Professional Documents
Culture Documents
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013
I. INTRODUCTION
TERATIVE learning control (ILC) is a methodology applicable to systems which repeatedly track a reference,
,
defined over a finite interval
. The aim is to use past
experience to sequentially improve tracking performance over
repeated trials of the task. Over the last 25 years it has been
an area of intense research interest in both theoretical and application domains, for recent overviews of the literature see [1]
and [2]. However, rather than follow a motion profile defined at
all points, in many applications the system output is only critical at a finite set of prescribed time instants. Examples include
production line automation, crane control, satellite positioning,
and robotic pick and place tasks in which the critical points
correspond to the location of the payloads. Furthermore, ILC
has recently been used to great effect within stroke rehabilitation [3], where motion control is naturally specified in terms of a
point-to-point optimization problem in order to correspond with
results from human motor learning [4].
The standard ILC framework is able to tackle the
point-to-point problem simply by employing an arbitrary
Manuscript received August 11, 2011; revised November 17, 2011; accepted
January 22, 2012. Manuscript received in final form February 08, 2012. Date
of publication March 14, 2012; date of current version nulldate. This work was
supported by Australian Research Council Future Fellow Grant FT0991385.
Recommended by Associate Editor S. S. Saab.
C. T. Freeman is with the School of Electronics and Computer Science, University of Southampton, Southampton SO17 1BJ, U.K. (e-mail:
cf@ecs.soton.ac.uk).
Y. Tan is with the Electrical and Electronic Engineering Department, University of Melbourne, Parkville VIC 3010, Australia (e-mail: yingt@unimelb.edu.
au).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TCST.2012.2187787
reference,
, which passes through the desired points.
However superior results follow if this is coupled with strategies such as Input Shaping in order to suppress vibrations that
occur between the critical points. This approach is taken in [5]
for a high-acceleration positioning table. An alternative is to
and to employ
use a simpler feedback controller to track
ILC to update parameters within the input shaping filter applied
to the reference, as proposed by [6] for control of an industrial
robot. Another approach is to develop ILC algorithms which
have two separate components; one which ensures tracking of
, and another which reduces the amplitude of residual
vibrations occurring after the point-to-point location is reached
[7]. The drawback to all these methods is that they fail to utilize
the extra freedom available in ILC design to address additional
is designed a
performance demands. Furthermore, if
priori to meet such performance objectives, these will not be
met in practice due to the presence of model uncertainty and
noise.
Other approaches to point-to-point motion control have
broken away from the standard ILC framework of tracking a
, but have only constatic reference defined over
sidered the case where a specified position must be reached at
time
, as in [8][13], or the case of a movement between
two equilibrium points [14]. While these approaches dispense
with tracking unnecessary output points, they do not use the
resulting freedom to tackle additional performance objectives
which may be of critical concern. A further limitation is that
they only consider a single point-to-point movement, rather
than a sequence of actions needed to build up complex movements, such as is required in robotic automation and production
line assembly.
This paper addresses current drawbacks by providing
a framework that can deal with an arbitrary number of
point-to-point movements, while also addressing a general
form of performance objective which encompasses a wide
variety of practical performance concerns. Along with soft
performance constraints, this includes hard constraints which
are needed to address actuator saturation, physical workspace
limitations, or imposed safety restrictions. This framework
significantly increases the flexibility and functionality of
point-to-point ILC compared with approaches currently available. Moreover, the action of embedding both the performance
and tracking objectives within the framework of iterative
learning yields algorithms which are capable of reaching optimal solutions in the presence of model uncertainty and noise.
To achieve this, ILC is employed as an iterative optimization
paradigm which uses experimental data to tackle a general form
of cost function involving the input, output and states. A similar
FREEMAN AND TAN: ILC WITH MIXED CONSTRAINTS FOR POINT-TO-POINT TRACKING
(1)
defined over the finite time interval
where the number of samples
. Here
,
,
are the state, input and output vectors,
respectively, and the input and output sequences are given by
605
where
is the unknown desired input sequence corresponding
to . This leads to
Over the
trial the relationship between the input and output
time-series can be expressed by
where the
matrix is
..
.
..
(3)
..
.
Here
is the response to initial conditions whose effect can
be absorbed into the reference trajectory, so that without loss of
generality it is assumed
, or equivalently
.
For some
, an ILC update of the form
(4)
can be considered as an iterative numerical method to solve the
tracking problem, and the derivation of a suitable matrix has
been the focus of significant research effort. Since
(5)
the update (4) is convergent to a solution satisfying (2) if and
only if
(6)
The convergence speed is determined by the magnitude of
and is maximum when
.
A. Point-to-Point ILC Formulation
Now suppose that the th plant output is only required to track
a reference trajectory at a fixed number,
, of sample
instants along the trial duration. These sample instants are given
by
. To define the point-topoint tracking problem it is first necessary to remove the points
that do not need to be tracked from the original reference .
This yields a reduced reference vector
whose length
is given by
(7)
It is then necessary to define a matrix transformation
such that
. This is achieved by
first introducing a row vector
whose th element
is 1 if the th element of
is required to be tracked, and 0
otherwise. The formal definition for is
Let
and
be the input and output vectors respectively on
the th trial, with
the tracking error. Then it is
necessary to find a sequence of control inputs that satisfies
(2)
..
.
if
otherwise
where
tion. The matrix
(8)
and
denotes the floor funcis then produced as follows: 1) set
,
606
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013
(9)
otherwise.
, when any output vector is
As seen by the relation
pre-multiplied by , it extracts the components that correspond
to prescribed point-to-point locations, while retaining the order
in which they appear.
Remark 2: Suppose that at each point-to-point location each
component of the output is required to track a reference point,
that is if
. In this case the matrix
has a simpler form given by the block-wise components
if
,
otherwise
(10)
standard ILC update (4) cannot force the plant to track an arbitrary reference trajectory . The point-to-point update (13) can
enforce tracking of an arbitrary reference
if and only if the
tracked points are chosen such that
(16)
Proof: A necessary and sufficient condition for an operto exist satisfying the convergence condition (15) is that
. For the standard ILC case
,
and hence
, leading to
having eigenvalues at unity. Now the
row of
is the
row of , hence if
and the
point-to-point samples are chosen to correspond to any subset
of linearly independent rows of , the convergence condition
(15) can be satisfied. If
then the additional condition
is imposed.
Remark 3: Let system (1) be written as discrete transfer-function matrix
with component
the transfer-function linking the th output with the th
input. If the relative degree of
is
, then we have
.
The ability of point-to-point ILC to employ a modified standard reference to recover feasibility is extremely important, especially as
in practice due to the delay action of a zeroorder hold. However many tasks are naturally defined only at a
small number of points, and hence additional benefits may also
be expected by not enforcing tracking of unnecessary points.
The next lemma shows how the space of feasible inputs expands
as the number of tracked points, , reduces.
Lemma 1: Assuming
, the feasible input
space which forces the system (1) to track is of dimension
, and is given by
.
The nullspace of
has an orthogonal basis given by the rows
of
, where
is such that the
ator
is full rank.
matrix
It is next illustrated how this enlarged space of feasible inputs
can be used to increase performance. In particular, the practically relevant case is addressed in which a weighted input norm
is required to be small. However, before this can be considered
a preliminary proposition is required.
Proposition 1: Let
comprise point-to-point locations
satisfying
. Let equal but with the
row removed, and hence correspond to tracking all but the
point-to-point location. Let the
eigenvalues of the matrix
be denoted
,
which also equal the singular values since is Normal. Similarly, let the
eigenvalues of the matrix
be denoted
, which also equal
the singular values since
is Normal. Then the following
relationship holds:
(17)
In particular, let equal the th column of with the th element removed. Then if the eigenvalues of are distinct and no
eigenvector of is orthogonal to then
(18)
FREEMAN AND TAN: ILC WITH MIXED CONSTRAINTS FOR POINT-TO-POINT TRACKING
607
(19)
(23)
is full rank.
subject to
is the norm of the orthogonal projection of
of
, and it follows that:
(22)
Now insert the relationship
The relationship
leads to the weighted input bound (20). It follows that the input
norm is small when point-to-point locations are selected which
maximize the smallest eigenvalue of
. Application
608
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013
(30)
with the input
and terminal state
.
Use of (30) therefore avoids calculation of the large matrix
appearing in (25) and the algorithms which follow.
Remark 6: It is shown in [20] that the gradient point-to-point
algorithm (25) applied to a linear system always converges to a
solution which minimizes
. Hence, using Theorem 2, (25)
converges to a solution with a norm satisfying
(29)
Application of Proposition 1 guarantees that each point removed
increases
and reduces
.
from
Hence the convergence rate (29) increases.
Having shown point-to-point ILC increases the convergence
rate, robustness margins are next established.
Theorem 4: Let there exist a multiplicative uncertainty on each element of the plant model
, such that
. Here
is the actual plant and the
model
corresponds to the matrix
used in the update
law (25). A sufficient condition for monotonic convergence is
that each
lies in the open interval
,
demonstrating an allowable phase margin uncertainty of 90 .
Proof: This is an extension of robustness analysis for
the standard gradient algorithm (
) in [19] for the
SISO case. Suppose that the uncertainty can be expressed in
the matrix form
, and that point locations are such
that
. Then using (25) the point-to-point error
satisfies
(31)
whose upper bound decreases as the number of tracked points
is reduced.
A. Inequality Constrained Gradient Point-to-Point ILC
Consider vector inequality constraints on the system input of
the form
, where
and
, where is
the number of imposed constraints. The point-to-point problem
(24) now becomes
subject to
(32)
This can be tackled using an interior-point approach to inequality constrained minimization, termed the barrier function
[21]. A logarithmic barrier function is employed, producing the
auxiliary problem
(33)
where
where
. If is positive, the first term on the right-hand
side is strictly positive for an arbitrary non-zero
and
,
and of
. Similarly the second term is of
and strictly
negative, and hence there always exists a
which ensures monotonic reduction in error norm. This also holds if the
components of
are reordered so that the elements corresponding to the same input are grouped, resulting in a reordering
of the matrix such that
. The stipulation that the components of associated with the same input
have the same uncertainty then results in
having the block
diagonal structure
, where
corresponds to the th input. A sufficient condition for to be positive definite is that each
is positive definite. This is the same
condition as that given in [19] which goes on to show that a
sufficient condition is that each
is positive-real. Therefore a sufficient condition for monotonic convergence is that
, are the
rows of , , respectively. The scalar
is used to weight the action of the barrier and should
be gradually increased to result in a solution which satisfies the
tracking requirement. The solution via the gradient method is
(34)
are given by
,
where the elements of
and
is the value of on trial .
With appropriately chosen scalars
and , this converges
to the zero error solution as
provided there exists
which satisfies
[21], [22]. In the context
of ILC the increase in must not be too rapid in order to ensure
that the barrier component effectively engages with the ILC update. Conversely it must be fast enough not to significantly reduce overall convergence speed. The selection of and must
therefore ensure:
1) the input (34) remains feasible (
);
FREEMAN AND TAN: ILC WITH MIXED CONSTRAINTS FOR POINT-TO-POINT TRACKING
2) the constraint term in (34) comprises a significant proportion of the input over samples which are required to adapt
to the imposed hard constraint;
3) 1) and 2) are satisfied without reducing .
It follows that an appropriate update strategy is to select the
highest value of that results in a feasible input without the
constraint term, that is, on trial choose a value
which satisfies
s.t.
and then update
(35)
609
according to
if
otherwise
..
.
(36)
..
.
..
.
..
..
.
(39)
and
.
For notational simplicity, and without loss of generality, it is
assumed that
.
Now take as any solution satisfying
and
. Providing it is feasible, that is
, such an input is found
in practice through application of the approach of Section III-B.
Also introduce
as a matrix with columns
that form an orthogonal basis of the nullspace of
. From
Lemma 1, a suitable candidate is
Denote
, then the minimization (37)
is equivalent to the inequality constrained problem
s.t.
(40)
with
This has split the solution into components and that minimize the soft and tracking constraints, respectively. The use
of means that updating does not affect the plant output at
the point-to-point locations which have already been forced to
follow the prescribed reference. Applying the barrier function
method to solve (40) yields the auxiliary problem
(37)
is a weighting matrix. It is
where
worthwhile highlighting that the formulation of point-to-point
tracking as an equality constraint that must be satisfied at each
iteration is a much stronger performance requirement compared
with that of standard ILC. Moreover, it is a requirement that
must be satisfied in the face of additional performance demands.
(42)
are given by
where the elements of
. With suitable updating of the step-sizes
and
, (42) is guaranteed to reach a local solution of
(37). As discussed in Section III-B, for the barrier function to
610
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013
s.t.
(43)
(48)
(44)
where the elements of
are given by
. To ensure that the step-sizes
and
engage productively, the procedure (35), (36) is again
employed, now with elements
.
The final update sequence on each trial is
(45a)
(45b)
(45c)
is the experimentally obtained performance function
where
.
Remark 8: In the absence of inequality constraints the soft
constraints reach a global minimum with convergence criterion
(49)
In the case where vibrations are suppressed at the point-to-point
locations,
is given by
, where is defined in (8).
2) Example 2Energy Constraints: Suppose instead a combination of the input and output signal norms are required to be
minimal, giving rise to the cost
subject to
(50)
This form of constraint may be used to reduce either the input
or output norm, by setting the other weight to zero. This may
lead to an excessively impulsive action, however, which can be
addressed by instead using a small, non-zero value multiplied
by the identity matrix (or alternatively minimizing the output
derivative). The cost (50) corresponds to
and
. This gives
, and the
update (45b) becomes
(46)
(51)
within the ILC framework. Here the signals can be read directly
or observed using a suitable estimator.
IV. NEWTON METHOD-BASED ILC
While providing a high level of robustness to plant uncertainty, the gradient descent approach has only a linear convergence rate. This section shows how the previous algorithms can
be extended to deliver quadratic convergence. Consider again
the point-to-point tracking problem
(52)
FREEMAN AND TAN: ILC WITH MIXED CONSTRAINTS FOR POINT-TO-POINT TRACKING
and now apply the Newton method [18] to solve it, yielding
611
(53)
where
is the Hessian matrix. From (15) the necessary and sufficient condition for convergence to zero error in a
single trial is now
Proof: Application of
produces the signal
(60)
(54)
which is satisfied if and only if
. Since its computation involves inverse and derivative operations, the update
(53) is difficult to implement, especially for large values of . It
may contain excessive amplitudes and high frequencies which
increase learning transients, and, depending on point-to-point
locations, it may be singular. However, it is shown in [23] that
is the solution, , to
subject to
(55)
where
operator which replaces
. Since
in (53) is
, the resulting
(61)
If this relation is applied
Algorithm 1
(a) apply input
to the real plant and record output
(b) solve (56) through repeated application of (57) to the
plant model
to obtain a suitable approximation to
(c) use the resulting input to form the next descent direction
in the Newton update (53). Go to (a)
The next theorem establishes how the number of inter-trial
updates influences the convergence rate of the overall ILC law.
Theorem 5: For some
, if inter-trials updates of (57)
are performed, the error evolution is given by
(58)
612
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013
(63)
This can be solved via the Newton method by imposing the inequality constraint in the inter-trial calculation of the descent direction,
, given by the solution to (55). Within
this inter-trial problem, the inequality constraint translates to
and the descent direction is thus generated using
subject to
(64)
subject to
s.t.
via the gradient method. Hence to solve (40) via the Newton
method, now impose the inequality constraint on (70). In terms
of the control input, on trial
, the constraint enforces
, which, assuming
has not yet been updated,
can be written as
. In terms of the Newton
descent direction, this translates to
.
Hence, (70) becomes
s.t.
Algorithm 2
(a) apply input
to the real plant and record output
(b) between trial and
, construct suitable
approximation to
satisfying
when used in Newton update (53), through
repeated application of (65) to the simulated plant
(c) use the resulting input to form the next descent direction
in the Newton update (53). Go to (a)
The number, , of inter-trial updates is chosen heuristically
to affect a compromise between the amplitude of the descent
direction , and the overall convergence of the ILC scheme,
dictated by (58). In practice this is application specific, and is
achieved by decreasing in response to excessive amplitudes/
frequencies in the input signal, levels of fluctuation in the error
norm, and the overall convergence rate achieved.
B. Incorporation of Additional Objectives
As previously considered, having satisfied the point-to-point
tracking requirement using the algorithms of Section III-B or
Section IV-B, an additional objective function may be introduced. This is required to be minimized while continuing to satisfy the point-to-point tracking requirement with an inequality
constrained input. The problem is given by
s.t.
(66)
This is equivalent to
s.t.
with corresponding update
(71)
applied to the plant
, where the elements of
are
given by
. Similar analysis to
that used in Theorem 5 relates the number of inter-trial updates
of (71) to the convergence of the Newton update (68) whose
descent direction it approximates.
Although separation of the soft constraint and tracking error
objective is ensured by the inequality constraint in (66), in practice
must also be updated to ensure it continues to be satisfied
in the presence of model uncertainty and noise. Therefore
in
(67) is also updated using the Newton ILC update
(72)
where it has been
with the constraint
assumed that
has just been updated via (68) as discussed.
The unconstrained Newton ILC descent direction,
, in (72) is the solution, , to
s.t.
(73)
FREEMAN AND TAN: ILC WITH MIXED CONSTRAINTS FOR POINT-TO-POINT TRACKING
613
(76)
Algorithm 3
to the system
(78)
to the simulated system
V. EXPERIMENTAL RESULTS
The ILC approaches developed have been tested on a six degree of freedom anthropomorphic robotic arm whose five rotary
joints are composed of PowerCubes (Schunk GmbH & Co.) incorporating brushless servomotors with integrated power electronics and transmission. These communicate with a dSPACE
ds1103 control board via a CAN bus at a rate of 500 kbit/s.
Results are presented for the first joint which is aligned in the
horizontal plane as shown in Fig. 1. Each servomotor includes
cascaded current and velocity control loops, and frequency response tests have established that the linear model (79), shown
at the bottom of the next page, adequately represents the system
dynamics, with input and output in degrees. A sampling time of
200 Hz has been used in all experimental tests.
614
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013
(79)
FREEMAN AND TAN: ILC WITH MIXED CONSTRAINTS FOR POINT-TO-POINT TRACKING
615
Fig. 4. Experimental results for point-to-point ILC with hard input constraints:
a) output; b) input; and c) point-to-point tracking error. Final trial output and
. The output produced
input are shown for the gradient update with
in simulation is also shown in a) and denoted standard ILC reference. When
this is used as a reference in standard ILC, the input produced violates the hard
constrains, as shown in b).
output is shown in Fig. 4(a)]. This output is then used as the reference for standard ILC (
) applied to the experimental
system using the optimal gain choice (27). The input produced
over 150 trials is shown in Fig. 4(b) and clearly does not
satisfy the hard constraint. This confirms that the standard ILC
framework is unable to satisfy constraints since a predefined
reference is not robust to model uncertainty and noise.
Finally the case of a soft constraint in conjunction with
inequality constraints is considered. The soft constraint
comprises minimizing the output derivative over the period
, as may be required when fragile payloads or
open top containers containing liquid are handled. This corresponds to the weight
and the function
in (37) or (66). Inequality
constraints of
have also been employed
through selection of
and
. For the gradient case, updates (45a) and (45c) are used in conjunction with (49). Results
are shown in Fig. 5 using
,
. Results using the
Newton update with mixed constraints are also shown, where
Algorithm 3 of Section IV-C has been implemented using 10
iterations of (76) with
to construct the descent direction
in (75), and 20 iterations of (74) with
to approximate
the descent direction in (72). As with previous results, the
Newton approach converges to very similar input and output
signals as the gradient approach, but requires significantly
fewer trials. To facilitate comparison with the standard ILC
framework, the constrained point-to-point tracking problem
has been solved in simulation using the nominal plant [with
616
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 3, MAY 2013
point-to-point tasks to be achieved while simultaneously tackling both hard and soft constraints of wide relevance. Experimental results confirm the practical utility and performance of
the proposed approaches and illustrate the benefit gained over
using the standard framework with an a priori generated reference. They also clearly show the benefit of point-to-point ILC
over the standard ILC framework.
Future work will consider the inclusion of prescribed
variation in the temporal point-to-point locations to provide
more flexibility and faster convergence properties. Constraints
linking two or more outputs will also be considered, allowing
coordinated movements to be performed while relaxing unnecessary temporal constraints.
REFERENCES
[1] D. A. Bristow, M. Tharayil, and A. G. Alleyne, A survey of iterative learning control a learning-based method for high-performance
tracking control, IEEE Control Syst. Mag., vol. 26, no. 3, pp. 96114,
2006.
[2] H. S. Ahn, Y. Chen, and K. L. Moore, Iterative learning control: Brief
survey and categorization, IEEE Trans. Syst., Man, Cybern. C, Appl.
Rev., vol. 37, no. 6, pp. 10991121, Nov. 2007.
[3] C. T. Freeman, E. Rogers, A. M. Hughes, J. H. Burridge, and K. L.
Meadmore, Iterative learning control in healthcare: Electrical stimulation and robotic-assisted upper limb stroke rehabilitation, IEEE Control Syst. Mag., vol. 32, no. 1, pp. 1843, Feb. 2012.
[4] D. M. Wolpert, Z. Ghahramani, and J. R. Flanagan, Perspectives and
problems in motor learning, Trends in Cognitive Sci., vol. 5, no. 11,
pp. 487494, 2001.
[5] H. Ding and J. Wu, Point-to-point control for a high-acceleration positioning table via cascaded learning schemes, IEEE Trans. Ind. Electron., vol. 54, no. 5, pp. 27352744, Oct. 2007.
[6] J. Park, P. H. Chang, H. S. Park, and E. Lee, Design of learning input
shaping technique for residual vibration suppression in an industrial
robot, IEEE/ASME Trans. Mechatron., vol. 11, no. 1, pp. 5565, Feb.
2006.
[7] J. van de Wijdeven and O. Bosgra, Residual vibration suppression
using hankel iterative learning control, Int. J. Robust Nonlinear Control, vol. 18, pp. 10341051, 2008.
[8] G. Gauthier and B. Boulet, Robust design of terminal ILC with
mixed sensitivity approach for a thermoforming oven, J. Manuf. Sci.
Eng., vol. 2008, 2008, Article ID 289391.
[9] Y. Wang and Z. Hou, Terminal iterative learning control based station
stop control of a train, Int. J. Control, vol. 84, no. 7, pp. 12631274,
Jul. 2011.
[10] J.-X. Xu and D. Huang, Initial state iterative learning for final state
control in motion systems, Automatica, vol. 44, pp. 31623169, 2008.
[11] Y. Chen and J.-X. Xu, A high-order terminal iterative learning control
scheme, in Proc. 36th Conf. Decision Control, 1997, pp. 37713772.
[12] J.-X. Xu, Y. Chen, T. Lee, and S. Yamamoto, Terminal iterative
learning control with an application to RTPCVD thickness control,
Automatica, vol. 35, pp. 15351542, 1999.
[13] G. Gauthier and B. Boulet, Terminal iterative learning control design with singular value decomposition decoupling for thermoforming
ovens, in Proc. Amer. Control Conf., 2009, pp. 16401645.