Automatica 40 (2004) 1195 – 1203

www.elsevier.com/locate/automatica

Brief paper
Adaptive iterative learning control for robot manipulators

Abdelhamid Tayebi∗
Department of Electrical Engineering, Lakehead University, 955 Oliver Road, Thunder Bay, Ont., Canada P7B 5E1
Received 11 September 2002; received in revised form 2 June 2003; accepted 23 January 2004

Abstract

In this paper, we propose some adaptive iterative learning control (ILC) schemes for trajectory tracking of rigid robot manipulators,
with unknown parameters, performing repetitive tasks. The proposed control schemes are based upon the use of a proportional-derivative
(PD) feedback structure, for which an iterative term is added to cope with the unknown parameters and disturbances. The control design
is very simple in the sense that the only requirement on the PD and learning gains is the positive de3niteness condition and the bounds
of the robot parameters are not needed. In contrast to classical ILC schemes where the number of iterative variables is generally equal to
the number of control inputs, the second controller proposed in this paper uses just two iterative variables, which is an interesting fact
from a practical point of view since it contributes considerably to memory space saving in real-time implementations. We also show that
it is possible to use a single iterative variable in the control scheme if some bounds of the system parameters are known. Furthermore,
the resetting condition is relaxed to a certain extent for a certain class of reference trajectories. Finally, simulation results are provided to
illustrate the e7ectiveness of the proposed controllers.
? 2004 Elsevier Ltd. All rights reserved.

Keywords: Iterative learning control; Adaptive control; Robot manipulators

1. Introduction sating for the gravity forces, the PD control scheme leads
to a steady-state error, which can eventually be reduced by
For the sake of implementation simplicity, the use of increasing the proportional and derivative gains (high-gain
classical linear controllers such as PD and PID in robotics feedback) or by introducing an integral action. The draw-
applications has attracted researchers and industrials for back of the high-gain feedback solution is related to the
many decades. In fact, a PD controller with gravity compen- fact that it may saturate the joint actuators or/and excite
sation is able to globally asymptotically stabilize the joint high-frequency modes. On the other hand, with the PID
positions of rigid robot manipulators at a given set-point control scheme only local asymptotic stability was proven
as long as the gravity forces can be instantaneously eval- under some relatively complex conditions until the introduc-
uated or at least known at the desired 3nal con3guration tion of the passivity property for robot manipulators, which
(Takegaki & Arimoto, 1981; Spong & Vidyasagar, 1989; allowed to design globally asymptotically stabilizing PID
Tomei, 1991). This condition is not easy to satisfy in prac- controllers without gravity compensation (Arimoto, 1996).
tical situations because the gravity terms generally depend Again this result requires the control gains to satisfy some
on the unknown and possibly time-varying payloads manip- relatively complex relations involving particularly the robot
ulated by the robot during a given task. Without compen- manipulator dynamics. Besides that, owing to the physical
property that the robot parameters appear linearly in the
Lagrange equation, another interesting approach has been

This paper was not presented at any IFAC meeting. This paper was derived for trajectory tracking instead of set-point regulation
recommended for publication in revised form by Associate Editor Antonio (Slotine & Li, 1987; Slotine & Li, 1991). This approach
E. de Barros Ruano under the direction of Editor Robert R. Bitmead. consists basically of a PD controller with an additional term
This research was supported by the Natural Sciences and Engineering generated by an appropriate adaptive rule in order to cope
Research Council of Canada (NSERC).
∗ Tel.: +1-807-343-8597; fax: +1-807-343-8928. with the unknown parameters assumed to be time-invariant.
E-mail addresses: [email protected], [email protected] Since robot manipulators are generally used in repeti-
(A. Tayebi). tive tasks, one should take advantage of the fact that the

0005-1098/$ - see front matter ? 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.automatica.2004.01.026
1196 A. Tayebi / Automatica 40 (2004) 1195 – 1203

reference trajectory is repeated over a given operation time. rule. In this framework, the acceleration measurements and
In this context, ILC techniques can be applied in order to the bounds of the robot parameters are not needed and the
enhance the tracking performance from operation to oper- only requirement on the control gains is the positive de3-
ation. Since the early works of Arimoto, Kawamura, and niteness condition. On the other hand, one of the proposed
Miyazaki (1984), Casalino and Bartolini (1984) and Craig controllers uses just two iterative variables, which is an in-
(1984), several ILC schemes for robot manipulators have teresting advantage from a practical point of view. We also
been proposed in the literature, see for instance Arimoto show that it is possible to bring down the number of iter-
(1996), Bondi, Casalino, and Gambardella (1988), De Luca, ative variables to one, at the expense of the knowledge of
Paesano, and Ulivi (1992), Horowitz (1993), Kavli (1992), some bounds of the system parameters. Furthermore, the
Kawamura, Miyazaki, and Arimoto (1988) and Moon, resetting condition is relaxed to a certain extent for a cer-
Doh, and Chung (1997). These ILC algorithms, whether tain class of reference trajectories. Finally, simulation results
developed for the linearized model or the nonlinear model, are provided to illustrate the e7ectiveness of the proposed
are generally based upon the contraction mapping theory controllers.
and require a certain a priori knowledge of the system
dynamics.
On the other hand, another type of ILC algorithms has 2. Problem formulation
been developed using Lyapunov and Lyapunov-like meth-
ods. In fact, in French and Rogers (2000) a standard Lya- Using the Lagrangian formulation, the equations of mo-
punov design is used to solve ILC problems. The idea con- tion of a n degrees-of-freedom rigid manipulator may be
sists of using a standard adaptive controller and to start the expressed by
parameter estimates with their 3nal values obtained at the
preceding iteration. In the same spirit, Choi and Lee (2000) M (qk (t))qLk (t) + C(qk (t); q̇k (t))q̇k (t) + G(qk (t))
proposed an adaptive ILC for uncertain robot manipulators, = k (t) + dk (t); (1)
where the uncertain parameters are estimated along the time
horizon, whereas the repetitive disturbances are compen- where t denotes the time and the nonnegative integer
sated along the iteration horizon. However, as in standard k ∈ Z+ denotes the operation or iteration number. The sig-
adaptive control design, this technique requires the unknown nals qk ∈ Rn , q̇k ∈ Rn and qLk ∈ Rn are the joint position,
system parameters to be constant. In Kuc, Nam and Lee joint velocity and joint acceleration vectors, respectively,
(1991), Ham, Qu, and Kaloust (1995), Ham, Qu, and at the iteration k. M (qk ) ∈ Rn×n is the inertia matrix,
Johnson (2000), Xu (2002), Xu, Badrinath, and Qu C(qk ; q̇k )q̇k ∈ Rn is a vector resulting from Coriolis and
(2000), Xu and Tan (2001), several ILC algorithms have centrifugal forces. G(qk ) ∈ Rn is the vector resulting from
been proposed based upon the use of a positive de3nite the gravitational forces. k ∈ Rn is the control input vector
Lyapunov-like sequence which is made monotonically containing the torques and forces to be applied at each
decreasing along the iteration axis via a suitable choice joint. dk (t) ∈ Rn is the vector containing the unmodeled
of the control input. In contrast to the standard adap- dynamics and other unknown external disturbances.
tive control, this technique is shown to be able to handle Assuming that the joint positions and the joint velocities
systems with time-varying parameters since the adapta- are available for feedback, our objective is to design a control
tion law in this case is nothing else but a discrete inte- law k (t) guaranteeing the boundedness of qk (t), ∀t ∈ [0; T ]
gration along the iteration axis. Based on this approach, and ∀k ∈ Z+ , and the convergence of qk (t) to the desired
Kuc et al. (1991) proposed an ILC scheme for the lin- reference trajectory qd (t) for all t ∈ [0; T ] when k tends to
earized robot manipulator model, while in Ham et al. in3nity.
(2000), Xu et al. (2000) nonlinear ILC schemes have been Throughout this paper, we will use the Lpe norm de3ned
proposed for the nonlinear model. Again these control as follows:
laws require a certain a priori knowledge of the system   t 1=p

 p
dynamics.  x( ) d if p ∈ [0; ∞);
In this paper, we present three simple ILC schemes for x(t)pe , 0


the position tracking problem of rigid robot manipulators.  sup x( ) if p = ∞;
The control schemes are built around a classical PD feed- 06 6t

back structure, for which an iterative term is added in order where x denotes any norm of x, and t belongs to the 3nite
to cope with the unknown parameters and disturbances. The interval [0; T ]. We say that x ∈ Lpe when xpe exists (i.e.,
convergence of the tracking error to zero, over the whole when xpe is 3nite).
3nite time-interval, is guaranteed when the iteration num- We assume that all the system parameters are unknown
ber tends to in3nity. The proof of convergence is based and we make the following reasonable assumptions:
upon the use of a Lyapunov-like positive de3nite sequence,
which is made monotonically decreasing through an ade- (A1) The reference trajectory and its 3rst and second
quate choice of the control law and the iterative adaptation time-derivative, namely qd (t), q̇d (t) and qLd (t), as well
 A. Tayebi / Automatica 40 (2004) 1195 – 1203 1197

as the disturbance dk (t) are bounded ∀t ∈ [0; T ] and Proof. Part 1: Let us consider the following Lyapunov-like
∀k ∈ Z+ . composite energy function 1 :
(A2) The resetting condition is satis3ed, i.e., q̇d (0) −
q̇k (0) = qd (0) − qk (0) = 0, ∀k ∈ Z+ . Wk (q̃˙k (t); q̃k (t); ˜k (t)) = Vk (q̃˙k (t); q̃k (t))

We will also need the following properties, which are com- 1 t ˜T
+  ( )−1 ˜k ( ) d ; (4)
mon to robot manipulators 2 0 k

(P1) M (qk ) ∈ Rn×n is symmetric, bounded, and positive

with ˜k (t) = (t) − ˆk (t), where (t) = [T (t) ]T ∈ Rm and
de3nite. ˆ
k (t) = [ˆTk (t) ˆk (t)]T is the estimated value of (t). The
(P2) The matrix Ṁ (qk ) − 2C(qk ; q̇k ) is skew symmetric,
unknown vector (t) is de3ned in (P3) and the unknown
hence xT (Ṁ (qk ) − 2C(qk ; q̇k ))x = 0; ∀x ∈ Rn .
parameter  is obtained according to (P1) and (A1) such
(P3) G(qk ) + C(qk ; q̇k )q̇d (t) = (qk ; q̇k )(t), where
that M (qk )qLd − dk  6 , ∀t ∈ [0; T ] and ∀k ∈ Z+ .
(qk ; q̇k ) ∈ Rn×(m−1) is a known matrix and (t) ∈
The term Vk (q̃˙k (t); q̃k (t)) in (4) is chosen as follows
Rm−1 is an unknown continuous vector over [0; T ].
(P4) C(qk ; q̇k ) 6 kc q̇k  and G(qk ) ¡ kg , ∀t ∈ [0; T ]
and ∀k ∈ Z+ , where kc and kg are unknown positive Vk (q̃˙k (t); q̃k (t)) = 12 q̃˙Tk M (qk )q̃˙k + 12 q̃Tk KP q̃k : (5)
parameters.
The di7erence of Wk is given by
Note that (P1–P3) are needed to derive the result in The-
orem 1 and (P1; P2; P4) are needed to derive the result in SWk = Wk − Wk−1 = Vk − Vk−1
Theorems 2 and 3. Assumption (A2) will be relaxed to a 
certain extent in Theorem 4. 1 t ˜T −1 ˜
+ (  k − ˜Tk−1 −1 ˜k−1 ) d
2 0 k

1 t TT −1 T
3. Adaptive ILC design = Vk − Vk−1 − (  k + 2TTk −1 ˜k ) d ; (6)
2 0 k

Now, one can state the following result

where Tk = ˆk − ˆk−1 . On the other hand, one can rewrite
Vk as follows
Theorem 1. Consider system (1) with properties (P1–P3)
under the following control law:
Vk (q̃˙k (t); q̃k (t)) = Vk (q̃˙k (0); q̃k (0))
k (t) = KP q̃k (t) + KD q̃˙k (t) + (qk ; q̇k ; q̃˙k )ˆk (t); (2)  t 
˙T L̃ 1 ˙T ˙ ˙T
+ q̃k M qk + q̃k Ṁ q̃k + q̃k KP q̃k d : (7)
0 2
with

ˆk (t) = ˆk−1 (t) + T (qk ; q̇k ; q̃˙k )q̃˙k (t); (3) Now, using (1) and (P2; P3), Eq. (7) leads to

where ˆ−1 (t) = 0, q̃k (t) = qd (t) − qk (t) and q̃˙k (t) = Vk (q̃˙k (t); q̃k (t)) = Vk (q̃˙k (0); q̃k (0))
q̇d (t) − q̇k (t). The matrix (qk ; q̇k ; q̃˙k ) ∈ Rn×m is de4ned  t
as (qk ; q̇k ; q̃˙k ) , [(qk ; q̇k ) sgn(q̃˙k )], where sgn(q̃˙k ) + q̃˙Tk [M (qk )qLd − dk + C(qk ; q̇k )q̇d
is the vector obtained by applying the signum func- 0

tion to all elements of q̃˙k . The matrices KP ∈ Rn×n , +G(qk ) + KP q̃k − k] d

KD ∈ Rn×n and  ∈ Rm×m are symmetric positive def-
inite. Let assumptions (A1–A2) be satis4ed, then 6 Vk (q̃˙k (0); q̃k (0))
q̃k (t) ∈ L∞e , q̃˙k (t) ∈ L∞e , k (t) ∈ L2e for all k ∈ Z+ and  t
limk→∞ q̃k (t) = limk→∞ q̃˙k (t) = 0, ∀t ∈ [0; T ]. + q̃˙Tk ((qk ; q̇k ) + KP q̃k
0

The proof of this theorem is in three parts. The 3rst part + sgn(q̃˙k ) − k) d
consists of taking a positive de3nite Lyapunov-like compos-
ite energy function (Xu, 2002; Xu & Tan, 2001), namely Wk ,
and show that this sequence is non-increasing with respect to
k and hence bounded if W0 is bounded. In the second part, we 1 Throughout this paper, we will use the abusive notation Wk (t) and
show that W0 (t) is bounded for all t ∈ [0; T ]. Finally, in the Vk (t) instead of Wk (q̃˙k (t); q̃k (t); ˜k (t)) and Vk (q̃˙k (t); q̃k (t)). We will also
third part, we show that limk→∞ q̃k (t) = limk→∞ q̃˙k (t) = 0, use Wk and Vk instead of Wk (t) and Vk (t) where this does not lead to
∀t ∈ [0; T ]. any confusion.
1198 A. Tayebi / Automatica 40 (2004) 1195 – 1203

 t
Part 3: Note that Wk can be written as follows Wk = W0 +
6 Vk (q̃˙k (0); q̃k (0)) + q̃˙Tk ((qk ; q̇k ; q̃˙k ) k
0 j=1 SWj . Hence, using (10), one has

+KP q̃k − k) d : (8) k

Wk 6 W0 − Vj−1
Now, substituting (2) in (8) we obtain j=1

k k
Vk (q̃˙k (t); q̃k (t)) 6 Vk (q̃˙k (0); q̃k (0)) 1 1
6 W0 − q̃Tj−1 KP q̃j−1 − q̃˙Tj−1 M (qj−1 )q̃˙j−1 ;
 t 2 2
j=1 j=1
+ q̃˙Tk ((qk ; q̇k ; q̃˙k )˜k − KD q̃˙k ) d :
0
(14)
(9)
which implies that
Using (3), (9) and (A2), Eq. (6) leads to
k k

1 t ˙T q̃Tj−1 KP q̃j−1 + q̃˙Tj−1 M (qj−1 )q̃˙j−1
SWk 6 −Vk−1 − q̃ ((qk ; q̇k ; q̃˙k )T (qk ; q̇k ; q̃˙k )
2 0 k j=1 j=1

+2KD )q̃˙k d 6 0: (10) 6 2(W0 − Wk ) 6 2W0 :

Hence, Wk is a non-increasing sequence. Thus if W0 is Hence, limk→∞ q̃k (t) = limk→∞ q̃˙k (t) = 0, ∀t ∈ [0; T ], since
bounded one can conclude that Wk is bounded. In Part 2 Wk (t) is bounded ∀k ∈ Z+ , ∀t ∈ [0; T ].
of the Proof we will show that W0 (t) is bounded for all
t
t ∈ [0; T ]. Hence, q̃k (t), q̃˙k (t) and 0 ˜Tk ( )−1 ˜k ( ) d are Note that under properties (P1–P3) the control law (2)–
bounded for all k ∈ Z+ and all t ∈ [0; T]. Since (t) is con- (3) involves m iterative parameters, where m is generally
t
tinuous over [0; T ], the boundedness of 0 ˜Tk ( )−1 ˜k ( ) d larger than the number of degrees-of-freedom n. It also re-
t T quires the knowledge of the matrix (qk ; q̇k ). However, by
implies the boundedness of 0 ˆk ( )−1 ˆk ( ) d . Conse-
using (P4) instead of (P3), the knowledge of the matrix
quently, one can conclude that k (t) ∈ L2e for all k ∈ Z+ .
(qk ; q̇k ) is not required anymore and the number of itera-
Part 2: Now, we will show that W0 (t) is bounded over
tive parameters is reduced to two as stated in the following
the time interval [0; T ]. In fact, considering (4) with k = 0,
theorem.
the time-derivative of W0 can be bounded as follows
1 Theorem 2. Consider system (1) with properties (P1; P2; P4)
Ẇ 0 6 q̃˙T0 ((q0 ; q̇0 ; q̃˙0 )˜0 − KD q̃˙0 ) + ˜T0 −1 ˜0 : (11) under the following control law:
2

Since ˆ−1 (t)=0, one has ˆ0 (t)=T (q0 ; q̇0 ; q̃˙0 )q̃˙0 (t). Hence k (t) = KP q̃k (t) + KD q̃˙k (t) + %(q̃˙k )ˆk (t) (15)
 
1
Ẇ 0 6 −q̃˙T0 KD q̃˙0 + ˆT0 + ˜T0 −1 ˜0 with
2
1 ˜T −1 ˜ ˆk (t) = ˆk−1 (t) + %T (q̃˙k )q̃˙k (t); (16)
6 −q̃˙T0 KD q̃˙0 −   0 +  T −1 ˜0 : (12)
2 0
where ˆ−1 (t) = 0. The matrices KP ∈ Rn×n , KD ∈ Rn×n
Using Young’s inequality, we have and  ∈ R2×2 are symmetric positive de4nite. The ma-
trix %(q̃˙k ) is de4ned as %(q̃˙k ) , [q̃˙k sgn(q̃˙k )]. Let
1
 T −1 ˜0 6 K−1 ˜0 2 + 2 assumptions (A1–A2) be satis4ed, then q̃k (t) ∈ L∞e ,
4K q̃˙k (t) ∈ L∞e , k (t) ∈ L2e for all k ∈ Z+ and limk→∞ q̃k (t)=
for any K ¿ 0. Hence, limk→∞ q̃˙k (t) = 0, ∀t ∈ [0; T ].

1 Proof. Here we use the same Lyapunov-like composite en-

Ẇ 0 6 − "1 q̃˙0 2 − "2 ˜0 2 + 2 ; (13)
4K ergy function (4), with  ∈ R2×2 and ˜ ∈ R2 . The vector 
is de3ned as  = [&; ']T ∈ R2 . The unknown parameters &
with "1 = #min (KD ), "2 = 12 #min (−1 ) − K#max 2
(−1 ) and and ' are de3ned as follows:
−1 2 −1
0 ¡ K 6 #min ( )=2#max ( ), where #min (∗)(#max (∗))
denotes the minimal (maximal) eigenvalue of (∗). Since  q̃˙Tk (M (qk )qLd + C(qk ; q̇k )q̇d + G(qk ) − dk )
is continuous over [0; T ], it is clear that it is bounded over
[0; T ], i.e., ∞e 6 max . Hence, one can conclude from 6 q̃˙k ( + kg + kc q̇d q̇k )
2
(13) that Ẇ 0 (t) 6 max =(4K), which implies that W0 (t) is
uniformly continuous and thus bounded over [0; T ]. 6 q̃˙k ( + kg + kc q̇d 2 + kc q̇d q̃˙k ); (17)
 A. Tayebi / Automatica 40 (2004) 1195 – 1203 1199

where kc and kg are de3ned in (P4), and  is obtained ac- q̃k (t) ∈ L∞e , q̃˙k (t) ∈ L∞e , k (t) ∈ L2e for all k ∈ Z+ and
cording to (A1) and (P1). Since q̇d is bounded, inequality limk→∞ q̃k (t) = limk→∞ q̃˙k (t) = 0, ∀t ∈ [0; T ].
(17) leads to
Proof. Here we use the following Lyapunov-like composite
q̃˙Tk (M (qk )qLd + C(qk ; q̇k )q̇d + G(qk ) − dk )
energy function
6 q̃˙Tk (&q̃˙k + ' sgn(q̃˙k )) 6 q̃˙Tk %(q̃˙k ); (18)
 t
1
where & = kc Supt∈[0; T ] q̇d  and ' =  + kg + kc Supt∈[0; T ] Wk = Vk (q̃˙k (t); q̃k (t)) + *−1 '˜2k ( ) d ; (24)
q̇d (t)2 . 2 0
Following the steps of the Proof of Theorem 1, and using
(18), we obtain with '˜k (t) = ' − 'ˆk (t), where the unknown parameter ' is
Vk (q̃˙k (t); q̃k (t)) 6 Vk (q̃˙k (0); q̃k (0)) de3ned in (18). The term Vk is de3ned in (5) and can be
 t written as (19), by virtue of (18).
Now, substituting (22) in (19) we obtain
+ q̃˙Tk (%(q̃˙k ) + KP q̃k − k) d : (19)
0

Substituting (15) in (19) we obtain Vk (q̃˙k (t); q̃k (t)) 6 Vk (q̃˙k (0); q̃k (0))
 t
Vk (q̃˙k (t); q̃k (t)) 6 Vk (q̃˙k (0); q̃k (0))
+ q̃˙Tk ('˜k sgn(q̃˙k ) + &q̃˙k − KD q̃˙k ) d :
 t 0
+ q̃˙Tk (%(q̃˙k )˜k − KD q̃˙k ) d : (20) (25)
0

Using (16) and (20) in (6), in view of (A2), we obtain

In the same way as in the Proof of the Theorem 2, using
SWk 6 −Vk−1 (23) and (25), in view of (A2), we obtain

1 t ˙T ˙ 
− q̃ (%(q̃k )%T (q̃˙k ) + 2KD )q̃˙k d 6 0: (21) 1 t
2 0 k SWk 6 −Vk−1 − q̃˙Tk (* sgn(q̃˙k ) sgn(q̃˙k )T
2 0
The remaining of the proof is omitted since it is similar to
the proof of Theorem 1. +2(KD − &I ))q̃˙k d ; (26)

Remark 1. Generally, in contraction-mapping-based ILC which is non-increasing if the matrix (KD − &I ) is positive
schemes, the number of iterative parameters is equal to the semi-de3nite. The remaining of the proof is omitted since it
number of the control inputs which is equal to the number is similar to the proof of Theorem 1.
of degrees of freedom n. In our approach (Theorem 2), we
use only two iterative parameters ˆk ∈ R2 , which is an inter- Note that to carry out the results in Theorems 1–3, we
esting fact from a practical point of view since it contributes assumed that the resetting condition is satis3ed. In fact, this
considerably to memory space saving. Now, the question is: condition, i.e., (A2), can be relaxed to a certain extent if
Can we bring the number of iterative parameters from two the alignment condition is satis3ed (Xu et al., 2000; Xu and
to one and at what expense? The answer is yes, but to the Tan, 2001), i.e., q̃˙k (0)= q̃˙k−1 (T ) and q̃k (0)= q̃k−1 (T ) that is
expense of a certain knowledge of the system dynamics. In Vk (0) = Vk−1 (T ). In other words, if the reference trajectory
fact, in contrast with the result in Theorem 2, the knowl- satis3es qd (0) = qd (T ) and q̇d (0) = q̇d (T ), we can start
edge of the parameter kc de3ned in (P4) is needed to guar- the system from where it was stopped at the last operation
antee the convergence of the tracking error to zero as stated instead of bringing the system to the same initial position at
bellow in Theorem 3. each operation. According to this discussion one can state
the following result.
Theorem 3. Consider system (1) with properties (P1; P2;
P4) under the following control law: Theorem 4. If the alignment condition, i.e., Vk (0) =
˙ ˆ ˙
k (t) = KP q̃k (t) + KD q̃k (t) + 'k (t) sgn(q̃k (t)) (22) Vk−1 (T ), is used in theorems 1–3 instead of (A2), then

with • q̃k (t) ∈ L∞e , q̃˙k (t) ∈ L∞e , k (t) ∈ L2e for all k ∈ Z+ .
'ˆk (t) = 'ˆk−1 (t) + *q̃˙Tk (t) sgn(q̃˙k (t)); (23) • limk→∞ q̃k (t) = ,, ∀t ∈ [0; T ], where , is 4nite and tends
to zero when #min (KP ) tends to in4nity.
where 'ˆ−1 (t) = 0. The matrices KP ∈ Rn×n and KD ∈ Rn×n
are symmetric positive de4nite, and * is a positive scalar. Proof. Since the proof is similar for Theorems 1–3, we will
Let assumptions (A1–A2) be satis4ed. If (KD − &I ) is prove this result just for Theorem 2. In fact, using (20) in
positive semi-de4nite, with & = kc Supt∈[0; T ] q̇d (t), then (6), and using the alignment condition instead of (A2), for
1200 A. Tayebi / Automatica 40 (2004) 1195 – 1203

t = T , we obtain  T
6T q̃˙Tk ( )q̃˙k ( ) d
SWk (T ) 6 Vk (0) − Vk−1 (T ) 0
  T
1 T ˙T ˙ T
− q̃ (%(q̃k )%T (q̃˙k ) + 2KD )q̃˙k d 6 q̃˙Tk ( )KD q̃˙k ( ) d : (35)
2 0 k #min (KD ) 0
 From (30) and (35), one can conclude that
1 T ˙T ˙
6− q̃ (%(q̃k )%T (q̃˙k ) + 2KD )q̃˙k d
2 0 k lim q̃k (t) − q̃k (0) = lim q̃k (t) − q̃k−1 (T ) = 0; (36)
k→∞ k→∞
6 0: (27) for all t ∈ [0; T ]. Hence limk→∞ q̃k (t) = ,, ∀t ∈ [0; T ], where
In the same way as in the Proof of Theorem 1, one can show , is a 3nite value. Finally, using (1) and (15), one can show
that Wk (T ) is bounded for all k ∈ Z+ and that , tends to zero when #min (KP ) tends to in3nity.
k
Remark 2. In all theorems proposed in this paper, one can
Wk (T ) = W0 (T ) + SWj (T ) show that the tracking error and its time derivative, at the
j=1
3rst iteration, can be made arbitrarily small, over the 3nite
k  T time interval [0; T ], by increasing the minimal eigenvalues
6 W0 (T ) − q̃˙Tj KD q̃˙j d ; (28) of the control and learning gains KP , KD and ( or *).
j=1 0
Remark 3. It is worth noting that the proposed control strat-
which leads to egy can be used in a straightforward manner for industrial
k  T robot manipulators already functioning under a PD con-
q̃˙Tj KD q̃˙j d 6 W0 (T ) − Wk (T ) 6 W0 (T ): (29) troller by just adding the iterative term to the control input
0
j=1 in order to enhance the tracking performance from operation
Hence, to operation.
 T
lim q̃˙Tk KD q̃˙k d = 0: (30) Remark 4. Note that the problem of set-point regulation is
k→∞ 0 included in Theorem 4 since a constant reference trajectory
Since Wk (T ) is bounded for all k ∈ Z+ , one can con- satis3es the alignment condition. Basically, in this particu-
T lar case, our algorithms will ensure a step-by-step conver-
clude that 12 0 ˜Tk ( )−1 ˜k ( ) d , $k (T ) 6 $ ¡ ∞.
gence to the position set-point. In fact, after one operation
Since $k (t) 6 $k (T ) 6 $, ∀k ∈ Z+ , ∀t ∈ [0; T ], one has
we arrive to a certain position q1 , from which we start the
Wk (t) = Vk (t) + $k (t) 6 Vk (t) + $. Thus,
next operation to arrive to q2 and so one, we arrive to the
Wk−1 (t) 6 Vk−1 (t) + $: (31) position qk after k iterations. The tracking error is reduced
progressively and converges to the 3nite value , as stated in
On the other hand, one has Theorem 4.
SWk (t) = Wk (t) − Wk−1 (t) 6 Vk (0) − Vk−1 (t): (32)
Remark 5. Note that, in all theorems proposed in this paper,
From (31) and (32), one can conclude that one can show that the control input k (t) ∈ L∞e for any
3nite k. This follows from the fact that q̃k (t); q̃˙k (t) ∈ L∞e ;
Wk (t) 6 Vk (0) + $ = Vk−1 (T ) + $: (33)
∀k ∈ Z+ which implies that the parameters estimates are
Since Wk (T ) is bounded ∀k ∈ Z+ , it is clear that Vk (T ) is bounded for any 3nite k.
bounded ∀k ∈ Z+ . Hence, from (33), one can conclude that
Wk (t) is bounded
t ∀k ∈ Z+ , ∀t ∈ [0; T ]. Consequently, q̃k (t), Remark 6. In practical applications, the manipulator joints
q̃˙k (t) and 0 ˜Tk ( )−1 ˜k ( ) d are bounded for all t ∈ [0; T ] are equipped with incremental encoders to measure the joint
and all k ∈ Z+ . Since  is continuous, the boundedness of positions. In general, the joint velocities are not measured
t ˆT
0 k
 ( )−1 ˆk ( ) d is guaranteed. Therefore, k (t) ∈ L2e but estimated from the joint positions using a 3ltered deriva-
for all k ∈ Z+ and so is qLk (t). tive (e.g., sq̃k =(1 + Tc s)). However, the measurement noise
Now, using the fact that ampli3cation, due to the derivative action, will accumulate
 t 2 through the iterative process. A potential solution to this
q̃k (t) − q̃k (0)2 = q̃˙k ( ) d ; (34) problem is to design a P-type iterative parametric updating
0 law that does not require the joint velocities measurements.
and using Schwartz inequality, we obtain In this case, the noise e7ect will be reduced considerably,
 t  t but will not be totally eliminated since the measurement
q̃k (t) − q̃k (0)2 6 q̃˙Tk ( )q̃˙k ( ) d 12 d noise from the joint positions will accumulate from itera-
0 0 tion to iteration. Consequently, in practical applications, it is
 A. Tayebi / Automatica 40 (2004) 1195 – 1203 1201

important to stop the learning process after a certain num-

ber of iterations once the tracking error reaches a certain
acceptable level. This crucial point will be investigated in
our future research.

Remark 7. It is worth noting that the sign function used in

the proposed control laws might lead to the chattering phe-
nomenon. In practical applications, the sign function can be
replaced by a continuous approximation (e.g., saturation) in
order to smooth out the control input and reduce the chatter-
ing. In fact, if we substitute in our algorithms the sign func-
tion by the following function: f(q̃˙k ) = q̃˙k =max{,; q̃˙k },
, ¿ 0, one can show that the tracking error converges to a
certain domain around zero, which can be made arbitrarily
small by decreasing ,. It might also be possible to show the Fig. 1. Sup-norm of the tracking error versus the number of iterations for
convergence to zero of the time-weighted norm (also called links 1 and 2 with the control law (2)–(3), with the resetting condition.
#-norm) of the tracking error, i.e., e−#t q̃k , for a suXciently
large #, under the saturation function. This point needs fur-
ther investigation and will be part of our future research. In
fact, in Cao and Xu (2002), a saturation-type learning vari-
able structure control, guaranteeing the convergence of the
#-norm of the tracking error to zero over a given 3nite-time
interval, was proposed for a class of uncertain nonlinear
systems. However, it is well known that the #-norm leads
generally to low convergence rates.

4. Simulation results

Let us consider a two degrees-of-freedom planar manip-

ulator with revolute joints described by (1). The matrix
M = [mij ]2×2 is given by m11 = m1 l2c1 + m2 (l21 + l2c2 +
Fig. 2. Sup-norm of the tracking error versus the number of iterations for
2l1 lc2 cos q2 )+I1 +I2 , m12 =m21 =m2 (l2c2 +l1 lc2 cos q2 )+I2 , links 1 and 2 with the control law (15)–(16), with the resetting condition.
and m22 = m2 l2c2 + I2 . The matrix C = [cij ]2×2 is given
by c11 = hq̇2 , c12 = hq̇1 + hq̇2 , c21 = −hq̇1 and c22 = 0,
where h=−m2 l1 lc2 sin q2 . The vector G =[G1 ; G2 ]T is given
by G1 = (m1 lc1 + m2 l1 )g cos q1 + m2 lc2 g cos(q1 + q2 ) and,
G2 = m2 lc2 g cos(q1 + q2 ). The robot parameters are given
by m1 = m2 = 1 Kg, l1 = l2 = 0:5 m, lc1 = lc2 = 0:25 m,
I1 = I2 = 0:1 Kg:m2 , g = 9:81 m=s2 . The disturbances are
assumed to be time-varying and also varying from itera-
tion to iteration as follows d1 = d2 = rand(k) sin(t), where
rand(k) is a random function taking its values between 0
and 1. The matrix (q; q̇; q̃˙k ) = [ij ]2×5 is given by 11 =
q̇2 sin q2 , 21 = −q̇1 sin q2 , 12 = 11 − 21 , 22 = 23 = 0,
13 = cos q1 , 14 = 24 = cos(q1 + q2 ), 15 = sgn(q̃˙1 )
and 25 = sgn(q̃˙2 ).
The desired reference trajectories for q1 and q2 are chosen
as q1; d (t) = sin(24t) and q2; d (t) = cos(24t) over the time
interval [0; 1s]. Applying the control law (2)–(3), with the Fig. 3. Sup-norm of the tracking error versus the number of iterations for
resetting condition, with KP = KD = 10I2×2 and  = 10I5×5 , links 1 and 2 with the control law (22)–(23), with the resetting condition.
where Ii×i is an i × i identity matrix, we obtain the re-
sult shown in Fig. 1. Applying the control law (15)–(16),
with the resetting condition, with KP = KD = 10I2×2 and KP = KD = 10I2×2 and * = 10 we obtain the result shown in
 = 10I2×2 we obtain the result shown in Fig. 2. Applying Fig. 3. Applying the control law (15)–(16), with the align-
the control law (22)–(23), with the resetting condition, with ment condition, with KP = 500I2×2 , KD =  = 10I2×2 , and
1202 A. Tayebi / Automatica 40 (2004) 1195 – 1203

References

Arimoto, S. (1996). Control theory of non-linear mechanical systems.

Oxford, UK: Oxford Science Publications.
Arimoto, S., Kawamura, S., & Miyazaki, F. (1984). Bettering operation
of robots by learning. Journal of Robotic Systems, 1, 123–140.
Bondi, P., Casalino, G., & Gambardella, L. (1988). On the iterative
learning control theory for robotic manipulators. IEEE Journal of
Robotics and Automation, 4(1), 14–22.
Cao, W.-J., & Xu, J.-X. (2002). On functional approximation of the
equivalent control using learning variable structure control. IEEE
Transactions on Automatic Control, 47(5), 824–830.
Casalino, G., & Bartolini, G. (1984). A learning procedure for the
control of movements of robotic manipulators. IASTED symposium
on robotics and automation, Amsterdam (pp. 108–111).
Choi, J. Y., & Lee, J. S. (2000). Adaptive iterative learning control
for uncertain robotic systems. IEE-Proceedings Control Theory and
Fig. 4. Sup-norm of the tracking error versus the number of iterations Applications, 147(2), 217–223.
for links 1 and 2 with the control law (15)–(16), with the alignment Craig, J. J. (1984). Adaptive control of manipulators through repeated
condition. trials. Proceedings of the American control conference, San Diego
(pp. 1566–1574).
De Luca, A., Paesano, G., & Ulivi, G. (1992). A frequency domain
approach to learning control: Implementation for a robot manipulator.
q1; 0 (0) = q2; 0 (0) = q̇1; 0 (0) = q̇2; 0 (0) = 0, we obtain the result IEEE Transactions On Industrial Electronics, 39(1), 1–10.
shown in Fig. 4. French, M., & Rogers, E. (2000). Nonlinear iterative learning by
an adaptive Lyapunov technique. International Journal of Control,
73(10), 840–850.
Ham, C., Qu, Z., & Johnson, R. (2000). A nonlinear iterative learning
5. Conclusion control for robot manipulators in the presence of actuator dynamics.
International Journal of Robotics and Automation, 15(3), 119–130.
Ham, C., Qu, Z. H., & Kaloust, J. H. (1995). Nonlinear learning control
Three adaptive ILC schemes have been proposed for the for a class of nonlinear systems based on Lyapunov’s direct method.
position tracking problem of rigid robot manipulators with Proceedings of the American control conference, Seattle, USA
unknown parameters and subject to external disturbances. (pp. 3024–3028).
The proposed controllers are based upon a PD feedback Horowitz, R. (1993). Learning control of robot manipulators. ASME
structure plus an iterative term designed to cope with the Journal of Dynamic Systems Measurements and Control, 115,
402–411.
unknown parameters and disturbances. The proof of con-
Kavli, T. (1992). Frequency domain synthesis of trajectory learning
vergence is based upon the use of a Lyapunov-like posi- controller for robot manipulators. Journal of Robotic Systems, 9(5),
tive de3nite sequence, which is shown to be monotonically 663–680.
decreasing under the proposed control schemes. The con- Kawamura, S., Miyazaki, F., & Arimoto, S. (1988). Realization of robot
trollers are very simple to design and to implement since motion based on a learning method. IEEE Transactions on Systems,
the only requirement on the PD and learning gains is the man and cybernetics, 18(1), 126–134.
Kuc, T. Y., Nam, K., & Lee, J. S. (1991). An iterative learning control of
positive de3niteness condition. The result in Theorem 2 is
robot manipulators. IEEE Transactions on Robotics and Automation,
particularly interesting since it does not require any a pri- 7(6), 835–841.
ori knowledge of the system dynamics, and the number of Moon, J. H., Doh, T. Y., & Chung, M. J. (1997). An iterative learning
iterative parameters used in this scheme is just two. Further- control scheme for manipulators. In Proceedings of international
more, in Theorem 3, we propose a solution to the problem conference on intelligent robots and systems, Grenoble, France.
using a single iterative variable at the expense of the knowl- Slotine, J. J., & Li, W. (1987). On the adaptive control of robot
manipulators. International Journal of Robotics Research, 6, 49–59.
edge of some bounds of the system parameters. Finally, the
Slotine, J. J., & Li, W. (1991). Applied nonlinear control. Englewood
relaxation of the resetting condition for a certain class of Cli7s, NJ: Prentice-Hall.
reference trajectories has been discussed and formulated in Spong, M. W., & Vidyasagar, M. (1989). Robot dynamics and control.
Theorem 4. New York: Wiley.
Takegaki, M., & Arimoto, S. (1981). A new feedback method for dynamic
control of manipulators. Journal of Dynamic Systems, Measurement
and Control, 103, 119–125.
Acknowledgements Tomei, P. (1991). Adaptive PD controller for robot manipulators. IEEE
Transactions on Robotics and Automation, 7, 565–570.
Xu, J.-X. (2002). The frontiers of iterative learning control—Part II.
The author would like to thank Dr. Jian-Xin Xu for the
Journal of Systems, Control and Information, 46(5), 233–243.
discussion about the paper (Xu, 2002), which motivated this Xu, J.-X., Badrinath, V., & Qu, Z. (2000). Robust learning control for
work. He would like also to thank the Associate Editor and robotic manipulators with an extension to a class of nonlinear systems.
the anonymous reviewers for their valuable comments. International Journal of Control, 73(10), 858–870.
 A. Tayebi / Automatica 40 (2004) 1195 – 1203 1203

Xu, J.-X., & Tan, Y. (2001). A suboptimal learning control scheme for as a Research Associate. In December 1999, he joined the department
nonlinear systems with time-varying parametric uncertainties. Journal of Electrical Engineering at Lakehead University, Canada, where he is
of Optimal Control–Applications and Theory, 22, 111–126. actually Associate Professor. His research interests are mainly related
to linear and nonlinear control theory including adaptive control, robust
control and iterative learning control, with applications to mobile robots,
robot manipulators, vertical take-o7 and landing aerial robots and electrical
Abdelhamid Tayebi received his B. Sc. in machines.
Electrical Engineering from Ecole Nationale
Polytechnique d’Alger, Algeria in 1992, his
M.Sc. (DEA) in robotics from Universit[e
Pierre & Marie Curie, Paris, France in
1993, and his Ph.D. in Robotics and Au-
tomatic Control from Universit[e d’Amiens,
France in December 1997. He pursued at
the same university as an ATER (Assistant
Professor) until September 1998, where he
joined the department of Computer Science
at the University of Quebec, Hull, Canada,