.806, 0).
'..J I I
Such dynamics clearly do not satisfy the imple-
mentation constraints (S4) and thus must be ruled
out in this application. So, we only retained the first
H _ controller. The corresponding nonlinear simu-
lations using the realistic model (2) are compared
with records on the "true" experiment in Fig. 7. For
each figure, the (l,l)-subplot shows the command
and the time-response of 'pl, the (1,2)-subplot
shows cp2, the (2,l)-subplot shows the control sig- Y

0 5 10 15 20 0 5 10 15 20
nal U , and the (2,2)-subplot shows 0,. We first ob- Time [sec] Time [sec]
serve the consistency of the responses of the
simulation and the experiment, which to some ex-
tent validates the simplified model we have used
for synthesis. One can also see that the control sys- Fig. 10. LFT Control results (y = 0.4733, cp,, = 60").

February 1999 51

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on May 27,2024 at 10:00:14 UTC from IEEE Xplore. Restrictions apply.
 tem has very poor performance using this controller even on the Design of an LFT Controller
somewhat reduced target cp,, = 45". An LPV control system with LFT structure is depicted in Fig. 9.
An LFT-LPV controller
p Controller Based on an Uncertain Model
We decided to improvethese preliminary results usingp synthesis
by explicitly taking account of the changing dynamics of the ADP,
where <PI = 65". We have used first-orderp scaling and the parame-
ters were tuned to o, = 0.02,03, = 0.8 and areg = 0.1, rYe8= 50,
Oreg = 45". With the H _ synthesis steps enforcing closed-loop pole
constraints, we obtained some controllers with poles is designed using the LFT description of the ADIP, where
-
(-61.497 f 25.596j , - 47.080, - 33.753, 'pl = 65 '. The parameters w , and o,were set to 0.02 and 05, re-
spectively. The underlying LTI dynamics of the LFT controller
- 0.8993, - 0.3723, - 0.2206,0},
are easily obtained by instantiating the LFT controller at some
which are clearly satisfactory with respect to the implementation frozen values of 8,. For the extreme values of 8, the following
constraints. In our example, usual p synthesis algorithms would dynamics are obtained.
lead to unacceptable controller dynamics. As before, the nonlin- e, = 1:{-365.56, - 359.74,- 291.81,
ear simulations and the hardware experiment are shown in Fig. 8. - 24.064,-1.3786,O) ,
As expected, this second controller provides better perfor-
mance on the target (p,, = 45". However, the very same controller e, = -1:{-365.56,- 359.72,- 291.84,
but with the target 9,' increased to 60" does not even provide sta- -42.888, -1.4367,O).
bility and we did not find any simple way to guarantee the same
level of performance on this larger range. Such dynamics are again satisfactory in regard to implementa-
Summing up our results, the H , controller is able to stabilize tion constraints and have been derived by minimizing, through
the range 45"but provides very poor performance. The p control- an LMI formulation, the norm of the A matrix of the LPV con-
ler provides adequate performance on the range 45" but no longer troller in the construction procedure. The nonlinear simulations
works for ranges up to 60", illustratinga fundamental tradeoff be- and the hardware experiments are shown in Fig. 11.As expected,
tween performance and the size of the operating range. In many both performance and the the size of the operating range have
applications this tradeoff is very limiting and can only be negoti- been enhanced as compared to previous LTI controllers.
ated outside the set of fixed controller structures by exploiting
gain-scheduling strategies.This is considered in the next section. Design of a Polytopic Controller
An LPV control system with polytopic structure is depicted in
LPV Syntheses Fig. 11. A similar design is now conducted using the polytopic
In this section, the synthesis techniques discussed in previ- description of the ADIP, where (p, = 65 ", leading to polytopic
ously are exploited to improve both performance and the operat- LPV controllers
ing range of the ADIP.

Polytopic LPV Plant

In order to satisfy the implementation constraints (S4), we have

used a refined synthesis technique which can handle constraints
on the closed-loop dynamics. With the selection o, = 0.1,
a,= 05 and aieg = 1, rr, = 50, 0, = 45", the underlying LTI
controllers obtained at the extreme values of the parameter range
have the following dynamics

Q~ = 0":{-69.352 + -50.251j, -53.529,

- 20.775, - 3.6379,0),

Cp, = 65":{-76.0805, - 35.506 F 41.303j,

-35.159, -2.1349,O).

They are again satisfactory. The nonlinear simulations and

Polytopic LPV Controller
the hardware experiments with the polytopic controller are
shown in Fig. 12. It is again observed that very good performance
Fig. 11. LPV control with polytopic structure. is achieved over the same operating range.

52 IEEE Control Systems

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on May 27,2024 at 10:00:14 UTC from IEEE Xplore. Restrictions apply.
 Simulation
Concluding Remarks
We can summarize the results of the four kinds
of synthesis methods as shown in Table 1.
This shows that each synthesis method can the-
oretically stabilize the operating range [-?&,‘pI]
but experimentally achieves only [-(p,,, cp,,]. As
for they-synthesis, we observe some gap between
the two range, that is, Cp, =65” and cp,, =45”.
This comes from the model approximation used
in the syntheses. In Table 1, T s indicates the set-
tling times in the experiments.
A quick view of the merits and drawbacks of
each technique is presented in Table 2.
0 5 10 15 20 Finally, we have presented a comprehensiveap-
Time [sec] Time [sec]
plication of LPV control techniques to the control of
18,1<1 an m-driven inverted pendulum. The particular
1.5
I interest of this application lies in the fact that all in-
gredients of the design problem have to be taken
into account, from the specificationsup to the con-
straints inherent to real-world implementations. In
this context, it has been shown that currently avail-
able synthesis methodologies, such as p and LPV
techniques, may fail to provide acceptable answers.
A major obstacle is undoubtly the implementation
-2 ,-.......~~.-........~....,....~...,,.,.,. constraint that puts hard limitations on the control-
-3
0 5 10 15 20
-1.5
0 5 10 15
I
20
ler dynamics. These limitations are generally diffi-
cult to handle within the usual formulation of LPV
Time [sec] Time [sec] control techniques.
It has been shown that a suitable extension of
these techniques including LMI region con-
straints on the closed-loop dynamics can over-
come this difficulty. When implementable, it has
been observed that LPV controllers outperform
1OOr I 1 1 fixed p controllers both in robustness and perfor-
mance. These observations were confirmed by
simulations but more importantly by a number of
records on the physical experiment.

References
[l] M.W. Spong and M. Vidyasagar, Robot Dynamics and
Control,Wiley, 1989.
1001 1 -4 1
...
f [2]C. Canudas de Wit, B. Siciliano, and G. Bastin (eds.),
0 5 10 15 20 0 5 10 15 20 The Theory of Robot Control, Springer, 1996.
Time [sec] Time [sec]
[3] M.K.H. Fan, A.L. Tits, and J.C. Doyle, “Robustness in
18,1<1
1.5t I the Presence of Mixed Parametric Uncertainty and Un-
modelized Dynamics,” IEEE Trans. Automat. Contl:, vol.
1 AC-36, no. 1, pp. 25-38, 1991.

0.5 [4] J.C. Doyle, A. Packard, and K. Zhou, “Review of

LFT’s, LMI’s and p,” CDC Brighton, England, pp. 1227-
0 1232, 1991.

-0.5 [5] A. Packard and G. Becker, “Quadratic Stabilization of

Parametrically-Dependent Linear Systems using Paramet-
-1 rically-Dependent Linear, Dynamic Feedback,” Advances
in Robust and Nonlinear Control Systems, vol. DSC-43,
-1.5 pp. 29-36, 1992.
0 5 10 15 20 0 5 10 15 20
Time [sec] Time [sec]
[6] A. Packard, “Gain Scheduling via Linear Fractional
Transformations, Syst. Contl: Lett., vol. 22, no. 2, pp. 79-
Fig. 12. Polytopic control results (y = 0.4129, (pic = 60’). 92, 1994.

February 2999 53

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on May 27,2024 at 10:00:14 UTC from IEEE Xplore. Restrictions apply.
 [ 191M. Chilali and P. Gahinet, “ H _ Design with Pole Placement Constraints:
Table 1. Performance indexes for LTI and LPV methods
an LMI Approach,” IEEE Trans. Autom. Contl:, vol. 41, 3, pp. 358-367,
Synthesis 1996.
[20] G. Garcia, D. Arzelier, J. Daafonz, and J. Bernussou, “An [LMI} For-
2.57
mulation for Robust Disk Pole Assignment via Output Feedback,” IFAC
0.79 45” 2.5s World Congress, San Fransisco, 1996

LFT-LPV 0.47 65O 60” 2s [21] T. Iwasaki and R.E. Skelton, “All Controllers for the General H _ Control
Problem: LMI Existence Conditions and State Space Formulas,” Auto-
Poly-LPV 0.42 65 O 60” 2s matica, vol. 30, no. 8, pp. 1307-1317, 1994.
[22] G.J. Balas, J.C. Doyle, K. Glover, A. Packard, and R. Smith, ‘b-Analy-
sis and Synthesis Toolbox,” The Mathworks, 1993.
1231 P. Gahinet, A. Nemirovski, A.J. Laub, and M. Chilali, “LMI Control
1 Table 2. Merits and drawbacks in LTI and LPV methods 1 Toolbox,” The Mathworks, 1995.

I Synthesis I Comments I [24] H. Kajiwara, P. Apkarian, and P. Gahinet, “Wide-Range Stabilization of

an Arm-Driven Inverted Pendulum Using Linear Parameter-Varying Tech-

I less performance
needs pole constraint
niques,’’ http://momiji98.ces.kyutech.ac.jp/publication.html,
1998.

Hiroyuki Kajiwara received the B S degree in Engi-

IP often higher order
needs Dole constraint neenng from Kyushu Institute of Technology in 1975,
the M S and Ph.D degrees in Engineering from Tokyo
LFT-LPV no direct pole constraint Institute of Technology in 1977 and 1985, respectively
can handle general @dependence He was a research assistant at Tokyo Institute of Tech-
nology, 1977-1982, and an associate professor at Oka-
Poly-LPV pole constraint OK yama University, 1982-1990 Since 1990, he has been
I I restricted to affine Bdeuendence I an associate professor at Kyushu Institute of Technol-
ogy His research interest is computer-aided control
system design and he is currently workmg on LPV modeling and control of
[7] P. Apkarian and P. Gahinet, “A Convex Characterization of Gain- an autonomous underwater vehicle.
Scheduled H _ Controllers,” IEEE Truns. Automat. Contl:,vol. 40, no. 5, pp.
853-864, 1995. Pierre Apkarian received the Engineer’s degree from
[8] P. Apkarian, P. Gahinet, and G. Becker, “Self-scheduled H _ Control of the Ecole SupCrieure d’Informatique, Electronique,
Linear Parameter-Varying Systems: A Design Example,” Automatica, vol. Automatique de Paris, in 1985, the M.S. degree and
31, no. 9, pp. 1251-1261, 1995. DiplBme d’Etndes Appronfondies in Mathematics from
[9] P. Apkarian and R.J. Adams, “Advanced Gain-Scheduling Techniques for Un- the University of Paris VI1 in 1985 and 1986, and the
certain Systems,” IEEE Trans. on Control System Technology,to appear, 1998. Ph.D degree in control engineering from the Ecole
Nationale SupCrieurede 1’ACronantiqueet de 1’Espace
[lo] C. Scherer, “Mixed H , / H _ Control,” Trends in Control: A European (ENSAE) in 1988. Since 1988, he has been a research
Perspective, volume of the Special Contribution to the ECC 95. scientist at ONERA-CERT and associate Professor at
[ 111C. Scherer, “Mixed H , / H _ Control for Linear Parametrically Varying ENSAE. His research interests include robust and gain-scheduling control
Systems,” CDCNew Orleans, LA, pp. 3182-3187, 1995. theory, linear matrix inequality techniques, mathematical programming and
[ 121 G. Scorletti and L. El. Ghaoui, “Improved Linear Matrix Inequality Con- applications in aeronautics. He is a member of IEEE and SIAM.
ditions for Gain-Scheduling,” CDC New Orleans, LA, pp. 3626-3631, 1995.
Pascal Gahinet graduated from the Ecole Poly-
[13] A. Helmersson, “Methods for Robust Gain-Scheduling,” Ph. D. Thesis,
technique in 1984 and from the E N S.T A , Paris,
Linkoping University, Sweden, 1995.
France, in 1986, and received the M S degree (electri-
[14] Yu. E. Nesterov and A S . Nemirovski, “Interior Point Polynomial cal and computer engineering), and the Ph D degree
Methods in Convex Programming: Theory and Applications,” SIAMStudies (system theory), from the University of California,
in Applied Mathematics vol. 13, SIAM, 1994. Santa Barbara, in 1987 and 1989. respectively He was
[15] L. Vanderberghe and S. Boyd, “Primal-Dual Potential Reduction with the French National Research Institute in Com-
Method for Problems Involving Matrix Inequalities,” Math. Programming puter and Control Sciences (I N.R LA.) from 1990 to
Series B, no. 69, pp. 205-236, 1995. 1995, taught control at E.N S T.A from 1991 to 1994,
[ 161S . Boyd andL. El. Ghaoui, “Method of Centers for Minimizing General-
and also served as a consultant for THOMSON-CSF and Atrospabale He
ized Eigenvalues,” Lin. Alg. andApplic., vol. 188, pp. 63-1 11, 1992. has been with The MathWorks Inc. since April 1996, and is currently senior
project leader for the control and advanced control product developments
[17] P. Gahinet and A. Nemirovski, “General-Purpose LMI Solvers with His research interests include robust control theory, linear matrix inequali-
Benchmarks,” CDC San Antonio, pp. 3162-3165, 1993. ties, numencal linear algebra, and numerical software for control He is
[ 181P. Gahinet and P. Apkarian, “A Linear Matrix Inequality Approach to H _ co-author of the LMI Control Toolbox and the Control System Toolbox for use
Control,” Int. J. Robust and Nonlinear Contl:, vol. 4, pp. 421-448, 1994. with Matlab

54 IEEE Control Systems

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on May 27,2024 at 10:00:14 UTC from IEEE Xplore. Restrictions apply.