IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 44, NO.

2, FEBRUARY 1999 357

Lemma 3 [Worst Case (W2)]: Let max be the maximal solution constructively given based on finite-dimensional operations and has
to the transcendental equation a structure of LQ feedback with preview compensation. Then we

: det [01 M In ]eK ()h

0 = 0: (40)
provide explicit formulas on the worst case initial uncertainties, i.e.,
those of initial state and the shape of stored disturbance, which
In
maximize the cost functional of resulting closed-loop system. The
Then the worst case cost (23) is given by maxx^(0)6=0 worst case initial uncertainties clarify the weak point of optimal
(Jopt =kx^(0)k2 ) = 2max and the worst case initial disturbance preview control in the sense that the effect of preview is spoiled.
w0 is constructively given as follows:

w0 ( ) = [0 DT ]eK ( )( +h) 0 g~ REFERENCES

In [1] Y. Fujisaki and M. Ikeda, “Synthesis of two-degree-of-freedom optimal
(0h    0) servosystems,” in Japanese, SICE Trans., pp. 907–914, 1991.
[2] T. Hagiwara, Y. Ohtani, and M. Araki, “Plant-variable-optimal robust
g~ 2 Ker[0 In ]eK ( )h 0
max M :
1 servo system with two degree of freedom,” SICE Trans., pp. 77–86,
(41)
In 1992 (in Japanese).
[3] M. Hayase and K. Ichikawa, “Optimal servosystem utilizing future value
of desired function,” in Japanese, SICE Trans., pp. 86–94, 1969.
VI. ILLUSTRATIVE EXAMPLE [4] A. Kojima and S. Ishijima, “Robust controller design for delay systems
For a system 6 = 1=(s2 + 1): in the Gap-metric,” IEEE Trans. Automat. Contr., vol. 40, pp. 370–374,
Feb. 1995.
, “Explicit formulas for operator Riccati equations arising in 1
H
6: x_ (t) = 01 001 x(t) + 10 u(t)
[5]
control with delays,” in Proc. 34th IEEE Conf. Decision and Control,
New Orleans, LA, 1995, pp. 4175–4181.
+ 10 w(t 0 2) [6] A. J. Pritchard and D. Salamon, “The linear-quadratic control problem
for retarded systems with delays in control and observation,” IMA J.
y(t) = [0 1]x(t)
Math. Contr. Inform., vol. 2, pp. 335–361, 1985.
(42) [7] U. Shaked and C. E. de Souza, “Continuous-time tracking problems in
in which the disturbance w0 is supplied in 0  t  2; we will design
H
an 1 setting: A game theory approach,” IEEE Trans. Automat. Contr.,
vol. 40, pp. 841–852, May 1995.
an LQ-preview control based on Theorem 1, then compare the time- [8] N. Suda and M. Ikeda, “Feedback and feedforward control in optimal
domain performance with those of LQ-control which does not have servosystems,” in Proc. IFAC Workshop System Structure and Control:
feedforward action. Our objective here is to highlight the benefit of State-Space and Polynominal Methods, Prague, Czechoslovakia, 1989,
pp. 157–160.
preview compensation, which is represented in (22). [9] M. Tomizuka, “Optimal continuous finite preview problem,” IEEE
Based on the realization (42), we define the cost functional (2) with Trans. Automat. Contr., vol. AC-20, pp. 362–365, 1975.
Q = 30 03 ; R = 1: In this case, the resulting closed-loop system
has poles 01:12 6 j 0:87 and effectively attenuates the oscillating
behavior.
For the closed-loop systems constructed with LQ-preview con-
Decentralized Control of Robot Manipulators:
trol (Theorem 1) or with usual LQ-control, which does not have
feedforward action, we will compare the time-domain performance Nonlinear and Adaptive Approaches
against the following initial uncertainties (x0 ; w0 ) : 1) x0 = 0 with
w0 (1)  1:0; 2) x0 = 0 with worst case (W2); and 3) worst case
Ming Liu
(W1). In the initial uncertainties, case (3) is most critical for the
closed-loop system constructed with LQ-preview control. Abstract— For robot arm tracking, the authors propose a nonlinear
Figs. 3–5 show time-domain responses for the cases 1)–3), respec- and an adaptive approach based on decentralized system structure.
tively. The continuous line shows the response of closed-loop system Using the passive feature of robots and cubic feedback to treat the
constructed with LQ-preview control, while the broken line shows nonlinear couplings and quadratic interconnections, the decentralized
adaptation is achieved by applying the linear-in-parameters property of
the response of usual LQ-control, which does not have feedforward the motion equation. The nonlinear feedback improves the performance
action. For the case when the uncertainty 1) is applied (Fig. 3), the of PD control from local to global stability and the adaptation reduces
closed-loop system with LQ-preview control shows favorable time- its tracking errors. The practical significance of the approaches lies in
domain performance and the effect of disturbance w0 (1)  1:0 is well the fact that they can be implemented in most robots without hardware
alteration.
attenuated. For the case when the worst case uncertainties 2) and 3)
are applied, the normalized costs max(Jopt =kx^(0)k2 ) are evaluated Index Terms—Adaptive control, decentralized structure, nonlinear sys-
by 10.10 and 29.56, respectively (Lemma 3 and Theorem 2). In tems, robot control.
these cases, the responses of LQ-preview control come to be similar
to those of usual LQ-control. These facts imply that the worst
I. INTRODUCTION
case analysis provides initial state uncertainties which cause definite
damage to the time domain performance. Although significant achievements, marked by the development
of centralized adaptive and robust control schemes (see, e.g., [3],
VII. CONCLUSION Manuscript received March 6, 1997. Recommended by Associate Editor,
C. Canudas de Wit.
A generalized LQ-control problem, which treats a mixed atten- The author is with the Department of Electrical and Computer Systems
uation of stored disturbance and initial state uncertainty, is solved Engineering, Monash University, Caulfield East, VIC 3145, Australia.
in the infinite-horizon time setting. The optimal preview control is Publisher Item Identifier S 0018-9286(99)01307-0.

0018–9286/99$10.00  1999 IEEE

Authorized licensed use limited to: Daniel Alvarado. Downloaded on October 01,2021 at 15:28:34 UTC from IEEE Xplore. Restrictions apply.
358 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 44, NO. 2, FEBRUARY 1999

[22], [23], [1], [12], and their bibliographies), have been made to Let qd (t) 2 <n be the reference position trajectory, and we define
improve the tracking performance of robots, the decentralized (or a generalized tracking error [22]
so-called independent joint) controller structure is still adopted by
the majority of modern robots in favor of its computation simplicity r = e_ + 8e (3)

where e = qd 0 q and 8 = diagf1 ; 1 1 1 ; n g > 0,

and low-cost hardware setup. Therefore, how to best improve the
a constant
tracking performance of robots based on this decentralized controller
matrix. Applying (3) to (1) yields
structure is still a challenging task.
The decentralized robust and adaptive control approaches for linear D(q)_r + H (q; q_)r = 0 +  (4)
and linear dominant systems have been well developed (see, e.g.,
[19], [21], [9], [8], [20], [24], and their references). For a set of and
linear dominant subsystems whose interconnections are nonlinear
but linearly bounded by the norms of the overall system states,
 = D(q)(qd + 8e_ ) + H (q; q_)(q_d + 8e) + g(q) + f (q_): (5)
the robust and adaptive approaches reported in [9], [21], and [8] Equations (3)–(5) define the tracking error dynamics of (1) for a
guarantee the exponential convergence of the tracking errors and class of qd and 8.
parameter estimation errors to a bounded residual set. In [20], We assume that: 1) qd 2 Cn2 and 2) all joints of the robot arms
nonlinear feedback was introduced to handle the interconnections considered are revolute, which implies that there exist some constants
bounded by a higher order polynomial of the system state norms. max > min > 0; cH ; cg > 0 such that
For robot decentralized tracking, these approaches cannot be en-
tirely realized since we cannot split the overall system into several min I < D(q) < max I; kHv (q)k  cH ; kg(q)k  cg
subsystems whose states and input torques are both isolated from (6)
each other due to the coupling caused by nonlinear inertial. Among where k 1 k is a proper norm. Using (6) and the bounded feature of
the early work reported in [17], [7], [6], and [11], the scheme shown f (q_), it can be shown (see [13] and [14]) that the upper bound of
in [7] followed the linear system design approach. Due to the reasons  in (5) satisfies
mentioned above, the effectiveness of the decentralized adaptation is
doubtful as it also appears in the interconnection terms. Aimed at local
kk  1 + 2 kek + 3 ke_ k + 4 kekke_ k
asymptotic stability, the approach reported in [5] will automatically where 1 to 4 are positive constants. Particularly, if the truncated
result in very large PD gains. In [18], an assumption that the joint L1 norm is adopted, we have
dynamic models are slowly time-varying has to be made in the
stability analysis. All approaches only result in the local stability kkT 1  1 + 2 kekT 1 + 3 ke_ kT 1 + 4 kekT 1 ke_ kT 1 (7)
of the tracking errors.
In this paper, based on the system structure proposed in [22], a set where
of nonlinear decentralized tracking error subsystems is formulated
1 = cH kq_d kT2 1 + fv kq_d kT 1 + max kqd kT 1 + cg + fc
so that the passivity property of the robot dynamics can be used
in controller design. Utilizing the PD control to ensure the local 2 = cH kq_d kT 1 k8k1
stability, we employ a cubic feedback to compensate the quadratic 3 = cH kq_d kT 1 + max k8k1 + fv
term in interconnection. This improves the local stability of PD 4 = cH k8k1 :
control to global stability. Furthermore, we apply an adaptive scheme
for each subsystem using the linear-in-parameter feature of motion The following lemma gives the upper bound of k kT 1 in terms of
equation to reduce tracking error. Since both schemes are based on a krkT 1 .
decentralized controller structure they can be easily implemented in Lemma 1: For  in (5), there exist constants 1 ; 2 ; and 3 > 0
most modern robot systems without hardware modification. such that

II. DYNAMICS AND DECENTRALIZED SYSTEM STRUCTURE

kkT 1  1 + 2krkT 1 + 3 krkT2 1 (8)

Consider the Lagrange–Euler equation of a rigid robot arm with n where

DOF’s (n = f1; 2; 1 1 1 ng):
1 = 1 + 2 ke0 k1 + 3 k8k1 ke0 k1 + 4 k8k1 ke0 k1
2

D(q)q + H (q; q_)q_ + g(q) + f (q_) =  (1) 0

2 = 2 k8k1 + 3 + 24 ke0 k1
1

n
where q 2 < is the joint angle vector;  2 < , the input n 01 :
3 = 4 k8k1
torque vector; D(q ) = DT (q ) 2 <n2n , the positive definite inertia
matrix; H (q; q_ )q_ = (In
q_ T )Hv (q )q_ 2 <n in which Hv (q ) = Proof: Since the input–output relations of two proper stable
linear time-invariant maps T1 : r 7! e and T2 : r 7! e_ defined
[H1T (q) 1 1 1 HnT (q)]T 2 <n 2n , the centrifugal and Coriolis torque;
g(q) 2 <n , the gravity torque; and f (q_) = [f1 (q_1 ) 1 1 1 fn (q_n )]T 2 by (3) satisfy
<n , the friction torque with kekT 1  k8k1 01krkT 1 + ke0 k1
fi (q_i ) = si e0d kq_ k + ci sgn(q_i ) + vi q_i
(9)
ke_ kT 1  krkT 1 + k8e0 k1
in which si ; di > 0; ci > 0; vi > 0 are constants related to the where ke0 k1 and k8e0 k1 are bounded zero input responses in
striction, Coulomb, and viscous friction, respectively. Obviously, which e0 = e(t0 ), substituting (9) into (7) yields (8).
there exist positive constants fc = maxi2n fsi + ci g and fv = We define the ith component of (4) as the ith subsystem of the
maxi2n fvi g such that kf (q_)k  fc + fv kq_k. overall error dynamics
It is well known that map  7! q_ defined by (1) is passive [2], [4], n
[16], [22], which implies that dij (q)_rj + q_T Hi (q)r = 0i + i ; i2n (10)
rT (D=
_ 2 0 H )r = 0; 8r 6= 0: (2) j =1

Authorized licensed use limited to: Daniel Alvarado. Downloaded on October 01,2021 at 15:28:34 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 44, NO. 2, FEBRUARY 1999 359

where dij (q) is the i-j th entry of D(q) and the ith component of (5) Proof: Consider the Lyapunov function candidate
n
i = dij (q)(qdj + j e_ j ) + q_T Hi (q)(_qd + 8e) V (r) = 1 rT D(q)r:
2
j=1 Its total derivative along the solution of (4) and (5), after taking (2)
+ gi (q) + fi (q_) (11) and (12) into account, is
n
V_ (r) = 0rT Kr 0 i ri2 + i ri4 + rT 
as the interconnections. The decentralized feedback control is then
to find a class of local state feedback and reference trajectory
feedforward laws i = i (ri ; qd ; q_d ; qd ) so that r, and therefore e
i=1
n n
and e_ , according to (9), are as small as possible.  0rT Kr 0 min ri2 0 min ri4 + rT :
i=1 i=1
III. PD AND NONLINEAR DECENTRALIZED CONTROL Applying the L1 norm for rT and T which are obtained by
The well-known decentralized linear PD law is given by truncating them at time T  t0 and recalling Lemmas 1 and 2,
we obtain
 = Kr = K e_ + K 8e
V_ (rT )  0kmin krkT2 1 0 min krkT2 1 0 min krkT4 1
where K = diagfk1 ; 1 1 1 ; kn g > 0. This control law, together with + n1 krkT 1 + n2 krk2T 1 + n3 krk3T 1 :
its various versions, has found a broad application in robot systems
due to its computation simplicity and independent joint controller Since the last term on the right-hand side satisfies 3 krkT3 1 
structure. It is well known that the PD law guarantees local L 1 3 (krkT2 + krkT4 )=2, we obtain
1 1
stability provided that K is sufficiently large (see, e.g., [10] and [13]).
V_ (rT )  0k krkT 1 + n 2 + 3
2 krkT 1 0 min krkT 1
2 2 2
It can be shown (see [13]) that the local rather than global stability min

of the PD control is caused by the 3 krkT2 term in (8) inherited

1
from h(q; q_). To cope with this effect, inspired by [20], we propose + n1 krkT 1 0 min 0 n2 3 krkT4 1 : (15)
the following decentralized PD-plus-nonlinear (PD + NL) control law:

i2n
Completing the squares for the second and fifth terms, and the third
i = ki ri + i ri + i ri3 ; (12) and fourth terms, respectively, on the right-hand side of (15), and
then dropping the resultant two negative squares terms, we obtain
where ki ; i ; i > 0 are controller parameters to be designed. (The
V_ (rT )  0kmin krkT2 1 + (n(22 + 3 )) + (n1 ) :
2 2
PD portion is written as two terms is merely for the convenience of
stability proof). We need the following lemma. 8(2min 0 n3 ) 4min
Lemma 2: For vector functions v; y 2 Cn [t0 ; T ], the following
In view of 0krkT2  0V (rT )=kDk let T ! 1 so that T can
1 1
be dropped, then we have V_ (r)  0c V (r) + c , where c and
inequalities hold:
n c are given by (14). This implies that V (r) is globally ultimately
ri2k  krkTk1 ;
2
k = 1; 2; 1 1 1 ; bounded by Vfc = c =c . In addition, since min (D)krk2 
i=1 T 1 2V (r)  2Vfc , where krk is the Euclidean norm of r, we have
jxT yjT 1  nkxkT 1 kykT 1: krk2  2Vfc =min (D) which is the set given by (14) and the theorem
is proven.
1
Proof: Apply the definition of truncated L norm (see [13]). Remark 3.1: The significance of cubic feedback in (12) is not only
The following theorem refers to the global stability of the control to achieve the global stability but also to improve robustness because
law (12). the gain conditions (13) are independent of 1 and 2 . (In PD control
Theorem 1: Consider tracking error dynamics (3)–(5), in which the gain condition depends on 1 ; 2 ; and 3 ; see [14]). Recalling
the input torque is given by (12). If ki ; i ; and i in (12) satisfy 3 and 4 given in Section II, the third inequality in (13) becomes
the conditions min > ncH =2, which indicates that as far as (13) is true, the L 1
stability can be guaranteed regardless of how large the sizes of D(q );
kmin > 0; min > 0; min > n3 =2 (13) g(q); and f (q_) are, even though smaller kmin may result in large
tracking errors.
where kmin = mini2n fki g; min = mini2n fi g and min = Remark 3.2: In the case where the PD gains are already large
mini2n fi g, then the tracking error r globally converges to a residual enough, the effect of the cubic feedback in (12) in steady state will
set be rather minor. This can be seen from (8); if ri is already quite
small, ri3 will be even smaller and become insignificant compared

fc = fr : krk2  2c =(c min (D))g (14) with the effect of the PD portion.

with a rate not slower than exp(0c t), where

IV. DECENTRALIZED FEEDBACK ADAPTIVE CONTROL

c = (n(22 + 3 )) + (n1 )
2 2
In the control design shown in the last section, as far as stability is
8(2min 0 n3 ) 4min concerned, no a priori knowledge of the system parameters is required
c = 2kmin
except an estimate of 3 . Virtually, if the nonlinear functions related
 (D)
max to local state in i are known, we can then introduce a decentralized
adaptive feedback to reduce the tracking errors.
max(min) (1) is the maximum (minimum) eigenvalue of (1) and k 1 k We introduce notations q  (q; q_); qi  (qi ; q_i ); and qd 
is the Euclidean norm. (qd ; q_d ; qd ). Based on the linear-in-parameter property of robot

Authorized licensed use limited to: Daniel Alvarado. Downloaded on October 01,2021 at 15:28:34 UTC from IEEE Xplore. Restrictions apply.
360 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 44, NO. 2, FEBRUARY 1999

model, we split i in (10) into two parts so that it becomes Proof: Consider the Lyapunov function candidate
n

dij r_j + q_T Hi (q)ri = i + wi (

qi ; qd )i + ai 1
V (r; ') = rT D(q)r + 'T 001 ':
1 (22)
j =1
2 2
where wi (qi ; qd ) 2 <12k is a row regression vector of subsystem i The total derivative of V (r; ') along the solution of (16), after taking
and i = [i1 ; 1 1 1 ; ik ]T 2 <k is an unknown constant parameter (2), (18), and (19) into account, is
vector related to subsystem i. (Its elements do not have to be the n
dynamic parameters of subsystem i.) Thus, the overall system (4) V_ (r; ') = 0rT Kr 0 i ri2 + i ri4 + rT a
and (5) becomes i=1

D(q)r_ + H (q; q_)r = 0 + W (

n
q; qd ) + a (16) + i (i 0 'i )'i : (23)
where  = [iT ; 1 1 1 ; nT ]T 2 <
i=1
; W (
k
q; qd ) + a = ;
and W ( q; qd ) = diagfw1 ( qn ; qd )g 2 <n2
q1 ; qd ); 1 1 1 ; wn ( k
. For the first three terms on the right-hand side of (23), we follow
Obviously, since  is bounded as shown in (8), both W ( q; qd ) and a the same procedure of proving Theorem 1. Apply the following
will be bounded in the same fashion, i.e., there exist some constants inequality:
0 <  0 i  i ; i 2 3, such that
n

k k 1   0 +  0 krk 1 +  0 krk 1 :
a T 1 2 T 3
2
(17)
T i (i 0 'T i )'T i  max kk1 k'kT 1 0 min k'kT2 1
i=1
q ; q ); i 2 n, are known, we propose
Under the assumption that w (
 2 kk1 0  2 k'k 1
i i d 2
max 2 min 2
the following decentralized PD-plus-nonliear-plus-adaptive (PD + min
T

NL + AD) control law:

in which the truncated L1 norm has been adopted, to the last term
qi ; qd )^i ;
i = ki ri + i ri + i ri3 + wi ( i2n (18)
we obtain
where ^i is an estimate of i . Define the parameter estimation error
30
vector as 'i = i 0 ^i , the decentralized adaptive control law is V_ (rT ; 'T )  0kmin krkT2 1 + n 20 + krk 1
2

given by 2 T

0 min krkT2 1 + n10 krkT 1

'_ i = 0^_ i = 00i wiT (
qi ; qd )ri + i 0i (i 0 'i ) 0
0 min 0 n2 3 krkT4 1
(19)

where i , a positive constant, and 0i = diagf i1 ; 1 1 1 ; ik g2

< 2
k k
, a positive definite constant matrix, are the adaptive controller + max kk1
2
2
0 min k'kT2 1 :
parameters whose affects will be given shortly. 2min 2
Let ' = ['T1 1 1 1 'Tn ]T , we can then have a compact version of
(19) for the overall system Thus, by means of completing squares we have

'_ = 00W T (
q; qd )r + 60( 0 ') min 0
V_ (rT ; 'T )  0kmin krkT2 1 0
2
k'k 1 + 4n
2
T

1

k 2
where 0 = diagf01 ; 1 1 1 ; 0n g 2 <
min
k
0 0 2
+ (n(22 + 3 ))0 :
whose di-
+ max kk1
2
agonal elements are 0i , and 6 = diagf1Ik ; 1 1 1 ; n Ik g 2
2

k 2
2min 8(2min 0 n3 )
< with Ik 2 <k 2k identical matrices, are posi-
k

tive definite constant matrices. As

We now present the following theorem.
Theorem 2: Under the assumptions that wi ( qi ; qd ); i 2 n are
known, consider system (16) and (17) and its input torque (18) and 0k krk 1 0  2 k'k 1
min
2
T
min 2
T

(19). If controller parameters in (18) satisfy that

 0 k2Dk k1 12 r Dr 0 
min T

min min
1 'T 001 '
2 T
min > n30 =2
T T T

kmin > 0; min > 0; (20)

 0 V (r ; ' )
a T T

then the tracking error r and parameter estimation error ' of the
overall system globally converge to a residual set we have V_ (r; ')  0a V (r; ') + a after T is dropped. This
implies that V is globally ultimately bounded by Vfa = a =a and

fa = f(r; ') : V (r; ')  Vfa = a =a g (21)
the theorem is proven.
at a rate not slower than exp(0at), where Remark 4.1: In view of (22) and (21), we have that
minfmin (D); min g (krk2 + k'k2 )  2V (r; ')  2Vfa = 2a =a .
a = min
2kmin Thus, the upper bound of the tracking error and parameter
kDk1 ; 
min min
estimation error can then be obtained as krk2 + k'k2  2a =
(n(220 + 30 ))2 + n10 + max
2 (minfmin (D); min ga ).
a =
8(2min 0 n3 ) 4min 2min 1
0 kk2 : Remark 4.2: To best reduce the tracking errors, a global feed-
forward and local feedback law (18) was applied. Since the joint
In the equations above, max(min) = max(min)i2n fi g; min =
mini2n;j2 k f ij g (the minimum element of matrix 0), and 10 ;
servo feedback loops are isolated from each other, it does not
violate the independent joint controller structure of most modern
20 and 30 are given by (17). robots. However, if necessary, the purely decentralized feedback

Authorized licensed use limited to: Daniel Alvarado. Downloaded on October 01,2021 at 15:28:34 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 44, NO. 2, FEBRUARY 1999 361

Fig. 1. Position tracking errors of joint 1.

Fig. 2. Position tracking errors of joint 2.

Fig. 3. Velocity tracking errors of joint 1.

and feedforward adaptation can be obtained by replacing wi ( qi ; qd ) control actions will excite the high-frequency unstable modes of
in (18) and (19) by wi (qi ; qdi ; q_di ; qdi ). In this case, the effect of joint compliance which have been unmodeled. In the case of de-
adaptation may be reduced as the size of the wi ( qi ; qdi ; q_di ; qdi )i centralized adaptive control, the estimated parameters, performing
is not as large as before. like variable gains, will neutralize the interconnections and tracking
Remark 4.3: Even though, theoretically, smaller tracking errors errors to ensure the stability [8]. In the following section it will be
can be obtained by increasing the PD gains, the resultant larger seen that the adaptive scheme not only leads to smaller tracking

Authorized licensed use limited to: Daniel Alvarado. Downloaded on October 01,2021 at 15:28:34 UTC from IEEE Xplore. Restrictions apply.
362 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 44, NO. 2, FEBRUARY 1999

Fig. 4. Velocity tracking errors of joint 2.

Fig. 5. Input torques of joint 1.

Fig. 6. Input torques of joint 2.

errors but also smaller control signals compared with those of PD are, respectively
control. d11 = 11 + 222 cos q2 ; d12 = d21 = 21 + 22 cos 2
d22 = 21
V. SIMULATION h1 = 022 2
2q1 q2 + q2 sin q2 ; h2 = 022 q12 sin q2
Consider (1) of a two-DOF planar robot arm, in which f (q_) = 0 g1 = 12 cos q1 + m2 glc2 cos(q1 + q2 )
(zero damping) for simplicity. The entries of D(q ); h(q; q_ ) and g (q ) g2 = m2 glc2 cos(q1 + q2 )

Authorized licensed use limited to: Daniel Alvarado. Downloaded on October 01,2021 at 15:28:34 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 44, NO. 2, FEBRUARY 1999 363

TABLE I [5] L. C. Fu and T. Y. Fu, “Robust adaptive independent joint control of

PHYSICAL PARAMETERS OF THE ROBOT ARM robot manipulator,” in Proc. 12th Symp. Automatic Control, Taiwan,
1988, pp. 363–371.
[6] , “How effective can a local PID feedback control really be for
trajectory tracking of robot motion?,” in Proc. 12th Symp. Automatic
Control, Taiwan, 1988, pp. 135–144.
[7] D. T. Gavel and T. C. Hsia, “Decentralized adaptive control of robot
manipulators,” in Proc. IEEE Int. Conf. Robotics Automation, 1987, pp.
1230–1235.
[8] D. T. Gavel and D. D. Siljak, “Decentralized adaptive control: Structural
where g is the acceleration of gravity, and ij are constant parameters conditions for stability,” IEEE Trans. Automat. Contr., vol. 34, pp.
given by 413–426, 1989.
[9] P. A. Ioannou, “Decentralized adaptive control of interconnected sys-
11 = m1 lc22 + m2 l12 + lc22 + I1 + I2 tems,” IEEE Trans. Automat. Contr., vol. AC-31, pp. 291–297, 1986.
12 = g (m1 lc1 + m2 l1 ) (24) [10] S. Kawamura, F. Miyazaki, and S. Arimoto, “Is a local linear PD
feedback control law effective for trajectory tracking of robot motion?,”
21 = m2 lc22 + I2 ; 22 = m2 l1 lc2 : in Proc. IEEE Int. Conf. Robotics Automation, 1988, pp. 1335–1340.
[11] F. Khorrami and U. Ozguner, “Decentralized control of robot manipu-
In the equations above, physical parameters and their numerical lators via state and proportional-integral feedback,” in Proc. IEEE Int.
values are shown in Table I. The reference trajectories specified are Conf. Robotics Automation, 1988, pp. 1198–1203.
qd1 (t) = 0:2 + 2 sin 2t and qd2 (t) = 01:7 + 1:8 cos 2t. [12] F. L. Lewis, C. T. Abdallah, and D. M. Dawson, Control of Robot
For simulation demonstrating the global stability of the PD + NL Manipulators. New York: Macmillan, 1993.
control given by Section III and the local stability of the PD control, [13] M. Liu, “Decentralized PD and robust nonlinear control of robot arms,”
J. Intelligent and Robot. Syst., vol. 20, pp. 319–332, 1997.
see [13] and [15]. [14] , “Stability conditions and convergence properties of PD control for
For the PD + NL + AD control shown in Section IV, the inter- robot manipulator tracking,” Monash Univ. Tech. Rep. ECSE-TR-96-4,
connections are 1996.
[15] , “Decentralized PD and robust nonlinear control of robot manip-
w 0 1 a1
 = W  + a = 1 + ulators,” Monash Univ. Tech. Rep. ECSE-TR-96-5, 1996.
0 w2 2 a2 [16] R. Ortega and M. W. Spong, “Adaptive motion control of rigid robots:
where 1 = [11 12 21 ]T 2 <3 ; 2 = [21 22 21 ]T 2 <3 in
A tutorial,” Automatica, vol. 25, pp. 877–888, 1989.

which ij are given by (24), w1 = [qd1 + 1 e_ 1 cos q1 qd2 ] 2 <123 ;

[17] H. Seraji, “Adaptive independent joint control of manipulators: Theory

qd2 + 2 e_ 2 qd1 cos q2 qd1 ] 2 <123 . In simulation, we set

and experiment,” in Proc. IEEE Int. Conf. Robotics Automation, 1988,
w2 = [ pp. 854–861.
k1 + 1 = 40; k2 + 2 = 30; 1 = 2 = 5; 1 = 2 = 8; [18] , “Decentralized adaptive control of manipulators: Theory, simu-
1 = 2 = 0:01; and ij = 1; i 2 2; and j 2 3. The initial values lation, and experiment,” IEEE Trans. Robotics Automation, vol. 5, pp.
183–201, 1989.
of parameter estimation and tracking errors are ^1 (t0 ) = ^2 (t0 ) = 0;
[e1 (t0 ) e2 (t0 )] = [00:2 0:2](rad) and [e_ 1 (t0 ) e_ 2 (t0 )] = [0:25 0
[19] M. E. Sezer and D. D. Siljak, “On decentralized stabilization and struc-
ture of linear large scale systems,” Automatica, vol. 17, pp. 641–644,
0:2](rad/s), respectively. 1981.
We also run the PD control using the same gains and initial tracking [20] L. Shi and S. K. Singh, “Decentralized adaptive controller design for
large-scale systems with higher order interconnections,” IEEE Trans.
error conditions. Automat. Contr., vol. 37, pp. 1106–1118, 1992.
Figs. 1–4 show the tracking errors of two schemes. It can be [21] D. D. Siljak, Large-Scale Dynamical Systems: Stability and Structure.
seen that the significant improvement has been obtained by the New York: North-Holland, 1978.
PD + NL + AD scheme. The input torque signals are shown in Figs. 5 [22] J. J. Slotine and W. Li, “On the adaptive control of robot manipulators,”
Int. J. Robotic Res., vol. 6, pp. 49–59, 1987.
and 6, which indicate that, after a very short initial phase, the input
[23] M. W. Spong, J. S. Thorp, and J. M. Kleinmark, “Robust microprocessor
torques of the PD + NL + AD are smaller than that of the PD control. control of robot manipulators,” Automatica, vol. 23, pp. 373–379, 1987.
The bounded parameter estimations are also verified in simulation [24] M. W. Spong and M. Vidyasagar, Robot Dynamics and Control.. New
but are omitted here. York: Wiley, 1989.

VI. CONCLUSION
A nonlinear and an adaptive decentralized control schemes for
robot tracking were presented. By refining the tracking error dynamics
of [22], the design was carried out by introducing a cubic feedback
and applying the passive and linear-in-parameter properties of robots’
dynamics to achieve global stability and reduced tracking errors.
Importantly, due to the independent-joint controller structure and
the simplicity of algorithms, the approaches can be implemented in
most robots by simply recoding their joint servos without hardware
alternation.

REFERENCES

[1] C. Abdallah, D. Dawson, P. Dorato, and M. Jamshidi, “Survey of robust

control for rigid robots,” IEEE Contr. Syst. Mag., pp. 24–30, Feb. 1991.
[2] S. Arimoto and F. Miyazaki, “On the stability of PID feedback with
sensory information,” in Robotics Research, M. Brady and R. P. Paul
Eds. Cambridge, MA: MIT Press.
[3] J. J. Craig, Adaptive Control of Mechanical Manipulator. New York:
Addison-Wesley, 1988.
[4] C. A. Desoer and M. Vidyasagar, Feedback Systems: Input–Output
Properties. New York: Academic, 1975.

Authorized licensed use limited to: Daniel Alvarado. Downloaded on October 01,2021 at 15:28:34 UTC from IEEE Xplore. Restrictions apply.