This paper was not presented at any IFAC meeting.

This paper was

recommended for publication in revised form by Associate Editor Henk
Nijmeijer under the direction of Editor Hassan Khalil. This research
was supported in part by grants from the Natural Sciences and Engin-
eering Research Council of Canada (NSERC), the University of West-
ern Ontario Academic Development and Startup Funds, and the Fonds
pour la Formation de Chercheurs et l'Aide a` la Recherche (FCAR) of
the province of Quebec.
*Corresponding author.
E-mail address: rajni@eng.uwo.ca (R. V. Patel).
Automatica 37 (2001) 1825}1834
Brief Paper
Nonlinear tip-position tracking control of a #exible-link manipulator:
theory and experiments

M. Moallem, R. V. Patel*, K. Khorasani'

Dept. of Electrical & Comp. Engineering, University of Western Ontario, London, ON, Canada
'Dept. of Electrical & Comp. Engineering, Concordia University, Montreal, QC, Canada
Received 14 January 2000; revised 6 October 2000; received in "nal form 6 March 2001
Abstract
This paper presents an observer-based inverse dynamics control strategy that results in small tip-position tracking errors while
maintaining robust closed-loop performance for a class of multi-link structurally #exible manipulators. This is done by de"ning new
outputs near the end points of the arms as well as by augmenting the control inputs by terms which ensure stable operation of the
closed-loop system. As part of the control design, a nonlinear observer is introduced to estimate the rates of change of #exible modes.
Experimental results are given for the case of a two-link manipulator with a #exible link that further con"rm the theoretical and
simulation results. 2001 Elsevier Science Ltd. All rights reserved.
Keywords: Flexible arm; Non-minimum phase systems; Nonlinear decoupling control; Observers; Nonlinear control
1. Introduction
The tip position tracking control problem of struc-
turally #exible robot manipulators is challenging due to
the nonlinear and non-minimum phase characteristics of
the plant. Even for a single-link #exible arm, it is well
known that the transfer function from the torque input to
the tip position output is non-minimum phase (Wang
& Vidyasagar, 1989b). Most of the early experimental
work in this area has addressed the end-point regulation
problem (Cannon & Schmitz, 1984). For a causal con-
troller, the non-minimum phase property hinders perfect
tracking of a desired tip trajectory with a bounded con-
trol input. Thus to achieve perfect tracking using a causal
controller, the #exible system should be minimum phase.
The minimumphase property may be achieved by output
redexnition, as done in Wang and Vidyasagar (1989b); De
Luca, Lucibello, and Ulivi (1989); Madhavan and Singh
(1991); and Moallem, Patel, and Khorasani (1996a), or by
a rede"nition of the output into slow and fast outputs in
the context of integral manifold theory (Hashtrudi-Zaad
& Khorasani, 1995; Moallem, Khorasani, & Patel, 1997).
Other approaches have been proposed to deal with the
exact tracking problem which are mostly applicable to
the linearized system, e.g., Hillsley and Yurkovich (1991).
From an experimental point of view, accurate know-
ledge of state variables is required by many advanced
control algorithms for #exible multi-link robots. It is
possible to measure joint positions, velocities, and
#exible modes of manipulators using shaft encoders,
tachometers, and strain gauges (Nemir, Koivo,
& Kashyap 1988), respectively. However measuring rates
of change of #exible modes cannot be easily or accurately
accomplished. One method is to integrate the outputs of
accelerometers installed along the arms or to use analog
or numerical di!erentiation of the de#ection variables.
The former approach may be restricted from economic
considerations and the latter is not practical because of
noise problems. Thus a nonlinear state observer is desir-
able in these circumstances. Several authors have studied
the development of nonlinear observers in the general
context of nonlinear systems or speci"cally intended
mainly for rigid robot manipulators (Canudas-de-Wit
0005-1098/01/$- see front matter 2001 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 0 5 - 1 0 9 8 ( 0 1 ) 0 0 1 3 0 - 3
Fig. 1. The output of link i (y
G
).
& Slotine, 1991), and to a lesser extent for #exible-link
manipulators (Moallem, Patel, & Khorasani, 1996b). An-
other signi"cant issue in controlling of #exible structures
is the spillover e!ect (Balas, 1978) where the controller
can cause the system to become unstable due to
excitation of higher frequency modes. However, from
a practical point of view, actuators and sensors have
band-limited characteristics.
In this paper, a control strategy is developed based on
input}output linearization of the #exible-link system. The
output re-de"nition concept is employed to choose points
near the tip outputs such that stable zero-dynamics are
achieved. An observation strategy is developed and in-
corporated in the control scheme in order to estimate the
rates of change of #exible modes. The theoretical develop-
ments are further enhanced by experimental studies for
a two-link manipulator with the second link #exible.
2. Inversion-based control
The input-state map of #exible-link manipulators is
not feedback linearizable (Wang & Vidyasagar, 1989a).
However the system is locally input}output linearizable.
Input}output linearization in nonlinear systems theory
is essentially based on the developments described in
Hirschorn (1979) and Byrnes and Isidori (1985). In order
to apply this technique to #exible-link manipulators,
consider the dynamics of a multi-link #exible manipula-
tor (Book, 1984), i.e.,
M(q,o)

qK
o$
#

(q, q )#g

(q, q , o, oQ )
f
`
(q, q )#g
`
(q, q , o, oQ )#Ko
"

u
0
, (1)
where q is the n;1 vector of joint variables, o is the m;1
vector of de#ection variables, f

, f
`
, g

, and g
`
are the
terms due to gravity, Coriolis, and centripetal forces, M is
the positive-de"nite mass matrix, K is the positive-de"-
nite sti!ness matrix, and u is the n;1 vector of input
torques (clamped mode shapes have been assumed). Let
us de"ne
H(q, o)"M(q, o)"

H
`
H
`
H
``

Then, (1) can be written in the state-space form

x "f (x)#g(x)u, where x'"[q' o' q ' o
Q '] and
f (x)"

q
oQ
!H

( f

#g

)!H
`
( f
`
#g
`
#Ko)
!H
`
( f

#g

)!H
``
( f
`
#g
`
#Ko)

,
g(x)"

O
'K>L'`L
H

(q, o)
H
`
(q, o)
.
Now referring to Fig. 1, since the beam de#ection is
usually small in relation to the link length, it is concluded
that
y
G
"q
G
#:
G
d
GC
/l
G
, i"1, 2,
2
, n. (2)
In (2), y
G
is the output of link i, d
GC
is the elastic de#ection
from an undeformed link, l
G
is the length of the link, and
:
G
is a variable which takes values between !1 and #1,
with :
G
"1, 0,!1 corresponding to the tip, joint angle,
and re#ected tip output positions respectively. Now, we
de"ne the output vector (see De Luca et al., 1989; Moal-
lem et al., 1996a) as
y"q#
L`K
(:)o, (3)
where (:) is a matrix depending on modal shape func-
tions and the vector :'"[:

2
:
L
] de"nes physical
output locations on the links for achieving stable zero-
dynamics (Moallem et al., 1996a). The input}output
description of (1) with the output described by (3) is then
obtained by di!erentiating the output vector y with re-
spect to time until the input vector appears, which is
given by
yK "a(:, x)#B(:, q, o)u (4)
where B(:, q, o)"H

#
L`K
H
`
and
a(:, x)"!(H

#H
`
)( f

#g

)
!(H
`
#H
``
)( f
`
#g
`
#Ko). (5)
Now let us de"ne a "nite domain around the desired
reference trajectory q
P
, q
P
given by

P
"x: q!q
P
(

, q !q
P
(
`
, o(
`
, o
Q
(
"

(6)
where
G
(i"1,
2
,4) are some positive bounds. Also
assume that B(:, q,o) is nonsingular in
P
. This is a con-
trollability-like assumption (Slotine & Li, 1991) and is
guaranteed to hold when for instance :"0. In this case,
B"H

is positive-de"nite and therefore invertible.

Thus in a neighborhood of :"0, B is guaranteed to be
invertible since it is continuously dependent on :. Now,
let u take the following form
u"B(:, q, o)(v!a(:, x))#K
B
(q)o#K
B
Q
(q)oQ (7)
1826 M. Moallem et al. / Automatica 37 (2001) 1825}1834
A
B
I
"

I
!
`
0
, b
B
I
(x,oK
`
)"

0
H
`

!
cg

co
`

B
K
`
oI
`
#O

(oI`
`
)

#H
``
(!
cg
`
co
`

B
K
`
oI
`
#O
`
(oI`
`
))

. (12)
where K
B
(q) and K
B
Q
(q) are gain matrices that a!ect
#exural dynamics and will be discussed later. It then
follows from (4) that
yK "v#BK
B
(q)o#BK
B
Q
(q)o
Q
(8)
which is an input}output linearization of the system
when K
B
(q) and K
B
Q
(q) are zero. As discussed later, these
terms are added to enhance robustness and should usu-
ally be selected su$ciently small. In the above formula-
tions, it is assumed that oQ is available. This assumption is
relaxed later on by replacing oQ with its estimation from
a nonlinear observer.
3. Observer design
3.1. Full-order observer
Consider the dynamics of a multi-link #exible manipu-
lator given by (1). De"ning o

"o and o
`
"o
Q
, the dy-
namics of #exible modes are written as
oQ

"o
`
oQ
`
"!H
`
(q, o

)( f

(q, q )#g

(x))
!H
``
(q, o

)( f
`
(q, q )#g
`
(x))#Ko

)#H
`
(q, o

)u
(9)
where x'"[q' q ' o'

o'
`
] as before. Let us choose the
following structure for the observer dynamics
oK
Q

"oK
`
#

(o

!oK

)oK
Q
`
"!H
`
(q,o

)( f

(q, q )#g

(x

))
!H
``
(q, o

)( f
`
(q, q )#g
`
(x

)#KoK

)#H
`
(q, o

)u
#(!H
``
(q, o

)K#
`
)(o

!oK

) (10)
where x'

"[q' q ' oK'

oK'
`
] is the available state vector
that may be used for control purposes and

and
`
are
observer gain matrices to be selected. De"ning the es-
timation errors oI

"o

!oK

, oI
`
"o
`
!oK
`
, the ob-
server error dynamics can be obtained by subtracting (10)
from (9), i.e.,
oI
Q
"A
B
I
oI#b
B
I
(x, oK
`
) (11)
where oI'"[oI'

oI'
`
] and
The terms in (12) have been obtained by using the Taylor
series expansions of g

and g
`
and noting that
oK
`
"o
`
!oI
`
, i.e.,
g

(x

)"g

(q, q , o

, o
`
)!
cg

co
`

B
K
`
oI
`
#O

(oI`
`
),
g
`
(x

)"g
`
(q, q , o

, o
`
)!
cg
`
co
`

B
K
`
oI
`
#O
`
(oI`
`
). (13)
It should be noted that since the components of o
`
in
g

and g
`
are of second order (centrifugal terms), the
above Taylor series expansions terminate after the
squared terms. Moreover, considering a "nite region

around the desired point (x, oK

`
), it follows from (12)
that b
B
I
(x, oK
`
)(koI. If
`
is of the same order of
magnitude as H
`
and H
``
, the coe$cient k is typically
small. This follows by noting that (cg

/co
`

B
K
`
) and
(cg
`
/co
`

B
K
`
) are O(ooK
Q
). Moreover, the matrix A
B
I
can be
made Hurwitz by a proper choice of the gains

and
`
,
e.g., by selecting

and
`
to be any positive-de"nite
matrices. Thus, choosing a Lyapunov function candidate
<

"oI'P
B
I
oI, where P
B
I
is the solution of the Lyapunov
equation
A'
B
I
P
B
I
#P
B
I
A
B
I
"!Q
B
I
(14)
it follows that <Q

)!(z
'"
(Q
B
I
)!2z
`
(P
B
I
)k)oI`.
Hence, provided that z
'"
(Q
B
I
)'2kz
`
(P
B
I
) it can be
concluded that the error dynamics are locally asymp-
totically stable. Note that if parametric uncertainties are
included, the error dynamics will converge to a residual
set around the origin as shown in the sequel. In this case,
assuming a bounded input vector u, it follows that
b
B
I
(koI#k
"
, where k
"
is an upper bound on all the
terms due to parametric uncertainties. Then assuming
that z
'"
(Q
B
I
)'2kz
`
(P
B
I
),
<Q

)!(z
'"
(Q
B
I
)!2z
`
(P
B
I
)k)oI`#2z
`
(P
B
I
)k
"
oI.
(15)
Furthermore, <

can be written as
z
'"
(P
B
I
)oI`)<

)z
`
(P
B
I
)oI`. (16)
It then follows from (15) and (16) that for all oI3

we
have
<Q

)!

<

#
`
<`

, (17)
where

"(z
'"
(Q
B
I
)!2z
`
(P
B
I
)k)/z
`
(P
B
I
) and
`
"
2z
`
(P
B
I
)k
"
/(z
'"
(P
B
I
). Considering (17), we can con-
clude that if <

(0)'`
`
/`

, then <Q

(0. Thus the

smallest residual set can be de"ned as
''
"
oI <

)`
`
/`

for which `
`
/`

is minimum.
M. Moallem et al. / Automatica 37 (2001) 1825}1834 1827
3.2. Reduced-order observer
Consider the following observation law
oK
Q
`
"!H
`
(q, o

)( f

(q, q )#g

(x
A
))
!H
``
(q, o

)( f
`
(q, q )#g
`
(x
A
)#Ko

)
#H
`
(q, o

)u#G(oK
`
!o
`
), (18)
where G is a Hurwitz gain matrix and x

is as de"ned
previously. The above equation re#ects the fact that
o
`
"oQ

is not measurable. However, by taking Go

`
to
the left-hand side of the equation and de"ning the auxili-
ary state variable z( "oK
`
#Go

, it can be shown that

z(

"Gz( !H
`
(q, o

)( f

(q, q )#g

(x

))
!H
``
(q, o

)( f
`
(q, q )#g
`
(x

)#Ko

)
#H
`
(q,o

)u!G`o

(19)
Similarly, de"ning an auxiliary state variable for the
system dynamics as z"o
`
#Go

, and subtracting the

second equation in the resulting dynamics from (19), it
follows that if G is a Hurwitz matrix, the error dynamics
are stabilized locally. When parametric uncertainties are
present a similar analysis as before can be carried out to
show asymptotic stability of the error dynamics.
4. Observer based inverse-dynamics control
In this section, we use the control law introduced in
Section 2 except that oQ is replaced by oK
Q
given by one of
the observer strategies discussed in Section 3. Since the
rede"ned output velocity is not available, we use its
estimated value from q #oK
`
. Thus, the estimated re-
de"ned output velocity tracking error can be de"ned as
e(

"y
'
!(q #oK
`
) and the rede"ned output tracking
error by e"y
'
!y. In these relationships y
'
and y
'
are
the reference trajectory and its velocity pro"le, respec-
tively. Now let v"yK
'
#K

e#K

e(

and replace oQ by oK
Q
in
(7). Then, substituting the resulting control law in (4) and
performing some algebraic manipulations yields
EQ "A
#
E#d
#
(:, x

, t) (20)
with
E'"[e' e '], A
#
"

0 I
!K

!K

,
d
#
(:, x

, t)"!BK
B
(q)o!BK
B
Q
(q)oK
Q
#O(K

oI)#O(ooK
Q
oI). (21)
Note that K

and K

are determined by the output

position error dynamics. Moreover, (9) can be written in
terms of u given by (7), that is
Q "A

(q)#

0
G

(x,oK
Q
, t)
, (22)
where
'"[o' oQ '],
A

(q)"

0 I
!P(q)K!H
`"
K
B
(q) !H
`"
K
B
Q
(q)
,
G

(x, o
Q
K
, t)"H
`
B(yK
'
#K

e#K

e )
!P(q)( f
`
(q, q )#g
`
(x

))#O(o)#O(ooI)
(23)
in which H
`"
represents H
`
(q,o) evaluated at o"0 (no
de#ection) and
P(q)"[H
``"
!H
`"
(H
"
#H
`"
)
;(H
`"
#H
``"
)]. (24)
Furthermore, A

(q) is made Hurwitz by proper selection

of matrices K
B
(q) and K
B
Q
(q). This can be guaranteed if the
pair (B

"
, A

"
) is locally controllable on the domain of
interest with
A

"
(q)"

0 I
!P(q)K 0
, B

"
(q)"

0
H
`"

. (25)
Noting that A
#
, A

, and A
B
I
are Hurwitz matrices, it can
then be shown (using a Lyapunov analysis) that the
trajectories of the closed-loop system converge to a resid-
ual set with small tracking errors e and e and bounded
o and o
Q
provided that certain conditions are satis"ed. The
following theorem summarizes the above results
Theorem. Let the control law (7) be applied to the original
nonlinear system (1) with B(q,o,:) nonsingular in
P
(see
(6)) and the pair (B

"
, A

"
) given by (25) controllable in
P
.
Provided that the desired trajectories and their time deriva-
tives (at least up to order 2) are continuous and bounded, it
then follows that the trajectories of E, c and oI (c is a small
scaling factor as discussed in Appendix 1), starting from
a bounded set converge to a residual set that can be made
small by proper choice of controller parameters if d
#
and
G

, given by (21) and (23), respectively, satisfy certain

inequality norm conditions (see Appendix 1, e.g., (29) and
(30)) in a bounded region of the state space (E, , oI).
Proof. The proof is given in Appendix 1.
5. Implementation of the control law
Fig. 2 shows the schematic diagram of our experi-
mental setup. The #exible link is a stainless-steel
1828 M. Moallem et al. / Automatica 37 (2001) 1825}1834
Fig. 2. Experimental setup for the #exible-link robot.
Fig. 3. Control module software implementation under Chimera.
60 cm;5 cm;0.9 mm rectangular bar with a 0.251 kg
payload attached to its end point. The mass of the bar is
0.216 kg that is comparable to its payload. The "rst link
is a 20 cm rigid aluminum bar. The "rst two #exible
modes of this system when linearized around zero joint
angles are 5.6 and 27.6 Hz. The roots corresponding to
the linearized zero dynamics (when the tip position is
taken as the output) are at $j76.9 and $16.0.
The sensory equipment consists of three strain gauge
bridges, two tachometers and two shaft-encoders that are
used to measure the #exible modes of the link, joint rates,
and joint positions, respectively. The signals from the
strain gauge bridges and the tachometers are then ampli-
"ed using low-drift ampli"er stages and further passed
through anti-aliasing "lters. These signals are then fed
into the XVME-500/3 analog input module from Xycom.
The actuators are 5113 Pittman DC brushless ser-
vomotors which are directly driven by 503 Copley PWM
servo-drive ampli"ers.
The digital hardware has been selected based on the
idea of a reconxgurable sensor-based control application.
The Chimera 3.1 Real Time Operating System (Stewart,
Schmitz, & Khosla, 1993) is used as a local operating
system in conjunction with a global UNIX operating
system. In Fig. 2, the Chimera 3.1 kernel runs on an
Ironics MC68030 processor with 33 MHz clock fre-
quency and a #oating-point co-processor. Fig. 3 repres-
ents di!erent software modules that were implemented
using Chimera 3.1 based on the idea of recon"gurable
subsystems. Five subsystems, or tasks, are used which
exchange information through a shared-memory mecha-
nism called global State Variable Table (SVT). The data
logging task (LogData) in Fig. 3 can be set to a value
lower than other tasks which are running at a frequency
of 350 Hz. The ReadSensors module obtains information
from sensors and performs necessary conversions to up-
date the shared memory with q, q , o. The conversion
factors are loaded into memory by reading a con"gura-
tion "le, which is performed once in the initialization
phase. The TrajGen module updates the SVT with de-
sired tip-position trajectory data. The Controller module
is a data producer and consumer task, and updates the
SVT with torques that have to be sent to the motors
through the Out2Actuators module. The tasks are sched-
uled according to the rate monotonic rule where the
priority of each task is proportional to its frequency.
The tip de#ection is constructed based on the measure-
ments obtained from the strain gauges. Towards this end,
the modal variables are "rst obtained and the tip position
is constructed by using (3) with all :
G
's equal to 1. The
modal variables are obtained by considering three points
M. Moallem et al. / Automatica 37 (2001) 1825}1834 1829
along the link and using the Moore-Penrose pseudo-
inverse formula as in Nemir et al. (1988).
The di!erential equations corresponding to a given
observation strategy were numerically solved using the
modixed midpoint method (Press, Flannery, Teukolsky,
& Vetterline, 1995). The implemented algorithm took
approximately 2 msec on the MC68030 Ironics processor
board. Thus, a sampling frequency of 350 Hz was used.
This rate was su$cient to allow computation of the
control law as well as the data acquisition and trajectory
generation tasks while maintaining closed-loop system
stability.
5.1. Experimental results
The control law for the two-link system was obtained
based on the dynamic model obtained by using the
symbolic manipulation software Maple (Redfern, 1993)
with only one #exible mode used for control design. It
was found that the model based on a single mode gave
better results than the model based on two #exible
modes. This behavior can be explained by two inter-
related factors: First, the extent to which #exural modes
are excited depends on factors such as the desired speed
and bandwidth of the actuators and the performance of
the controller in the closed-loop system. Second, the
decision about how many modes to use in the model for
control design depends on the degree of excitation of
modal variables. This issue becomes important speci"-
cally in regard to utilizing the mass matrix in the control
law. As the number of modal variables is increased in the
model, the mass matrix becomes more ill-conditioned
which can lead to loss of performance or even instability.
Further simulations were carried out for a single-link
#exible manipulator, without implementing an observer,
which con"rmed the above issue. Thus we utilized one
modal variable in the practical implementation of the
control law. As an alternative potential solution, one may
neglect the terms containing o in the inertia matrix while
keeping all the information about the rates (this solution
was proposed by a reviewer).
In all experiments, the redexned output of the second
link was chosen to correspond to the angle from the hub
to the point at 0.8l
`
(l
`
is the length of the second link).
This location was obtained to be as near the tip position
as possible while maintaining the marginally minimum
phase characteristic (see Moallem et al., 1996a for more
details).
Fig. 4 shows the results when a reduced-order observer
is used to estimate rates of change of the de#ection modes
and the control goal is to track a 3 s quintic polynomial
trajectory with a 2 s tail of zero velocity and acceleration.
The observer gain was G"!120 and the error dynam-
ics gains were K

"22.1I
```
, K

"9.4I
```
, and
K
B
"K
B
Q
"0 since the small damping of the #exible-
link was enough to guarantee stability. Similarly, the
results for the full-order observer are given in Fig. 5,
with

"100I
```
,
`
"19000I
```
, K

"30.3I
```
,
K

"11.0I
```
, K
B
"0, and K
B
Q
"0.2B. In all the "g-
ures, the control activity after the 3 s quintic trajectory is
over is due to the gravity e!ects that bend the link at the
stopping point. Initially, we also tested the use of analog
and numerical di!erentiators for obtaining modal rates.
These attempts were not successful due to noise problems
and additional delays associated with low pass "ltering of
the signals.
The system was also tested experimentally with a feed-
forward PD controller based on the rigid model of the
manipulator. This was done to assess the performance
improvement of the proposed scheme over a conven-
tional method. Maintaining the previous gains for
K

and K

resulted in instability of the closed-loop

system. On the other hand, reducing the gains to K

"1
and K

"2 resulted in large tracking errors and in-

creased the sensitivity of the closed-loop system to uncer-
tainties (see Moallem, Patel, & Khorasani, 1997). It
should be mentioned that the marginally stable poles of
the system coupled with factors such as the neglected
dynamics of the drive system are responsible for this
instability as the feedback gains are increased.
A comparison with other control schemes for #exible-
link manipulators is interesting and is given in the
following. In Moallem, Khorasani, and Patel (1997), we
performed trajectory tracking experiments on the (sec-
ond) #exible link of the system shown in Fig. 2 based on
the method of Integral Manifolds and Singular Perturba-
tion Theory, e.g. Siciliano and Book (1989). For similar
quintic trajectories, the maximum experimental tip posi-
tion error was 0.05 rad for the integral manifold ap-
proach, 0.12 rad for the singular perturbation method
(Moallem, Khorasani, & Patel, 1997), and 0.10 rad for
the full-order observer (see Fig. 5f ). Furthermore, the
overall tracking performance is better in Fig. 5. It should
also be pointed out that the former experiments were
performed on a single-link arm due to computational
limitations imposed by the algorithms. As a result, the
schemes presented in this paper are more e$cient for
real-time implementation. Finally, the feedback gains can
be increased to a greater extent while maintaining stabil-
ity of the closed-loop system by using the schemes pre-
sented in this paper.
6. Conclusion
A control strategy based on nonlinear inversion was
proposed for a class of structurally #exible robot
manipulators and experimentally tested on a two-arm
#exible-link setup. Full- and reduced-order observation
strategies were also introduced to estimate the rates of
change of #exible modes that are not conveniently or
economically accessible. Considerable e!ort was spent to
1830 M. Moallem et al. / Automatica 37 (2001) 1825}1834
Fig. 4. Experimental results using the reduced-order observer (x-axis: time (s)). (a) Strain measurements at points 1 (*), 2 (} }), and 3 (2) (m/m). (b)
Actuator inputs (Nm): 1 (*), 2 (} }). (c) Joint velocities (rad/s): 1 (*), 2 (} }). (d) First joint angle (*) and desired trajectory (2) (rad). (e) Second link tip
angle (*) and desired trajectory (2) (rad). (f) Trajectory tracking errors (rad): 1 (*), 2 (} }).
determine how many #exible modes should be used in
controller design to achieve accurate and robust perfor-
mance. Essentially, a trade-o! is required between im-
proving robustness with respect to spillover by using
a higher number of #exible modes and avoiding the loss
of accuracy and reliability due to increased numerical
ill-conditioning that is inherent in using a larger number
of #exible modes.
A comparison of the performance of the proposed
scheme with conventional methods and more advanced
methods for #exible-link manipulator control was given.
It is concluded that the proposed approach yields overall
superior performance with small tip position tracking
errors.
Appendix A: Stability analysis
In what follows, a closed-loop stability analysis is
performed for the case of a full-order observer. A similar
analysis can be given for the case where a reduced-order
observer is used. Consider the dynamic equations of the
closed-loop system given by (11)}(12), (20)}(21) and
(22)}(23). Since A
B
I
, A
#
and A

(q) are Hurwitz matrices,

the following Lyapunov equations are satis"ed
A'
B
I
P
B
I
#P
B
I
A
B
I
"!Q
B
I
,
A'
#
P
#
#P
#
A
#
"!Q
#
, (A.1)
A'

(q)P

(q)#P

(q)A

(q)"!Q

,
M. Moallem et al. / Automatica 37 (2001) 1825}1834 1831
Fig. 5. Experimental results using the full-order observer (x-axis: time (s)). (a) Strain measurements at points 1 (*), 2 (} }), and 3 (2) (m/m). (b)
Actuator inputs (Nm): 1 (*), 2 (} }). (c) Joint velocities (rad/s): 1 (*), 2 (} }). (d) First joint angle (*) and desired trajectory (2) (rad). (e) Second link tip
angle (*) and desired trajectory (2) (rad). (f) Trajectory tracking errors (rad): 1 (*), 2 (} }).
where P
B
I
, Q
B
I
, P
#
, Q
#
, P

(q), Q

are symmetric positive-

de"nite matrices. Let us choose the positive-de"nite
Lyapunov function candidate
<"E'P
#
E#K'P

(q)K#oI'P
B
I
oI, (A.2)
where K"c with c being a small positive scaling con-
stant (typically less than one). The scaling factor is intro-
duced as a result of the stability requirement that E and
oI converge to small values near zero while remains
bounded. Taking the time derivative of < yields
<Q "!E'Q
#
E!K'Q

K!oI'Q
B
I
#2d'
#
P
#
E
#2b'
B
I
P
B
I
oI#2cAKP

#AKPQ

K. (A.3)
Consider a continuous bounded reference trajectory (at
least C`) and a bounded region

containing the origin

of (E, , oI). Then for all (E, , oI)3

LR>"", from (21)

we have
d
#
)
c
"
c
K#c

oI, (A.4)
where c
"
depends on K
B
(q) and K
B
Q
(q) and should be
su$ciently small by proper choice of these matrices. The
c

term is a!ected by K

and the higher order terms in

(21). Similarly, for all (E, , oI)3
`
LR>"", where
`
is
a "nite region containing the origin of (E, , oI), from (23)
1832 M. Moallem et al. / Automatica 37 (2001) 1825}1834
we have
G

)l

#l
#
E#
c
`
c
K#c
`
oI, (E, , oI)3
`
,
(A.5)
where l

is mainly a!ected by the reference trajectory

and l
#
is a!ected by K

and K

. Constants c
`
and
c
`
correspond to O(o) and O(ooK) terms in (23), respective-
ly. A detailed analysis also reveals that for all
(E, , oI)3
`
LR>"", where
`
is a "nite region con-
taining the origin of (E, , oI), we have PQ

(q))l

`
. This
can be shown by noting that P

(q) has a "nite growth

rate with respect to q (see e.g. Khalil, 1992, Problem 5.21)
and the reference trajectory is bounded. Hence, letting

G
"

`
and substituting the previous inequal-
ities in <Q and rearranging the right-hand side of the
resulting inequality in terms of E, and oI, after
some algebraic manipulations, yields
<Q )![E K oI]

E
K
oI
#2'

E
K
oI
(E, , oI)3
G
,
(A.6)
where
"

z
C
c
"
c
#l
#
cP

P
#

c
"
c
#l
#
cP

!l

`
!2P

c
`
cc
`
P

P
#
cc
`
P

!2c
`
P
I

,
'"[0 cl

0], (A.7)
and z
C
"z
'"
(Q
#
), z

"z
'"
(Q

), z

"z
'"
(A
B
I
). Now
consider the matrix . This matrix will be positive-de"-
nite if z

'l

`
, z
M
'2c
`
P
B
I
and the c
G
's are su$ciently
small (which may be achieved by proper choice of the
gain matrices). Note however that c appears only in the
o!-diagonal terms of this matrix. A very small c will
increase the (2,1) element while a very large c increases
(2,3) and (2,1) elements. The element (2,1) is min-
imum for c"(c
"
/(l
#
P

) which should typically be

made small by design. As explained earlier, c
"
depends on
K
B
(q) and K
B
Q
(q). On the other hand, the element (2,1) is
directly a!ected by c
"
. Thus a reduction of c
"
will be
desirable since it will result in a smaller o! diagonal term.
Now, suppose that for a given system and controller
parameters, is positive de"nite. Let us de"ne the sets
R"(E, K, oI) [E K oI][E K oI]'
)2'[E K oI]'
G
,
S"(E, K, oI) <(E, K, oI))c

L
G
, (A.8)
T"(E, K, oI) <(E, K, oI)"c
`
L
G
where 0(c
`
)c

and c
`
is the smallest constant such
that RLT. Since the trajectory de"ned by y
'
,y
'
,yK
'
is
bounded, and is positive-de"nite, R is uniformly
bounded. If the initial state is outside S!R, where
<Q )0, it follows that there exists a "nite time t
'
such that
any solution starting from S!R, at t'0, will enter
T at t
'
, and remain in T thereafter for all t*t
'
. The
residual set Tencompasses an ellipsoid in [E K oI]
coordinates with its bounds being ('/z
'"
() and
('/z
`
() which are both of order c. Since
K"c, it follows that is of the order of 1/c on this
residual set. As discussed earlier, controller parameters
should be chosen such that is positive-de"nite. To this
end, some qualitative robustness measures may be ob-
tained. In particular, making the norms of
P
#
, K
B
(q), K
B
Q
(q), P
B
I
, P

small yields a better margin of

robustness.
References
Balas, M. (1978). Active control of #exible systems. Journal of Optimiza-
tion Theory and Applications, 25, 415}436.
Book, W. J. (1984). Recursive Lagrangian dynamics of #exible manipu-
lator arms. International Journal of Robotic Research, 3, 87}101.
Byrnes, C. I., & Isidori, A. (1985). Global feedback stabilization of
nonlinear systems. Proceedings of the 24th IEEE Conference on
Decision and Control, pp. 1031}1037.
Cannon, R. H., & Schmitz, E. (1984). Initial experiments on the end-
point control of a #exible one-link robot. International Journal of
Robotic Research, 3, 62}75.
Canudas-de-Wit, C., & Slotine, J.-J. E. (1991). Sliding observers for
robot manipulators. Automatica, 27, 859}864.
De Luca, A., Lucibello, P., & Ulivi, G. (1989). Inversion techniques for
trajectory control of #exible robot arms. Journal of Robotic System,
6, 325}344.
Hashtrudi-Zaad, K., & Khorasani, K. (1995). Control of nonminimum
phase singularly perturbed systems with applications to #exible link
manipulators. In H. K. Khalil, J. H. Chow, & & P. A. Ioannou
(Eds.), Lecture Notes in Control and Information Sciences. Berlin:
Springer.
Hillsley, K. L., & Yurkovich, S. (1991). Vibration control of a two-link
#exible robot arm. IEEE International Conference on Robotics and
Automation (pp. 2121}2126). Sacramento, CA.
Hirschorn, R. M. (1979). Invertibility of nonlinear control systems.
SIAM Journal on Control and Optimization, 17, 282}297.
Khalil, H. K. (1992). Nonlinear Systems. New York: Macmillan.
Madhavan, S. K., & Singh, S. N. (1991). Inverse trajectory control and
zero-dynamics sensitivity of an elastic manipulator. International
Journal of Robotics and Automation, 6, 179}191.
Moallem, M., Khorasani, K., & Patel, R. V. (1997). An integral mani-
fold approach for tip position tracking of #exible multi-link
manipulators. IEEE Transactions on Robotics and Automation, 13,
823}837.
Moallem, M., Patel, R. V., & Khorasani, K. (1996a). An inverse dynam-
ics control strategy for tip position tracking of #exible multi-link
manipulators. International Federation of Automatic Control 13th
World Congress. Vol. A, San Francisco, CA (pp. 85}90).
Moallem, M., Patel, R. V., & Khorasani, K. (1996b). An observer-based
inverse dynamics control strategy for #exible multi-link manipula-
tors. IEEE International Conference on Decision and Control
(pp. 4112}4117). Kobe, Japan.
M. Moallem et al. / Automatica 37 (2001) 1825}1834 1833
Moallem, M., Patel, R. V., & Khorasani, K. (1997). Experimental results
for nonlinear decoupling control of #exible multi-link manipulators.
IEEE International Conference on Robotics and Automation (pp.
3142}3147). Albuquerque, NM.
Nemir, D. C., Koivo, A. J., & Kashyap, R. L. (1988). Pseudo-links and
self-tuning control of a non-rigid link mechanism. IEEE Transac-
tions on Systems Man and Cybernetics, 18, 40}48.
Press, W. H., Flannery, B. P., Teukolsky, S. A., & Vetterling, W. T.
(1995). Numerical recipes in C: The art of scientixc computing. NY:
Cambridge University Press.
Redfern, D. (1993). The Maple handbook. New York: Springer.
Siciliano, B., & Book, W. J. (1989). A singular perturbation approach to
control of lightweight #exible manipulators. International Journal of
Robotic Research, 7, 79}90.
Slotine, J.-J. E., & Li, W. (1991). Applied nonlinear control. Englewood
Cli!s, NJ: Prentice-Hall.
Stewart, D. B., Schmitz, D., & Khosla, P. K. (1993). Chimear 3.1: The
real-time operating system for reconxgurable sensor-based control
systems. The Robotics Institute, Dept of Electrical and Computer
Engineering, Carnegie Mellon University.
Wang, D., & Vidyasagar, M. (1989a). Feedback linearizability of multi-
link manipulators with one #exible link. IEEE International Confer-
ence on Robotics and Automation. Tampa, FA.
Wang, D., & Vidyasagar, M. (1989b). Transfer functions for a single
#exible link. IEEE International Conference on Robotics and
Automation, Tampa, FL (pp. 1042}1047).
K. Khorasani (M 85) received the B.S.,
M.S., and Ph.D. degrees in Electrical and
Computer Engineering from the Univer-
sity of Illinois at Urbana-Champaign in
1981, 1982 and 1985, respectively.
From 1985 to 1988 he was an Assistant
Professor at the University of Michigan at
Dearborn and since 1988, he has been at
Concordia University, Montreal, Canada,
where he is currently a Professor in the
Department of Electrical and Computer
Engineering, and since 1998 an Associate
Dean in the Faculty of Engineering and Computer Science. His main
areas of research are in stability theory, nonlinear and adaptive control,
modeling and control of #exible link/joint manipulators, neural net-
works applications to pattern recognition, control, and robotics, adap-
tive structure neural networks, and distributed and collaborative force
feedback of haptic interfaces in virtual environments. He has author-
ed/co-authored over 175 publications in the above areas.
Professor Patel received the B.Eng. degree
in Electronics with First Class Honors
from the University of Liverpool, England,
in 1969, and the Ph.D. degree in Electrical
Engineering from the University of Cam-
bridge, England, in 1973.
From 1973 to 1998 he held postdoctoral
and faculty positions at the University of
Cambridge, England, Lund Institute
of Technology, Sweden, NASA Ames
Research Center, USA, University of
Waterloo, Canada, Delft University of
Technology, Holland, and the Control Systems Centre, UMIST, Eng-
land, and Concordia University, Canada. At present, he holds the
position of Professor and Chair in the Department of Electrical and
Computer Engineering at the University of Western Ontario, Canada.
Professor Patel has a broad research background with expertise in
such areas as design and control of electromechanical systems; practical
applications of modern control theory; neural network applications;
real-time control, numerical techniques for robotics and control;
modeling, simulation and 3-D graphics animation of robotic systems;
application of virtual-reality and haptics in human-computer interac-
tion. He has published over 190 technical papers, and has co-authored
a control systems textbook and three research monographs on robotics.
In 1994, he edited (with A. J. Laub and P. M. Van Dooren) an IEEE
Press Reprint Book on Numerical Linear Algebra Techniques for
Systems and Control.
Professor Patel was elected a Fellow of the IEEE in 1992 for his
contributions to the numerical analysis and design of control systems.
He has served as an Associate Editor of several leading journals in the
areas of control systems and robotics, including the IEEE Transactions
on Automatic Control and Automatica.
M. Moallem received the Ph.D. degree in
Electrical & Computer Engineering from
Concordia University, Montreal, Canada,
in 1996. He received the B.Sc. (with highest
ranking) and M.Sc. degrees in Electrical
and Electronic Engineering from Shiraz
University, and Sharif University of
Technology, Iran, in 1986, and 1988,
respectively.
Dr. Moallem is currently an Assistant
Professor in the Department of Electrical
& Computer Engineering, at the Univer-
sity of Western Ontario. He has held postdoctoral and research posi-
tions at Concordia University and Duke University, Durham, NC,
USA prior to joining the University of Western Ontario. Dr. Moallem's
research interests include algorithm development and software design
for advanced robotic and mechatronic applications.
1834 M. Moallem et al. / Automatica 37 (2001) 1825}1834