Sie sind auf Seite 1von 6

Auto-tuning PID Controller for Robotic Manipulators

Youngjin Choi,' Yonghwan Oh,' Sang Rok Oh,3 and Wan Kyun Chung4
Intelligent System Control Research Center,
1,2,3
Korea Institute of Science and Technology (KIST), KOREA
Department of Mechanical Engineering,
Pohang University of Science and Technology (POSTECH), KOREA

Abstract tions for gains. Recently, the noticeable 'square tuning'


and 'linear tuning' rules were proposed using an inverse
This paper suggests an auto-tuning method of PID optimal PID controller. They were derived from the
trajectory tracking controller for robotic manipulators. performance prediction equation suggested in [Z]. Ac-
In general, the PID trajectory tracking controller for tually, the performance tuning by the name of square
mechanical systems shows the performance limitation. law was suggested for the first times in 131. In this pa,
Since the control system including performance limita- per, an auto-tuning method of an inverse optimal PID
tion can not have equilibrium points, we define newly trajectory tracking controller will be proposed by mak-
the quasi-equilibrium region as an alternative for equi- ing use of the direct adaptive control scheme based on
librium point. Also, the quasi-equilibrium region is the extended disturbance input-to-state stability.
used as the target performance of the auto-tuning PID
trajectory tracking controller. The auto-tuning law is Recently, the direct adaptive control scheme for nonlin-
derived from the direct adaptive control scheme, based ear systems was developed in [4,5]. The direct adaptive
on the extended disturbance input-to-state stability. control is different from the indirect adaptive control
Finally, experimental results show that the control per- in that the control parameters are estimated directly
formance is enhanced by an auto-tuning method assist- without intermediate calculations involving plant pa-
ing the achievement of target performance. rameter estimates. Strictly speaking, the conventional
adaptive motion (trajectory tracking) control methods
given in [SI can be classified as the indirect adaptive
control for robotic manipulator systems because the
1 Introduction plant parameters are estimated t o construct a dynami-
cal model compensator. In this paper, the auto-tuning
Although many anto-tuning methods for PID con- method of an inverse optimal PID trajectory track-
troller were proposed in [l],they were for the chemical ing controller will be proposed according to the direct
process control systems. Since most process systems adaptive control scheme.
show very slow responses with time-delay effect, the
auto-tuning algorithms developed for process control
systems can not be directly applied t o mechanical sys-
2 Automatic Performance Tuning
tems. Though the performance tuning by gain changes
has brought one's interest with wide acceptance, there
Since the PID controller has the performance limitation
still exist no generally applicable auto-tuning laws for
for trajectory tracking of robotic manipulator as proved
mechanical systems.
in [Z], the auto-tuning method is devised so that it
The PID controller for mechanical systems, especially can accomplish the target performance of PID control
robotic manipulators, has been widely used with vari- system. To begin with, let us obtain the state-space
ous usages. As a matter of fact, the importance of PID description for trajectory tracking system model in the
control comes from the easy applicability and clear ef- following section.
fects of each proportional, integral and derivative con-
trol. Inspired by the extended disturbance input-to- 2.1 T r a j e c t o r y Tracking System Model
state stability of PID control for Lagrangian systems The robotic manipulator system is described by using
including robotic manipulator, the inverse optimality Lagrangian equation of motion with configuration co-
of PID controller was proved in [Z] with some condi- ordinates q = [ q l ,qz, . . ,qnlT E $3" as follows:

'Post Doctoral Fellow, KIST, yjchoiQpostech.edu + S ( Q )+ d ( t ) = 7,


M ( q ) 9+ c(q,q)q (1)
'Research Engineer, KIST, oyh~amadeus.kist.re.kr
3Principal Research Engineer, KIST, sroh~amadeus.kist.re.kr where M ( q ) = M T ( q ) E Pnx"is Inertia matrix,
4Professor, POSTECH, wkchungQpostech.edu C ( q ,q ) q E L" Coriolis and centrifugal torque vector,
Proceedings of the American Control Conference
02003 IEEE
0-7803-7896-2/03/$17.00 350 Denver, Colorado June 4-6. 2003
g ( q ) E P" gravitational torque vector, T E P" the where
control input torque vector and d ( t ) unknown external
disturbances. In fact, the system model (1) obtained A ( I ,t ) =

1
by Lagrangian equation of motion is for the set-point 0 I 0
regulation control. Here, let us define the extended 0 r
disturbance as following form: -M-'CKI -M-'CKp - KI - M - -'C- K p
and
+ +
w ( t ,e , e, S e ) = M ( q )( i i d K p e K f e ) + 0
+ +
C(q,0)( Q d K p e K I J e ) g(n) W + + , (2)

where K p , K I are diagonal positive constant matri-


ces, e = qd - q is the configuration error vector and
desired configurations (qdrq d ,q d ) are the function of Second, let us investigate the characteristics of ex-
time, hence, the extended disturbance zu is the function tended disturbance defined a s (2). The extended dis-
of time, configuration error, its derivative and integral, turbance w can be divided into the linear parameteriza-
since q(= qd - e ) and q ( = qd - e ) are the function tion part and external disturbances as following form:
of time, configuration error and its derivative. If the
extended disturbance defined above is applied to the +
w ( x , t )= Y(s,t)O d ( t ) , (6)
set-point regulation system model (l), then the trajec-
tory tracking system model can be obtained as where the regressor matrix Y ( m ,t ) can be found by
separating a real constant parameter vector O from
M(q)i+C(q,q)s=w t,e,e, e +U,
( J> (3) +
Y ( z ,t ) O = M(q)(ijd K p e K l e ) +
+C(q,@ ) ( q d + K p e + K I s e ) + g ( 9 ) ,
where the control input and composite error vector are
defined by and the real constant parameters consist of masses and
moments of Inertia of each link. Using the extended
U = -T disturbance of (6), the state-space representation of ( 5 )
s = e + K p e f K ~ edt
J can be modified to

+ B ( z ,t ) d + B ( z t, ) u .
X = A ( z , t ) z+ B ( z t,) Y ( z , t ) O
The difference and common point between the set-point (7)
regulation system model (1) and trajectory tracking Here, the regressor Y ( z , t )is not a zero matrix at
system model (3) were explained by Remark 1-2 in [Z]. x = 0 except the case of set-point regulation control
and gravity free motion. In other words, if g ( q ) # 0
B y using the composite error vector, the inverse opti- or qd # 0,qd # 0 , then Y ( x ,t ) # 0 at I = 0. Hence,
mal PID controller suggested in [Z] can be expressed as z = 0 is not an equilibrium point of (7) even when
following compact form: d ( t ) = 0 , because one term B ( 0 ,t)Y(O,t ) O has the
T = (K + 7-21) s, (4)
time-varying characteristics according t o desired tra-
jectories. Actually, since the equilibrium point can not
with three conditions: be found for the system (7), we will introduce the con-
cept of quasi-equilibrium region in the following sec-
tion.
(Cl) K , K p , K I > 0 , constant diagonal matrices
(CZ) K $ >~ K I , 2.2 Quasi-Equilibrium Region
To begin with, we s s u m e that there exist no external
(C3) y > 0 . disturbances for system model (7), namely d = 0 . If
the inverse optimal PID controller (4) described by a
First, let us obtain the state space description for tra- state vector as following form:
jectory tracking system model (3). If we define 3n-
dimensional state vector as follows: U = -(K+y-21)[K~,Kp,I]~

[ 21 = [ ] edt
EP3n,
is applied to (7), then the closed-loop control system is
obtained as follows:

X = A,($,t)a: + B ( m , t ) Y ( x ,t ) O (8)
then the state space representation of trajectory track-
ing system model (3) can be obtained by where

X = A ( z ,t ) x + B ( z ,t)w + B ( I ,t)u (5) A, = A -B (K + Y 2 1 )~ K KI ,p , I ] .


Proceedingsof lhe American Control Conference
351 Denver, Colorado June 4-6, 2w3
Since above closed-loop system has no equilibrium xel is determined as follows:
points as explained in previous section, we will de-
fine the quasi-equilibrium region and find it for above
closed-loop system in following Theorem.
Third, the quasi-equilibrium region is obtained by its
definition as follows:
Tlieorem 1 If the quasi-equilibrium region is defined as
the interior region of ball with the largest radius among
state vectors satisfying x = 0 , then it is obtained as
following form:
where Iz.(t)i = IK,' [K+y-21]-1Y,(t)OI. 0

The size of quasi-equilibrium regioii is inversely propor-


where x.(t) means the state vector satisfying x = 0 in
tional to the integral gain K I as shown in an equation
(8) and its Euclidian norm is as follows:
(9) in Theorem 1. Also, lsrgc K and small y make the
quasi-equilibrium region small. If we are t o approach
Iz.(t)l = IK;' [ K + ~ ~ 1 ] - ~ Y , ( t ) O (9)
l
I
the quasi-equilibrium region t o the point x. = 0 using
an inverse optimal PID controller, irrespective of the
+
and Y e ( t ) e= M(qd)@d C ( q d , Q d ) Q d f g ( q d ) .
constant parameter vector 0 and desired configurations
(qd, Q d , q d ) ,then one of three conditions should be sat-
isfied: the one is that K I gain matrix goes to infinity,
another is that &gain y to zero and the other is that
the gain K to infinity. This explains indirectly why an
inverse optimal PID controller for robotic manipulators
can not bring the global asymptotic stability(GAS).
In fact, the quasi-equilibrium region of Theorem 1 has
very close relation with performance limitation of PID
controller. Ailother expression for quasi-equilibrium re-
gion will be suggested using n-dimensional composite
error in following Rcmark.

R e m a r k 1 L e t s = e + K p e + K I J e d t . Ifwemultiply
[ K IK , p , I ] b y I, in Theorem 1:
A
s.(t) = [KI,KP,~IG(~)
= [K + y - 2 1 ] - ' Y , ( t ) t l , (11)

then the quasi-equilibrium region expressed b y the com-


posite error vector is obtained as follows:

In above Remark, the quasi-equilibrium region was d o


fined using the composite error as shown in Figure 1.
Its size is dependent upon the gain K rndtrix, &gain
y and the inverse dynamics Y , ( t ) O according to desired
configurations. Actually, the quasi-equilibrium region
can be used as a criterion for target performance cho-
sen by user. Also, it indirectly proves the existence
of gains which can achieve the target performance. In
following section, we will propose an automatic per-
formance tuning method of PID controller assisting to
accomplish the target performance.
Proceedings of the American Control Conference
352 Denver. Colorado June 4-6, 2003
Quasi-Eouiiibrium Ts2 Proof.
ing form:
First, we take Lyapunov function as follow-

1
A

V ( s ,K ,t ) = -sTMs
2
+
1
-2t T [(??(t) - K n ) r-' (??(t) - K O ) ] ,

where tr[.]means the trace of given matrix. If the auto


tuning inverse optimal PID controller (12) is applied t o
the trajectory tracking system model (3), then we can
get the time derivative of Lyapunov function along the
solution trajectory of closed-loop system as follows:

F i g u r e 1: Quasi-equilibrium region

2.3 A u t o - t u n i n g Law
Since the quasi-equilibrium region is determined by
the size of gains of PID controller, we should
know the inverse dynamics Y.(t)t9 depending on de-
sired configurations(qd, qdrqd) and dynamic parame-
ter vector(@)t o calculate the quasi-equilibrium region.
However, it is difficult to exactly identify the dynamic
parameters of a general robotic manipulator. Actually,
if they are known exactly, then the model-based con-
troller using inverse dynamics will show better perfor- Here, if the following matrix trace property is applied
mance than a PID controller. Hence, we use the con- t o above equation:
cept of target performance instead of quasi-equilibrium
region. If we determine the target performance, then
the size of quasi-equilibrium region should be adjusted
t o achieve the target performance by using the auto- then above time derivative of Lyapunov function is ar-
tuning law for gains. Here, we choose the gain matrix ranged as follows:
K in (11) as an auto-tuning parameter. The auto-
tuning law is derived from the direct adaptive control
scheme and ISS characteristics of trajectory tracking
v = -2 (Kn+ ? I )1 s- r2'J
Y2
is-m(2+
system in following Theorem.
YZ
-lW12
2
+ t T [(g(t)
- K O )(r-'k(t)- SS')] , (14)
Theorem 2 Assume that there exists the smallest con-
stant diagonal gain matrix Kn > 0 of PID controller Also, if the diagonal elements of ( I - ' k ( t ) -)'ss are
(4) guaranteeing the target performance (Cl) as follows: zeros, then the trace term of (14) becomes zero because
(g(t)- K n ) is a diagonal matrix. In other words,
the auto-tuning law (13) is derived from the following
If the auto-tuning inverse optimal PID controller: relation:

dki
if dt = r&),
- for i = 1 , . . . , n
using the auto-tuning law as following form:
then [ - Kn)
tr (g(t) (I-'f?(t) - ss')] =0

Therefore, if the auto-tuning inverse optimal PID con-


is applied to the trajectory tracking system (3) when troller (12) is applied to trajectory tracking system,
then we can get the following relation from (14):
K n > K ( t ) ,then the closed-loop control system is ex-
tended disturbance input-to-state stable(lSS), where si
is i-th element of composite error vector s, 2; and r,
are i-th diagonal elements of the diagonal time-uarying
gain matrix g(t)> 0 and the update gain matrix Since the right hand side of above inequality (15) is
r > 0 , respectively. unbounded functions for s and w , respectively, hence,
Proceedings of the American Control Conference
353 Denver, Colorado June 4-6.2003
the trajectory tracking system with an auto-tuning in-
verse optimal PID controller is extended disturbance
input-to-state stable(1SS). 0

Actually, when PID controller does not achieve the tar-


get performance, namely K n > g(t), the auto-tuning
law suggested in Theorem 2 will help t o achieve the tar-
get performance. On the contrary, when K(t) K n , >
the target performance can he accomplished without an
1

auto-tuning law, namely K = 0 for K 2 Kn. Also,


since the time derivative of Lyapunov function (14) can :
he arranged as follows:
Figure 2: Target performance and non-tuning region

the ISS is also assured for 2 K n . As a matter


of fact, t,he controller (12) without an auto-tuning law
(13) is equal to the conventional inverse optimal PID
controller. In following section, we will discuss about
the criterion on whether the auto-tuning law is neces-
sary t o achieve the target performance or not.

2.4 C r i t e r i o n for Auto-tuning


To apply an auto-tuning law, we should exactly know F i g u r e 3: 3-DOF Robotic Manipulator
the gain matrix K n guaranteeing target performance
(n).However, we can not know the matrix Kn till the law has the property of a nonlinear damping. Strictly
experimental result satisfies the target performance as speaking, the first term of anta-tuning PID controller
follows: (12) means the nonlinear damping which helps t o sta-
bilize the control system against extended disturbance
and the second term is a linear controller.

hut we can calculate the size of composite error Is(r)I


according to time progress. Moreover, since the auto-
3 E x p e r i m e n t a l results
tuning law was composed of the decentralized type in
Theorem 2, we suggest the decentralized criterion for
To show the effectiveness of an auto-tuning inverse op-
auto-tuning as follows:
timal PID controller, we utilized three link robot ma-
nipulator as shown in Figure 3. The desired configura-
tion profiles of Figure 4.(a) are obtained hy solving in-
verse kinematics for 6 line trajectories. Also, the given
where n is the number of configuration coordinates. trajectories require the fast motion (maximum velocity
As soon as the error arrives at the tuning region of % 3 r a d / s ) of robotic manipulator as shown in Figure
(16), the anto-tuning law should he implemented to 4.(b). First, we determine the static gains of auto-
assist the achievement of target performance. Hence, tuning inverse optimal PID controller as K p = 201,
the target performance is somewhat different from the Kr = 1001 and y = 0.5 satisfying the design guide-
non-tuning region as shown in Figure 2. lines (Cl),(C2),(C3) in section 2. Then the controller
has the following form: for i = 1,2,3,
Though we do not use the constant gain matrix K n
which offers the target performance, an auto-tuning in- 7, = (f?,(t) + 4 ) S ,
verse optimal PID controller (12) results in the effect of
gain K n for lsil > LL
Jz;; as shown in an equation (15).
On the contrary, if the composite error stays in non-
s, = e, +20e, + l o 0
J e,dt

tuning region of Figure 2, namely lsil 5 &, then the


auto-tuning process is stopped. For this case, we expect
that the gain updated by an auto-tuning law (13) where ri is i-th element of input torque vector r ,
will be larger than the matrix K n which brings the tar- K;(O)= 0.1 and n = 3. Second, since the composite er-
get performance. As a matter of fact, the'auto-tuning ror is approximately proportional to the configuration
Proceedings of the American Control Conference
354 Denver. Colorado June 4-6. 2003
TimeIsl Iimelsl

(a) Desired configuration (h) Desired configuration (a) Composite error: s (h) Auto-tuned gains:
velocity : Qd

(E)
(c) Configuration error : (d) Configuration velocity
e error : e
-
(c) Composite error (s) (d) Auto-tuned gains
for 0 1s for 0 1s -
Figure 5 : Composite error and auto-tuned gains
Figure 4: Performance of auto-tuning PID controller
4 C o n c l u d i n g Remarks

error with proportional constant K p . the target per- In this paper, the quasi-equilibrium region was defined
formance can be approximately determined as follows: to guarantee the existence of controller gain achieving
target performance. Also, we proposed the auto-tuning
inverse optimal PID controller assisting the achieve-
ment of target performance. Finally, we showed the
validity of auto-tuning law through the experiment.
where lsijt and le,Jt are the target composite error and
configuration error, respectively. For instance, if we are
t o obtain the performance of le,/t < 0.02 r a d for each
References
driving axis, then the target performance should be
determined as R = 1.0 by (17). Also, the update gain [l] c. C. Yu, Autotuning of PID COntTolkTS: Relay
r = 1000 was used for the experiment. Figure 4.(c) Feedback Approach, Springer, 1999.
and (d) show experimental results such as the config- [2] Y . Choi, W. K. Chung, and I. H. Suh, “Perfor-
uration error and its velocity error. In figures, we can mance and H m optimality of PID trajectory tracking
see that the errors are large at initial time and errors controller for Lagrangian systems,” IEEE Trans. on
are reduced till target performance can be achieved by Robotics and Automation, vol. 17, no. 6, pp. 857-869,
an automatic gain tuning. As a matter of fact, the Dec. 2001.
automatic gain tunings are executed at the exterior of 131 J. Park and W. K. Chung, “Design of a robust
two dotted lines of Figure 5.(a). After the auto-tuning H, PID control for industrial manipulators,” naris.
ic finished, the tuned gains arrived at 1?1 = 136.98, ASME J. of Dyn. Syst., Meas. and Contr., pp. 803-
K z = 65.60 and K3 = 6.83 as we can see in Figure 812, 2000.

-
5.(b). To see the tuning process in detail, the horizon-
tal ranges of 0 1 second of Figure 5.(a) and (b) are
enlarged as shown in that figure (e) and (d). The auto-
[4] W. M. Haddad and T. Hayakawa, “Direct adap-
tive control for nonlinear systems with bounded energy
C2 disturbances disturbances,” IEEE Conf. on Decz-
tuning of first axis is started at 0.11 second and ended sion and Control, pp. 2419-2423, 2000.
at 0.26 second because the error goes over the dotted [5] V. Chellaboina, W. M. Haddad, and
line ( l / & = 0.408) for the first time as shown in Fig- T. Hayakawa, “Direct adaptive control for non-
ure 5.(c). Also, the error of second axis goes over the linear matrix second-order dynamical systems with
dotted linc from 0.13 to 0.24 second. Finally, since the state-dependent uncertainty,” PTOC.of the Amencan
error of third axis goes over the dotted line downward Control Conference, pp. 4247-4252, 2001.
twice, the auto-tuning is implemented twice as shown (61 R. Ortega and M. W. Spang, “Adaptive motion
in Figure 5.(d). The experimental result of Figure 5.(a) control of rigid robots: A tutorial,” IEEE Conf. on
shows that the target performance is achieved after 0.6 Decision and Control, pp. 1575-1584, 1988.
second when the auto-tuning is finished.
Proceedings of the American Control Conference
355 Denver, Colorado June 4-6.2W3

Das könnte Ihnen auch gefallen