Anda di halaman 1dari 11

IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 11, NO.

1, FEBRUARY 2006 55

Design of Learning Input Shaping Technique for


Residual Vibration Suppression in an Industrial Robot
Juyi Park, Member, IEEE, Pyung-Hun Chang, Member, IEEE, Hyung-Soon Park,
and Eunjeong Lee, Member, IEEE

Abstract—In this paper, a practical method is proposed to sup- in their structure and robust against disturbances or parameter
press residual vibrations of industrial robots without a real-time variations, thus very useful for practical applications.
estimation of vibration frequencies. Through theoretical analysis Based on the observations above, we have considered the in-
and experiments, we designed an input shaping technique (IST)
for the first three axes of a six-degrees-of-freedom industrial robot. put shaping technique (IST) proposed by Singer and Seering [3]
Iterative learning IST (LIST) is applied to the first axis to suppress as the feedforward scheme to be combined with a feedback
its time-varying nonlinear residual vibration, while conventional controller. Since the beginning, the IST has attracted attention
IST is applied to the second and third axes. Experimental results owing to its effectiveness and simplicity. Its effectiveness has
show that LIST can suppress residual vibrations to a level similar been proven in practical systems such as ship cranes [5], chip
to that of a time-varying IST which requires complicated real-time
estimation of a dynamic model. The LIST is an attractive method mounters [6] and open container, of liquid [7]. Nevertheless,
for suppression of nonlinear and time-varying residual vibrations since the IST was proposed originally for linear time-invariant
in industrial robots which perform repetitive tasks because most systems [3], it is not so effective for systems with nonlinear
industrial robots have limited computing power and memory space and time-varying characteristics such as the multilink robots
in their controllers. we are interested in. Even robust IST [3], which compensates
Index Terms—Flexible structure, industrial robot, input shap- for inaccuracies of frequency estimation, is not of much help
ing, iterative learning algorithm, residual vibration. for these systems.
To overcome this difficulty, many attempts have been made
to improve IST for nonlinear and time-varying systems. As for
I. INTRODUCTION
nonrobotic systems, online adaptive schemes were developed by
N controlling industrial robots, fast and precise motions are
I required for better productivity. Such motions, however, are
often restricted by residual vibrations in the end-effector, which
Tzes and Yurkovich [8] and Bodson [9]. As for robotic systems,
Rappole [10] applied a time-varying IST (TVIST) to a two-link
flexible manipulator using a lookup table. Magee and Book [11]
tend to be time varying and nonlinear due to the configuration- modified the IST to eliminate the first two modes of vibration in
dependent friction, inertia variation, and nonlinear stiffness of a large and flexible manipulator with a configuration-dependent
joints. inertia. Cho and Park [12] proposed a method to determine the
For suppression of residual vibrations in flexible systems, exact time-varying impulse sequence and applied it to a two-link
there are two distinct approaches: open-loop feedforward [1] flexible robot. Additionally; there have been similar attempts to
and closed-loop feedback [2]. In terms of performance, the latter apply the IST to various robots [13].
scheme is more attractive than the former because it is inherently But these schemes require intensive computing power for
more robust against disturbances and parameter variations. In real-time computation or memory space for a frequency map.
terms of practical implementation, however, the closed-loop Moreover, some schemes require an exact dynamic model of the
approach makes overall systems more complex and expensive. plant in real time. Because of the limited computing power and
More specifically, the increased states due to vibrational modes memory space of industrial robots, however, it is difficult to meet
increase the order of the control systems, thereby requiring more these requirements. Furthermore, it is a major work to obtain
computation and more sensors as well as the measurability of the an exact model for an industrial robot. These practical issues
additional states. Because of these difficulties, many researchers have made us develop a scheme to adopt IST for an industrial
have developed feedforward schemes integrated with feedback robot.
controllers [3], [4]. They reported that the schemes are simple The IST we employed in this paper is an iterative learning
IST (LIST) [14], which iteratively updates the parameters of
IST based on data from previous trials. The idea central to
Manuscript received December 24, 2003; revised April 18, 2005. Recom-
mended by Technical Editor R. V. Patel. The work of P.-H. Chang was supported this approach came from our observation that most tasks of
by the Korea Science and Engineering Foundation (KOSEF) through the Center industrial robots are planned in advance and they are repet-
for Human-Friendly Welfare Robotic Systems (HWRS). itive by nature. The robot to which we apply iterative LIST
J. Park is with Daewoo Shipbuilding and Marine Engineering Company Ltd.,
Koje City, 656-714 Korea. is an industrial robot which carries heavy payloads on plant
P.-H. Chang and E. Lee are with the Korea Advanced Institute of Science and floors. Since IST and LIST were developed originally for single-
Technology (KAIST), Daejon 305-701, Korea. input single-output (SISO) systems, this paper deals with tech-
H.-S. Park is with the Rehabilitation Institute of Chicago, Chicago, IL 60611
USA. niques to extend the application of IST and LIST to industrial
Digital Object Identifier 10.1109/TMECH.2005.863365 robots that are multiple-input multiple-output (MIMO) systems.

1083-4435/$20.00 © 2006 IEEE


56 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 11, NO. 1, FEBRUARY 2006

TABLE I
PARAMETERS OF THE ROBOT

Fig. 1. Schematic diagram of the industrial robot.

Although the basic idea of LIST was already presented in [14]


with simple simulations and experiments, auxiliary techniques
are needed for LIST to be more practical. This paper, there-
fore, represents techniques for practical application of LIST
including an application strategy for MIMO systems with cou-
pling, parameter updating scheme for nonlinear time-varying
systems, and a selection method for initial guesses of IST pa-
rameters. In addition, this paper presents analytical verification
for convergence of LIST. Since vibration in the first joint is sub-
stantially large and time varying, LIST is applied only to the first Fig. 2. Model of joint flexibility.
joint and conventional IST to the second and third joints. We
use the input/output value of the first joint for the learning pro-
cess so that its nonlinear and time-varying residual vibrations Therefore, the dynamic equations are derived only for the first
can be suppressed without a dynamic model of the robot. The three joints.
experimental results of iterative LIST are compared to those of Since joint flexibility is the primary cause of robot vibration,
TVIST, and their merits and demerits are evaluated. This proce- it is included in the model. The flexibility of the ith joint is
dure can be used as a practical guideline for various industrial modeled as a torsional spring with stiffness KT i [13]. Accord-
robots other than ours. ingly, each axis also includes two joint variables representing
This paper is organized as follows. Section II describes the both motor angle (θm i ) and link angle (θi ), as shown in Fig. 2.
dynamics of the industrial robot and discusses IST application The dynamic equations for this robot are derived by using
strategy for the robot. Section III shows its nonlinear and time- Lagrange’s method and described as follows:
varying vibrations and presents the main idea and algorithm of bk 1 KT 1
the iterative LIST. Section IV describes our experimental results, M11 θ̈m 1 + bm 1 θ̇m 1 + θ̇r 1 + θr 1 = τ1 (1)
r1 r1
and discussion is followed by conclusions in Section V.
M22 θ̈1 + b1 θ̇1 − bk 1 θ̇r 1 − KT 1 θr 1 + N1 (θ, θ̇) = 0 (2)
II. UNDERSTANDING OF A ROBOT DYNAMICS FOR
bk 2 KT 2
LIST APPLICATION M33 θ̈m 2 + bm 2 θ̇m 2 + θ̇r 2 + θr 2 = τ2 (3)
r2 r2
In this section, the dynamic properties of the robot are de-
scribed in order to show the nonlinear and time-varying charac- M44 θ̈2 + M46 θ̈3 + b2 θ̇2 − bk 2 θ̇r 2 − KT 2 θr 2
teristics of the robot, and a strategy to apply IST for the robot is +N2 (θ, θ̇) + G2 (θ) = 0 (4)
presented.
bk 3 KT 3
M55 θ̈m 3 + bm 3 θ̇m 3 + θ̇r 3 + θr 3 = τ3 (5)
A. Robot Dynamics r3 r3

The industrial robot of our interest has a parallelogram- M64 θ̈2 + M66 θ̈3 + b3 θ̇3 − bk 3 θ̇r 3 − KT 3 θr 3
linkage structure with six degrees of freedom (6 DOF). The +N3 (θ, θ̇) + G3 (θ) = 0 (6)
schematic diagram of the robot is shown in Fig. 1 and its esti-
mated parameters are presented in Table I. As can be expected where θr i = ( θrmi i − θi ), θ = [θ1 θ2 θ3 ]T ; and Ni (θ, θ̇) denotes
from the numerical values of these parameters, the robot is in- the Coriolis and centrifugal torque and Gi (θ) the gravity torque
tended for handling heavy payloads or spot welding with the at the ith joint; Ni (θ, θ̇)s and Gi (θ)s are shown in Appendix I.
maximum payload of 120 kg. In the equations above, it is assumed that viscous damping with
To examine its frequency variation properties, the dynamic a coefficient of bm i is also present at each motor, bi at each joint,
equation of the robot was derived based on the schematic di- and bk i at each torsional spring.
agram depicted in Fig. 1. Note that these last three axes have Under the assumption that each link has a uniform and
little effect on inertia variation and, thus, on frequency variation. symmetric cross section, the elements of inertia matrix Mij (θ)s
PARK et al.: INPUT SHAPING TECHNIQUE FOR AN INDUSTRIAL ROBOT 57

can be expressed as the following: are drawn from the analysis of dynamic equations of the robot
and experimental verification.
M11 = Jm 1 (7)
1) For axes 1, 2, and 3, the variation of IST parameters in
2 2 2 2
M22 = (ma lg a + Ja + Jc )c2 + (mb lg b + Jb + Jd )c3 one axis has little effect on the residual vibration of the
other two. For example, even though the IST parameters
+ mc (lb c3 − lg c c2 )2 + md (lc c2 + lg d c3 )2
of axis 1 are changed, the residual vibration in axes 2 and 3
+ mP {lc c2 + (ld − lb )c3 }2 (8) are affected only slightly (Appendix II).
2) Residual vibration in axis 1 is dominant while those in
M33 = Jm 2 (9)
axes 2 and 3 are relatively small (see Fig. 10).
M44 = ma lg a 2 + Ja + mc lg c 2 + Jc + md la 2 + mP la 2 (10) Based on these observations, we have developed a strategy to
determine IST parameters as follows.
M46 = {md la lg d − mc lb lg c + mP la (ld − lb )} cos(θ2 − θ3 )
1) Apply conventional IST to axes 2 and 3. Although the
(11) residual vibrations in axes 2 and 3 are not eliminated
entirely with conventional IST, the remaining vibrations
M64 = M46 (12)
are negligible so that conventional IST can be used in these
M55 = Jm 3 (13) axes without fine tuning.
2) Tune IST parameters for axis 1 by applying iterative learn-
M66 = mb lg b + Jb + mc lb + md lg d + Jd + mP (ld − lb )2
2 2 2
ing algorithm.
(14)
where lk and lg k (k = a, b, c, d), as shown in Fig. 1, denote the III. ITERATIVE LEARNING INPUT SHAPING TECHNIQUE
length of the corresponding link and the length to the center of
A. IST and Measures of Residual Vibrations
gravity of the link, respectively. In the same manner, mk and
Jk represent the mass and moment of inertia of the link, respec- The two-impulse sequence for suppression of the residual
tively. In addition, mP denotes the payload, attached to the end- vibration is given in the following general form:
effector, and ci and si symbolize cos θi and sin θi , respectively.  
u(s) = A1 e−T 1 s + A2 e−T 2 s r(s) (16)
Note in (8) and (11) that M22 and M46 vary with both θ2 and
θ3 , whereas (10) and (14) show that M44 and M66 are constant. where r(s) is a reference trajectory. The two-impulse sequence
Owing to M22 and M46 , hence, the robot has time-varying u(s) has four parameters: the magnitudes of impulses A1 and
vibrations [15]. A2 , and the application time T1 and T2 . Without loss of general-
The vibration is not only time varying, but also nonlinear ity, T1 may be fixed to 0 for a faster response, and A2 may also
due to robot dynamics such as nonlinear spring characteristics be fixed to (1 − A1 ) in order to maintain a unit gain. Then, there
of harmonic drive and friction. Moreover, nonlinear terms in remain only two independent parameters to be determined, A1
dynamic equations, such as Coriolis, centrifugal, and gravity and T2 .
torque, cause nonlinear residual vibration of the robot. As an According to Singer and Seering’s study [3], the residual
example, let us examine N1 , which can be expressed in the vibrations due to the two-impulse sequence vanish if A1 and
following form: T2 are properly selected by the following equations:
N1 (θ, θ̇) = f1 (θ2 , θ3 , θ̇2 , θ̇3 )θ˙1 . (15) 1
A1n = √ 2 (17)
Since f1 (θ2 , θ3 , θ̇2 , θ̇3 ) is a nonlinear function, N1 (θ, θ̇) can 1+e −ζ π /
1−ζ
be regarded as a nonlinear damper with a nonlinear damping π
coefficient. Therefore, we can see that the nonlinear terms in T2n =  (18)
N1 cause nonlinear residual vibration of the robot. ωn 1 − ζ 2
Due to such time-varying and nonlinear characteristics, resid- where ( · )n denotes the exact values of ( · ), and ζ and ωn mean
ual vibrations of industrial robots are not well suppressed by the damping ratio and the natural frequency of the residual
conventional IST, which was derived for linear time-invariant vibration to be suppressed, respectively.
(LTI) systems [8]–[10], [12]. To this end, we may apply mod- When A1 = A1n or T2 = T2n , however, the residual vibra-
ified versions of the IST technique for the robot. The authors, tion does not vanish after the application time T2 . For exam-
however, found that, with properly tuned IST parameters, con- ples, we examined the responses for four cases: T2 > T2n and
ventional IST can suppress residual vibration in nonlinear and T2 < T2n for A1 = A1n , and A1 > A1n and A1 < A1n for
time-varying systems [14], [16]–[18]. Since conventional IST T2 = T2n . Fig. 3 shows the responses to the two-impulse se-
does not require intensive computational time or large memory quence for these four cases. In the figure, the solid lines rep-
space, we employed IST to suppress the residual vibration of resent the original vibration, the unit impulse response without
the industrial robot. IST, and the dotted lines represent the resultant responses to
the two-impulse sequence. Fig. 3(a) shows the response for the
B. IST Application Strategy for the Robot
case T2 > T2n , and Fig. 3(c) for the case T2 < T2n ; Fig. 3(b)
The dynamic equations above clearly show that the robot is a for the case A1 > A1n , and Fig. 3(d) for the case A1 < A1n .
multivariable system. Nevertheless, the following observations Both responses show that the residual vibrations still persist and
58 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 11, NO. 1, FEBRUARY 2006

Fig. 3. Responses to the two-impulse sequence with inaccurate parameters. Solid lines represent unit impulse responses; dashed lines represent the resultant
responses to the two-impulse sequence with T 2 = T 2n or A 1 = A 1n .

TABLE II with T2 varying while A1 is fixed as in Fig. 4(a); and with A1


OBSERVED RANGES OF φ FOR NINE POSSIBLE CASES
varying while T2 is fixed as in Fig. 4(b).
Fig. 4(a) shows that the magnitude and the phase vary
in nonlinear fashions, with M having local minima at T2 =
T2n , 3T2n , 5T2n , . . . (or TT2n2 = 1, 3, 5, . . .) [19]. Accordingly,
we have restricted the limit 0 ≤ T2 ≤ 2T2n throughout the
learning process. In this region, M has only one minimum with-
that there exist phase differences from the original vibration
out any additional local minima, thereby enabling us to apply a
without IST.
function minimization method for parameter update. The range
In order to adopt a learning scheme for IST, we have used two
of A1 , 0 ≤ A1 ≤ 1, is determined by (17), and Fig. 4(b) shows
measures: the magnitude and phase difference of residual vibra-
that M and φ vary linearly for the range. For this reason, the
tion. More specifically, the magnitude measure M is defined as
convergence of A1 has posed little difficulty.
 tf
M≡ |ν(t)| dt (19)
ti B. Iterative Learning Scheme
where ν(t) is any signal representing residual vibration such as The iterative learning scheme updates the IST parameters A1
position, velocity, or acceleration. ti ≤ t ≤ tf is the duration and T2 by using M and φ obtained from the experiment (or
when the residual vibration is significant. simulation) with parameters from the previous execution. The
Another measure, φ, is the phase difference between the resid- overall algorithm is summarized as follows: With an initial guess
ual vibrations with and without IST. Examples of φ for improper of A1 and T2 , update T2 first while keeping A1 fixed, until M
sets of T2 are shown in Fig. 3. When T2 is larger than the proper reaches its minimum; then update A1 with keeping T2 fixed;
value, φ has a positive value or phase lead as shown in Fig. 3(a). alternate the updating of A1 and T2 until a satisfactory M is
On the other hand, when T2 is smaller than the proper value, obtained. The phase difference φ is examined only to determine
φ then has a negative value or phase lag as shown in Fig. 3(c). the update rule as explained in the next section.
Hence, by observing φ, we can determine whether an IST pa- 1) Parameter Update: For convenience, let p denote a pa-
rameter is larger or smaller than the proper value. Table II shows rameter to be updated, A1 or T2 , with pk denoting p at the kth
the observed ranges of φ for nine possible cases of improper A1 iteration and Mk denoting M obtained from an experiment (or
and T2 . simulation) by using pk .
In order to put the observation above into a broader perspec- The iterative learning scheme is to find pn , the proper param-
tive, we draw the simulation plots of M and φ for two cases: eter to minimize M , which has a shape shown in Fig. 5(a). Thus,
PARK et al.: INPUT SHAPING TECHNIQUE FOR AN INDUSTRIAL ROBOT 59

Fig. 4. Variation of M and φ: (a) for T 2 varying with A 1 fixed, (b) for A 1 varying with T 2 fixed.

Fig. 5. The shape of M and M̃ .


Fig. 6. (a) A minimum is in the interval [b, c] if f (d) < f (b) or (b) a minimum
is in the interval [a, d] if f (b) < f (d).
updating the IST parameters may be regarded as a function min-
imization and various function minimization techniques can be
used to this end.
From a different perspective, however, the parameter updating
may be viewed as a root finding problem and various root finding
methods can be utilized for this purpose. To elaborate, define a
e M̃ as
M̃ = M sign(pn − p) (20)
where
 1, if x > 0
Fig. 7. If M̃ is discontinuous near p n , p k +1 obtained by (21) may be too far
from p n .
sign(x) = 0, if x = 0
−1, if x < 0.
a < b < d, the interval in which f (x) has a minimum can be
Fig. 5(b) shows M̃ as a function of p. Note that the condition, narrowed down.
sign(x) = 0 if x = 0, makes M̃ = 0 when p = pn , and thus we In the golden section search, the new point d in a given interval
can use root finding methods for parameter updates. is called the golden mean point since it is normally selected as
In parameter updating, we have made use of a well estab- a golden mean of the given triplet [21]. Instead of using the
lished function minimization method called the golden section golden mean point method, however, the secant method is used
search [20] in conjunction with a root finding method called the to increase the speed of convergence. This is possible because
secant method [21]. The latter was used to increase the speed of we know the sign of ek from φ, where ek = pn − pk .
convergence for the former. In the secant method, pk +1 is calculated from
Briefly summarizing, the golden section search is based on the
fact that a minimum can be found when there is a triplet of points, M̃k
pk +1 = pk + (pk − pk −1 ) (21)
a < b < c, such that f (b) is less than both f (a) and f (c). In this M̃k −1 − M̃k
case, the function has a minimum in the interval [a, c]. With this
idea, one can narrow down the interval by choosing a new point where M̃j is the M̃ obtained from experiment or simulation
d between a and b or between b and c. Let us suppose that d is using pj .
chosen between b and c. Then we know that a minimum is in The secant method, however, does not work well when M̃ is
the interval [b, c] if f (d) < f (b) as in Fig. 6(a); or a minimum not continuous near pn as shown in Fig. 7, because it assumes the
is in the interval [a, d] if f (b) < f (d) as shown in Fig. 6(b). As function to be smooth near the root [21]. If M̃ is discontinuous
the process above is repeated with a new triplet b < d < c or near pn , (21) determines pk +1 far from pn as shown in Fig. 7.
60 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 11, NO. 1, FEBRUARY 2006

Fig. 8. p k +1 is not in the interval where M has a minimum when the parameter
is updated improperly.

Fig. 9. Experimental setups used for the LIST. The learning process is done
on the IBM-PC 486.
As a remedy to overcome this difficulty, the following con-
straint is added: (23), we can obtain the following equation to calculate pk +1 :

|pk +1 − pk | ≤ |pk − pk −1 |. (22) Mk


pk +1 = pk − (pk − pk −1 )sat . (25)
As a result, when M̃ is discontinuous near pn , the following Mk −1 + Mk
equation is used instead of (21): Although this method increases the iteration numbers and
  delays the convergence of M , it enables us to obtain exact pk +1
M̃k without φ.
pk +1 = pk + (pk − pk −1 )sat (23)
M̃k −1 − M̃k 3) Selection of Initial Parameters: The initial guesses of A1
and T2 can be calculated using Singer’s method [3]. Although
where sat(x) is a saturation function such that this method is derived from and intended for LTI systems, we
 find that they still provide good initial guesses for nonlinear
x, if |x| < 1
sat(x) = time-varying systems.
sign(x), otherwise.
At the second iteration, p1 cannot be obtained by using the
We have proved that the updating scheme described thus far updating rule mentioned above because it requires two previous
makes pk converge to pn as iteration proceeds. The proof is experimental results. The new updating rule for p1 is made in
presented in Appendix III. the following form:
2) Updating Rule for Varying Frequencies: For some non-
linear systems, the phase difference φ is not reliable any more. p1 = p0 − sign(e0 )p0 δ (26)
For example, a spring-mass system with a nonlinear spring vi- where δ denotes a heuristically determined constant less than
brates with decreasing frequencies as the amplitude becomes one. In our learning process, we use δ = 0.1 and sign(e0 ) is
smaller, and it is of no use to measure φ between vibrations that obtained from the initial phase difference φ0 .
have different frequencies. 4) Condition for Minimum M : As mentioned above, we can
In the case of nonlinear systems with vibrations of varying narrow down the interval where M has a minimum—or M̃ has a
frequencies, therefore, we use a simple trial-error method root—through the learning process. We can say that M has con-
such as the following instead of using φ for parameter update: verged to its minimum if the interval is sufficiently small so that
Assume that both pk and pk −1 are either larger or smaller than
pn , that is, ek ek −1 > 0. Let us apply this assumption to (23) pU − p L < (27)
and obtain the following equation to calculate pk +1 : where pU and pL denote the upper and lower bound of the
interval, respectively and the predefined error bound.
pk +1 = pk + (pk − pk −1 )

Mk sign(ek ) IV. EXPERIMENT


× sat
Mk −1 sign(ek −1 ) − Mk sign(ek )

A. Experimental Setup
Mk
= pk + (pk − pk −1 )sat . (24) Experiments were performed to verify that iterative LIST
Mk −1 − Mk
is effective for suppression of residual vibrations in a robot
If pk +1 obtained from (24) is in the interval in which M has a conducting repetitive tasks. For the experiments, the reference
minimum, 1 our assumption is confirmed to be correct, and the trajectory with its trapezoidal velocity profile was given, and
learning proceeds to the next cycle. On the contrary, if the pk +1 the results of iterative LIST were then compared to those of
is not in the interval in which M has a minimum (as shown in both conventional IST as well as TVIST based on Rappole’s
Fig. 8), it indicates that the pk +1 was not properly updated, that method [10], [15].
is, the assumption is wrong. Hence, applying ek ek −1 < 0 to The schematic diagram of the experimental setup is shown
in Fig. 9. As position controllers, proportional controllers with
inner PI velocity control loops run in a stand-alone robot con-
1 It was explained with the golden section search. troller at sampling frequency 2 kHz. An accelerometer, from
PARK et al.: INPUT SHAPING TECHNIQUE FOR AN INDUSTRIAL ROBOT 61

Fig. 10. Experimental results: residual vibration without IST. Note that the Fig. 11. Experimental results. Magnitude measure M decreases as the number
robot moved for 1.35 s and the tip response after 1.35 s is the residual vibration. of iterations increases.

PCB Piezotronics, Inc., is attached at the tip of the robot to


measure the residual vibrations of the tip.
As shown in Fig. 9, IST is implemented on the robot con-
troller, whereas the learning algorithm is carried out in the
outside controller, an IBM PC 486DX2-66. While the robot
is moving, the acceleration signals sensed by the accelerom-
eter are sent to the PC. Then the signals are read by an A/D
converter at sampling frequency 200 Hz and filtered through
a bandpass filter to filter out both drift signals and noise. The
filtered acceleration signals are stored in the PC memory. After
one cycle of motion, the PC calculates M and φ, and up-
dates the IST parameters. The values of the updated IST pa- Fig. 12. Tip acceleration measured in swing direction (axis 1). Both TVIST
and LIST have suppressed the residual vibrations. Note that because the robot
rameters are sent to the robot controller, and the next cycle moved for 1.35 s, the tip acceleration after 1.35 s is the residual vibration.
begins.
Prior to the application of IST filters for the robot, we ex-
amined the level of the residual vibrations without IST filters. B. Experimental Results
In these experiments, the reference trajectories were given as Fig. 11 shows the experimental results of the proposed iter-
S-curves with trapezoidal velocity profiles. Fig. 10 shows the ative LIST. With the initial parameters, the magnitude of the
tip acceleration signal processed through a filter with a band- residual vibration is quite large. As the number of iterations in-
width between 2–30 Hz, when only feedback control is used. creases, the residual vibrations become smaller and M reaches
The robot traveled for 1.35 s and the tip response after 1.35 s is its minimum at the sixth iteration. Note that tip acceleration is
regarded as the residual vibration. As is shown in the figure, the used as ν(t) to calculate M using (19).
residual vibrations in the swing direction, which were caused In Fig. 12, the results of iterative LIST are compared with
by the axis 1’s movement, are substantially larger than those in those of TVIST. LIST and TVIST reduced the peak residual
other directions. vibration to 14% and 13% of the original vibration without IST,
The experiment above made us observe that the frequency of respectively, whereas IST reduced it only to 40%. Therefore,
residual vibration decreases with smaller amplitude whereas the the comparison displays that both iterative LIST and TVIST
damping ratio remains nearly constant. The frequency variation, have well suppressed the residual vibrations, which implies that
in our study, is caused by both the time-varying inertia due to we can choose any of the two schemes according to application
change of the robot’s configuration and the nonlinear flexibility conditions. If the real-time estimation of the varying frequency
in the harmonic drive. is both possible and affordable, TVIST is applicable; if not, and
In order to demonstrate the effectiveness of iterative LIST, two the task is repetitive, iterative LIST is suitable.
other schemes of IST have been compared: conventional IST and Although higher frequency vibrations exist in the transient
the TVIST based on Rappole’s method [10], [15], respectively. region, their amplitudes are very small. In our application, such
They are described as follows. vibrations in the transient region were not important because
1) If For adopting conventional IST to nonlinear time-varying this robot was designed for point-to-point motion. Therefore,
systems, we have designed an IST filter using the mean we believe that LIST is useful for the repetitive tasks which
value of varying frequencies. consist of point-to-point motions.
2) If The TVIST is represented by
V. CONCLUSION
u(t) = A1 r(t) + A2 r(t − T2 (t)). (28)
For suppression of nonlinear and time-varying residual
The second impulse varies with time, and its application vibrations in a 6-DOF industrial robot, we have proposed an
time T2 (t) is estimated in real time from the position data iterative LIST which updates the parameters of IST by using
of each axis. data from previous trials. Based on theoretical analysis and
62 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 11, NO. 1, FEBRUARY 2006

experiments, LIST is applied to the first axis of the robot and


conventional IST is applied to the second and third axes. As
a result, the application of iterative LIST become convenient
because the iterative learning scheme is applied only for the
swing axis, which exhibits significant time-varying frequencies
and large vibrational amplitudes.
Experimental results show that iterative LIST reduces the
nonlinear and time-varying residual vibrations in the robot as
effectively as TVIST does. The LIST is useful for repetitive tasks
of industrial robots where real-time estimation is not available, Fig. 13. An example of reference trajectories and responses used in simula-
whereas TVIST is effective for nonrepetitive tasks by changing tions.
the IST parameters in real time according to the estimated fre-
quency. The iterative LIST is a promising technique to suppress G3 (θ)
the nonlinear and time-varying residual vibrations in industrial
robots, which have a limited computing power and memory = {−mb lg b − mc lb + md lg d + mP (ld − lb )}c3 g
space in most cases, because it does not require real-time esti- where g denotes the gravitational acceleration constant.
mation of vibration frequencies.
APPENDIX II
APPENDIX I SIMULATION
Ni (θ, θ̇) AND Gi (θ) Simulations were performed to validate the application strat-
Ni s and Gi s in Section II are shown in the following equa- egy proposed in Section II-B with the dynamic model of (1)–(6)
tions: and parameters listed in Table I and in the following:
Jm 1 = 0.015 kg · m2 Jm 2 = Jm 3 = 0.01 kg · m2
N1 (θ, θ̇)
 r1 = r2 = r3 = 100
= −2 ma lg2a c2 s2 + Ja c2 s2 − mc lg c s2 (lb c3 − lg c c2 )
KT 1 = KT 2 = KT 3 = 1.0 × 108 N · m
+ Jc c2 s2 + md lc s2 (lc c2 + lg d c3 )
b1 = b3 = 2000 N · m · s b2 = 200 N · m · s
+ mP lc s2 {lc c2 + (ld − lb )c3 }] θ˙1 θ˙2
 bm 1 = bm 2 = bm 3 = 50 N · m · s
− 2 mb lg2b c3 s3 + Jb c3 s3 + mc lb s3 (lb c3 − lg c c2 )
bk 1 = bk 2 = bk 3 = 2000 N · m · s
+ md lg d s3 (lc c2 + lg d c3 ) + Jd c3 s3
where motor inertias and gear reduction ratios were given as
+ mP (ld − lb )s3 {lc c2 + (ld − lb )c3 }] θ˙1 θ˙3 robot specifications. Spring and friction constants were esti-
mated from experimental results.
N2 (θ, θ̇)
Each joint was controlled by a proportional–integral–
= ma lg2a c2 s2 + Ja s2 − mc lb lg c c3 s2 + mc lg2c c2 s2 + Jc s2 derivative (PID) position controller, and a reference trajectory
for the PID controller was shaped by IST. Fig. 13 shows an
+ md lc2 c2 s2 + md lc lg d s2 c3 + mP lc2 c2 s2 example of reference trajectories and responses of joints in sim-
+ mP lc (ld − lb )s2 c3 } θ˙12 ulations. The inner-loop PID controller has the following form:

+ {−mc lb lg c + md la lg d + mP la (ld − lb )} τi = kP i (θdi − θi ) + kD i (θ̇di − θ̇i ) + kI i (θdi − θi ) dt
× sin(θ2 − θ3 )θ˙32
where their control gains were tuned as follows:
N3 (θ, θ̇)
kP 1 = 12 × 106 , kP 2 = kP 3 = 24 × 106
= mb lg2b c3 s3 + Jb s3 − mc lb2 c3 s3 − mc lb lg c c2 s3 + Jd s3
kI 1 = 8 × 107 , kI 2 = kI 3 = 16 × 107
+ md lg2d c3 s3 + md lc lg d s2 s3 kD 1 = 2000, kD 2 = kD 3 = 4000.

+ mP (ld − lb ) c3 s3 + mP lc (ld − lb )c2 s3 θ˙12
2
Sampling time for all simulation was fixed to 0.5 ms.
+ {mc lb lg c − md la lg d − mP la (ld − lb )} In the first simulation the robot moved without IST and re-
sulted in large residual vibrations as shown in Fig. 14. Since the
× sin(θ2 − θ3 )θ˙22 settling time was about 0.85 s for the given trajectories (Fig. 13),
vibrations after 0.85 s are regarded as residual vibrations. The
G2 (θ)
second simulation was performed with IST that was designed
= (ma lg a + mc lg c + md la + mP la )c2 g independently for each joint based on the results of the first
PARK et al.: INPUT SHAPING TECHNIQUE FOR AN INDUSTRIAL ROBOT 63

Fig. 15. Three different cases in updating parameters.

compared with those of residual vibrations without IST. The


simulation results exhibit that the peak amplitude of residual
vibrations in axes 2 and 3 varied only by 1% to 3%, while the
IST parameter in axis 1 had a negligible effect on the residual
vibration of the other two axes.

APPENDIX III
CONVERGENCE OF LIST

Fig. 14. Residual vibrations with various IST parameters of axis 1: θi − θm i For the purpose of verifying the convergence of the proposed
after settling time (a) without IST, (b) with IST, and (c) with IST parameter updating rule, the error (p − pn ) is examined. First, the updating
variation in axis 1. Note that dashed and thick solid lines overlap each other in condition is divided into the following two cases for examina-
axes 2 and 3.
tion.
Case 1) Both pk −1 and pk are smaller or larger than pn as
TABLE III shown in Fig. 15(a).
TRAJECTORIES AND SIMULATIONS RESULTS
Case 2) Either pk −1 or pk is smaller than pn and the other is
larger than pn as shown in Fig. 15(b).
Without loss of generality, we considered a situation repre-
sented in Fig. 15(a) for Case 1. By subtracting pn from both
sides of (23), we obtain
 
M̃k
ek +1 = ek + (pk − pk −1 )sat (29)
M̃k −1 − M̃k

where ej = pj − pn . Since pk −1 < pk < pn and M̃k < M̃k −1 ,


we know that
ek < 0, pk − pk −1 > 0, M̃k −1 − M̃k > 0.
With these inequalities, (29) can be
 
M̃k
ek +1 = −|ek | + (pk − pk −1 )sat . (30)
M̃k −1 − M̃k
If pk +1 < pn as shown in Fig. 15(a-1), the following inequal-
ity is obtained from (30):
 
simulation. As shown in Fig. 14, residual vibrations in axes 2 M̃k
|ek +1 | = |ek | − (pk − pk −1 )sat < |ek |.
and 3 were reduced to 17.6% and 18.2% of the original vibration, M̃k −1 − M̃k
respectively. The next simulation was to examine the variation (31)
of residual vibrations in axes 2 and 3 with a change of IST On the other hand, if pk +1 > pn as shown in Fig. 15(a-2), the
parameters in axis 1. IST parameters in axis 1, therefore, were following inequality holds from (30):
decreased by 50% while keeping those in axes 2 and 3 constant.  
Comparing the results with those of the second simulation, we M̃k
|ek +1 | = −|ek | + (pk − pk −1 )sat
observed that there were very little differences. M̃k −1 − M̃k
More simulations were performed with various trajectories as
≤ −|ek | + (pk − pk −1 ) = −2|ek | + |ek −1 |
listed in Table III. The values summarized in the table represent
the maximum peak-to-peak amplitude of residual vibrations < |ek −1 |. (32)
64 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 11, NO. 1, FEBRUARY 2006

Equations (31) and (32) can be combined as follows: [5] M. Agostini, G. G. Parker, K. Groom, H. Schaub, and R. D. Robinett,
“Command shaping and closed-loop control interactions for a ship
|ek +1 | < max[|ek |, |ek −1 |]. (33) crane(I),” in Proc. 2002 American Control Conf., 2002, pp. 2298–2304.
[6] P. H. Chang and J. Park, “A concurrent design of input shaping tech-
Moreover, (33) can be represented as (34) by introducing a nique and a robust control for high-speed/high-precision control of a chip
mounter,” Control Eng. Practice, vol. 9, no. 12, pp. 1279–1285, Dec.
constant ck +1 which satisfies 0 < ck +1 < 1: 2001.
[7] J. T. Feddema et al., “Control for slosh-free motion of an open container,”
|ek +1 | < ck +1 max[|ek |, |ek −1 |]. (34) IEEE Control Syst. Mag., vol. 17, no. 1, pp. 29–36, Feb. 1997.
[8] A. Tzes and S. Yurkovich, “An adaptive input shaping control scheme for
vibration suppression in slewing flexible structures,” IEEE Trans. Contr.
In Case 2, we obtain the following equation from (21): Syst. Technol., vol. 1, no. 2, pp. 114–121, Jun. 1993.
[9] M. Bodson, “An adaptive algorithm for the tuning of two input shaping
methods,” in Proc. American Control Conf., vol. 3, 1997, pp. 1340–1344.
M̃k −1 ek − M̃k ek −1 [10] B. W. Rappole, “Minimizing residual vibrations in flexible systems,” Mas-
ek +1 = . (35) ter’s thesis, Dept. Mech. Eng., MIT, Cambridge, MA, 1992.
M̃k −1 − M̃k [11] D. P. Magee and W. J. Book, “Implementing modified command filtering
to eliminate multiple modes of vibration,” in Proc. 1993 American Control
Since ek ek −1 < 0 and M̃k M̃k −1 < 0 as shown in Fig. 15(b), Conf., Jun. 1993, pp. 2700–2704.
(35) becomes [12] J. K. Cho and Y. Park, “Vibration reduction in flexible systems using a
time-varying impulse sequence,” Robotica, vol. 13, pp. 305–313, 1995.
[13] R. Kinceler and P. H. Meckl, “Corrective input shaping for a flexible-joint
M̃k −1 ek | − |M̃k ek −1  manipulator performing point-to-point motion,” in Proc. IEEE Int. Conf.
|ek +1 | =
|M̃k −1 | + |M̃k | Control Applications, 1996, pp. 391–396.
[14] J. Park and P. H. Chang, “Learning input shaping technique for non-LTI
systems,” Trans. ASME, J. Dyn. Syst. Meas. Control, vol. 123, pp. 288–
M̃k −1 | − |M̃k 
≤ max(|ek |, |ek −1 |) 293, Jun. 2001.
|M̃k −1 | + |M̃k | [15] H. S. Park, P. H. Chang, and J. S. Hur, “Time-varying input shaping
technique for an industrial robot,” in Proc. IEEE/RSJ Int. Conf. Intelligent
< ck +1 max(|ek |, |ek −1 |) (36) Robots and Systems, vol. 1, 1999, pp. 285–290.
[16] J. Y. Smith, K. Kozak, and W. E. Singhose, “Input shaping for a simple
nonlinear system,” in Proc. American Control Conf., 2002, pp. 821–826.
where ck +1 is a constant which satisfies (max (|M̃k −1 |, |M̃k |))/ [17] J. Park, P. H. Chang, and E. Lee, “Can a time invariant input shaping
(|M̃k −1 | + |M̃k |) < ck +1 < 1. technique eliminate residual vibrations of LTV systems?,” in Proc. 2002
From (34) and (36), we can say that |ek +1 | has the following American Control Conf., 2002, pp. 2292–2297.
[18] J. Park and C. B. Schrader, “Can an input shaper with two impulses
inequality: suppress nonlinear residual vibrations?,” in Proc. 2003 American Control
Conf., 2003, pp. 3172–3177.
|ek +1 | < cM max(|ek |, |ek −1 |) (37) [19] T. Singh and S. R. Vadali, “Robust time-delay control,” Trans. ASME, J.
Dyn. Syst. Meas. Control, vol. 115, no. 2, pp. 303–306, 1993.
where cM is a maximum value among all cj s for j = [20] W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling, Numer-
2, 3, . . . , ∞. By using (37), |ej | for j = 2, 3, 4, . . . , k is rep- ical Recipes in C, The Art of Scientific Computing. Cambridge, U.K.:
Cambridge Univ. Press, 1988.
resented as follows: [21] L. W. Johnson and R. D. Riess, Numerical Analysis, 2nd ed. Reading,
MA: Addison-Wesley, 1982.
|e2 | < cM eM
|e3 | < cM max(|e1 |, |e2 |) < cM eM
Juyi Park (M’02) received the B.S. degree in pro-
.. duction engineering the M.S. degree in precision en-
. gineering, and the Ph.D. degree in mechanical engi-
neering from the Korea Advanced Institute of Science
|ek | < ckM−2 eM and Technology (KAIST), Daejon, Korea, in 1991,
1993, and 2001, respectively.
where eM = max(|e0 |, |e1 |). As a result, He was a Postdoctoral Fellow at the University
of Texas at San Antonio from 2001 to 2002. From
lim |ek | < lim ckM−2 eM = 0. (38) 2002 to 2004, he was a Research Associate in the
k →∞ k →∞ Department of Electrical Engineering and Computer
Science, Vanderbilt University, Nashville, TN. Since,
In other words, pk converges to pn as iteration proceeds. 2004, he has been working on intelligent robot controllers at the Robot R&D
Institute, Daewoo Shipbuilding and Marine Engineering Company Ltd., Koje
City, Korea. His research interests are suppression of residual vibration in flex-
ible structures, robots for disabled people, and real-time robot controllers.
REFERENCES
[1] H. Asada, Z.-D. Ma, and H. Tokumaru, “Inverse dynamics of flexible robot
arms of trajectory control,” in Proc. Modeling and Control of Robotic Pyung-Hun Chang (S’86–M’89) was born in Pusan,
Manipulators, 1987 ASME Winter Annu. Meeting, 1987, pp. 329–336. Korea, in 1951. He received the B.S. and M.S. de-
[2] P. T. Kotnik, S. Yurkovich, and U. Ozguner, “Acceleration feedback for grees from Seoul National University (SNU), Seoul,
control of a flexible manipulator arm,” J. Robot. Syst., vol. 5, no. 3, Korea, in 1974 and 1977, respectively, and the Ph.D.
pp. 181–196, Jun. 1988. degree from the Massachusetts Institute of Technol-
[3] N. C. Singer and W. P. Seering, “Preshaping command inputs to reduce ogy (MIT), Cambridge, MA, in 1987, all in mechan-
system vibration,” Trans. ASME, J. Dyn. Syst. Meas. Control, vol. 112, ical engineering.
pp. 76–82, 1990. From 1984 to 1987, he was involved in a research
[4] M. Ballesteros and W. Book, “Implementation alternatives for dual rate project in the field of robotics as a Research Assis-
control systems with command shaping,” in Proc. American Control Conf., tant at the Artificial Intelligence Laboratory of MIT.
2002, pp. 2285–2291. Since 1987, he has been on the faculty of and is now a
PARK et al.: INPUT SHAPING TECHNIQUE FOR AN INDUSTRIAL ROBOT 65

Professor in the Department of Mechanical Engineering, Korea Advanced Insti- Eunjeong Lee (S’88–M’94) received the B.S. de-
tute of Science and Technology (KAIST), Daejon, Korea. His research interests gree from Ajou University, Suwon, Korea, in 1986,
are high accuracy/speed control with application to mechanical systems, robust the M.S. degree from Northwestern University,
control of nonlinear plants such as robot manipulators, and task-oriented design Evanston, IL, in 1989, and the Ph.D. degree from
of robot manipulators. Case Western Reserve University, Cleveland, OH, in
1994, all in mechanical engineering.
She is currently a Visiting Professor in the Depart-
ment of Electrical Engineering and Computer Sci-
ence at the Korea Advanced Institute of Science and
Hyung-Soon Park was born in Seoul, Korea, in Technology (KAIST), Daejon, Korea. She was an
1972. He received the B.S. degree in precision en- Assistant Professor in the Department of Mechanical
gineering and the M.S. and Ph.D. degrees in mechan- Engineering at the University of Texas at San Antonio from 2000 to 2002. Her
ical engineering from the Korea Advanced Institute research interests are biomimetics at the micro/nanoscale, MEMS, intelligent
of Science and Technology (KAIST), Daejon, Korea, control, and superconducting renewable evergy systems.
in 1994, 1996, and 2004, respectively. Dr. Lee received a University of Houston/NASA Johnson Space Cen-
He is currently a Research Associate at the Re- ter Aerospace Postdoctoral Fellowship (1996–1999) and the Workshop Ma-
habilitation Institute of Chicago, Chicago, IL. His terial/Product Performance Award from the 1997 International Workshop on
research interests include robust control algorithms, Superconductivity. She served the IEEE Control Stytems Society as an Asso-
design and control of rehabilitation robots, and tele- ciate Editor for the Conference Editorial Board from 2000 to 2002. She is also
operation systems. a member of ASME.

Anda mungkin juga menyukai