Professional Documents
Culture Documents
Abstract-In this paper we give some initial results on the adaptive A number of applications of these techniques have been made:
control of minimum-phase nonlinear systems which are exactly input-their chief drawback, however, appears to arise from the fact that
they rely on an exact cancellation of nonlinear terms in order to
output linearizable by state feedback. Parameter adaptation is used as a
get linear input-output behavior. Consequently, if there are errors
technique to robustify the exact cancellation of nonlinear terms, which is
or uncertainty in the model of the nonlinear terms, the cancel-
called for in the linearization technique. Only the continuous-time case is
lation is no longer exact. In this paper we suggest the use of
discussed in this paper; extensions to the discrete-time and sampled-data
case are not obvious. parameter adaptive control to help robustify, i.e., make asymptot-
ically exact the cancellation of nonlinear terms when the uncer-
tainty in the nonlinear terms is parametric. Some other attempts in
I. INTRODUCTION
this regard have been made by Marino and Nicosia [14] and
T is well known that, under rather mild assumptions, the Nicosia and Tomei [15], using a combination of high gain, sliding
I input-output response of a nonlinear system can be rendered modes, and adaptation. Some previous work in this spirit is in
linear by means of state feedback. This was implicitly or explicitly Nam and Arapostathis [13]. Our development is, we believe,
pointed out in several papers dealing with the study of noninteract- considerably more general and straightforward than theirs (specif-
ing control of nonlinear systems, like those of Porter [22], Singh ically, no error augmentation and stronger stability theorems) and
and Rugh [16], Freund [17], and Isidori, Krener, Gori-Giorgi, was in tum motivated by our work in the adaptive control of a
and Monaco [ 181. Independently, a substantially identical synthe- specific class of linearizable systems-rigid link robot manipula-
sis technique was successfully implemented in some relevant tors (see Craig, Hsu, and Sastry [5] for details, including
practical applications, like the control of flight dynamics (Meyer implementation of the scheme on an industrial robot arm).
and Cicolani [lo]) and the control of rigid-link robot manipulators We would also like to mention the work done in parallel by
via the so-called computed torque method, to mention a few. Taylor, Kokotovic, and Marino [20] on the adaptive control of
Parallel to these developments, beginning with the work of fully-state linearizable single-input, single-output systems. While
Brockett [23], several authors studied the problem of when the our scheme specializes to their scheme in the instance that the
differential equation relating the input to the state can be rendered system is state (rather than input-output) linearizable, their paper
linear via state feedback and coordinates transformation. The also considers the effect of parasitic dynamics on the adaptation
problem was completely solved by Jakubczyk and Respondek [24] scheme. Taylor, Kokotovic, and Marino prove the robustness of
and, independently, by Hunt, Su, and Meyer [25]. The former their scheme to parasitics; we have, however, not undertaken such
design technique is often referred to as exact input-output a study here.
linearization, while the latter one as exact state-space lineariza- The paper is organized as follows: we give a brief review of
tion. The bridge between the two techniques lies in the fact that input-output linearization theory for continuous-time systems
the design of a state-space linearizing control is equivalent to the along with the concept of a minimum-phase nonlinear system as
design of output functions for which input-output linearization developed in Bymes and Isidori [3] in Section II. We discuss the
is possible. The theory is now well developed and understood adaptive version of this control strategy in Section I11 along with
(see, for instance, expository surveys in Isidori [26], Isidori [8], its applications to the adaptive control of robot manipulators. In
and Claude [4] for the continuous-time case. For the discrete-time Section IV, we collect a few comments about the discrete-time
and sampled-data versions of the theory, see Monaco and and sampled-data cases along with some future directions.
Normand-Cyrot [ 1 13, Monaco, Normand-Cyrot, and Stomelli
[ 121, and Jakubczyk [27]. The class of systems is described (in the 11. REVIEWOF EXACTLINEARIZATION TECHNIQUES
continuous-time case) by
P
A . Basic Theory
X = f ( x ) + gi(x)ui A large class of nonlinear control systems can be made to have
i= I
linear input-output behavior through a choice of nonlinear state
feedback control laws. We review the theory here in order to fix
Y1= hl ( x ) notation. Consider the single-input, single-output system
i=f(x)+g(x)u
yp=hp(x)
Y =h(x) (2.1)
with x E Wn, U, y E WP, and f,g;, h, smooth functions.
with x E R; f,g , h smooth. Differentiating y with respect to
time, one obtains
Manuscript received March 4, 1988; revised March 10, 1989. Paper
recommended by Past Associate Editor, P. Crouch. This work was supported
in part by NASA under Grant NAG2-243 and by the U.S. Army Research j = L,h + L,hu. (2.2)
Office under Grant DAAG29-85-K-0072. The work of the second author was
supported in part by a McKay Visiting Professorship held at the University of Here, Lfh, L,h stand for the Lie derivatives of h w.r.t. f,g ,
California, Berkeley. respectively. If ( L , h ) ( x ) # 0 tlx E Wn,then the control law of
S. S. Sastry is with the Electronics Research Laboratory, University of
California, Berkeley, CA 94720. the form a ( x ) + P(X)LJ,namely
A. Isidori is with the Dipartimento di Informatica e Sistemistica, Universik
di Roma, La Sapienza, Rome, Italy.
IEEE Log Number 8930790.
yields the linear system If A(x) E RpxP is bounded away from singularity, the state
feedback control law
U= -A(x)-' = [ ]
L;l hl
i
L y h,
+A(x)-'U (2.12)
j i =L j h + (L,Lfh)u. (2.4)
yields the closed-loop decoupled, linear system
In (2.4) above, L j h stands for Lf(Lfh) and L,Lfh stands for
L,(Lfh). As before, if L,Lfh # 0 Vx E R", the law
(2.13)
linearizes the system (2.4) to yield Once linearization has been achieved, any further control objec-
tive such as model matching, pole placement, tracking may be
y=v. easily met. The feedback law (2.12) is referred to as a static-state
feedback linearizing control law.
More generally, if y is the smallest integer such that L,L h = 0 If A (x) defined in (2.10) is singular, linearization may still be
f o r i = O;-.,y - 2andL,L;-'h(x) # O v x E R",&enthe achieved using dynamic state feedback. The development may be
control law followed by using integrators before some of the inputs; exact
conditions under which linearization may be achieved by dynamic
state feedback are given, for instance, in Descusse and Moog [6].
In (2.8) Lfhj stands for the Lie derivative of hi with respect to f , defined as
similarly Lgihj. Note that if each of the (Lgihj)(x) = 0, then the Ti(x) = Z I 1 = h (x)
inputs do not appear in (2.8). Define y j to be the smallest integer
such that at least one of the inputs appears in y p ) , i.e.,
T2 (X) = 212 = Lfh (X)
y p ) = L ; j h , + g Lgi(L?-'hj)ui (2.9) T, (X) = ZI,= L; ~ I h (x)
i= 1
with
with at least one of the Lgi(LY-'hj) # OV x. Define the p x p dTi(x)g(x)=O for i = y + l , a . . , n
matrix A (x) as
A(x)= [ Lg1(L;'-'h1)
...
Lgl(LT-'hp)
* * *
.* .
LgP(L;'-'hl)
Lgp(L)-'hp)
. (2.10)
is a diffeomorphism onto its image. If we set z2 = (T,, 1,. . .,
T,) =, it follows that the equations (2.1) may be written in the
normal form as
i l l = 212
yy
=
Lyh,
+A(x)
UP
. (2.11)
(2.20)
SASTRY AND ISIDORI: ADAPTIVE CONTROL OF LINEARIZABLE SYSTEMS 1125
L={x E W":h(X)=Lfh(X)=...=L;-'h(x)=O}
vs0
and hereby defined by the vector field
Using this along with the bounds (2.25) it is easy to establish that
z2 is bounded.
Remarks:
(note that this is a vector field of L becausef(x) is tangent to L ) . 1) Proposition 2.1 establishes that a bounded input to the
If L is connected and the zero dynamics is globally asymptotically exponentially stable, unobservable dynamics yields a bounded
stable (i.e., if the system is globally minimum-phase, then the state trajectory z2.
normal forms are globally defined if the vector fields 2) The hypothesis of Proposition 2.1 calls for a strong form of
stability-exponential stability; in fact, counterexamples to the
g ( x ) , adj-g(x), . . ., ad;-'g(x) Proposition exist if the zero-dynamics are not exponentially
stable, for example, if some of the eigenvalues of d/dz2$(0, z2)lie
are complete, where on the jw-axis.
3) The hypotheses of Proposition 2.1 can, however, be
weakened substantially by requiring only that all trajectories of
(2.22) are eventually attracted to a compact set, for instance, by
requiring that
Note that this condition can be guaranteed by requiring that the
vector fields in question are globally Lipschitz continuous, for zc$(O, z 2 ) -~4 z 2 1 2 for I z ~ ~ z R . (2.28)
example. In this paper we systematically assume global minimum-
phase property and the existence of globally defined normal Condition (2.28) can be thought of as being an attractivity
forms. condition-a simple contradiction argument involving
An interesting application of the notion of normal form and
minimum-phase property is the following one. Assume the control
U in (2.23) is chosen so that y ( t ) tracks y M ( t ) ,i.e.,
y ' y ~ ) + a l ( yM( Y - ' )Y-( ? ? I ) ) + . .. +a,(yM-y) (2.24) should convince the reader that (2.28) guarantees that all
trajectories of (2.22) eventually enter a ball of radius R . With
with a],. * * , ay chosen SO that S Y a1sY-I * * * +
ay is a + + condition (2.28) replacing the exponential stability hypotheses of
Hurwitz polynomial. It is easy to see that this control results in Proposition 2.1 and the Lipschitz dependencies as before, we see
asymptotic tracking and bounded state zr (or equivalently y , that the proof goes through. Condition (2.28) itself can be restated
9,. * y7-l) provided y M ,j ~. . , y s - l ) are bounded.
e , e , in a form reminiscent of (2.28) involving a more general
1126 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 34, NO. 11, NOVEMBER 1989
Lyapunov function V ( z , ) with the weakening that in Isidori and Moog [9], depending on which linear definition one
chooses to generalize: 1) the dynamics associated with the
dV subsystem that becomes unobservable when a certain state
- . $(O, Z2)S - W I Z 2 1 only for lz21r R . (2.29) feedback is implemented (to the extent of maximizing unobserva-
dz2
bility); 2) the internal dynamics consistent with the constraint that
Thus, bounded tracking only requires that the conditions (2.25) the output is zero; 3) the dynamics of the inverse system. In
only hold outside a ball of radius R. We refer to this condition as particular, the notion that is behind 2) has some interesting
exponential boundedness of the zero dynamics. features that render it particularly suitable for the design of
2) The Multiinput, Multioutput Case: Definitions of mini- stabilizing feedback (see Byrnes and Isidori [28]) and to study
mum phase for the square multiinput, multioutput nonlinear asymptotic tracking.
systems parallel the development on the SISO case above only if
the matrix A ( x ) defined in (2.10) is nonsingular for all x E Wn. III. ADAPTIVE
CONTROL
OF LINEARIZABLE
SYSTEMS
In this case, locally around any point xo of R", a diffeomorphism
(zl, z2) = T ( x ) can be defined, with In practical implementations of exactly linearizing control laws,
the chief drawback is that they are based on exact cancellation of
zT=(hl, Lfhl, * * e , L j ' - ' h l , h2, * . a , nonlinear terms. If there is any uncertainty in the knowledge of
the nonlinear functions f and g , the cancellation is not exact and
Ly-lhz, e * * , hp, * * a , LT-'hp). the resulting input-output equation is not linear. We suggest the
In these coordinates the equations (2.1) read use of parameter adaptive control to get asymptotically exact
cancellation. The following simple example makes our philosophy
i l l =z12
clear.
"I
f( x i = effi(x) (3.1)
i= 1
(2.30)
"2
Yl =zll g ( x )= Bfgj(x) (3.2)
j=I
Y2 = Zh,+ 1
with e,!, i = 1, e , nl; Of, j = 1, -
n2 unknown parameters
a ,
g m ( Z 1 , 22)
fI(Z1, z2)
fm(Z1r
i
z2)
. (2.32) with e,!, ef standing for the estimates of the parameters e;, Of,
respectively, at time t . Consequently, the control law U is
replaced by
Using this control in the equations for z2 and assuming as before
that 0 E W" is the equilibrium point of the undriven system (i.e.,
f ( 0 ) = 0) and hI(0) = = hm(0) = 0 , we see that the
* e -
-
mode control law
(3.9)
1 -
U=- ( - L f h + j M + k sgn ( y M - y ) ) . (3.15)
L J
and L,h
The error equation (3.11) is then replaced by one of the form
e + k sgn e = d ( t ) (3.16)
- U=^
L,h
1
(-~+,,+ff(y,-y)).
$= -ew. (3.13)
For y 2 2, (3.18), (3.19) are not linear in the unknown
The Lyapunov function u(e, 4) = 1 /2e2 +
1/24 4 is decreasing parameters 6;.For example,
along trajectories of (3.11), u(e, 4) = a e 2 5 0; thereby
establishing bounded e , 4. Also 5, e2 dt < W . However, to
establish that e -P 0 as t -P 00 we need to verify that e is uniformly
continuous (or alternately that e is bounded). This e r n needs w ,
-
a continuous function of x (well defined since L,h is bounded
away from zero) to be bounded. Now, note that bounded e,
bounded y M y is bounded. From this and the minimum-phase
assumption (cf. Proposition 2.1) it follows that x is bounded.
Hence, w is bounded and e is uniformly continuous and so tends to
and
zero as t -+ 03.
Remarks:
and so on. The development of the preceding section could easily
1) Prior bounds on the parameters O2 are frequently sufficient to
guarantee that L T is bounded away from zero. Several standard be repeated if we defined each of the parameter products to be a
new parameter in which case the B f B 1 and efe: of (3.20) and
techniques exist in the literature for this purpose (see Sastry and
(3.21) are parameters. Let 8 E W kbe the k-(large !) dimensional
Bodson [ 191).
2) Theorem 3.1 makes no statement about parameter conver- vector of parameters e;, Of, Ote;, 0; e:, -
.. Thus, for example,
gence. As is standard in the literature, one can conclude from i f y = 3 then 8 contains e;, e;, e;O:, e;e;e;, Ole2, e;e;e:. NOW
(3.1 l ) , (3.13) that e, 4 both converge exponentially to zero if w is
implemented is
for the purposes of tracking, the control law to be J.
U=-
1
-
The overall adaptive control law then is
L,L;-lh
A
( - LJh + ;). (3.23)
Note that Wl and W2 can be added to get a new regressor vector Equation (3.35) is referred to as a normalized gradient-type
W . It is of interest to note that 0' cannot be explicitly identified in algorithm [unlike the unnormalized update law (3.13)]. Some
this case since the regressor m_ultiplying it is zero. Also note that properties of @, el follow immediately with no assumptions on the
W is a function of x,y M ,and 8. Note that terms involving only O2 boundedness of E . In what follows we will use the following
or even any products of the 0' are absent and so may be dropped notation.
from the vector 8. i) 0 is a generic L2 n L , function which goes to zero as t +
We keep the form (3.24) of the error equation. Note that SY + 00.
alsY-' + +
* * ayis Hurwitz by choice of tracking control. For ii) y is a generic L2 fl L , function.
the purposes of adaptation we need a signal of the form iii) K is a (large) bound.
iv) llzllt will refer to the norm suprr, I z ( T ) ~ , the truncated L,
e l = p l e ( , - l ) +.. . + P,e (3.27) norm.
Proposition 3.2 (Properties of the Identifier): Consider the
with the transfer function error equation
plsY-l+ . . *+ p , P / a , s Y - I + * . .+a, (3.28) el=I@=t (3.35)
strictly positive real. Indeed, if such a signal el were measurable, with the update law
the basic tracking theorem would follow immediately from
arguments similar to the linear arguments. The difficulty with
constructing the signal (3.27) is that e, e @ ) ,*. are not (3.36)
measurable since
Then
e = YM- Lf h
I@ E L,, & E L2 f l L, and
L;=yM-L
P I I @ T t ( t ) l ~ ~ ( l + l l E l l tV) t . (3.37)
and so on, with the L h not explicitly available.
f
Motivated by the linear case, where a so-called augmented Proof: Consider the Lyapunov function
error scheme is necessitated, we define the augmented error.
Some notation is needed at this point. Define V@)=+TI@.
Then we have
(3.38)
Then the equation (3.29) may be written as
so that we get that 9 E L,. Since [ Vdt < w , we also have that
e = M ( s ) . I@TW (3.30) e l / ( l + $, T,$) E L2. Further, since
with the convention (standard in the adaptive control literature)
(3.39)
that the hybrid notation (3.30) refers to the convolution between
the inverse Laplace transform of M ( s ) and a T W . Also the
exponentially decaying initial condition terms are dropped since we have that 4 E L,. Also since
they do not alter the stability proof (for these points and a review
of linear adaptive control, we refer Sastry and Bodson [ 191, or the (3.40)
survey paper Sastry [21]). Define the augmented error
el = e + ( O T M ( s )w-M(s)OTw). (3.31) we see that 4 E L2 (the first term in (3.40) is integrable, see
SASTRY AND ISIDORI: ADAPTIVE CONTROL OF LINEAFUZABLE SYSTEMS 1129
(3.38) above and the second bounded). Finally, define Combining ( 3 . 4 9 , (3.46) and noting that x is a diffeomorphism of
zI, z2, we see that
IIx IIr 5 K I1 W I1r + K
and
The first term is in L2 f l L , and the second bounded. Thus, y is
indeed in L2 fl La. Hence, (3.37) follows. I11IIr 5 K I1 WII, + K - (3.48)
Remarks: The conclusions of Proposition 3.2 are generic Using the facts that 118 W/axll and )Ia W/d8 1) are bounded and
identifier properties. Other identifiers such as the normalized (3.48), we get
least-squares identifier also yield these properties; for details, see
Sastry and Bodson [19]. II rt-IIrsKII WIIr+K (3.49)
We are now ready to state and prove the main theorem.
Theorem 3.3 (Basic Tracking Theorem for SZSO for thus W is regular * E = M W is regular as well (since M is
Relative Degree Greater than I): Consider the control law of stable). A similar calculation yields 9 ' W to be regular as well.
(3.22), (3.23) applied to an exponentially minimum-phase nonlin- For consider,
ear system with parameter uncertainty as given in (3.1), (3.2). d
-
If y, yM, * ,y(Y- I ) are bounded, -si bounded away - (9'W)=Q'W+9TW.
dt
(3.50)
from zero, f, g, %, Lib, L,L:h are Lipschitz continuous
functions of x , and W ( x , 0) has bounded derivatives in x , 8, then Using (3.49) and 9, 4 E L , we get
the parameter update law
(3.51)
(3.36)
But from (3.44), (3.47)
with
II~IIr~KII*'WIIr+K (3.52)
= L - ' ( s )W (3.35)
so that
yields bounded tracking (i.e., x is bounded and y --t y , as t +
03).
II WII sKII@'WIIr+K. (3.53)
Proof: The proof uses three technical lemmas proven in
Sastry-Bodson [ 191 and summarized in the Appendix. Combining (3.53) with (3.51) yields the regularity of a' W . From
Step I-Bound on the Error Augmentation: By the Swap- the regularity of E, 9'W, one can establish that 9't/1 + 11511r
ping Lemma (Lemma 3 of the Appendix) with M (playing the role has bounded derivative and so is uniformly continuous. Since by
of W )= c(sZ - A ) - l b (3.37)
[ is related to W by stable filtering. Hence, linearizing control laws for a class of continuous-time systems
decouplable by static state feedback. The extension to continuous-
11m=~l1 WIIf+K. (3.59) time systems not decouplable by static state feedback is not as
obvious for two reasons.
Using the estimate (3.53) followed by (3.56) in (3.55) we see that i) The different matrices involved in the development of the
(3.58) may be written as control laws in this case depend in extremely complicated fashion
on the unknown parameters.
lel~Pllellf+P. (3.60) ii) While the true A ( x ) may have rank q < p , its estimate
Since L3 + 0 as t -+ 00 we see from (3.60) that e goes to zero as t
d(x) during the course of adaptation may well be full rank in
--* W. This in turn can be verified to yield bounded W , x , etc.
which case the procedure of Section U-A cannot be followed.
Remarks: The discrete-time and sampled-data case are also not obvious
1) The parameter update law appears not to take into account for similar reasons.
prior parameter information such as the mutual existence of Oi, i) The nonadaptive theory, as discussed in Monaco, Normand-
Oj, and so on. It is important, however, to note that-&: best Cyrot, and Stornelli [ 111 is fairly complicated since
estimate of OiOj in the transient period may not be OiOj. If,
however, the parameters are close to their correct values, such Y k + l = h O (f(Xk)+&?(Xk)Uk) (4.1)
information is useful. But, since parameter convergence is not
guaranteed in the Proof of Theorem 3.3, it may$$ be a good idea is not linear in uk in the discrete-time case and a formal series for
to constrain the estimate of OiO, to be close to OiOj. (4.1) in uk needs to be obtained (and inverted!) for the
The preceding remarks are not designed, however, to amelio- linearization. Consequently, the parametric dependence of the
rate completely ones concerns that the number of parameters control law is complex.
increases very rapidly with y. ii) The notions of zero-dynamics are not as yet completely
2) In several problems it turns out that L;h and L,L;-h developed. Further, even in the linear case, the zeros of a sampled
depend linearly on some unknown parameters. It is then clear that system can be outside the unit disk even when the continuous-time
the development of the previous theorem can be carried through. system is minimum-phase and the sampling is fast enough
3) Thus far, we have only assumed parameter uncertainty in f (Astrom, Hagander, Stemby [ 11).
and g ; but not in h. It is not hard to see that if h depends linearly Thus, we feel that the present contribution is only a first step in
on unknown parameters, then we can mimic the aforementioned the development of a comprehensive theory of adaptive control
procedure quite easily. for linearizable systems.
4) Parameter convergence can be guaranteed in Theorem 3.3 if
W is sufficiently rich in the sense stated after Theorem 3.1. APPENDIX
5) If some of the 0;s are known, they can be replaced in the TECHNICAL LEMMAS
algorithm by these true values and adaptation for them turned off.
6) The parameter update law is a state feedback law as before. We state three important lemmas which are proven in Sastry-
Bodson [19]. The notation y, /3, K from the text is widely used in
C. Adaptive Control of MIMO Systems Decouplable by Static the Appendix.
State Feedback Lemma 1 (BOBI Stability):
Let y = H(s)u be the output of a proper, minimum-phase
From the preceding discussion it is easy to see how the linear system with input U. If U, zi E L,,, and U is regular, i.e.,
linearizing, decoupling static state feedback control law for square
systems (minimum-phase) can be made adaptive-by replacing
the control law of (2.12) by
Then
1= I l u l l f ~ K l l Y l l f + ~ .
Remark: If the input is regular and the plant is minimum-
phase, then bounded system output implies bounded system input.
Note that if A ( x ) is invertible, then the linearizing control law is Lemma 2: Let y = H(s)u be the output of a proper stable
also the decoupling control law. Thus, if A ( x ) and the L;ihi system H(s) driven by U.
depend linearly on certain unknown parameters, the schemes of
the previous sections (those of Section 111-A if yl = y2 = . * = lf
y, = 1 and those of Section III-B in other cases) can be readily
adapted. The details are more notationally cumbersome than
II U IIf S Y (IQ111+ Y9