You are on page 1of 6

8th IFAC Symposium on Mechatronic Systems

8th IFACAustria,
Vienna, Symposium
Sept. on
4-6,Mechatronic
2019 Systems
8th
8th IFAC Symposium
IFACAustria,
Symposium on Mechatronic Systems
Vienna,
8th IFAC Sept. on
Symposium 4-6,Mechatronic
on 2019
Mechatronic
Systems online at www.sciencedirect.com
Available
Systems
Vienna, Austria, Sept. 4-6, 2019
Vienna, Austria, Sept. 4-6, 2019
Vienna, Austria, Sept. 4-6, 2019
ScienceDirect
IFAC PapersOnLine 52-15 (2019) 175–180
Adaptive
Adaptive Iterative
Iterative Learning
Learning Control
Control of
of
Adaptive
Robot Iterative
ManipulatorsLearning
for Control
Friction of
Adaptive
Robot Iterative
ManipulatorsLearning
for Control
Friction of
Robot Manipulators
Compensation
Robot Manipulators for Friction
for Friction
Compensation
Compensation
Compensation
Richard Lee, Liting Sun, Zining Wang, Masayoshi Tomizuka ∗
Richard
Richard Lee, Liting Sun, Zining Wang, Masayoshi Tomizuka ∗∗∗
Richard Lee, Lee, Liting
Liting Sun, Sun, Zining
Zining Wang, Wang, Masayoshi
Masayoshi Tomizuka Tomizuka ∗
∗Richard Lee, Liting Sun, Zining Wang, Masayoshi Tomizuka
∗ University of California Berkeley, Berkeley, CA 94720 USA (e-mails:
∗rUniversity

University of
of California
lee1@berkeley.edu, California Berkeley,
Berkeley, Berkeley,
litingsun@berkeley.edu,
CA
CA 9472094720 USA
Berkeley,wangzining@berkeley.edu, USA (e-mails:
(e-mails:
∗rUniversity of
lee1@berkeley.edu, California Berkeley,
litingsun@berkeley.edu,Berkeley, CA 94720 USA
wangzining@berkeley.edu, (e-mails:
University of California Berkeley, Berkeley, CA 94720 USA (e-mails:
rr lee1@berkeley.edu,
lee1@berkeley.edu, litingsun@berkeley.edu,
litingsun@berkeley.edu,
tomizuka@me.berkeley.edu).
tomizuka@me.berkeley.edu).
wangzining@berkeley.edu,
wangzining@berkeley.edu,
r lee1@berkeley.edu, litingsun@berkeley.edu,
tomizuka@me.berkeley.edu). wangzining@berkeley.edu,
tomizuka@me.berkeley.edu).
tomizuka@me.berkeley.edu).
Abstract: Iterative learning control (ILC) has been well recognized for its ability to improve
Abstract:
Abstract: Iterative
Iterative learning
learning control
control that (ILC)
(ILC) has
has been
been well
well recognized
recognized for
for its
its ability
ability to
to improve
improve
the trackingIterative
Abstract: performance learningof systems
control (ILC) perform
has been repetitive
well tasks. Itsfor
recognized achievable
its ability performance,
to improve
the
the tracking
Abstract:
tracking performance
Iterative learningof systems
control that
(ILC) perform
has been repetitive
well tasks.
recognized Its achievable
for its ability performance,
to improve
however,
the tracking canperformance
be significantly
performance of
of systems
degraded
systems that
that byperform
the presence
perform repetitive
repetitive tasks.
of non-repetitive
tasks. Its
Its achievable
disturbances
achievable performance,
performance,which
however,
the tracking
however, can
can be
be significantly
performance
significantly of may degraded
systems
degraded that by
by the
perform
the presence
repetitive
presence of
of non-repetitive
tasks. Its manipulators
non-repetitive disturbances
achievable
disturbancesperformance,which
which
vary
however, every iteration.
caniteration. Such
be significantly be
degraded a case
by thefor high-precision
presence robot
of non-repetitive disturbances which whichare
vary
however,
vary everyevery can
every subject be
iteration. Such
significantly
Such may may
mayand be
degraded a case
by for
the
be velocity-dependent
a case
case for high-precision
presence
for high-precision
high-precision of robot
non-repetitive
robot manipulators
disturbances
manipulators which
which whichare
are
commonly
vary iteration. to nonlinear
Such be a friction robotforces.manipulators
With a traditional which PD-
are
commonly
vary
commonly every subject
iteration.
subject to nonlinear
Such mayand and velocity-dependent
bechange
a casealong for iterations,
high-precisionfriction forces.
robot With
manipulatorsa traditional
which PD- PD-
are
type
commonly ILC scheme,
subject asto
to nonlinear
joint velocities
nonlinear and velocity-dependent
velocity-dependent friction
friction
friction forces.
forces
forces. With
With aa traditional
can vary substantially
traditional PD-
type
commonly
type ILC
ILC scheme,
subject
scheme, as
as tojoint
joint velocities
nonlinear
velocities and change
change along
velocity-dependent
along iterations,
iterations, friction
friction
friction forces
forces.
forces Withcan
can vary substantially
a traditional
vary substantially PD-
and deteriorate
type deteriorate
ILC scheme,the the performance
as joint velocitiesof of ILC.
change To
along address
iterations, this problem,
friction forces we propose
can an adaptive
vary substantially
and
type ILC scheme,in performance
aswhich
joint velocities ILC.
change To
along address
iterations, this problem,
thisfriction we propose an
an adaptive
and deteriorate
ILC
and deteriorate
algorithm, the
the performance
an adaptive
performance of
of ILC. To compensation
friction
ILC. To address
address this signalforces
problem,
problem, is wewe can
introduced vary substantially
propose
propose with
an adaptive
ILC to
adaptive
ILC
and algorithm,
deteriorate
ILC algorithm,
algorithm, in
thewhich
in which
which an
performance adaptive
an adaptive
adaptive of friction
ILC. To
friction compensation
address
compensation this signal
problem,
signal is introduced
we propose
is introduced
introduced with ILC
ILC to
an adaptive
with to
adaptively
ILC identify in the friction
an model over multiple
friction iterations.signal
compensation Theoretical
is convergence with analysis
ILC to
adaptively
ILC identify
algorithm, the friction model over multiple iterations. Theoretical convergence analysis
adaptively
is provided
adaptively withinsimulation
identify
identify
which
the
the
an verification
friction
friction
adaptive
model
model
friction
over
over
compensation
onmultiple iterations.
a 3-degree-of-freedom
multiple iterations.
signal is introduced
Theoretical
(DOF)
Theoretical convergence withanalysis
planar manipulator.
convergence
ILC to
analysis
is providedidentify
adaptively
is with simulation
the friction verification
model overon a 3-degree-of-freedom
multiple iterations. Theoretical (DOF) planar
convergence manipulator.
analysis
is provided
provided with
Experimental
Experimental with simulation
verification
simulation
verification is verification
is also performed
verification
also performed
on
on aaon 3-degree-of-freedom
aa 5-DOF
5-DOF robot
on3-degree-of-freedom robot for
(DOF)
silicon
silicon wafer
for (DOF)
planar manipulator.
planarhandling.
wafer manipulator.
handling. The
The
is provided
Experimental
verification with
results simulation
verification
show that verification
is also
also performed
theperformed on
proposed on a
on 3-degree-of-freedom
a 5-DOF
adaptive5-DOF ILC robot (DOF)
for silicon
approach silicon planar
wafer
can wafer manipulator.
handling.
achievehandling.significantlyThe
Experimental
verification verification
results show is
that the proposed a
adaptive ILC robot for
approach can achieve significantlyThe
Experimental
verification
better tracking verification
results show that
performance is also
that thantheperformed
proposed on
traditional a 5-DOF
adaptive
ILC methods.ILC robot for silicon
approach can wafer
achievehandling.significantlyThe
verification
better tracking results show
performance the proposed adaptive ILC approach can achieve significantly
verification
better tracking results show thatthan
performance thanthetraditional
traditional ILC
proposed adaptive
ILC methods.
methods.ILC approach can achieve significantly
better tracking performance than traditional ILC methods.
© 2019,tracking
better IFAC (International
performance Federation of Automatic
than traditional ILC Control)
methods. Hosting by Elsevier Ltd. All rights reserved.
Keywords: Iterative learning control, Model reference adaptive control, Nonlinear friction
Keywords:
Keywords: Iterative learning control, Model reference adaptive control, Nonlinear friction
Keywords: Iterative
identification,
identification,
Robots
Iterative
Robots
learning
learning control,
control, ModelModel reference
reference adaptiveadaptive control,
control, Nonlinear
Nonlinear friction friction
Keywords: Iterative
identification, learning control, Model reference adaptive control, Nonlinear friction
identification, Robots Robots
identification, Robots
1. INTRODUCTION
1. INTRODUCTION istics (frequency
(frequency characteristics
characteristics or or temporal
temporal characteris-
characteris-
istics
1.
1. INTRODUCTION
INTRODUCTION istics
tics
istics or(frequency
both),
(frequency the characteristics
design
characteristics of robust or
or temporal
learning
temporal characteris-
filters spans
characteris-
1. INTRODUCTION tics
isticsor both), thecharacteristics
(frequency design of robust learning filters
or temporal spans
characteris-
Iterative Learning Control (ILC) is a well-established con- tics
from
tics or
or both),
regular
both), the
low-pass
the design
design filters,of
of robust
to band-pass
robust learning
learning or filters
notch
filters spans
filters,
spans
Iterative Learning Control (ILC) is a well-established con- from tics or
from
and
regular
both),
regular
to
low-pass
the design
low-pass
time-frequency
filters,
filters,of to
filters. to
band-pass
robust
band-pass
For learning
instance,
or notch
or filters
notch
as
filters,
spans
filters,
discussed
Iterative
trol scheme
Iterative Learning
for
Learning Control
systems
Control (ILC)
with
(ILC) is aa well-established
repetitive
is tasks.
well-established Its con-
funda-
con- from
and regular
to low-pass
time-frequency filters,
filters. to band-pass
For instance, or notch
as filters,
discussed
trol scheme
Iterative for
Learning systems
Controlwithwith
(ILC) repetitive tasks.
is a awell-established Its funda-
con- and from regular
to low-pass filters, to band-pass or notch filters,
trol
trol scheme
mental idea
scheme isfor
to systems
for iteratively
systems with repetitive
construct
repetitive tasks.
control
tasks.inputIts
Its funda-
input based
funda- in
in to time-frequency
Bristow
andBristow et al.
time-frequency
et (2006),filters.
al. (2006), low-pass
filters.
low-pass
For instance,
filters
For filters
instance, are as
are
discussed
widely
as
widely used
discussed
used
mental
trolerrorsidea
scheme is to
for iteratively
systems withconstruct
repetitivea control
tasks. based
Its funda- andBristow
in
to to time-frequency
handle et
model al. (2006), filters.
uncertainties low-pass Forat instance,
filters
high are as discussed
widely
frequencies, used
and
mental
on
mental idea
idea is
from to iteratively
previous
is toprevious
iteratively construct
trails so
construct that a control
the input
performance based of in Bristow et al. (2006), low-pass filters are widely used
on errors
mental idea from
is toprevious
iteratively trails so thata
construct
control
the
adisturbance
control
input based
performance
input based of to to
in
handle model
in Bristow
handle
Sun et et (2014),
model
al.
uncertainties
al. (2006),
uncertainties
a low-pass
notch Q
at filters
at
high frequencies,
high
filter and area widely
frequencies,
disturbance
and
used
and
on
the errors
system
on errors from
(e.g., trajectory trails so
tracking, that the performanceattenua- of to handle model uncertainties at high frequencies, and
the systemfrom
on errors (e.g.,previous
from trajectory
previous
trails
trails
so thatdisturbance
tracking,
so that
the performance
theapplied
performanceattenua- of in
of in to Sunhandle
Sun
observer
et al.
et
are model
al.
(2014),
(2014),
introduced
a notch Qatfilter
uncertainties
aa tonotch
ILC Q
to filter
block
and
highexternal
and
a disturbance
frequencies,
aa disturbance
vibrationsand
the
tion)
the system
can
system be (e.g.,
(e.g., trajectory
optimized.
trajectoryIt hastracking,
been
tracking, disturbance
widely
disturbance attenua-
on many
attenua- in Sun
observer et
areal. (2014),
introduced tonotch
ILC Q
to filter
block and
external disturbance
vibrations
tion) can
the system be optimized.
(e.g., trajectoryIt has been
tracking, widely applied
disturbance on many
attenua- in Sun et al. (2014), a notch Q filter and a disturbance
tion)
tion) can
can be
industrial be optimized.
optimized. It
manipulators due
Itduehas
hasto been
its widely
been simplicity
widely applied
and on
applied many observer
many
superior
on observer are
concentrated are introducedcertainto
at certain
introduced ILC
ILC to to block
frequencies.
tofrequencies. block Inexternal
Lin et
external vibrations
et al.
al. (2015),
vibrations
industrial
tion) can be
industrial
performance
manipulators
optimized.
manipulators
in reducing Itduehas to
to
unknown
its
been
its simplicity
widely
simplicity
but
and
applied
and
repetitive many concentrated
superior
on
superior
distur- aobserver
concentrated
time-frequency
concentrated
at
are introduced
at
at certain
numerical
certain
tofrequencies.
ILC Q tofilter
frequencies.
block In
In
was
In
Lin
external
Lin
Lin et
proposed
et al.
al.
(2015),
vibrations
(2015),
based
(2015),
industrial manipulators
performance in reducingdue to its simplicity
unknown but repetitive and superior
distur- aaon time-frequency
concentrated numerical
at certain Q filter In
frequencies. wasLin proposed
et al. to based
(2015),
industrial manipulators due to its uncertainties
simplicity and(Takegaki
superior a time-frequency
robust
time-frequencyprincipal numerical
component
numerical Q
Q filter
analysis
filter was
was proposed
(RPCA)
proposed based
filter
performance
bances
bances such
performance
such as
performance
asin
in
in
reducing
gravity
reducing
gravity
reducing
and unknown
model
unknown
andunknown
but
but
model uncertainties
but
repetitive
repetitive
repetitive
distur-
distur-
(Takegaki
distur- on robust principal
a time-frequency component
numerical analysis
Q filter was (RPCA)
proposed tobased
filter
based
on
out robust
the
on robust principal
non-repetitive component
principal component disturbances, analysis
analysis and (RPCA)
in Peng to filter
et al.
bances
and
bances
and
bances
such
such as
Arimoto
Arimoto
such as(1981)
gravity
as(1981)
gravity
gravity
and
and
and
and
and
model
Bondi
model
Bondi
model
et uncertainties
et al. (1988)). (Takegaki
al. (1988)).
uncertainties
uncertainties
(Takegaki out
(Takegaki out the non-repetitive
on robust principal disturbances,
component analysis and(RPCA)
in Pengto
(RPCA)
filter
to et al.
filter
and Arimoto
and Arimoto (1981) and Bondi et al. (1988)). (2017),
out the
the anon-repetitive
robust
non-repetitive Q filter disturbances,
in ILC
disturbances, was and
designed
and in
in Peng
along
Peng et
et al.
with
al.
However,
and Arimoto
However, the (1981)
the
and Bondi
disturbances
(1981) and Bondi
disturbances and
et al. (1988)).
and dynamic
dynamic
et al. (1988)). uncertainties are
uncertainties are (2017),
(2017),
a(2017),
the aanon-repetitive
outiteratively
a
robust Q filterdisturbances,
robust
robust tuned Q
Q filter
feedback
filter
in ILC was designed
in
in ILC
ILC was
controller
was
and
designed
to
designed
in alongetwith
Peng
along
suppress
along
al.
with
both
with
However,
hardly the
exactly disturbances
repetitive and
in dynamic
practice. Inuncertainties
most systems,are a(2017),
iteratively
a robust tuned feedback
Q filter in ILC controller
was designed to suppress
alongofboth
with
However, the disturbances and dynamic uncertainties are a iteratively
repetitive
a iteratively and tuned feedback
non-repetitive controller
disturbances. to suppress
Methods both
the
hardly
However, exactly repetitive in
the disturbances and practice.
dynamicInuncertainties
most systems, are repetitive
a iteratively andtuned feedback disturbances.
non-repetitive
tuned
controller to Methods suppressofboth the
hardly
both
both
hardly
exactly
exactly repetitive
iteration-varying
hardlyiteration-varying
exactly
repetitive
repetitive
(i.e.,in
(i.e., in
practice.
practice. In
non-repetitive)
innon-repetitive)
practice.
In most
In and
most
systems,
and iteration-
most iteration-
systems, repetitive
systems, first groupand
repetitive
first group andwork
work wellfeedback
non-repetitive
well
controlleroftoType
disturbances.
for disturbances
non-repetitive
for disturbances
disturbances. of
suppress
Methods
Methods
Type I, but
I, butofboth
the
can
of can
the
both iteration-varying
invariant (i.e.,
both iteration-varying repetitive)(i.e., non-repetitive)
disturbances
(i.e., disturbances
non-repetitive) and
exist. iteration-
Iteration-
and Iteration-
iteration- first repetitive
hardly group and
handle work non-repetitive
well
disturbances for disturbances.
disturbances
of Type II of Methods
Type
because I, but
their of the
can
fre-
invariant
both (i.e.,
iteration-varying repetitive)(i.e., non-repetitive) exist.
and iteration- first
hardly group
handle work well
disturbances for disturbances
of Type II of Type
because I, but
their can
fre-
invariant
varying
invariant (i.e.,
disturbances
(i.e., repetitive)
can
repetitive) bedisturbances
categorized
disturbances exist.
into
exist. Iteration-
two types:
Iteration- first group work well for disturbances of Type I, but can
varying
invariant disturbances
(i.e., repetitive) can be categorized exist.
disturbances into two types: hardly
Iteration- hardly handle
quency or temporal
handle disturbances
temporal characteristics
disturbances of
of Type
Typeoften II
II because
often change their
because with fre-
their the
fre-
varying
varyingII --disturbances
Type
Type external
disturbances
external
can
can be be categorized
state-independent
state-independent categorized into
into two
disturbances
disturbances (most quency
two types:
types:
(most
hardly handle
quency
iteratively
quency
or
or
or temporal
updated
temporal
characteristics
disturbances
characteristics
control of Type
inputs
characteristics dueoften
often
change
IItobecause
change
their
change
with
their
with
dependence
with
the
fre-
the
the
varyingI -disturbances
Type external can betocategorized into
and two types:
II -- iteratively
quency updated
or states.
temporal control inputs due
characteristics to their
often change dependence
withjoint
the
external
external external state-independent
vibrations
Type I - vibrations belong
state-independent
belong this type),
to this
disturbances
type),
disturbances
and Type
Type
(most
(most
II iteratively
on system
iteratively updated
updated Forcontrol
example,
control inputs
inputs due
due to
in robot
robot to their
their dependence
manipulators,
dependence
Type
external
internalI - external
vibrations
state-dependent state-independent
belong to this
disturbances. disturbances
type),
Joint and Type
friction (most
II
force - on system
iteratively states.
updated For example,
control inputsin due tomanipulators,
their dependence joint
external state-dependent
internal vibrations belong to this type),
disturbances. Jointand Typeforce
friction II - on on system
velocities states.
states. For
change
systemchange fromexample,
For iterationin
example, torobot
into manipulators,
iteration
robot with changing
manipulators, joint
changing
joint
external
internal
in vibrations belong
robot state-dependent
manipulators to this example
is disturbances.
a typical
typical type),
Joint and Typestate-
offriction
such II - velocities
force on system states. from
For iteration
example, iteration
intorobot with
manipulators, joint
internal
in robot state-dependent
manipulators is disturbances.
a exampleJointof friction
such force velocities
state- control
velocities change
inputs.
change from
Consequently,
from iteration
iteration this to iteration
changes
iteration thewith
joint
with changing
friction
changing
internal
in robot
dependent state-dependent
manipulators
disturbances. is disturbances.
aa typical
To make example
ILC Jointoffriction
schemes such force control
state-
applicable velocities inputs.
change Consequently,
from iteration this tochanges
iteration the jointchanging
with friction
in robot manipulators is typical example of such state- control
forces inputs.
in nonlinear
controlininputs. Consequently,
and
Consequently, this
inconsistent changes
this changes ways. the joint friction
dependent
in robot
dependent
in the presence
disturbances.
manipulators
disturbances.
presence of these is To
these Toatwomake ILC
typical
make typesILC
schemes
example
ofschemes
applicable
of such state- forces
applicable
iteration-varying nonlinear
controlininputs. and inconsistent
Consequently, this changes ways. the joint friction
the joint friction
dependent
in the disturbances.
of Totwo make typesILCof schemes applicable forces
iteration-varying forces in nonlinear
nonlinear and
and inconsistent
inconsistent ways.
ways.
dependent
in disturbances. To make ILC schemes applicable To
forces deal with
in with
nonlinearthe non-repetitive
non-repetitive
and inconsistent disturbances
ways. of Type
Type II, II,
in the
the presence
disturbances,
presence
disturbances, many
many
of these
these two
approaches
of approaches two types
have been
types
have
of
of iteration-varying
been proposed. Such
iteration-varying
proposed. Such To deal the disturbances of
in the presence
disturbances,
approaches can many of approaches
also these
be two types
grouped haveinto oftwo
been iteration-varying
proposed.
categories: Such1) To
an
To deal
deal with
adaptive
with the
ILC
the non-repetitive
(AILC)
non-repetitive approach disturbances
was
disturbances firstof
of Type
proposed
Type II,
II,
disturbances,
approaches can many approaches
alsoapproaches
be grouped have been
intobeen proposed.
twoproposed.
categories: Such1) an To dealadaptive ILC (AILC) approach was firstof proposed
disturbances,
approaches
ILC with
approaches can
robust
can
manyalso be
learning
also be grouped
filters
grouped
haveinto
(i.e.,
into two
Q
two categories:
filter),
categories:andSuch1)
2)
1) an
by
by Kuc with
an adaptive
adaptive
Kuc and ILC
and
the
Leenon-repetitive
ILC
Lee
(AILC)
(1991). approach
(AILC)
(1991). Baseddisturbances
approach
Based on was
on was first
Lyapunov
Type
first proposed
Lyapunov theory,
proposed
theory,
II,
ILC with
approaches robust
canThe learning
also begroup filters
grouped (i.e.,
into Q
two filter),
categories:and 2)
1) an adaptive ILC (AILC) approach was first proposed
ILC
ILC with
adaptive
withILC. robust
ILC.
robust learning
first
Thelearning
filters
focuses
filters (i.e.,on
(i.e.,on Q filter),
filter), and
Qdesigning robust
and 2) by
2) by Kuc
AILC and
and Lee
Kucnaturally fits(1991).
Leefits Based
Based on
in situations
(1991). situations onwhereLyapunov theory,
the unknown
Lyapunov unknown
theory,
adaptive
ILC withfilters
adaptive
learning robust
ILC. The
to
first
learning
first
separate
group
group
focuses
filters
focuses(i.e.,onQdesigning
iteration-varying filter),
designing
and
robust
and
robust
invariant 2) AILC by Kucnaturally
AILC
disturbances
AILC
andare
naturally
naturally
Lee fits
not
fits
in
(1991).
in
only
in
Based on
situations
iteration-varying,
situations
where
where
where
the
Lyapunov
the
but
the
theory,
unknown
also state-
unknown
adaptive filters
learning ILC. The first group
to separate focuses on designing
iteration-varying robust disturbances
and invariant AILC naturally are not fitsonly iteration-varying,
in situations where but
the also state-
unknown
adaptive ILC. The first group focuses on designing robust disturbances
dependent.
disturbances Theare
are not
key
not only
idea
only iteration-varying,
of AILC,
iteration-varying, as but
discussed
but also
in
also state-
Park
learning
disturbances
learning
disturbances
learning
filters
filters
filters
so to
to
sotothat
separate
that iteration-varying
iteration-varying
separate iteration-varying
iteration-varying
separate iteration-varying
ones
ones will
and
will
and
and
invariant
not enter
invariant
not enter dependent.
invariant disturbances The key
are Tayebi
not only idea iteration-varying,
of AILC, as discussed butanalsoin state-
Park
state-
disturbances
the iteration so
disturbances that
loop.
so iteration-varying
Depending
thatDepending
iteration-varying the ones
on the utilized
ones will not
not enter
character-
willcharacter- dependent.
et
enter dependent. al. (1996)
dependent. The
and
The key
key idea
idea of
(2004),
of AILC,
is
AILC, to as
as discussed
introduce
discussed in
in Park
adaptive
Park
the
the
iteration
disturbances
iteration
loop.
so thatDepending
loop. iteration-varying on
on the
utilized
ones
utilized not enter et
willcharacter- et
et
al. (1996)The
al.
al. (1996)
(1996)
and Tayebi
and
and
key idea
Tayebi
Tayebi
(2004),
of AILC,
(2004),
(2004),
is to as
is
is to
to
introduce
discussed
introduce
introduce
an in
an
an
adaptive
Park
adaptive
adaptive
the iteration loop. Depending on the utilized character- et al. (1996) and Tayebi (2004), is to introduce an adaptive
the iteration loop. Depending on the utilized character-
2405-8963 © 2019, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Copyright © 2019 IFAC 578
Peer review
Copyright © under responsibility of International Federation of Automatic
2019 IFAC 578 Control.
Copyright
Copyright ©
© 2019
2019 IFAC
IFAC
10.1016/j.ifacol.2019.11.670 578
578
Copyright © 2019 IFAC 578
2019 IFAC MECHATRONICS
176
Vienna, Austria, Sept. 4-6, 2019 Richard Lee et al. / IFAC PapersOnLine 52-15 (2019) 175–180

signal which iteratively identifies and compensates for the 2.2 Friction Compensation via Standard ILC
unknown disturbances and uncertain system parameters.
Commonly, it requires the parameters to be constant Update Law in Standard ILC In traditional ILC algo-
within one iteration such as in Choi and Lee (2000), Hsu rithms, the control input is iteratively updated as:
et al. (2003) and Yang et al. (2002). u j+1 = Q ∗ [u
uj + L ∗e j ], (3)
As discussed above, joint friction forces in robot manipu- e j
where superscript j denotes the iteration index, and is
lators, are state-dependent and time-varying disturbances. the tracking error at j-th iteration defined as ej qq d −qq j
Hence, to achieve good tracking performance, we propose with q d and q j representing, respectively, the desired
a novel AILC to compensate for the joint friction forces. joint trajectory and the actual joint trajectory at j-th
Compared to previous formulations (e.g., Choi and Lee iteration. L ∗ represents the learning filter/matrix that
(2000)), we have proposed a new adaptation law based on varies depending on the specific control algorithm and Q∗
the work in Slotine and Li (1987), which can significantly represents a generic Q-filter/matrix that aims to improve
improve, not only the accuracy of parameter estimates, the robustness of the learning process. For instance, in
but also the tracking performance in joint space. Theoret- PD-type ILC, L ∗ is a gain matrix, and Q ∗ is typically
ical analysis and proof of error convergence are provided, a low-pass filter to prevent high-frequency measurement
along with simulation and experimental verification. Per- noises from deteriorating learning performance.
formance comparisons between a traditional AILC design
in Choi and Lee (2000) and our proposed AILC are also Insufficient Friction Compensation via Standard ILC To
provided, demonstrating that our proposed approach can demonstrate the limitations of Standard ILC, we applied
achieve better tracking performance with improved param- a PD-type ILC on a simulated planar 1-DOF manipulator
eter estimates when nonlinear joint friction forces exist in subject to Stribeck friction. The tracking performance
robot manipulators. in joint space and the friction forces over 10 iterations
are shown in Fig. 1. We can see that standard ILC
2. PROBLEM STATEMENT fails to effectively compensate for significant changes in
friction between iterations. Particularly at later iterations,
2.1 Dynamic Model of Manipulators the non-smoothness of the frictional force become more
significant, resulting in a degradation in ILC performance.
In general, the dynamic model of an n-DOF robot manip- 6
Position Error (deg)

First Iteration
4 Final Iteration
ulator can be expressed as 2
Intermediate Iterations

0
M (qq )q̈q̈ + C (qq , q̇ )q̇ + G (qq , q̇q ) + d = u (1) -2

where q , q̇q , q̈q ∈ IRn denote the vectors of positions, veloc-


-4 First Iteration
6 Final Iteration
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 Intermediate Iterations
ities, and accelerations of all joints, respectively. M (qq ) ∈
Friction Force (N m)
4
Velocity Error (deg/s)

15
2
IRn×n is the mass inertia matrix, C (qq , q̇ ) ∈ IRn×n is the 10
5 0
centripetal Coriolis force matrix, and G (qq , q̇q ) ∈ IRn is the 0 -2
-5
gravitational and friction force vector on the system. Un- -10
-4
-6
certainty in the dynamic model may result in uncertainty -15
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

of any of these three terms. d ∈ IRn is the repetitive Time (s) Time (s)

disturbance exerted on the system and u ∈ IRn is the (a) Evolution of errors (b) Frictional forces
input torque. In (1) and throughout this paper, all vectors
and matrices are time dependent, but the time variable t Fig. 1. Joint error and friction forces over ten iterations
is omitted for brevity.
Considering the time-varying nature of friction and its
In planar manipulators, such as the one used in this paper, nonlinear dependence on system states, it is difficult to
there is no gravitational force and the term G (qq , q̇ ) in (1) compensate for this iteration-varying disturbance through
can be reduced to G (q̇q ) = Ff r (q̇q ) ∈ IRn , representing the standard ILC. Hence, we need to introduce an additional
friction force vector on all n joints. While many different control signal which can identify the parameters that can
friction models are applicable, we use the Stribeck model efficiently represent the friction model, and incorporate
since it can best characterize the stiction phenomenon those parameter estimates into the design of ILC. In
observed in our experiments with the robot manipulators. the following section, a novel adaptive ILC algorithm is
According to the model in Hlouvry et al. (1994), the proposed to acheive this.
Stribeck friction force on each joint can be individually
expressed as 3. THE PROPOSED ADAPTIVE ITERATIVE
√ −( i )2 q̇i
q̇ LEARNING CONTROL
Ff ri (q̇i ) = 2e(Fbrki − Fci )e vsti
vsti
(2) 3.1 Control Law Design
q̇i
+ Fci tanh( ) + fi q̇i ,
vcouli The control input u j in our proposed adaptive ILC consists
where i = 1, 2, · · · , n. Fbrki and Fci represent, respectively, of three components as follows:
the breakaway friction force and the Coulomb friction u j = u jf + u jc + u jl , (4)
force. fi is the viscous friction coefficient. vsti and vcouli
are thresholds for the Stribeck velocity and the Coulomb where is the feedback input, u jf
is the adaptive control u jc
j
velocity, respectively. input, and u l is the iterative learning input.

579
2019 IFAC MECHATRONICS
Vienna, Austria, Sept. 4-6, 2019 Richard Lee et al. / IFAC PapersOnLine 52-15 (2019) 175–180 177

The feedback control input is of PD-type, defined as Theorem 1.


u jf = L (ė + aee )  Lz .
j j j
(5)  T  T

The matrix L and scalar a are pre-defined constants to 1) lim e j dt = lim ėej dt = 0,
j→∞ 0 j→∞ 0
stabilize the system. j
2) lim θ̃θ exists and is bounded,
The iterative control input is given by j→∞

= + βL u j+1
Lz ,
l u jl (6) j
j
where 0 < β ≤ 1 is a constant, scalar learning gain selected where θ̃θ  θ − θ̂θ represents the error associated with a
based on the desired convergence rate. mismatch in parameter estimates from their true values.
The above formulation of feedback input (5) and iterative In order to prove the validity of this theorem, we first
input (6) share the same formulation as in Park et al. consider the application of all control inputs to rearrange
(1996) and Choi and Lee (2000). As for the adaptive con- the dynamic model of our system. Combining (1) with (4,
trol input ujc , we propose a new control law to compensate 5, 8) results in the following,
for the time-varying friction forces as follows:
j
Ĉ(qq j ,q̇q j )q̇q d +Ĝ
u jc =M̂ (qq j )q̈q d +Ĉ Ĝ(q̇q j )a(M̂ (qq j )ėej +Ĉ
Ĉ(qq j , q̇q j )eej ), M (qq j )q̈q̈ j +C
C (qq j , q̇ j )q̇ j +G
G(q̇ j )+dd=Lz
Lz j +Y ujl .
Y j θ̂θ +u (11)
(7)
Notice then that
where (·)ˆ denotes the estimate of a variable.
M (qq j )żz j = M (qq j )q̈q̈ d + aM
M (qq j )ėej − M (qq j )q̈q j ,
Again, for planar manipulators, we have Ĝ G(q̇q j )=F̂f r (q̇q j ).
Note that in this work, we focus on the scenarios where C (qq j , q̇q j )zz j = C (qq j , q̇q j )q̇ d + aC
C (qq j , q̇q j )eej − C (qq j , q̇q j )q̇q j ,
all the parameter uncertainties are associated with the Y j θ j = M (qq j )q̈q̈ d + C (qq j , q̇q j )q̇ d + G (q̇ j )
friction forces, i.e., the mass, inertia, and Coriolis matrices
are assumed to be exactly known. This allows us to rewrite M (qq j )ėej + C (qq j , q̇q j )eej ).
+ a(M
(7) into the linear form as in (8) for the purposes of the
adaptation law: Rearranging these three equations, we can rewrite the
input to the system in a different form,
j j
u jc = Y θ̂θ , (8)
j
j
where Y and θ̂θ represent the regression matrix and C (qq j , q̇q j )q̇ j + G (q̇ j ) + d =
M (qq j )q̈q̈ j +C
(12)
the estimated parameter vector of all joints, respectively. Y j θ j − M (qq j )żz j − C (qq j , q̇q j )zz j + d .
Detailed definitions of Y j and θ j with the Stribeck friction
model are provided in Section 4.1. ujl  d −u
If we now define the following term ũ ujl and equate
the right-hand sides of (11) and (12), we arrive at the
3.2 Adaptation Law following,
j
To update the estimation of θ , the following update law is ujl .
M (qq j )żz j + C (qq j , q̇q j )zz j + Lz j = Y j θ̃θ + ũ (13)
proposed, which operates in both the time and iteration
domain and is of the same form as in Choi and Lee (2000): The term ũujl represents error due to the repetitive distur-
bance in the presence of an iterative input. Assuming that
˙ T
d is bounded (A1), we can see that the goal of the iterative
time-domain update: θ̂θ j = −Γ
ΓY j z j , (9a)
j+1 j input is to minimize ũujl .
iteration-domain update: θ̂θ (0) = θ̂θ (T). (9b)
The remainder of this proof comes in two parts. First, we
Γ is a constant, positive-definite learning matrix, and T j
denotes the duration of an iteration. z j is defined in (5), show that both θ̃θ and z j remain bounded in the time
j
i.e., z j =ėej +aeej . The time-domain update in (9a) is a domain. Second, we will show that the bound on θ̃θ will
gradient-type update law as in Feemster et al. (1998). either remain the same or decrease in each subsequent
T T
The iteration-domain update in (9b) means the initial iteration and that 0 z j z j dt will converge to zero.
parameter estimate of the next iteration is set as the final
parameter estimate of the previous iteration such that
parameter estimates remains continuous across iterations.
Boundedness of Error Signals Errors in the time domain
will be bounded if inputs to the system are bounded.
3.3 Stability and convergence analysis
Proof. Consider, the following Lyapunov candidate func-
In this section, we will prove the stability and error tion in the time domain,
convergence of the proposed control scheme as given by
(4)-(9). We introduce the following two assumptions: 1 T 1 jT j
W j = z j M (qq j )zz j + θ̃θ Γ −1θ̃θ . (14)
A1) The iteratively repetitve disturbance input d in (1) 2 2
is bounded ∀t ∈ [0, T]; Taking its time derivative results in (15). Here, note that
A2) The resetting condition holds, e j (0) = ėej (0) = 0 ; ∀j M (qq j )  M j , and C (qq j , q̇q j )  C j for brevity,

580
2019 IFAC MECHATRONICS
178
Vienna, Austria, Sept. 4-6, 2019 Richard Lee et al. / IFAC PapersOnLine 52-15 (2019) 175–180

1 T j j jT
 T
T ˙
Ẇ j = z j M j żz j + z j Ṁ M z + θ̃θ Γ −1θ̃θ j
T T

2 ∆V1j = uj+1
ũ l uj+1
L −1ũ l ujl L −1ũ
− ũ ujl dτ
0
T
= z j (Y
j
ujl − C j z j − Lz j )
Y j θ̃θ + ũ  T T

1 T j j jT
= ujl + ũ
(∆ũ ujl )T L −1 (∆ũ
ujl + ũ
ujl ) − ũ ujl dτ
ujl L −1ũ
˙
+ z j Ṁ M z + θ̃θ Γ −1θ̃θ j 
0
2 T T T
1 T j T j (15) = ujl L −1 ∆ũ
∆ũ ujl + 2∆ũ
ujl L −1ũ
ujl
= z j (ṀM − 2C C j )zz j + z j Y j θ̃θ 0
2  T
jT ˙ T T T T
ujl
+ θ̃θ Γ −1θ̃θ j − z j Lz j + z j ũ = β 2z j Lz j − 2β(Ẇ j + z j Lz j )dτ
T T 0
ujl .
= −zz j Lz j + z j ũ j+1T j+1 jT j
∆V2j = β(θ̃θ (0)Γ Γ−1θ̃θ (0) − θ̃θ (0)Γ Γ−1θ̃θ (0))
j
L)||zz || +
≤ ||zz ||( −λmin (L j
ujl ||
||ũ ) From (14), it can seen that
 T T 1  j T
where ||•|| denotes the 2-norm of a vector and λmin (•) the 1 T  j
Ẇ j dτ = z j M (qq j )zz j  + θ̃θ (T)Γ Γ−1θ̃θ (T)
minimum eigenvalue of a matrix. We’ve taken advantage 0 2 0 2
j 
of the fact that ṀM − 2C C j is a skew symmetric matrix jT j
− θ̃θ (0)Γ Γ−1θ̃θ (0)
(a general property of the robot manipulators, Lin et al.
(1995), to conclude that the term 12 z j (Ṁ
T
M − 2C C j )zz j = 0. 1 T
= z j (T )M
M (qq j )zz j (T )
T j j T
˙ 2
Moreover, (zz j Y j )θ̃θ + θ̃θ Γ −1θ̃θ j = 0 due to (9a). 1 j+1T j+1 jT j
+ θ̃θ (0)ΓΓ−1θ̃θ (0) − θ̃θ (0)Γ Γ−1θ̃θ (0),
Given this result, we can now establish the passivity of 2
the system through the following proof by contradiction: because it is prescribed that θ j+1 (0) = θ j (T) and by (A2),
Assume z j (t) is not bounded, such that limt→∞ ||zz j (t)|| = the initial errors of each iteration z j (0) = 0 . When we
∞, and that the time interval [0, T] is sufficiently long substitute this result into our equation for ∆V1j and add
enough such that z j (T) is very large. Then there exists a ∆V1j + ∆V2j = ∆V j . The following equation follows,
time,  T
T T
∆V j =−βzz j (T)M
M (qq j )zz j (T)−β(2−β) z j Lz j dτ. (17)
ujl 
ũ
tb  sup{t | 0 < t ≤ T and zz j (t) ≤ } 0
λmin (LL)
 tb Due to the form of ∆V j , the value of V j will continue
j j ∗T
θ̃θ (tb )=θ̃θ (0)−Γ
Γ Y j z j (t)dτ  to decrease with each subsequent iteration. Since V 1
0
 tb is positive, and V j is bounded below by zero, then it
j j must converge to some fixed value, limj→∞ V j ≤ V 1 .
≤θ̃θ (0)+λmax (ΓΓ) sup (λmax (Y Y )) zz j (t)dτ j
t∈[0,tb ] 0 uj and θ̃θ will also approach zero
Consequently, its terms ũ
as j → ∞ and converge to some fixed value. Additionally,
j since V j must converge, then ∆V j must eventually equal
Here, we assume ||θ̃θ (0)||<∞. Since both ||zz j (tb )|| and zero. In order for this to occur, it must be true that
j  T jT j
||θ̃θ (tb )|| may be large but finite values, the function 0
z Lz dτ = 0 as j → ∞.
ujl ||
||ũ
W j (tb ) is also finite. Additionally, since ||zz j || > ∀t ∈ Lastly, if we consider the relationship between z j and e j ,
λmin (L L)
j j
[tb , T], it is also the case that Ẇ (t) < 0, and accordingly we can establish the transfer function Ej (s) = 1 under
j j Z (s) s+a
W (t) ≤ W (tb ) < ∞ ∀t ∈ [tb , T]. However, it is not pos-
(A2). Hence, as long as a > 0, it is guaranteed that z j =00
sible that W j (t) be bounded and ||zz j (t)|| not be bounded, j j
and this creates a contradiction. As a result, ||zz j || must be leads to e =00, which further indicates that ėe = 0 .
ujl || is bounded.
bounded when ||ũ 4. SIMULATION

4.1 Manipulator Configuration


Convergence in the Iteration Domain We now analyze
ujl changes as j increases, and prove the error z j
how ũ In this section, we study by simulating an application of
converges as j → ∞. the proposed adaptive ILC algorithm to a 3-DOF planar
manipulator with joint friction, as shown in Fig. 2(a). Due
Proof. Consider the following discrete-time Lyapunov to the difficulty of directly identifying all unknown param-
candidate function, eters, we first identify vsti and vcouli offline. Hence, the
 T unknown friction force for each joint i can be represented
T jT j
V j (t)= ujl L −1ũ
(ũ ujl )dτ +βθ̃θ (0)Γ
Γ−1θ̃θ (0)=V1j +V2j . as
τ =0 Ff ri (qi , q̇i ) = Y iθ i ,
(16) √
θ i = [ 2e(Fbrk,i − Fc,i ) Fc,i fi ]T  [bi1 bi2 bi3 ]T ,
where V1j and V2j equate to the first and second term of q̇i q̇i 2 q̇i
Y i = [ e− vst tanh( ) q̇i ].
V j (t), respectively. We now define the terms ∆ũ uj+1
ujl  ũ l − vst vc
ujl = (dd −u
ũ ujl −βL ujl ) = −βL
Lz j )−(dd −u Lz j and ∆V j  V j+1 − Note that Y i is the i-th row of the regression matrix Y .
j
V j . We then compute the following differences, For the system of interest, the vector θ̂θ becomes a vertical

581
2019 IFAC MECHATRONICS
Vienna, Austria, Sept. 4-6, 2019 Richard Lee et al. / IFAC PapersOnLine 52-15 (2019) 175–180 179

j j j

Joint 1 (deg)
10 baseline AILC
concatenation of θ̂θ 1 , θ̂θ 2 , θ̂θ 3 and Y j becomes a matrix, the 5 proposed AILC

rows of which are Y j1 , Y j2 , Y j3 . The sampling time is 1ms.


0
-5
0 2 4 6 8 10 12 14 16
200

Joint 2 (deg)
5

150

Angular Position (deg)


0
Joint 1
100 joint 1
r1 joint 2 -5
l3 joint 3
0 2 4 6 8 10 12 14 16
l1 50
20

Joint 3 (deg)
Joint 2 r3
0
0
Joint 3
-50
r2 0 2 4 6 8 10 12 14 16 -20
l2 0 2 4 6 8 10 12 14 16
Time (s)
Time (s)
(a) The 3-DOF manipulator (b) Desired trajectories in joints
Fig. 4. Comparison result in position errors (final iteration)
Fig. 2. The planar manipulator and desired trajectories in
simulation Feedback

Joint 1 (N m)
20 20 Feedback

Joint 1 (N m)
Iterative Iterative
10 Adaptive 10 Adaptive
0 0
-10 -10
0 2 4 6 8 10 12 14 16
4.2 Simulation Results 10
0 2 4 6 8 10 12 14 16

Joint 2 (N m)
5

Joint 2 (N m)
5
0
0
The desired joint trajectories for the manipulator are -5
-5

shown in Fig. 2(b). These are typical trajectories for


0 2 4 6 8 10 12 14 16
0 2 4 6 8 10 12 14 16

Joint 3 (N m)
2
the extension and retraction operations of a silicon wafer
Joint 3 (N m)
5
0
0
handling robot to carry wafers to desired locations. Five -5
-2

iterations of the proposed AILC are run and position errors 0 2 4 6 8 10 12 14 16


-4
0 2 4 6 8 10 12 14 16
Time (s) Time (s)
are shown in Fig. 3, demonstrating the effectiveness of the
(a) Proposed AILC (b) Baseline AILC
proposed adaptive ILC scheme.

20
First Iteration
Fig. 5. Torque inputs in the 5th iteration (feedback u jf ,
Joint 1 (deg)

Final Iteration
10 Intermediate Iterations

0 iterative u jl , and adaptive u jc )


-10
0 2 4 6 8 10 12 14 16

10
5. EXPERIMENTS ON A WAFER HANDLING
ROBOT MANIPULATOR
Joint 2 (deg)

5
0
-5
-10
0 2 4 6 8 10 12 14 16
5.1 Experimental Setup
50
Joint 3 (deg)

0 We utilize a 5-DOF Silicon Wafer Handling robot for


experimentation. The joints are configured as shown in
-50
0 2 4 6 8 10 12 14 16 Fig. 6: Joint 1 refers to the base rotation joint, Joint 2
Time (s)
refers to the translation joint in z direction, and Joints 3, 4
Fig. 3. Position errors of the proposed AILC scheme over and 5 refer to the remaining rotation joints. Each joint can
five iterations be controlled independently. Among all five joints, Joint
1 has the largest moment of inertia, and Joint 5 suffers
from static friction most significantly. These properties
We compare the proposed approach with the one in Choi make the control of these two joints challenging. Hence,
and Lee (2000) (noted as “baseline”), with results shown we focused our experiment on Joints 1 and 5. The desired
in Fig. 4. The proposed algorithm can achieve much better trajectories of these two joints are shown in Fig. 6(b).
tracking performance than the baseline AILC. Moreover,
the torque inputs in Fig. 5 show that the adaptive input
component contributes less to the overall control effort of
the baseline AILC than in the proposed AILC.
0

The performance improvement of the proposed AILC in -10


Angular Position (deg)

Fig. 4 is a result of the adaptive input design. We can see -20

in (7) that the proposed adaptive input accounts for model -30

discrepancies by producing a model-based feedforward in- -40

put and a feedback compensation signal based on model -50 Joint 1


Joint 5

estimates and tracking error. As the error decreases, the -60

adaptive input gradually reduces to a model-based feed- -70


0 0.5 1 1.5 2 2.5 3
Time (s)
forward control signal. In contrast, the baseline adaptive
input in (A.1) also accounts for model discrepancies, but (a) The robot (side view) (b) Desired trajectories
in contrast, its control effort will approach zero as model
estimates improve and error decreases. Fig. 6. The experiment setup and desired trajectories

582
2019 IFAC MECHATRONICS
180
Vienna, Austria, Sept. 4-6, 2019 Richard Lee et al. / IFAC PapersOnLine 52-15 (2019) 175–180

For both joints, we test both the baseline and proposed REFERENCES
AILC approaches. 23 iterations and 13 iterations of the Bondi, P., Casalino, G., and Gambardella, L. (1988). On the iterative
both algorithms were conducted on Joint 1 and Joint learning control theory for robotic manipulators. IEEE Journal
5, respectively. The number of iterations was set such on Robotics and Automation, 4(1), 14–22.
that neither algorithm showed any significant performance Lin, C.Y., Sun, L., and Tomizuka, M. (2015). Robust principal
improvement after that many iterations. component analysis for iterative learning control of precision
motion systems with non-repetitive disturbances. In American
Control Conference (ACC), 2015, 2819–2824. IEEE.
5.2 Experimental Results Park, B.H., Kuc, T.Y., and Lee, J.S. (1996). Adaptive learning
control of uncertain robotic systems. International Journal of
Fig. 7 gives the position errors in the final iteration and Control, 65(5), 725–744. doi:10.1080/00207179608921719.
the performance improvements over iterations by using Bristow, D.A., Tharayil, M., and Alleyne, A.G. (2006). A survey of
the two AILC schemes. Compared to the baseline, the iterative learning control. IEEE Control Systems, 26(3), 96–114.
proposed AILC algorithm results in small performance Choi, J.Y. and Lee, J.S. (2000). Adaptive iterative learning control
improvements in Joint 1 but significant improvements in of uncertain robotic systems. IEE Proceedings - Control Theory
and Applications, 147(2), 217–223. doi:10.1049/ip-cta:20000138.
Joint 5, as shown in Fig. 7(a). The friction forces in both
Slotine, J.J.E. and Li, W. (1987). On the adaptive control of robot
cases can be better compensated. manipulators. The International Journal of Robotics Research,
5 10
7 6(3), 49–59. doi:10.1177/027836498700600303.
0.5
4.5 Joint 1 baseline Hsu, C.T., Chien, C.J., and Yao, C.Y. (2003). A new algorithm
Joint 1 proposed
4 of adaptive iterative learning control for uncertain robotic sys-
Position Error (deg)

Joint 5 baseline
3.5
Joint 5 proposed
0
3 tems. In 2003 IEEE International Conference on Robotics and
2.5 Automation (Cat. No.03CH37422), volume 3, 4130–4135 vol.3.
J

2
Joint 1 baseline 1.5
doi:10.1109/ROBOT.2003.1242232.
-0.5
Joint 1 proposed 1 Kuc, T. and Lee, J.S. (1991). An adaptive learning control of
Joint 5 baseline 0.5
Joint 5 proposed
0
uncertain robotic systems. In [1991] Proceedings of the 30th
-1
0 0.5 1 1.5 2 2.5 3
2 4 6 8 10 12 14 16 18 20 22
IEEE Conference on Decision and Control, 1206–1211 vol.2. doi:
Iteration number
10 10.1109/CDC.1991.261560.
(a) Tracking errors (b) Error evolution Yang, S., Fan, X., and Luo, A. (2002). Adaptive robust iter-
5 ative learning control for uncertain robotic systems. In Pro-
Velocity Error (deg/s)

Fig. 7. Tracking errors in final iteration and along itera- ceedings of the 4th World Congress on Intelligent Control and
0
tions Automation (Cat. No.02EX527), volume 2, 964–968 vol.2. doi:
-5
10.1109/WCICA.2002.1020719.
As a compact metric of performance, we define the follow- Feemster, M., Vedagarbha, P., Dawson, D.M., and Haste, D.
ing
-10 cost associated with tracking error, (1998). Adaptive control techniques for friction compensation.
0 0.5 1 1.5
Time (s)
2
 tf3
2.5
In Proceedings of the 1998 American Control Conference. ACC
J= z TLz j dτ. (18) (IEEE Cat. No.98CH36207), volume 3, 1488–1492 vol.3. doi:
0 10.1109/ACC.1998.707076.
This metric is a direct result of (16) and (17), and can Hlouvry, B.A., Dupont, P., and Wit, C.C.D. (1994). A survey of
models, analysis tools and compensation methods for the control
be expected converge to zero as j → ∞. Fig. 7(b) shows
of machines with friction. Automatica, 30(7), 1083 – 1138. doi:
the evolution of J along iterations using the baseline and https://doi.org/10.1016/0005-1098(94)90209-7.
proposed AILC. We can see that the proposed algorithm Lin, H., Lin, T., and Yae, K. (1995). ”on the skew-symmetric
leads to great improvements in all iterations of Joint 5 and property of the newton-euler formulation for open chain robot
later iterations of Joint 1. manipulators. Proc. American Control Conf.
Peng, C., Sun, L., and Tomizuka, M. (2017). Constrained iterative
6. CONCLUSION learning control with pso-youla feedback tuning for building
temperature control. IFAC-PapersOnLine, 50(1), 3135–3141.
Sun, L., Chen, X., and Tomizuka, M. (2014). Selective iterative
In this paper, we discussed the fundamental limitation learning control with non-repetitive disturbance rejection. In
of standard ILC with application to robot manipulators Proceedings of 2014 International Symposium on Flexible Au-
subjected to iteration-varying joint friction forces. A new, tomation.
adaptive iterative learning control algorithm is proposed Takegaki, M. and Arimoto, S. (1981). A new feedback method for
to identify an unknown Stribeck friction model and com- dynamic control of manipulators. Journal of Dynamic Systems,
pensate for it. Proof of error convergence, simulation and Measurement, and Control, 103(2), 119–125.
experimental studies with the proposed and a baseline Tayebi, A. (2004). Adaptive iterative learning control for robot
manipulators. Automatica, 40(7), 1195–1203.
AILC were provided. The results showed that the proposed
algorithm performed better than the baseline in terms of
reducing tracking performance over iterations while main- Appendix A. BASELINE ADAPTIVE INPUT
taining the same parameter adaptation scheme.
The proposed approach can be widely applied to other The baseline adaptive input proposed in Park et al. (1996)
systems subject to iteration-varying friction models. More- and Choi and Lee (2000) is,
over, estimation of unknown parameters can be extended u jc = M e (qq j )q̈q̈ d + C e (qq j , q̇ j )q̇ d + Ĝ e (q̇q j )
to include those associated with mass inertia M (qq ) and (A.1)
Coriolis C (qq , q̇ ) terms. By introducing a novel, adaptive M (qq j )ė j + C (q̇ j )eej )
+ a(M
input component, Type II non-repetitive disturbances can where M e (qq j ) = M (qq j ) −M M (qq d ) , C e (qq j , q̇ j ) = C (qq j , q̇ j ) −
j j
be better addressed using the proposed approach than Ĝ(q̇ jd ).
Ĝ(q̇ j ) − Ĝ
C (qq d , q̇ d ), and Ĝ e (q̇ j ) = Ĝ
previous AILC approaches.

583

You might also like