Professional Documents
Culture Documents
2000 - Longman - Iterative Learning Control and Repetitive Control For Engineering Practice PDF
2000 - Longman - Iterative Learning Control and Repetitive Control For Engineering Practice PDF
RICHARD W. LONGMAN{
This paper discusses linear iterative learning and repetitive control, presenting general purpose control laws with only a
few parameters to tune. The method of tuning them is straightforward, making tuning easy for the practicing control
engineer. The approach can then serve the same function for learning/repetitive control, as PID controllers do in classical
control. Anytime one has a controller that is to perform the same tacking command repeatedly, one simply uses such a
law to adjust the command given to an existing feedback controller and achieves a substantial decrease in tracking error.
Experiments with the method show that decreases by a factor between 100 and 1000 in the RMS tracking error on a
commercial robot, performing a high speed trajectory can easily be obtained in 8 to 12 trials for learning. It is shown that
in engineering practice, the same design criteria apply to learning control as apply to repetitive control. Although the
conditions for stability are very di erent for the two problems, one must impose a good transient condition, and once
such a condition is formulated, it is likely to be the same for both learning and repetitive control.
International Journal of Control ISSN 0020± 7179 print/ISSN 1366± 5820 online # 2000 Taylor & Francis Ltd
http ://www.tandf.co.uk/journals
Iterative learning control and repetitive control 931
action. Storing this information, retrieving it, and per- and any practicing control system designer will feel very
haps processing it, requires the use of a digital computer comfortable working with them. Then the ILC/RC
on line. Hence, it is best to start with a discrete time should be designed to work based as directly as possible
ILC/RC formulation. on this description of system behaviour. This avoids
the usual process of trying to ® t the data with a math-
2.3. Use the existing feedback controller ematical model (e.g. a higher order di erence equation
model, or a state space di erence equation model). Such
Much of learning control theory chooses to simul-
models fail to exactly ® t data introducing extra discre-
taneously create the feedback control law and the learn-
pancy in performance by comparison to the real world,
ing control law. Such an approach will highly constrain
and they constrain the type of behaviour, for example by
the number of practical applications. Here we work with
choice of the order of the model. By staying as close as
methods that can apply to any existing feedback control
possible to the data in the frequency response plot these
system, and simply learn to improve its performance. In
sources of error are eliminated.
the case of robots or other purchased hardware, using a
law that simultaneously creates both, requires that the
2.7. Guarantee good learning transients
manuf acturer be the one to implement ILC/RC. Here, it
can simply be the user of the robot that applies ILC/RC. It is shown that simply satisf ying the true stability
There are many more users than manuf acturers. condition for ILC is of little value in ensuring a practical
learning process. Having direct control over the learning
transients in ILC/RC is fundamental. The ILC/RC laws
2.4. Adjust the command to the feedback controller, not
described here, address this by aiming for monotonic
the manipulated variable
decay of all frequency components of the error (up to
Often in ILC and RC the manipulated variable, e.g. a chosen cut-o frequency). The designer knows the rate
the torque applied to a robot link, is adjusted by the of decay for each frequency, and can make adjustments
learning process. This requires going into the existing to in¯ uence the rate.
controller and modifying the signal it sends to actuators.
Here such complexity is avoided. The ILC/RC simply
2.8. Guarantee long term stability
adjusts the command given to the existing feedback con-
trol system. One can show that these two approaches are It is demonstrated below that linear ILC/RC can
essentially equivalent mathematically (Solcz and easily exhibit very long term instabilities, not evident
Longman 1992) so here the approach that is by far the for perhaps thousands of repetitions. The ILC/RC
easiest to apply is chosen. laws give the designer methods to kill such instabilities,
by adjusting the appropriate parameter(s).
2.5. Make simple IL C/RC laws with a small number of
2.9. Make an impedance match with the knowledge base
param eters to adjust
of practicing control engineers
The largest classes of feedback control laws in prac-
To maximize the use of ILC/RC in engineering prac-
tice are, proportional (P), integral (I), proportional plus
tice, making such an impedance match allows the con-
derivative (PD) and PID. These controllers are simple,
trol system designer to grasp the concepts fast and know
require the control system designer to adjust only one,
how to apply them immediately. By connecting to the
two, or three parameters, and the manner of doing the
Bode plot information, we create such a match.
adjustment is relatively straightforward. The approach
here aims to do something analogous for ILC/RC, pre-
senting laws with only a few parameters to adjust in well 3. Two test beds for experimental demonstrations
de® ned ways.
To demonstrate the e ectiveness of ILC/RC laws,
experimental results are presented from two experi-
2.6. Make use of typical knowledge about the feedback mental test beds. ILC tests are on the robot in ® gure
control system behaviour 1, a Robotics Research Corporation K-Series 8071HP
One should not try to make a universal learning seven degree-of-f reedom robot. The same desired trajec-
controller that (in theory) works on all systems. To get tory is speci® ed for each link, a cycloidal path increasing
good performance in practice, the ILC/RC design the joint angle from zero to 908, and then the cycloidal
should make use of easily obtainable information path is reversed to come back to zero again (the end-
about the existing feedback control system dynamics. points are connected by a polynomial having zero ® rst
It is reasonable that the designer ® rst perform a fre- and second derivatives at each end). When all seven
quency response test to produce Bode plots going up joints are executing this curve at the same time, it creates
to the Nyquist frequency. These plots are easy to obtain, a large motion through the workspace. The timing of the
Iterative learning control and repetitive control 933
u j ˆ ‰ uj …0† uj …1† ¢¢¢ uj … p ¡ 1†ŠT …2† 4.3. Di erent forms of the learning gain matrix
The simplest form of linear iterative learning control
y j ˆ ‰ yj …1† yj …2† ¢¢¢ yj … p†ŠT …3† can be termed integral control based learning. It uses a
discrete version of an integral, running in repetitions,
Underbars are used analogously on other variables. The one `integral’ for each time step, and adds this to the
subscript j indicates the jth repetition. At repetition 0 command. As in integral control, it will not allow the
one normally applies the desired command as the input error at that time step to remain constant, building up
u0 . Then, a rather general linear learning control law the corrective action until the error goes away (or the
makes a change in the command each repetition that process goes unstable). The control law is
is a linear combination of the error e j¡1 measured in
the previous repetition, where e j ˆ y ¤ ¡ y j , i.e. j
X
uj‡1 …k † ˆ uj …k† ‡ ¿ej …k ‡ 1† ˆ u0 …k† ‡ ¿ e` …k ‡ 1†
u j ˆ u j¡1 ‡ L e j¡1 …4† `ˆ0
state frequency behaviour. The partial isometry diagonal blocks raised to the jth power. A generic
- -
law produced the same monotonic decay on such block has the form …¶I ‡ J † j where J is all zero
trajectories that are so short that steady state except for a diagonal of ones immediately above the true
- -
frequency response thinking is not appropriate. diagonal. Note that if J is an r by r matrix, then J r is
(v) The linear phase lead learning control law is like zero, and
integral control based learning, but the non-zero - - -
…¶I ‡ J† j ˆ ¶ j I ‡ a1¶ j¡1 J ‡ ¢ ¢ ¢ ‡ ar ¶ j¡r‡1 Jr¡1 …11†
diagonal is shifted toward the upper right of the
matrix (Wang and Longman 1996). The linear for appropriate combinatorial values for the coe cients
phase lead acts as a simple compensator to pro- a1 and ar . Again, as j goes to in® nity, we need that the
duce monotonic decay up to higher frequencies. eigenvalue ¶ to be less than one in magnitude, in order
for all possible initial errors to converge to zero as the
(vi) The linear phase lead with triangular window-
repetitions progress.
ing has this same shift, but instead of having
only one non-zero diagonal, it has a non-zero
band (Wang and Longman 1996, Hsin et al. 5. The promise and failure of a universal learning
1997 b). The triangular window is used to pro- controller
duce a frequency cut o with no phase shift. All lower triangular forms of learning control law L ,
(vii) Linear phase lead with causal Butterworth low have a very powerful, robust stability result. Suppose
pass ® ltering ® lls up everything in the matrix the components on the diagonal of L are given by ¿.
below the upper shifted diagonal (Hsin et al. Then …I ¡ PL † is lower triangular with the entries on the
1997 a). The low pass ® lter again serves as a diagonal equal to 1 ¡ CB¿. The diagonal elements of a
cuto , and this time the linear phase lead is lower triangular matrix are the eigenvalues. Therefore,
adjusted to also compensate for the phase lag (10) says the ILC controller will be stable, and converge
of the Butterworth ® lter. to zero tracking error for all initial errors, if and only if
the learning gain ¿ satis® es
The ILC/RC laws suggested below for practical 0 < CB¿ < 2 …12†
application, modify these various forms of ILC by
including a frequency based cut-o of the learning. The following comments suggest that this is a very easy
This can be a ® lter on the error itself, or a ® lter of the condition to satisfy.
command just before it is applied to the system, or a (i) Typical systems have CB 6ˆ 0. When (1) comes
® lter of just the learning control part of the command from discretizing a continuous time system, it
before it is applied. will not normally be zero. The value CB is the
response at time step one for a unit pulse input
4.4. A necessary and su cient condition for at step zero. During this time step, the input ±
convergence to zero tracking error in IL C output relationship of the continuous time
From (8), a chosen learning gain matrix L will cause system is the unit step response, which for
all possible initial tracking errors to converge to zero, if typical asymptotically stable systems will not
and only if all eigenvalues of the matrix …I ¡ PL † are less cross zero no matter how long you wait before
than one in magnitude the next time step starts. When the control
system is digital, in most cases one can make
j¶ i …I ¡ PL †j < 1 8i …10† the same conclusion. The possible exception to
In order to be complete, it is shown here that this de® nes CB 6ˆ 0 occurs when there are extra delays in
the true stability boundary for linear ILC laws (4). All the system, for example from having more
square matrices are either diagonalizable by a similarity than one digital part in the loop. In this case
transf ormation, or they can be put in Jordan canonical one can determine what the true time delay is
form. Suppose ® rst that M diagonalizes …I ¡ PL † ˆ in the system, and then reformulate the problem
M ¡1 LM, and de® ne e¤j ˆ Me j . Since M is invertible, so that the desired trajectory is speci® ed starting
e j is zero if and only if e¤j is zero. Writing (8) in terms from the ® rst time step for which an input at
of this transf ormed error gives e¤j ˆ L j e0¤, from which it zero can in¯ uence the output. Then CAi B for
is clear that all eigenvalues on the diagonal of L must be the appropriate i takes the place of CB in (12),
less than one in magnitude for convergence to zero from and again we have the same strong result.
all initial errors. When a Jordan canonical form applies, (ii) Normally you know the sign of CB, since it is
…I ¡ PL † ˆ M ¡1 JM. The J j in e¤j ˆ J j e0¤ is block the unit step response after one sample time.
diagonal, and raising a square block diagonal matrix For a typical feedback control system, when a
to the jth power is the same as the matrix with its positive step input is commanded, the response
936 R. W . L ongman
step process. Represent the location of the `wave front’ , where ¿…0† is de® ned as the entry on the diagonal of L ,
the time step for which the errors at all previous time and the other ¿ are on the subdiagonals successively.
steps have become essentially zero, using Region Substitute this into (19), and use a very small " so that
3 in ® gure 7. Thus Region 3 de® nes a transient region one can neglect terms that depend on z ¡1 compared to
for the learning process that grows, and once Region 3 terms independent of z, and then (19) says that we need
is larger than Region 1, the steady state frequency to satisfy j1 ¡ CB¿…0†j < 1, which is the same as (12) (or
response modelling apples only in Region 4. What hap- (10) specialized to lower triangular L ). We conclude that
pens in Region 4 is irrelevant to convergence of the equations (10), (12) and (19) de® ne the true stability
learning process, and hence condition (16) which only boundary, and the frequency response condition (16) is
applies in Region 4 is irrelevant to convergence. Region a su cient condition for stability. We conclude : by the
3 will eventually grow to include everything. development of (16), satisf ying (16) means that there will
In learning control the di erence between the fre- be monotonic decay in Region 4. In addition, the above
quency response condition for convergence (16) and discussion says that satisfying (16) indicates that Region
the true stability boundary given by (12), is very large, 3 grows to ® ll the whole time interval, although the error
with one depending on the system dynamics in A and history in this region need not be monotonic. Thus,
the other not. In fact, it is not obvious that satisf ying there are two mechanisms for convergence operating
(16) necessarily implies convergence. For example, per- simultaneously, when (16) is satis® ed.
haps the settling time is longer than p time steps, so that
no part of the trajectory can be modelled by steady state 6.3. Relationship to the exact Euclidean norm
frequency response. To determine the status of the fre- monotonic decay condition
quency response condition (16) as a stability condition
The frequency response condition (16) indicates
(Elci et al. 1994 b), sum the geometric series in (15) to
monotonic decay in Region 4, but not for the entire
obtain
trajectory. In this section we consider what is required
9
Ej …z† ˆ ‰ 1 ¡ zF…z†G…z†Š j E0 …z† > to guarantee monotonic decay for the entire trajectory,
>
>
= in terms of some error norm. One considers the
j
Y j …z† ˆ ‰ 1 ¡ zF…z†G…z†Š Y 0 …z† …18† Euclidean norm as the most natural (Jang and
>
>
j >
; Longman 1994, 1996 a). One can of course examine
‡ ‰ 1 ¡ …1 ¡ zF…z†G…z†† ŠY ¤ …z† the maximum component norm as in Oh et al. (1997),
The z-transform of a function is a power series in the or other norms. De® ne the matrix H ˆ I ¡ PL , and
complex variable z¡1 . If the error Ej …z† and the response suppose that it has a singular value decomposition
Y j …z† are going to relate to a convergent function in the H ˆ USV T where U and V are unitary matrices and S
limit as repetitions j goes to in® nity, then there must be is the diagonal matrix of non-negative singular values ¼i .
a non-zero circle of convergence, call its radius ". Then equation (8) can be written as U T e j ˆ SV T e j¡1 .
Therefore, in order for (18) to de® ne convergent func- Multiplication of an error by a unitary matrix does
tions in the limit, it is necessary that not change the Euclidean norm of the error, so that
the largest the norm of e j can be is the norm of ej¡1
j1 ¡ zF…z†G…z†j < 1 8jz ¡1 j < " for some " > 0 multiplied by the largest singular value, i.e.
…19† ke j k µ max …¼i †ke j ¡1 k …21†
i
Once the functions are de® ned in such a region, then
they can be extended to the rest of the complex plane Various iterative learning control laws have been
by analytic continuation. Note that if (16) is satis® ed, developed to satisfy this condition. This includes the
then (19) is automatically satis® ed using " ˆ 1 corre- contraction mapping learning control law of Jang
sponding to the unit circle in the z plane. Therefore, et al. (1994, 1996 a) and Lee-Glauser et al. (1996) that
one concludes that satisf ying (16) also guarantees con- uses the learning matrix L ˆ sPT with the scalar s
vergence, even though the development given in the pre- su ciently small (note that the comments on robustness
vious section based on steady state frequency response of this approach are not valid). It also includes the
thinking only applies to Region 4. learning control law based on a partial isometry
Now examine how (19) is related to (10) and (12), all (Jang and Longman 1996 b), with L ˆ sV P UPT where
of which describe the true stability boundary. Expand V P and UP are from the singular value decomposition
G…z† and F…z† in power series of z¡1 to obtain of P.
Now consider the relationship between the true
)
G…z† ˆ CBz¡1 ‡ CABz¡2 ‡ CA2 Bz¡3 ‡ ¢ ¢ ¢ monotonic decay condition (21) and the approximate
…20† monotonic decay condition (16) developed based on
zF…z† ˆ ¿…0†z ‡ ¿…¡1†z0 ‡ ¿…¡2†z¡1 ‡ ¢ ¢ ¢ steady state frequency response (Jang and Longman
Iterative learning control and repetitive control 939
1996 b). Again consider matrices L that are lower trian- ej … p† ˆ …I ¡ CB¿†e j¡1 … p† ¡ CAB¿ej¡1 … p ¡ 1† ¡ ¢ ¢ ¢
gular with all entries in each individual diagonal being
¡ CAp¡1 B¿ej¡1 …1† …22†
identical, so that they correspond to causal time invar-
iant ® lters. Then H ˆ I ¡ PL is lower triangular and has
Unlike the error at step 1 in equation (17), this equation
the same identical entry property. Analogous to the ® rst
has p terms summed together. The number of terms that
of (20), there is associated with this matrix, a Laurent
are signi® cant is limited by how long it takes for the unit
series H …z† ˆ h0 ‡ h1 z¡1 ‡ h2 z ¡2 ‡ ¢ ¢ ¢, where we intro-
pulse response to become negligible, i.e. what power i is
duce the z argument to distinguish the series from the
needed before Ai is negligible. With a fast sample time,
matrix. The transpose of matrix H is upper triangular,
this can easily be a very large number of terms.
and has associated with it a Laurent series H …z ¡1 † ˆ
Increasing the sample rate can make the behaviour
h0 ‡ h1 z1 ‡ h2 z2 ‡ ¢ ¢ ¢. The singular values of H are the
worse. With a settling time of one quarter of a second,
square roots of the eigenvalues of HH T . This matrix
and a sample time of 1000 Hz, there would be 250 terms.
does not have the identical entry property along each
So one can easily imagine the sum of all of these terms
diagonal. However, since the hi are Markov parameters
producing a very large error at the end of the trajectory,
of an asymptotically stable system (when G…z† and F…z†
while waiting for the wave of convergence to arrive. In
are asymptotically stable transf er functions) they Region 3 the error decreases, but in Region 4 the
become negligible eventually. Supposing that the matrix phenomenon of (22) can make large errors. What
is very large compared to the number of time steps happens in Region 4 has no in¯ uence on whether the
needed for the Markov parameters to become negligible system ultimately converges.
(e.g. one settling time), then the identical entry property Now study the convergence from a frequency
on diagonals is essentially met, and HH T becomes a response point of view, looking at Region 4. To satisfy
Toeplitz matrix. Toeplitz and SzegoÈ have proved prop- ®
(16), one needs to have the Nyquist plot of z F…z†G…z†
erties of Toeplitz matrices (Grenander and SzegoÈ 1958) for z going around the unit circle, stay inside a circle of
that the eigenvalues when the matrix is in® nite in size radius one centered at ‡1. (Note that this time we write
coincide with the set of complex values that the associ- ®
z . This explicitly indicates the linear phase lead ® ¡ 1,
ated Laurent series assumes at evenly spaced points and it is now explicit rather than buried in the de® nition
around the unit circle kzk ˆ 1. The Laurent series is of F.) The radial distance from the centre of this circle at
H …z†H …z¡1 †, and for z on the unit circle z¡1 corresponds ‡1 to a point on the Nyquist plot for frequency ! is
to the complex conjugate. Hence the singular values, as M …!†, and this is the factor by which the amplitude of
the matrix gets large, converge to the square root of the the error component at this frequency is multiplied each
magnitude of H …z†H …z¡1 † at the associated points on repetition. Hence, violating (16) for some frequencies
the unit circle. In other words, they converge to the will create monotonic growth of error in Region 4 for
magnitude M …!† of the frequency response of these components, and this error only goes away by
1 ¡ zF…z†G…z† at the evenly spaced discrete frequencies. having Region 3 grow to include all p time steps.
The conclusion is that as the number of time steps p gets Figure 8 considers the third order model of the feedback
large, the condition for monotonic decay of the controllers for a robot link described earlier. The curves
Euclidean norm of the error (21) converges to the con- show the amplitude growth or decay with repetitions for
dition (16). Stated in a di erent way, equation (21) is the
® nite time version of (16), where (16) only guarantees
monotonic decay in Region 4, while (21) guarantees it
throughout the p time steps.
condition for good learning transients The ® is one in (23), and this accounts for the one time
Since it is so much harder to satisfy the su cient step delay in a typical digital control system between a
stability condition (16) than it is (10) or (12), one may change in the input and its ® rst in¯ uence on the sampled
ask, do I really need to satisfy (16) in my ILC design? In output. The ® is set to higher values to produce linear
engineering practice one not only needs convergence, phase lead repetitive controllers. A shift of a chosen
but one needs reasonable transients. Condition (16) number of time steps is a larger percentage of a period
indicates monotonic decay of each frequency compon- as the frequency is increased, making this ® produce a
ent of the error, when the settling time of the feedback linear phase lead in the frequency response. The i0 is
controller is not long compared to the p time steps of the usually negative. The summation is a convolution sum,
desired trajectory. Monotonic decay is good behaviour. but an alternative is to process the error through a linear
But could I still get good transients by satisfying some di erence equation, which might be a compensator or
Iterative learning control and repetitive control 941
a causal Butterworth ® lter. There is of course an real time. One can compute between repetitions, or
equivalence between a di erence equation solution and simply apply the result at the next repetition once it is
its convolution sum solution, but the second option can ready.
be written explicitly as Sometimes the time domain multiplication of matrix
L with the error contains many multiplies. An alternative
u…k† ˆ u…k ¡ p† ‡ e^…k ¡ p ‡ ® † …25† computation for matrices of the form (28), is to perform
e…k ¡ p ‡ ® ¡ 1† ‡ ¢ ¢ ¢
e^…k ¡ p ‡ ® † ˆ ¬1 ^ the product in the z-transf orm domain. Since the
product represents a convolution sum, one simply
‡ ¬n^e^…k ¡ p ‡ ® ¡ n^†
takes the transf orms of the error and the ¿…k†, multiplies
¡ 1 e…k ¡ p ‡ ® ¡ 1† ¡ ¢ ¢ ¢ them together and takes the inverse transform (Lee-
Glauser et al. 1996). The advantage is that the computa-
¡ n^e…k ¡ p ‡ ® ¡ n^† …26† tion can be much faster. The disadvantage is that it
The z-transf orm of the discrete time sequence of gains in ignores edge e ects from the ® nite size of the matrix.
(24) is
7.3. Necessary and su cient conditions for convergence
F…z† ˆ ¿…i0 †z¡i0 ‡ ¢ ¢ ¢ ‡ ¿…0†z0 ‡ ¢ ¢ ¢ ‡ ¿…if †z ¡if …27†
to zero tracking error in RC
In the case of (25) and (26), the corresponding F…z† is
7.3.1. T he error di erence equations. In order to
the z-transf er function of the compensator or ® lter (26).
understand the error history for RC laws, we create
When needed, we denote F…z† in terms of its ratio of
the z-transf er functions from the desired trajectory and
polynomials, F…z† ˆ FN …z†=FD …z†.
the repetitive disturbance to the resulting tracking
error, and interpret the result in terms of the
7.2. Comparison between IL C and RC formulations corresponding di erence equation. The output is
This section examines the relationship between the Y …z† ˆ G…z†U…z† ‡ V …z†, where V …z† ˆ C…zI ¡ A†¡1 W …z†
repetitive control laws discussed here, and the learning is the output from the periodic disturbance. The V …z†
gain matrix of (4). The matrix L analogous to the RC can be considered periodic, since any non-periodic
law (24) is terms are part of the transient response of the presum-
2 ¿…1 ¡ ® † ¢ ¢ ¢ ¿…if † ¢¢¢ 0
3 ably stable feedback control system, and thus go to
6 zero. The transf orm of (24) or (25) and (26) gives
6 .. .. .. .. .. 7
7
6 . . . . . 7 z® F…z†
6 7 U …z† ˆ E …z† …29†
L 6 ¢ ¢ ¢ ¿…1 ¡ † ¢ ¢ ¢ ¿…if † 7 …28†
ˆ 6 ¿…i0 † ®
7 zp ¡ 1
6 7
6
4
.. .. .. .. .. 7
5 where the error is de® ned as E …z† ˆ Y ¤…z† ¡ Y …z†. Write
. . . . .
G…z† in terms of its numerator and denominator
0 ¢¢¢ ¿…i0† ¢¢¢ ¿…1 ¡ ® † polynomials as G…z† ˆ N …z†=D …z†. Combining these
The values of i0 and if might be so large that the matrix produces
is ® lled. Or it may start ® lled, but when the gains get ‰ …zp ¡ 1†D …z†FD …z† ‡ z® N…z†FN …z†ŠE …z†
small enough they are truncated. When (25) and (26) is
used, the triangle of zeros in the upper right remain, but ˆ ‰ D …z†FD …z†…zp ¡ 1†Š‰ Y ¤ …z† ‡ V …z†Š …30†
the rest of the matrix will be ® lled by the pulse response This can be converted to a di erence equation by noting
history of the ® lter (26), which again might be truncated that the terms in square brackets are polynomials, and
if needed. Thus, the general repetitive control laws that a term such as zp E …z† represents e…k ‡ p† in the time
discussed above can correspond to general learning domain. For periodic commands Y ¤ …z† and disturb-
control gain matrices L , with the stipulation that the ances V …z† of period p, the right-hand side of (30) is
gains in the matrix within each diagonal are all the zero, since …zp ¡ 1†Y ¤…z† is the command shifted one
same. Without this property, the corresponding RC period ahead, minus the command at the present time.
version would produce time varying repetitive control The same is true of the disturbance term once it reaches
laws. Most ILC laws have this property, for example, steady state. Hence the error history is the solution of a
learning laws based on frequency response thinking have homogeneous di erence equation.
the property automatically. One exception to this rule is
the learning control law based on a partial isometry 7.3.2. Conclusions from the time domain viewpoint. The
(Jang and Longman 1996 b). Hence, most RC laws tracking error will converge to zero for all initial con-
have a direct ILC counterpart, and vice versa. The ditions (all initial values of the error during period/
main di erences are the ® nite time nature of ILC, and repetition zero), if and only if all roots of the charac-
the fact that in learning control one need not compute in teristic polynomial
942 R. W . L ongman
…zp ¡ 1†D …z†FD …z† ‡ z® N…z†FN …z† ˆ 0 …31† in P¤ …z† will usually be much larger than the order of
the numerator because for causal systems G…z† and F…z†
are inside the unit circle. Unfortunately, the high order
will generically have one more pole than zero, and then
of this polynomial produced by the number of time steps
p is usually a large number. Hence, only ® that are so
p in a period, usually makes it impractical to use this
large that they send one into the next period would
condition directly to determine stability, e.g. by actually
violate this condition. With the denominator order
® nding the roots, or by use of the Jury test or the Routh
larger, the part of the contour at in® nity maps into
tests with bilinear transf ormation. It is somewhat easier
zero. Also, the part of the contour along the branch
to use the departure angle condition from root locus
cut cancels.
plot rules to determine if su ciently small repetitive
Under the above stated conditions, one simply needs
control gains produce a process that converges to zero
to determine the number of times the plot of P ¤ …z†
tracking error. Longman and Solcz (1990) start from a
encircles ‡1 for z going around the unit circle in order
base case of a system with a single pole at the origin (on
to determine stability of the repetitive control system.
top of the zero of the repetitive controller). This case has
Simply look at the frequency response of P ¤ …z† by sub-
a root locus where all roots starting on the unit circle
stituting z ˆ exp …i!T † with T being the sample time,
move radially inward. Then the pole at the origin is
and letting ! go up to Nyquist frequency. If this plot
moved to the location of a true pole of the system,
together with its complex conjugate re¯ ection, does not
and the remaining poles and zeros are introduced. It is
encircle the point ‡1, then the repetitive control system
reasonably simple to evaluate the change in departure
is asymptotically stable, and converges to zero tracking
angle from radially inward that occurs from moving the
error for all possible initial errors.
one pole and adding the remaining poles and zeros. If
this change is less than §908 for all poles on the unit
circle (and the feedback control system is stable) then 7.4. Approximate condition for monotonic decay in RC
su ciently small repetitive control gains produce con- This section and the next section are analogous to
vergence to zero tracking error. }} 6.1 and 6.2 and follow Huang and Longman (1996).
7.3.3. Conclusions from frequency domain stability One can rearrange the terms in equation (31) to read
theory. Now consider the somewhat more practical zp E …z† ˆ ‰ 1 ¡ z F…z†G…z†ŠE …z†
®
…32†
use of frequency response conditions based on Nyquist
stability theory. Divide (23) by zp D …z†FD …z†, to obtain One might think of E …z† as the transf orm of the error at
P…z† ² 1 ¡ z¡p …1 ¡ z® F…z†G…z†† ˆ 0. The numerator of one repetition, and then zp E …z† shifts one period giving
this expression after being put over a common the error in the next repetition. By setting z ˆ exp …i!T †
denominator is the actual characteristic equation of in the square brackets in (32) it becomes the frequency
interest, and for asymptotic stability we want no roots transf er function from one repetition to the next. The
outside or on the unit circle. The denominator equation suggests that requiring
zp D …z†FD …z† contains the characteristic equation of the j1 ¡ ei!T ® F…ei!T †G…ei!T †j < 1 …33†
closed loop feedback control system, and the charac-
teristic equation of any compensator we choose to use for all ! up to Nyquist frequency, would then establish
in the repetitive control law. Both should correspond convergence to zero tracking error. This is the same
to asymptotically stable systems, so that we know a condition for RC as (16) is for ILC (although this
priori there are no roots outside the unit circle. Form time derived with any linear phase lead ® pulled out of
a contour in the complex z plane that includes all of the F…z††.
the plane outside the unit circle, by going around the This thinking is of course quite heuristic (the next
unit circle in a counterclockwise direction, out to in® - section gives a precise conclusion). A frequency transf er
nity along a branch cut on the negative real axis, then function applies only to steady state response, so
circle at in® nity in a clockwise direction, and return technically, one cannot use it except after the error has
along the branch cut. Then for each root of the converged. E …z† is not the transf orm of the error for one
numerator characteristic equation outside the unit period, but the transform of the error for all time, as is
circle, the angle of P…z† will decrease by 2º after tra- zp E …z†. Hence, the monotonic decay argument only
versing the contour. If there is no change in the angle makes sense in a quasi steady state situation. If one
of P…z† then the learning process is convergent. makes signi® cant changes in the repetitive control signal
Now consider instead plotting P ¤ …z† ˆ z¡p …1¡ during each period, then it is certainly not steady state.
®
z F…z†G…z†† for z going around the contour. Then In addition, altering the repetitive control signal toward
the number of times this plot encircles the point ‡1 in the end of the previous period changes the initial
a clockwise direction is equal to the number of roots conditions for the next. Referring to ® gure 7 as a
outside the unit circle. The order of the denominator representation of a period rather than a repetition, it is
Iterative learning control and repetitive control 943
clear that one again needs Region 2 to dominate over world, that initially the error decreases substantially,
Region 1 so that the overall behaviour can be viewed in and then it starts to grow, as if it were unstable.
terms of steady state frequency response thinking. Figure 6 demonstrates that this growth is not likely to
Hence, the hidden assumptions are to learn su ciently be an indicator of mathematical instability, but rather it
slowly and have a system with a settling time short com- is just a poor transient while waiting for Region 3 in
pared to a period. Under such conditions, satisfying (33) ® gure 7 to grow. In repetitive control, the same substan-
should produce monotonic decay of all frequency com- tial initial decay of the error is often observed. But this
ponents of the error. time, when the error starts to grow, it normally is an
indicator that the repetitive control system is unstable.
7.5. Relationship to the true stability boundary in RC The resetting of the initial condition in ILC creates
Region 3 and the stability that it produces. Repetitive
As in the previous ILC sections, an if and only if
control has no such mechanism, and hence, once it starts
condition de® ning the borderline of convergence to
to diverge, it will continue to do so.
zero tracking error has been developed, and a heuristic
Initially ILC and RC seemed very di erent, with
condition for monotonic decay of the repetitive control
very di erent formulations. The stability properties
transients has been developed. One now asks, does satis-
seemed unrelated. Here it is claimed that in practice
fying this heuristically derived monotonic decay con-
there is really no di erence between linear ILC and
dition actually guarantee convergence to zero tracking
RC. To get practical results, one wants to satisfy the
error, independent of whether the su ciently slow learn-
same condition, (16) or (33), and hence there is really
ing and relative region size conditions described above
no distinction in designing and tuning the laws. A repe-
are satis® ed? Above it was determined that (under
titive control law for a given system can be used as a
appropriate assumptions) the repetitive control system
learning control law as well. Generally, if one is stable,
is asymptotically stable and converges to zero tracking
then the other is stable. On the other hand, if a learning
error for all initial errors, if and only if the plot of
control law has transients that are reasonable in engin-
P ¤…z† ˆ z¡p …1 ¡ z® F…z†G…z†† does not encircle the
eering practice, then it most likely satis® es the condition
point ‡ 1 for z going around the unit circle. For such
(16) or (33) and hence could serve as a repetitive control
z, jz¡p j ˆ 1. If one requires that (33) be satis® ed, then
law.
jP ¤ …z†j < 1 for all z on the unit circle contour, and one
Acknowledging this similarity, there are some
concludes that it would be impossible to encircle the
distinctions between ILC and RC: (1) A seemingly big
point ‡1. Therefore, condition (33) is a su cient con-
distinction is that in learning control one can consider
dition for asymptotic stability of RC.
the computation to be a batch computation performed
In ILC the true stability boundary (12) is unrelated
between repetitions. If there is not enough time between
to the system dynamics in matrix A, and the monotonic
repetitions, one can always apply the resulting control
decay condition (16) is heavily dependent on A. Hence,
update at the next repetition after the computation is
there is a large discrepancy between conditions for stab-
ready. This produces extra design freedom by compar-
ility and for monotonic decay. Repetitive control is
ison to typical repetitive control that operates in real
totally the opposite, the approximate monotonic decay
time. However, the distinction disappears when one uses
condition is essentially the same as the true stability
batch process repetitive control, i.e. repetitive control
boundary for typical values of p. Violating (33) can
laws that use batch computations applied when ready
very easily send the system unstable. Suppose that (33)
(Hsin et al. 1997 a). (2) The above equivalence is based
is violated for a range of frequencies D !. Then the phase
on satisfying (16) or (33), conditions that apply to time
change in P ¤…z† produced by z¡p over this interval is
invariant control laws. Learning control is free to make
D !pT . The number p is usually a large number, for ex- learning gain matrices that vary the entries going down
ample in a one second trajectory at a sample rate of
each diagonal. The ILC based on a partial isometry has
1000 Hz, p is 1000. If D !pT corresponds to more that
this form which is needed for monotonic convergence in
2º for some frequency below Nyquist, then ‡1 is
short trajectories. (3) A third item is that in repetitive
encircled, and the repetitive control process is unstable.
control there is an issue of keeping the corrective signal
Hence, satisfying (33) in repetitive control, is not just
in phase with the period of the error that is to be
important for good transients, it is almost a requirement
cancelled. There are three issues. One repetitive control
for convergence.
problem with no counterpart in ILC is that the period
may not be an integer number of sample times, making
7.6. A comment on the relationship between IL C and it hard to do the updates from the error at the corre-
RC in engineering practice sponding time in the previous period. When the period is
Section 6.4 described and explained the commonly determined by a periodic command, then one is likely to
observed property of learning control applied to the real be able to have an integer number of samples per period,
944 R. W . L ongman
but when it is determined by an external disturbance, case of a pure linear phase lead compensator
this becomes less likely. The second is that the period zR F…z† ˆ ¿z® where ¿ is the learning gain. One
may not be perfectly known and hence the repetitive adjusts the ® and the ¿. Larger values of ¿
control signal has to track the true period to keep in attempt to learn faster, but this e ect is limited.
phase. The third is that there are many repetitive control Lower values create smaller ® nal error levels in
problems where the period is determined by an external the presence of noise, and will allow a somewhat
source that drifts with time, creating a need for the higher frequency cuto . For any chosen ¿ one
repetitive control algorithm to track changes of period. adjusts ® to keep the plot satisfying (33) to as
Generally, it is best if one can have a signal that indi- high a frequency as possible. If one decides to
cates to the repetitive controller when the next period use a causal low pass ® lter in F…z† to improve the
starts. In the double reduction timing belt drive, an cuto (see } 11.4), then the ® is adjusted to com-
index pulse was used related to the output shaft angle pensate for phase lags from this ® lter as well as
encoder. In work on eliminating 60 Hz and harmonics G…z† to keep the product inside the unit circle
on the Je erson National Accelerator Facility 4 GeV (33) to as high a frequency as possible. Section
continuous electron beam, a trigger signal from the 10 discusses this process.
60 Hz supply is available (Longman et al. 2000). When (d) In practice it is unreasonable to demand that
no such signal is available, then one can aim to use (33) be satis® ed for all frequencies up to Nyquist.
pattern matching to keep things synchronized as in Hence, a cuto is made, to stop the learning
Longman and Huang (1994) . Another approach is to process for frequencies that violate condition
develop disturbance identi® cation methods as demon- (33). This means that we do not aim to get
strated in Edwards et al. (2000), but this starts to go zero tacking error, but rather aim for zero track-
outside the class of approaches being discussed here. ing error for the frequency components of the
The basis function repetitive control in Longman et error up to the cuto , aiming to leave frequency
al. (2000) is a useful way to deal with the period not components above the cuto unaltered. Section
being an integer number of sample times. 11 describes a set of choices of how to make a
cuto , and what to apply the cuto to.
8. The approach to designing practical linear learning
and repetitive controllers
The remainder of this paper makes use of the con- 9. Experiments to perform on the feedback control
clusions above in order to present a road map for system
designing simple and e ective learning and repetitive Experiments can easily be performed to produce
controllers. The design process works as follows: data for generating the frequency response G…ei!T † of
the feedback controller. Typically one applies a white
(a) Aim to satisfy the approximate monotonic decay
noise command to the system, takes the discrete
condition for the frequency components of the
Fourier transf orm of the output, divides it by the dis-
error (33). It is assumed that the system settling
crete Fourier transf orm of the input, to obtain the fre-
time is su ciently short that this condition is a
quency transf er function of the system G…ei!T †. This is
reasonable indicator of monotonic decay of the
the result one needs. Normally in obtaining the fre-
frequency components.
quency response one goes one step further and ® nds
(b) It is assumed that there is an existing feedback the magnitude M …!† and phase angle ³…!† and makes
controller, and that one can conduct frequency the Bode plots. Since performing frequency response
response tests on it. The manner of conducting tests is a standard procedure, there is considerable
these experiments is described in } 9. expertise available to produce clean results when the
®
(c) A suitable compensator z F…z† is chosen. This is data is particularly noisy. One can use long data records
accomplished by multiplying the chosen com- to improve the accuracy, and one can turn o the input
pensator transf er function, converted to a fre- after some time and allow the output to decay, in order
quency transf er function by substituting to decrease leakage e ects. Often an average of shifted
z ˆ ei!T , times the experimental frequency subsets of the data is used to prod-uce clean results.
response data G…ei!T †, to produce a plot of the Ideally one should ensure getting reproducible results
®
frequency response of z F…z†G…z†. This is done up to Nyquist frequency, but if this is di cult, the cuto
for various choices of compensator parameters, frequency can be chosen to cuto below where the
and then those parameters that keep the plot results become questionable. For more precise results
within the unit circle up to the highest frequency, one can do a sine sweep instead of putting in white noise
or up to a desired frequency, are chosen. In the which produces the most precise results.
Iterative learning control and repetitive control 945
response characteristics. There are ® lter initial conditions computation of a convolution sum ILC law as a prod-
for both applications, ® rst at the start of the data set, and uct in the transf ormed space (Lee-Glauser et al. 1996).
in the second application they correspond to the ® nal
11.2.3. Other options. There are some other options of
time in the data. In ILC this issue is addressed by extend-
how to prevent di culty from frequencies that go out-
ing the signal being ® ltered at both ends of the trajectory
side the unit circle. One is to stabilize by quantization,
in an appropriate way, using a su cient number of
limiting the number of digits used in the ILC/RC
points (Plotnik and Longman 2000). For example one
updates (Hsin et al. 1997 b,c), but a ® lter of the above
can simply repeat the endpoint for many time steps, or
type can still be helpful to cut out components above
pad with zeros. One can also do an extension that tries to
the cuto that accumulated from transient e ects
maintain continuity and continuity of the ® rst derivative
(Chen and Longman 1999). Another alternative is to
across the endpoints, by suitable re¯ ections of points
use experimentally determined matched basis functions
near the end. Note that using the default extensions in
(Oh et al. 1997, Wen et al. 1998, 2000, Frueh and
MATLAB may not be su cient, and the user should be
Phan 1999), where the choice of basis functions implies
prepared to perform his own extensions. In repetitive
a cuto . In the ILC problem this approach prevents
control, this extension problem is solved by using more
divergence, and in RC with slow enough learning it
real data, as described below.
also prevents divergence (Longman et al. 2000). A
The fact that this is a non-causal ® lter that cannot be
third alternative, with ILC laws such as phase cancel-
performed in real time, is not a problem in ILC where
lation, one can use on-line identi® cation of the phase
one can perform the computation between repetitions,
lag. When the phase lag information used in the learn-
or apply the result to the next repetition when ready. In
ing law is su ciently wrong to cause growth of error,
RC one can easily distribute the ® ltering computations
it also produces precisely the data needed to know
through the real time computations, and then a batch
how to ® x the problem (Elci et al. 1994 c, Longman
update of the RC action is made when the computation
and Wang 1996).
is ready. One approach is as follows (Hsin et al. 1997 b).
Start ® ltering in real time as the data arrives, ® ltering
11.3. W hat signal to cuto /predicting the steady state
from the beginning of one period and continuing
error levels
through three periods. Then continue ® ltering in real
time, but ® ltering the previous stored result backward Initially, one might simply think to use the cuto on
until ® nished at the end of the 6th repetition. Starting the error signal, cutting out the unwanted frequency
with the 7th repetition, apply the zero-phase ® ltered components before the ILC/RC law sees the error. If
result corresponding to the middle of the three periods there are no components left outside the unit circle,
® ltered. The other two periods just served to allow decay then they cannot grow. However, in the case of a zero-
of the ® lter transients both forward and backward (the phase IIR ® lter, the ® lter does not produce a perfect
smoothing procedures in Chen and Longman (1999) can cuto , but lets a small amount of signal above the cuto
also be used to give improved results) . If the compensa- through. Eventually this will accumulate, causing the
tor works well, and if F…z† contains a low pass ® lter as tracking error to grow. In applications, the learning pro-
discussed in } 11.4, then the time between batch ® ltering cess is used for only eight or 12 repetitions to ® nd what
updates can be made long, and between these updates, command to apply to a control system to give good
one continues to apply the RC law. For example, in one tracking performance. In this case long term stabiliz-
application on the timing belt drive testbed using a 12th ation is not an issue and ® ltering only the command is
order ® lter in F…z†, using a zero phase IIR ® lter every appropriate (Longman and Huang 1994).
2000 repetitions would be su cient (Hsin et al. 1997 b). When the learning is left on, one must be sure that
there is no accumulation of signal outside the unit circle.
11.2.2. Zero-phase cli ® lter. A batch computation There are two options (Longman and Songchon 1999).
that in theory forms an ideal cut-o is based on dis- One is to apply ® lter F to the total signal that is given as
crete Fourier transf orms. Transform the signal into its a command to the feedback controller. For simplicity,
frequency components, eliminate those above the cut- assume that this ® lter is applied to every repetition, as
o , and then transf orm back. Call this a cli ® lter, one would do in ILC with su cient time between repeti-
since theoretically it produces a perfect cli at the tions for the computations. The command for each repe-
desired frequency. Handling leakage e ects is done the tition is given by
same way as transients in the IIR ® ltering above (Plot- ®
Uj‡1 …z† ˆ F‰ Uj …z† ‡ z F…z†Ej …z†Š; U0 …z† ˆ Y ¤ …z†
nik and Longman 2000), either extending the end-
points of the signal in the case of ILC, or using more
…34†
data in the case of RC. There can be advantages in The second option is to leave out the desired trajectory
ILC to be in the frequency domain, speeding up the from the repeated ® ltering process and only ® lter the
950 R. W . L ongman
learning or repetitive control modi® cations to the com- attenuation of the ® lter. Once (38) is satis® ed, then we
mand that is produced by the ILC/RC learning process wish to know what the ® nal error level will be, as a result
9 of the forcing function. Once steady state error is
Uj …z† ˆ Y ¤…z† ‡ UL ;j …z† >
>
= reached, Ej …z† ˆ Ej¡1 …z† which we can call Ess …z†. The
UL ; j‡1 …z† ˆ F‰ UL ; j …z† ‡ F…z†Ej …z†Š; …35† transf er function from Y ¤…z† ¡ V …z† to Ess …z† with
>
>
; z ˆ exp …i!T † is
U0 …z† ˆ 0
1 ¡ MF2 …!†
In the case of a cli ® lter, F is unity up to the cuto H1 …!† ˆ …39†
frequency and zero above the cuto . For the zero phase 1 ¡ MF2 …!†‰ 1 ¡ G…ei!T † F…ei!T †Š
Butterworth ® lter, suppose the magnitude versus fre- Multiply the amplitudes of the frequency components of
quency function of the chosen order ® lter is MF …!†. Y ¤…z† ¡ V …z†, determined above experimentally, by
When applied in zero phase form, this magnitude jH1 …!†j to get the amplitude of the steady state error
change appears twice, and there is zero phase, making at each frequency above the cuto .
the frequency transf er function of the ® lter equal to
F …exp …i!T †† ˆ MF2 …!†. Now develop the convergence Option 2: The error at repetition j becomes
condition and the ® nal error level for each option. Ej …z† ˆ …1 ¡ G…z††Y ¤ …z† ¡ G…z†UL ;j …z† ¡ V …z†. Use this
in Ej …z† ¡ FEj¡1 …z† in order to create a term
Option 1: The error at repetition j is equal to UL ;j …z† ¡ FUL ;j ¡1 …z† and replace it by FF…z†Ej¡1 …z†.
Ej …z† ˆ Y ¤ …z† ¡ G…z†Uj …z† ¡ V …z†, and we wish to The result analogous to (36) is
combine this with (34) and eliminate U dependence so
that we have an equation for the error in terms of the Ej …z† ¡ F …1 ¡ G…z†F…z††Ej ¡1 …z†
two inputs, Y ¤ …z† and W …z†. To do this, write the dif- ˆ …1 ¡ F†‰ …1 ¡ G…z††Y ¤ …z† ¡ V …z†Š …40†
ference Ej …z† ¡ FEj¡1 …z† in order to create a term
Uj …z† ¡ FUj¡1 …z† that can be replaced by FF…z†Ej¡1 …z† Examining this equation, the condition for monotonic
according to (34). The result is decay of the learning transients is the same as (38). Also,
the frequency transf er function from disturbance V …z†
Ej …z† ¡ F …1 ¡ G…z†F…z††Ej¡1 …z† ˆ …1 ¡ F†…Y ¤…z† ¡ V …z†† to its steady state particular solution error is again
…36† H1 …!†. But the frequency transf er function from com-
mand to its steady state particular solution is di erent
The right-hand side represents a forcing function, and
the left-hand side represents a di erence equation. If ‰ 1 ¡ MF2 …!†Š‰ 1 ¡ G…ei!T †Š
there were no cuto , the right-hand side would be
H3 …!† ˆ …41†
1 ¡ MF2 …!†‰ 1 ¡ G…ei!T †F…ei!T †Š
zero, but the left-hand side homogeneous equation
would be unstable. By using a cuto , one produces Hence, if the point of the ILC/RC is to eliminate e ects
stability for the left-hand side, at the expense of leaving of repeating disturbances, with the command being zero
some forcing function on the right. In particular, for an or a constant, as is common in repetitive control
ideal ® lter, the right is zero below the cuto and passes problems, then there is no di erence between the two
the components above the cuto unchanged. And this options. On the other hand, when the main objective of
remaining part produces a steady state particular sol- the learning process is to eliminate tracking errors of the
ution. Examining the homogeneous equation, it can be feedback controller following a command, which one is
written as better depends on the dominant frequency range in the
command. If the dominant frequency content of the
Ej …z† ˆ ‰ F …1 ¡ G…z†z® F…z††ŠEj¡1 …z† …37† command is in a range for which the ratio
Hence, the monotonic decay condition for the learning jH3 …!†j=jH1 …!†j ˆ j1 ¡ G…ei!T †j …42†
transients when there is a cuto ® lter becomes
is greater than unity, then it is best to ® lter the desired
jMF2 …!†‰ 1 ¡ G…ei!T †ei®!T F…ei!T †Šj < 1 …38† trajectory, Option 1. If the dominant error is in a fre-
for all ! up to Nyquist frequency. One picks the order of quency range where this plot is below one, then use
the Butterworth ® lter and its cuto in order to satisf y Option 2. One applies (39) and (41) to the two inputs
this condition. Equation (38) can be interpreted as say- Y ¤…z† and V …z† to predict the ® nal error levels associ-
ated with the measured disturbance and the desired
ing, if the growth of signal components outside the unit
command, in order to decide which approach to use.
circle in 1 ¡ G…ei!T †ei®!T F…ei!T † times the attenuation
from the Butterworth ® lter MF2 …!† is less than one,
then convergence will result. When the ® lter is not 11.4. Introducing a real time ® ltering cuto
applied to every repetition, the corresponding condition In repetitive control one cannot apply the zero phase
is that the growth between ® ltering will be less than the low pass ® lter every period. Between applications,
Iterative learning control and repetitive control 951
components outside the unit circle will grow. Thus, it is grow to destabilize the learning, they in¯ uence the ® nal
best if the Nyquist plot does not go far outside the unit error level. Use of smooth updating procedures helps to
circle so that this growth is small. This can be accom- prevent such accumulations, and to get to a low ® nal
plished by using a causal IIR low pass ® lter in F…z†. error level (Chen and Longman 1999), but zero phase
Although in repetitive control one has this extra moti- ® ltering can ® lter out any accumulation from transients
vation to keep the plot from extending far outside the and produce the low ® nal error level in a more robust
unit circle, it can be desirable in learning control as well. manner.
For example, it may make it unnecessary to apply zero
phase ® ltering except occasionally. It can even eliminate
the need for zero phase ® ltering as described below. 12. Summary and conclusions
The tuning of the linear phase lead ILC/RC design In classical control there are routine control laws:
described in } 10.2 is modi® ed slightly in this new situa- proportional (P), integral (I), PI, proportional plus deri-
tion. The linear phase lead must be adjusted to not only vative (PD), and PID. Using these laws involves tuning
compensate for the phase lag in G…z† but that in F…z† as either one, two, or three parameters by rather routine
well. A simple design approach is to split ® into two procedures. Three types of ILC/RC laws have been
parts ® 1 ‡ ® 2 . Also write F…z† ˆ ¿F ^ …z† where ¿ is the described here, which again involve tuning only a
^
gain and F…z† is the Butterworth ® lter. Then adjust ® 2 small number of parameters. The method of tuning is
just as before, to keep z® 2 ¿G…z† within the unit circle up more straightf orward than for classical control. Hence,
to as high a frequency as possible. Having determined it is suggested that these laws can ® ll the same niche for
the needed cuto , then pick F ^ …z† as a high order ILC/RC as the P, I, PI, PD and PID do for classical
® ^
Butterworth ® lter, and adjust 1 to keep z 1 F
® …z† inside control. Application of the methods will always decrease
the unit circle to as high a frequency as possible. Then the tracking error, and normally they result in very sub-
combine the two to make the associated Nyquist plot. stantial reductions. Often one can approach the repeat-
This gets one into the right range of values for the phase ability level of the hardware. In the case of the robot
lead and the cuto . Some adjustment of these values experiments a factor of 100 improvement in tracking
may be needed to make a ® nal consistent design. An error was obtained for a high speed manoeuvre using
alternative to the Butterworth ® ltering is to use linear the ® rst and simplest of these laws. Approaching a fac-
phase lead coupled with a zero-phase triangular or tor of 1000 is possible with the other versions.
Bartlett window (Wang and Longman 1996, Hsin et al. The ® rst law is integral control based with a fre-
1997 b). In this case the linear phase lead is only needed quency cuto , the second adds a linear phase lead and
for G…z†. Triangular windows are much further from the third adds a causal low pass ® lter to help produced a
producing an ideal cuto and have side lobes as well. cuto . The methods presume there is a functioning feed-
But both approaches can be implemented in real time back control system. The ® rst step in designing the ILC/
and can produce good results. RC is to perform a frequency response test on this
One can often make the amount by which the plot system. In picking the ILC/RC gain ¿, it is easiest to
goes outside the unit circle very small, so that the maxi- simply pick a value in the range 1 =4 µ ¿ µ 1. Lower
mum size of the error associated with all frequencies values in this range favour smaller ® nal error levels in
outside the circle is less than some very small number. the presence of noise, and slightly higher cuto frequen-
By using a quantization of the error signal seen by the cies, while higher values favour faster learning. Then
ILC/RC that is larger than this value, one produces long multiply the experimental G…ei!T † data by the frequency
term stabilization, and no longer needs a zero phase transf er function of the chosen ILC/RC law and plot the
® lter. Hsin et al. (1997 c) give the details of how to result, noting at what frequency the plot leaves the unit
make ILC/RC that are stabilized based on quantization circle centred at ‡1. In the case of the linear phase lead
instead of zero phase ® ltering. Experiments are reported ILC/RC law one makes this plot for a range of linear
there where the quantization level needed for stabiliz- phase leads of one, two, three, etc. time steps and
ation was below the number of digits carried by the chooses the lead that keeps the plot inside the unit circle
digital to analogue, and analogue to digital conver- up to the highest frequency. This information allows
tersÐ roughly seven digits accuracy. Long term stabil- you to choose the cuto frequency. When using a causal
ization was demonstrated both analytically and in an low pass ® lter, there is some extra adjustment done in a
experiment run for 10 000 repetitions. Thus, it is possible similar manner, because the phase lead needs to com-
that this design approach totally eliminates the need for pensate for the causal low pass ® lter as well. Then
zero phase ® ltering. However, it is still advantageous to choose the method of cuto , IIR (use (38)) or cli ,
apply zero phase ® ltering, because during the learning make the decision of what to ® lter using (39), (41) and
process there are transients that contain frequency com- (42), and the design is done. Equations (39), (41) and
ponents outside the unit circle. Although these do not (42) predict the ® nal error levels reached by the design.
952 R. W . L ongman
If the cuto was chosen too high, for example sis on handling the non-linear equations for robotics,
because of inaccurate data, then the signature of this narrowing the possible users to a small set. It made
condition is clear. When using a zero phase IIR, one control laws that required the robot manuf acturer to
monitors the RMS error to see if it goes above a thresh- replace his whole control approach. Concerning linear
old, violating the monotonic decay property. In the case learning laws, the bad transients of integral control
of a cli ® lter, one can actually see what frequency is based learning made this simple approach impractical.
growing and correct the cuto accordingly. Therefore, For both learning and repetitive control there was the
one can monitor the system behaviour, and make phenomenon of apparent initial convergence, and then
corrections. Since the tuning is easy, one can create apparent or actual divergence, perhaps after many repe-
self tuning ILC/RC as discussed in Longman and titions. The possibility of such divergence is unsettling to
Wirkander (1998) and Wirkander and Longman (2000). an engineer who needs to deliver a working product.
If the compensation is good and a causal low pass These di culties prevented widespread use. This paper
® lter is used, as in the third law, one may be able to presents ILC/RC laws that address these issues. They
apply the zero phase ® ltering only occasionally (e.g. are applicable to most any linear feedback control
every 2000 repetitions in some experiments on the timing system. They are not specialized to robotics and do
belt would be su cient). In fact, with some quantization not require the feedback controller to have a speci® c
introduced, either intentionally or for example in the form. The laws are simple and it is very easy to tune
A/D or D/A converters, it is possible to eliminate the them. Since the di culties that prevented wide applica-
zero phase ® lter altogether. But using a zero phase ® lter tion of ILC/RC are now addressed, the time is ripe for
occasionally will normally produce better ® nal error an explosion of applications in engineering practice.
levels by eliminating some high frequency accumulation
from transients in the data. There is another class of
application where a zero phase cuto is not needed. References
These are applications in which one simply applies A RIMOTO, S., K AWAMURA, S., and M IYAZAKI, F ., 1984,
the learning (without using a zero phase cuto ) until Bettering operation of robots by learning. Journal of
the error goes past a minimum, records the command Robotic Systems, 1, 123 ± 140.
that produced the minimum, and uses it without further C ASALINO, G ., and BARTOLINI, G ., 1984, A learning procedure
for the control of movements of robotic manipulators.
update. All long term stability issues are thus bypassed. IAST ED Symposium on Robotics and Automation,
This can be particularly appropriate for robots on Amsterdam, The Netherlands, pp. 108 ± 111.
assembly lines, simply tuning the commands while C HANG, C . K ., L ONGMAN, R . W ., and P HAN, M ., 1992,
setting up the line, and then using the commands Techniques for improving transients in learning control
thereaf ter. systems. Advances in the Astronautical Sciences, 76, 2035 ±
2052.
There are two situations where the above design pro- C HEN, H . J ., and L ONGMAN, R . W ., 1999, The importance of
cess may not yield the desired performance. The linear smooth updates in learning and repetitive control.
phase lead will not be able to get above a nearly Proceedings of the 38th IEEE Conf erence on Decision and
undamped resonance in the system producing close to Control, Phoenix, Arizona, USA, pp. 258 ± 263.
a 1808 cli in the phase. For this case, one can either C RAIG, J . J ., 1984, Adaptive control of manipulators through
repeated trials. Proceedings of the American Control
introduce a cancellation of these lightly damped poles Conference, San Diego, USA, pp. 1566 ± 1573.
into the F…z† or shift to the phase cancellation law (Elci E DWARDS, S. G ., A GRAWAL, B. N ., P HAN, M . Q ., and
et al. 1994 c, Longman and Wang 1996, Hsin et al. L ONGMAN, R . W ., 2000, Disturbance identi® cation and
1997 a). The second situation is when the desired trajec- rejection experiments on an ultraquiet platform. Advances
tory in learning control is so short that steady state in the Astronautical Sciences, 103, 633 ± 651.
E LCI, H ., P HAN, M ., L ONGMAN, R . W ., J UANG, J .-N ., and
frequency response thinking is not representative, in U GOLETTI, R . 1994 a, Experiments in the use of learning
which case one can shift to the ILC law based on a control for maximum precision robot trajectory tracking.
partial isometry (Jang and Longman 1996 b). For RC Proceedings of the 1994 Conf erence on Information Science
the ® x is easy, simply de® ne the period to be some and Systems, Department of Electrical Engineering,
integer multiple of the actual period, so that the settling Princeton University, Princeton, NJ, USA, pp. 951 ± 958.
E LCI, H ., L ONGMAN, R . W ., P HAN, M ., J UANG, J .-N ., and
time is now short compared to the new period. U GOLETTI, R ., 1994 b, Discrete frequency based learning
The ® eld of iterative learning control started to control for precision motion control. Proceedings of the
develop fast in 1984, and repetitive control originated 1994 IEEE International Conf erence on Systems, Man, and
some years earlier. Although both ® elds o er substantial Cybernetics, San Antonio, TX, USA, Vol. 3, pp. 2767 ± 2773.
improvements in the performance of control systems E LCI, H ., L ONGMAN, R . W ., P HAN, M ., J UANG, J .-N ., and
U GOLETTI, R ., 1994 c, Automated learning control through
doing repetitive tasks, they have seen relatively little model updating for precision motion control. Adaptive
application. We speculate on some of the reasons. In Structures and Composite Materials : Analysis and
the case of ILC, the literature placed much of its empha- Applications, ASME, AD-Vol. 45/MD-Vol. 54, pp. 299± 314.
Iterative learning control and repetitive control 953
F RUEH, J . A ., and P HAN, M .Q ., 1999, Model reference learning L ONGMAN, R . W ., A KOGYERAM, R . A ., H UTTON, A., and
control with basis functions. Proceedings of the 38th IEEE J UANG, J .-N ., 2000, Stability of matched basis function
Conference on Decision and Control, Phoenix, AZ, USA, repetitive control. Advances in the Astronautical Sciences,
pp. 251± 257. 105, in press.
F RUEH, J . A ., and P HAN, M .Q ., 2000, Linear quadratic L ONGMAN, R . W ., BEIGI, H . S. M ., and L I, C . J ., 1989,
optimal learning control (LQL). International Journal of Learning control by numerical optimization methods.
Control, 73, 832 ± 839. Modeling and Simulation, Instrument Society of America,
Grenander, U., and SZEGÖ, G ., 1958, T oeplitz Forms and 20, 1877 ± 1882.
Their Applications (Berkeley, Los Angeles : University of L ONGMAN, R . W ., and C HANG, C .-K ., 1990, Learning control
Calif ornia Press). for minimizing a quadratic cost during repetitions of a task.
H ARA, S., and Y AMAMOTO, Y ., 1985 a, Synthesis of repetitive AIAA/AAS Astrodynamics Conference, a Collection of
control systems and its applications. Proceedings of the 24th T echnical Papers, Portland, Oregon, USA, pp. 530 ± 537.
IEEE Conf erence on Decision and Control, Fort Lauderdale, L ONGMAN, R . W ., C HANG, C . K ., and P HAN, M ., 1992,
FL, pp. 326 ± 327. Discrete time learning control in nonlinear systems.
H ARA, S., O MATA, T ., and N AKANO, M ., 1985 b, Stability of Proceedings of the 1992 AIAA/AAS Astrodynamics
repetitive control systems. Proceedings of the 24th IEEE Specialist Conference, Hilton Head, South Carolina, USA,
Conference on Decision and Control, Fort Lauderdale, FL, pp. 501± 511.
pp. 1387 ± 1392. L ONGMAN, R . W ., and H UANG, Y .-C ., 1994, Use of unstable
H SIN,Y . P ., L ONGMAN, R . W ., SOLCZ ,, E . J ., and DE J ONG, J ., repetitive control for improved tracking accuracy. Adaptive
1997 a, Experiments bridging learning and repetitive control. Structures and Composite Materials : Analysis and
Advances in the Astronautical Sciences, 95, 671 ± 690. Applications, AD-Vol. 45/MD-Vol. 54, ASME, pp. 315± 324.
H SIN,Y . P ., L ONGMAN, R . W ., SOLCZ , E . J ., and DE J ONG, J , L ONGMAN, R . W ., and H UANG, Y .-C ., 1996, Analysis and
1997 b, Experimental comparisons of four repetitive control frequency response of the alternating sign learning and repe-
algorithms. Proceedings of the 31st Annual Conf erence on titive control laws. Advances in the Astronautical Sciences,
Information Sciences and Systems, Johns Hopkins 90, 1105 ± 1124.
University, Department of Electrical and Computer L ONGMAN, R . W ., and SOLCZ , E . J ., 1990, Small gain robust-
Engineering, Baltimore, Maryland, USA, pp. 854± 860. ness issues in the p-integrator repetitive controller. AIAA/
H SIN,Y . P ., L ONGMAN, R . W ., SOLCZ , E . J ., and DE J ONG, J ., AAS Astrodynamics Conference, a Collection of T echnical
1997 c, Stabilization due to ® nite word length in repetitive Papers, Portland, Oregon, USA, pp. 537 ± 551.
L ONGMAN, R . W ., and SONGCHON, T ., 1999, Trade-o s in
and learning control. Advances in the Astronautical Sciences,
designing learning/repetitive controllers using zero-phase
97, 817± 836.
® ltering for long term stabilization. Advances in the
H UANG, Y .-C ., and L ONGMAN, R . W ., 1996, The source of the
Astronautical Sciences, 102, 243 ± 262.
often observed property of initial convergence followed by
L ONGMAN, R . W ., and W ANG, Y ., 1996, Phase cancellation
divergence in learning and repetitive control. Advances in the
learning control using FFT weighted frequency response
Astronautical Sciences, 90, 555 ± 572. (To appear in the
identi® cation. Advances in the Astronautical Sciences, 93,
Journal of Intelligent Automation and Soft Computing.)
85± 101.
I NOUE, T ., N AKANO, M ., and I WAI, S., 1981, High accuracy L ONGMAN, R . W ., and W IRKANDER, S.-L ., 1998, Automated
control of a proton synchrotron magnet power supply. tuning concepts for iterative learning and repetitive control
Proceedings of the 8th W orld Congress of IFAC, XX, laws. Proceedings of the 37th IEEE Conf erence on Decision
216 ± 221. and Control, Tampa, Florida, USA, pp. 192 ± 198.
J ANG, H . S., and L ONGMAN, R . W ., 1994, A new learning M IDDLETON, R . H ., G OODWIN, G . C ., and L ONGMAN, R . W .,
control law with monotonic decay of the tracking error 1985, A method for improving the dynamic accuracy of a
norm. Proceedings of the T hirty-Second Annual Allerton robot performing a repetitive task. University of Newcastle,
Conference on Communication, Control, and Computing, Newcastle, Australia, Department of Electrical Engineering
Coordinated Sciences Laboratory, University of Illinois at Technical Report EE8546. Also, 1989, International Journal
Urbana-Champaign, USA, pp. 314± 323. of Robotics Research, 8, 67± 74.
J ANG, H . S., and L ONGMAN, R . W ., 1996 a, An update N AKANO, M ., and H ARA, S., 1986, Microprocessor-based
on a monotonic learning control law and some fuzzy repetitive control. Microprocessor-Based Control Systems,
logic learning gain adjustment techniques. Advances in the 279± 296.
Astronautical Sciences, 90, 301 ± 318. O H, S. J ., L ONGMAN, R . W ., and P HAN, M . Q ., 1997, Use of
J ANG, H . S., and L ONGMAN, R . W ., 1996 b, Design of digital decoupling basis functions in learning control for local
learning controllers using a partial isometry. Advances in the learning and improved transients. Advances in the
Astronautical Sciences, 93, 137 ± 152. Astronautical Sciences, 95, 651 ± 670.
L EE, S. C ., L ONGMAN, R . W ., and P HAN, M ., 1994, Direct O MATA, T ., N AKANO, M ., and I NOUE, T ., 1984, Applications
model reference learning and repetitive control. Advances of repetitive control method to multivariable systems.
in the Astronautical Sciences, 85, 2047 ± 2066. T ransactions of SICE, 20, 795 ± 800.
L EE-G LAUSER, G ., J UANG, J .-N ., and L ONGMAN, R . W ., 1996, P HAN, M ., and L ONGMAN, R . W ., 1988 a, A mathematical
Comparison and combination of learning controllers: com- theory of learning control for linear discrete multivariable
putational enhancement and experiments. Journal of systems. Proceedings of the AIAA/AAS Astrodynamics
Guidance, Control and Dynamics, 19, 1116 ± 1123. Conference, Minneapolis, Minnesota, USA, pp. 740 ± 746.
L ONGMAN, R . W ., 1998, Designing iterative learning and repe- P HAN, M ., and L ONGMAN, R . W ., 1998 b, Liapunov based
titive controllers. In Z. Bien and J. Xu (Eds) Iterative model reference learning control. Proceedings of the 26th
L earning Control : Analysis, Design, Integration and Annual Allerton Conf erence on Communication, Control and
Applications, (Boston: Kluwer Academic Publishers), pp. Computing, The University of Illinois Coordinated Sciences
107 ± 146. Laboratory, pp. 927 ± 936.
954 Iterative learning control and repetitive control
P HAN, M ., and L ONGMAN, R . W ., 1989, Indirect adaptive con- U CHIYAMA, M ., 1978, Formulation of high-speed motion
trol with guaranteed stability. Proceedings of the 1989 pattern of a mechanical arm by trial (in Japanese).
Conference on Information Sciences and Systems, Johns T ransactions of the Society for Instrumentation and Control
Hopkins University Department of Electrical Engineering, Engineers, 14, 706 ± 712.
Baltimore, MD, USA, pp. 125± 131. W ANG, Y ., and L ONGMAN, R . W ., 1996, Use of non-causal
Phan, M.Q., and F RUEH , J . A ., 1998, System identi® cation and digital signal processing in learning and repetitive control.
learning control. In Z. Bien and J. Xu (Eds) Iterative Advances in the Astronautical Sciences, 90, 649 ± 668.
L earning Control : Analysis, Design, Integration, and W EN, H .-P ., and L ONGMAN, R . W ., 1997, Conditions for con-
Applications (Norwell, MA : Kluwer Academic Publishing) vergence in indirect adaptive control based learning control.
pp. 285± 306. Advances in the Astronautical Sciences, 95, 623 ± 642.
P HAN, M . Q ., L ONGMAN, R . W ., and M OORE, K . L ., 2000, A W EN, H .-P ., P HAN, M . Q ., and L ONGMAN, R . W ., 1998,
uni® ed formulation of linear iterative learning control. Bridging learning control and repetitive control using basis
Advances in the Astronautical Sciences, 105, in press. functions. Advances in the Astronautical Sciences, 99, 335 ±
P LOTNIK, A ., and L ONGMAN, R . W ., 2000, Subtleties in learn- 354.
ing control using zero-phase and cli ® ltering. Advances in W EN, H .-P ., and L ONGMAN, R . W ., 1999, Floquet stability
the Astronautical Sciences, 103, 673 ± 692. analysis of disturbance cancellation using iterative basis
SOLCZ , E . J ., and L ONGMAN, R . W ., 1992, Disturbance rejec- function feedback. Advances in the Asronautical Sciences,
tion in repetitive controllers. Advances in the Astronautical 102, 783 ± 802.
Sciences, 76, 2111 ± 2130.
W IRKANDER, S.-L ., and L ONGMAN, R . W ., 1999, Limit cycles
T OMIZUKA, M ., T SAO, T .-C ., and C HEW , K .-K ., 1989,
for improved performance in self-tuning learning control.
Analysis and synthesis of discrete time repetitive controllers.
Advances in the Asronautical Sciences, 102, 763 ± 781.
Journal of Dynamic Systems, Measurement, and Control,
111, 353 ± 358.