Systems & Control Letters 18 (1992) 147-158
North-Holland
Explicit design of time-varying stabilizing
control laws for a class of controllable
systems without drift
Jean-Baptiste P o m e t *
Laboratotre d'Automattque de Nantes, U R A C.NR.S 823,
E C.N., 1 rue de la Noe, 44072 Nantes, France
Recewed 1 June 1991
Revised 26 September 1991
Abstract This paper gives a systematic way to destgn time-
variant feedback control laws for a class of controllable non-
linear systems. This class contains a lot of systems which
cannot be stabilized via a tlme-invariant feedback control law.
T h e interest of this work lies in the design method since a
general existence result ~s already available. The techniques
employed here are basic' they mainly involve classical Lya-
punov analys~s.
Keywords N o n h n e a r systems; asymptotic stabihzation; con-
trollabdity; t~me-varymg feedback.
I. Introduction
Let us consider a control system of the form
m ul =u~(x) . . . . ,u~ = u ~ ( x )
Yc = E ukf~ (1)
k=l
where x lives in ~" and the controls u k are real.
The fk'S are smooth vector fields.
Such affine control systems enjoy nice open-
loop properties due to their symmetry (any trajec-
tory can be followed in either direction by chang-
ing the sign of the controls). For instance, full
rank of the Control Lie Algebra (i.e. the vector
space spanned at any point by all the Lie brackets
of the vector fields fk having dimension n) is
sufficient for controllability in any usual sense. It
has however been known for a long time that
controllable nonlinear systems may fail to be sta-
bilizable by time-invariant static or dynamic feed-
* This work was done while the author was visiting at I N R I A
Sophla-Antipolls, 06565 Valbonne, France
0167-6911/92/$05.00 © 1992 - Elsevmr S o e n c e Pubhshers B.V. All rights reserved
147
back. The link between controllability and stabi-
lizability, or between open-loop and closed-loop
properties in general is a topic of ongoing re-
search, and actually one of the most challenging
problems in nonlinear systems. For a survey on
nonlinear stabilization, see for instance reference
Sontag [12].
As a matter of fact, one can appreciate the gap
between stabilizability and controllability for a
system like (1) in the following (negative) result,
which is a simple consequence of a necessary
condition for feedback stabilization given by
Brockett [1]. Note that condition (2) does not
contradict controllability, on the countrary. It is
satisfied by our examples (30) and (36).
Proposition 0. If m < n and
Rank{fl(0) . . . . . fro(0)} = m (2)
then there exists no continuous feedback law
makmg 0 a locally asymptotically stable equilib-
rium point of the closed-loop system. There does
not exist either any continuous dynamic feedback
law
U 1 = U l ( ~ , X ) . . . . . Um = U m ( ~ , X),
~ = g ( ~ , x)
making (~, x) = (0, O) a locally asymptotically sta-
ble equilibrium point of the closed-loop system.
Proof. Let e 0 be a nonzero vector linearly inde-
pendent from fl(0) . . . . . f,,(0). By continuity, there
is an e > 0 such that for all (x, Ul,..., u m) with
II x II < e, the vector
~, ukf~(x)
k=l
148 J -B Pomet / Ttme-t a~mg stabzhzmg control hTws
is &fferent from Aeo for any A in I~. Therefore,
the map
( x, u, ..... Urn) "-" ~ ukfk(x)
I,-1
does not map the neighborhood [ - e , e] "+m of 0
in ~,+m Into a neighborhood of 0 in ~ ' . This
proves the proposition because the necessary
condition for local asymptotic stabilizability of (1)
given in [1] is violated. This condition is necessary
for dynamic feedback stabilisation as well. []
This impossibility to stabilize by means of a
state feedback has been noticed on many occa-
sions in the area of mechanical systems with
nonholomic constraints; see, for instance, Samson
[9] and Campion, d'AndrEa-Novel and Bastln [2].
With this motivation, Samson tried, successfully,
time varying control laws to stabilize some non-
holonomic robots; see [9,10]. This showed the
power of thls class of control, and is the starting
point of some research on time-varying stabiliza-
non for nonlinear systems. It has recently been
proved by Coron [3] that system (1) is stabilizable
by means of a periodic time-varying feedback if
its Control Lie Algebra has rank n except possi-
bly at the origin, but the proof of existence does
not explicitly provide the control law. S6pulchre
[11] using a different approach, also gives some
time-varying control laws in some other situa-
tions
The present paper presents a constructive ap-
proach allowing to derive an explicit control law
and Lyapunov function under an assumption
which is more restrictive than full rank Lie Alge-
bra. By explicit, we do not necessarily mean that
the control law is an algebraic expression of some
functions fixed by the designer and the compo-
nents of the vector fields, although this is the case
in Section 2.2: in the more general situation con-
sidered here (Section 2.1), writing the control
laws involves the flow of an ordinary differential
equation. Note that this exhibits almost the same
degree of complexity as performing feedback hn-
earization for a feedback linearizable system (i.e.
in particular finding the linearizing coordinates),
which involves the flow of several ordinary differ-
ential equations; see [7,5].
The assumpnon we make is the following: we
require the Control Lie Algebra to have full rank
and to be generated by iterated L~e brackets
involving at most one f~ different from )'t, and
only once:
Assumption 1. For any x ~ 0,
Rank{f,(x),f,(x), ..Jm(X),
[fl, J2](x), • .[fl, fm](X) , ".
ad~, f2 ( x ) . . . . ad~,fm ( x ) . . . . } = n. (3)
Remark 1. Of course, fl does not play a privi-
leged role. If there is a linear combination of the
fk's such that the fk's and all the iterated Lie
brackets of this linear combination with them
span ~ ' , then one may satisfy Assumption 1 by
making a preliminary feedback transformation
which 'renames' this linear combination fl. We
take advantage of this possibility in the treatment
of the 4-dimensional example (36) in Section 3.2.
It is actually possible to extend the method
presented here from the case of Assumption 1 to
the more general case of full rank Control L e
Algebra for x 4: 0. This extension makes use of
results contained in [3], it is presented in the
future publication [4] and it is briefly outlined in
section 3.3.
The techniques employed here are based on
Lyapunov analysis and LaSalle's invarlance prin-
ciple [8] They are similar to these employed by
Jurjevic and Qulnn [6]; see Remark 5 in the proof
of Theorem 1 (Section 4.2).
The paper is organized as follows: in Section 2,
a class of time-varying stabilizing control laws for
system (1) is proposed, Section 3 illustrates the
method by applying it to two examples from non-
holonomic robots, and outlines a possible exten-
sion, Section 4 gives a proof of the stability re-
sults stated in Section 2.
2. The control laws and the main results
Here, we explicitly build a family of control
laws, with a degree of freedom represented by a
function a. Section 2.1 gives a general way of
building a time-varying stabilizing control law un-
der Assumption 1. This method involves solving a
hnear PDE, i.e. the expression of the control law
(and the Lyapunov function) involves the flow of
a certain ordinary differential equation. Section
2.2 gives, under the additional Assumpnon 2, a
 J -B. Pomet / Ttme-carymg stabthzmg control laws 149

way to derive some stabilizing control laws with-
out solving any differential relations, except pos-
sibly for finding some coordinates in which this
assumption is met: the control laws are actually
algebraic functions of the components of the vec-
tor fields in these coordinates and of some vari-
ables and functions to be chosen by the designer.
2.1. A general expression under Assumption 1
The following proposition gives a periodic
time-varying function, used in the design of the
control laws, and meant to be a Lyapunov func-
tion for the final closed-loop system.
For a given time-varying function a(t, x), we
may consider the time-varying ordinary differen-
tial equation
2 = a(t, x ) f , ( x ) , (4)
and define its 'general' solution ~b(t~, t2, x): it is
the value at time t I of the solution of (4) which
started from x at time t 2. In other words,
OA(tl, t2, X)
Ot 1
= a ( t , , ~b(tl, t2, x))fl(~b(t ,, t 2, x)), (5a)
~b(t, t, x) = x . (5b)
is defined, from classical results on ordinary
differential equations, on an open subset of ~ ×
x En containing the points (t 1, t 2, x) such that
t I = t 2.
Proposition 1. Let a be a time-varying functton
(i.e. a function from ~ × ~n to ~), and let us
suppose that:
(i) it is 2w-periodic and 'odd' with respect to
time, and vanishes for x = O:
a(t+2"rr, x)=a(t,x) V(t,x), (6)
a(-t,x)=-a(t,x) V(t, x), (7)
a(t,0)=0 Vt; (8)
(ii) it satisfies the following majoration, for a
certam K > 0:
where ~b is given by (5), /s well defined from
~( ~ n to [0, -~- 00) and enjoys the following prop-
erttes :
(a) It ts 2rr-periodic with respect to ttme:
V(t + 2w, x) = V(t, x). (11)
(b) V as well as its derivative with respect to x
vamsh if and only if x = O:
V(t, x ) = 0 ~ x = 0 , (12)
OV
-~x(t,x)=0 ¢~ x = 0 . (13)
(c) It is (the only) solution of the following
hnear PDE with inttial conditions at t = 0:
OV OV
- ~ ( t , x) + a ( t , x)-~x (t, x ) f l ( x ) =0, (14)
v(0, x) = ½ll xl[ 2, (15)
(d) For any positive real number K, the set
{ x, V( t, x) < K for some t ~ R} ts bounded.
(16)
Remark 2. (a) In the expression of V (10), the
squared norm might be replaced by any proper
positive function, vanishing only at 0, and with no
other singular point.
(b) Majoration (9) might be replaced by any
condition ensuring that all the solutions of .f =
a(t, x)fl(x) are defined for any time.
(c) Equation (7) might be replaced by
a(T-t, x) = - a ( t , x)
for a certain T ~ R, or by any condition ensuring
that all the solutions of 2 = a(t, x)f~(x) are 2-rr-
periodic.
This proposition will be proved in Section 4. It
turns out that, for any a meeting (6), (7), (8), (9),
the 2rr-periodic time-varying function 1/given by
(10) may be assigned to be non-increasing along
the closed-loop system by using the following
2w-periodic time-varying control law:

l a ( t , x) lllfl(x)ll < g ( l + Ilxll) V(t, x). Ul(t, x) = a ( t , x) - L f l V ( t , x),

(9) u2(t, x) = - L h V ( t , x),
(17)
Then, the funcnon V given by
1 2 urn(t, x) = -LfmV(t, x).
V(t, x ) = ~11~0(0, t, x)II , (10)
151) J -B Pomet / Ttme-l arymg stabthzmg control laws
Under Assumption 1 and a further reqmre-
ment on a, this control law asymptotically stabi-
lizes (1). An explicit a meeting all the require-
ments may very easily be gwen, but we keep some
freedom on a in order to be able, on some
examples, to derive the simplest control laws we
can. This is made precise by the following theo-
rem:
Theorem 1. (a) Under Assumption 1, and if a is
chosen such that conditwns (6), (7), (8), (9) of
Proposition 1 are satisfied and
LfV(t,x)=O Vk~[1,m]]
= x=0,
0Ja(t, x)=0 Vj>I
Otj
then 0 is a globally uniformly asymptottcally stable
equilibrium point of the time-varymg penodic
closed-loop system obtained by applying the feed-
back law defined by (17), and the function V ts
non-increasing along its solutions.
(b) It is always possible to find a function a
meeting the requtred condtttons. A possible chotce
lS
JI x II 2
x)=
(1 + [[x[12)(1 + I l f l ( x ) I [ 2)
R e m a r k 3. A more general choice for a is
a ( t , x) = (sin t ) p ( x )
where p vanishes only for x = 0 and is such that
(9) is met. Many more choices are possible, but
the expressions of LFkV are needed, and L F V
depends on a. For example (36) (Section 3.2), it
is a priori possible not to include ~4 in the
expression of a because, regardless of a, Lf2V =
~4 whenever a is zero.
2.2. A simpler expression of the control laws m a
less general case
The above design of the control laws makes
use of the function V constructed in Proposition
1. Its expression involves the flow of a differential
equation. It is therefore very natural to look for
some situations in which the expression of V can
be explicited without solving any such differential
relation. It turns out that if the vector field ]i is
globally redressable, it is actually possible to gwe
an explicit expression of V and the control laws;
it is the purpose of th~s section. As seen from
example (36) in Section 3.2, it may often be
simpler, ff possible, to make the necessary trans-
formations (i.e. a preliminary static feedback and
a change of coordinates) to meet Assumption 2
than to the use the general expressions given in
previous Section 2.1.
Assumption 2 (simplifies the expression of the
feedback law). There exists a system of coordi-
nates (x~ . . . . . x , ) such that
(18)
f, - (21)
0x i
Under this assumption, it is always possible to
choose V of the form
V ( t , x) = ~_(x
l 1 + h ( t , x 2 . . . . . x~)) 2
+ ~1x 22 + • - • + ~x;,,
1 ?
(22)
with some requirements on h, so that, instead of
taking a as a design p a r a m e t e r and solving the
sin t. P D E (14)-(15) for V, we are now able to take h
as a design p a r a m e t e r and compute a so as to
(19) satisfy (14)-(15):
Proposition 2. If Assumption 2 is met, then for
any function h(t, x 2. . . . . x , ) satisfymg
(20)
h ( t + 2~', x z . . . . ,xn)
= h ( t , x 2 . . . . . xn) V(t,x) (23)
h ( t , 0) - - 0 Vt, (24)
the function V gwen by (22) and the function a
gwen by
Oh
a ( t , x) = - ~t (t, x 2 . . . . . x , ) , (25)
are such that all the conclustons of Proposttion 1
(from (10) to (16)) are verified and m addition a
satisfies (6) and (8).
This proposition gives, from h, an a and a V
which are sufficient, as before, to construct the
control law according to (17). We have the follow-
 J-B Pomet /Ttme-varymg stabthzmg controllaws 151

ing theorem, which is a straightforward conse- Since [fl, f2] = a/ax2, we have
quence of Theorem 1:
Rank{f1, f2, [f~, f2]} = 3,
Theorem 2. (a) Under Assumption 1 and 2, and tf
so that Assumption 1 is satisfied. Assumption 2 is
h is chosen such that (23) and (24) are met and:
met as well in the original coordinates, so that we
=0 Vk ~ [2,m] may follow the method of Section 2.2 and take a
LfW(x) 1 ~ x=O. V of the form (22):
OJh
Vj> 1
0t'(t' x)=0
J (26)
V ( t , x ) = l ( x I + h ( t , x2,
1 2 1 2
x3)) 2

(32)
" - ~ X 2 q- ~X 3.

where Following (28), we may choose

W ( x ) = ' -~x2+
2 . . . +~x
, n,
2 (27) h ( t , x2, x3) = (x2z + x 2) cos t,
then 0 ts a globally uniformly asymptotically stable and, from (17), the control law
equdibrium point of the ttme-varying periodic
closed-loop system obtained by applying the feed- u l = ( x z z + x 2) s i n t
back law defined by (17), wtth a and V defined by -(x, + cos ,),
(22) and (25). The function V ts non-increasing
along its solutions. U2= --2(X, + ( x 2 + x 2 ) c o s t)
(b) It is always possible to chose a function h " ( X l X 2 q - X 3 ) COS t
meeting the requirements. A possible choice ts
--(XlX2q"X3).
h( t, x ) = (cos t ) ( x 2 + " " +x2), (28)
A simpler choice of h is possible:
a( t, x) = (sin t ) ( x ~ + " " +x2). (29)
h ( t , X2, X3) = X 2 COS t (33)
Remark 4. Of course, some other choices than
also satisfies (26) since
(28) are possible, as illustrated by the second
choice for example (31) in Section 3.1. L f l V = x I + x 2 cos t, (34a)
L h V = ( x I + x 2 cos t ) x 1 cos t + x2x 1 +x3,
3. T w o e x a m p l e s (34b)
3.1. A three-dimenstonal example ah
-- = - x 2 sin t, (34c)
0t
Let us consider the following 3-dimensional
example: ~2h
Otz = - x 2 cos t. (34d)
21 = ul, (30a)
2 2 =XIU 2 , (30b) With this choice of h, we find the following
simpler control law:
3f3 = U 2. (30c)
This is a reduced-order model for a three-wheeled u I = x 2 sin t - (x 1 + x 2 cos t), (35a)
cart with two independent rear motorized wheels. u 2 = - ( x 1 + x 2 cos t ) x 1 cos t - ( X l X 2 q - x 3 ) .
A time-varying feedback law has already been (35b)
proposed for such a system in [9]. We show here
that our method, and even the simplified one
given in Section 2.2, allows one to derive very 3.2. A four dimensional example
simple feedback laws for this system. System (30)
is of the form (1) with The following system in ~4:

(1 + i~2 = Ul COS ~:4e'~:3, (36a)

fl = 0X 1 , f2 =Xl OX2 + OX'-'-~ (31)
(3 = ul sin G , (36b)
152 J -B Pomet / Tmw-t arymg stabthzmg ~ontrol law;

~4:4= u2, (36c) The expressions of the Lyapunov lunctlon and

the control law involve the flow of the t~me-vanant
Is a reduced order model for a three-wheeled cart ordinary differential equation
with a steering front wheel. -61 and ~:z are the
cartesian coordinates of the middle of the rear (1 = 1~ = a( t. -6) cos _64 e '~, (40a)
axle, -63 is the angle between the axis of the cart
and a fixed direction, and ~:4 is the steering angle. ( = a ( t, -6) sin -64, (40b)
A time-varying stabilizing feedback law has al-
-64 = 0, (40c)
ready been proposed for such a system in [10].
Let us shown that: with a choice of a such that (6), (8), (7), (9), and
- Our general method (Section 2.1) apphes. It (18) are met. On this example, as on most exam-
is not easy however to write down the expression pies, one may find a simpler a than the one given
of the corresponding control law. in (19). For example,
- Our simplified method (Section 2.2) also
apphes, in some coordinates and after a prelimi- a(t,-6)=-61 slnt+~2sin2t+-63sin3t (41)
nary static feedback transformation. We write
may easily be proved to meet conditions (6), (7),
down completely the corresponding control law.
(8), (9), and (18) (the latter is met because LhV
- Our control law is different from the one
=~4 when a(t, ~ ) = 0). However, even m this
proposed by Samson [10], the expressions ob-
case, it is not easy to write explicitely the solu-
tained in [10] are slightly simpler, and may be
tions of (40), and therefore the controls u I and
re-obtained through a generahzatlon, outhned in
Secnon 3.3, of the method presented in Section
It turns out that it is simpler to perform a
2.1.
prehminary (static) feedback transformation in
System (36) is of the form (1):
order to meet Assumptions 1 and 2 together and
apply Section 2.2. If we define some new controls
~= ulgl(-6 ) + u,g2(-6 ) (37)
t' l and t' 2 by
(we use g instead of f on purpose) with
Ul=t'24-t'l, /,/2 = UI , (42)

system (36) reads

g l = COS -63 COS -64 ~-61

= r l f 1 4- t ' 2 f 2 (43)

+ sm -6~ cos -64 ~-6----~÷ sin -64 ~-6~-, (38a) with

f1=g2+gl, f2=gl. (44)

g 2 -~ - - (38b)
A very simple computation allows us to deduce
from (39) that these fl and f2 satisfy Assumption
A simple computation gives: 1. In addition, since the vector field fl has a
constant nonzero component on 0/0-64, ~t meets
Rank{gl, g2, [g,, gz], [g,, [g,, g21]} --4, Assumpnon 2, and as a matter of fact, the coor&-
(39a) nates x l, x 2, x 3, x 4 may be chosen as
[g2, [g,, g2]] = g , , (39b) Xl = -64, (45a)

so that Assumption 1 is met w~th f~ =g~ and x2 = -63 + cos so4 - 1, (45b)
f2 =g2 but not with f l =g2 and f2 =g~- Unfortu- X 3 4- i x 4 = -61 + i~:2
nately Assumption 2 is not met by f l = g~ since it
may be proven that the vector field gl is not - - e ~(~+c°s~4 l)f~'te-lC°~'r COS r dr.
globally redressable. It is therefore not possible /

~0
to use the method of Section 2.2 with fl =g~, (45c)
f2 =g2. The method of Section 2.1 does apply.
 J - B Pomet / Ttme-varymg stabthzmg control laws 153

Note that finding these coordinates involved the Samson [10] designed some different control
expression of the solutions of laws for the same system, using

(1 + i~2 = cos ~4 el~3, (46a) Z 1 3t- iz 2 = (Sc, + i¢2 ) e -'e' (51)

instead of ~cI and ~:z ((zt, z2, ¢3, so4)is (x, y O, a)
~3 = sin ~4, (46b)
in [10]). We may rewrite his control laws as
(4 = 1, (46c) u I = a j ( t , zl, z 2, so3, ~4)
whereas applying the method from Section 2.1 - L g j V ( t , z l, z2, ~3, ~4), j = l , 2, (52)
involved the solutions of (40), which is more
complicated due to the presence of a. In coordi- where, k being a function of t, z2, ~3 (k is a
nates ( x l , . . . , x4), the vector fields are degree of freedom),

0 V ~-~-l ( z I ~ - k ( t , z2, so3))2

= , (47a)
fl ~x 1 2 1 2 1 2 (53)
q'- 2Z2 "[- 7~3 q'- 2~4'
0 0 3k
f 2 = sin x l - +X3(xl, x z ) - - 0~1 - - Ot '
(54)
~X 2 OX3
sin ~:4 sin ~:4 ~ Ok
Zll (55)
+ X4( xl, x2) Ox4 , (47b)
012 = -- ~4 Z I Z 2 q- - -
¢4 ~31~-
l( t, zl, z2, ~3, ~4)
with
sin so4 [ Ok Ok ] cos £4 - 1
X3(X1, X2) d - i x 4 ( x l , x2) : ~4 ~ Z2 -- - - Z l
0Z 2
e'X2(e -lc°sx~ cos x 1 (56)

(l is h in [10]). These laws are proved to yield

- i sin x 1 j:e= cos ~" d~" ) (48)
stabilization for suitable choices of k, like
k ( t , z2, ~3) =z2 sin t+~: 3 cos t.
The Lyapunov function V is given by (22) with From a computational point of view, both (52)
h = (x22 + x 2 +x42) cost: and (50) give the controls as analytic fuctions of t,
~:1, ~2' ~3' ~4 and involve some 'unusual' analytic
V(t, x) = l ( x 1 -~- ( x 2 -[- x 2 -']-x42) cos t) 2
functions. However, one may find it more conve-
1 22 + ~X
+ ~X 1 32 + ~x
1 42 (49) nient to have to compute
sin ~4 cos ~4 -- 1
and we get from (17) with and
~4 ~:4
a(t,x)=(x 2 + x 2 + x 2) s i n t
as in (52) than to have to compute
the following control law: ~4
J cos(cos ~) cos dr
Ul(t, x ) = ( x 2 +x32 + x ~ ) sin t 0

- x,- + x3 + cos t,
and

u2(t, x) fo'~' sin(cos r) cos T dr,

= - 2 [ cos t(x I + ( x 2 +x 2 + x 2 ) c o s t ) + 11
• [x2 sin Xa nt-x3)(3(X1, X2)
-k-X4X4(Xl, x2) ].
as in (50).
From a methodological point of vtew, the
method leading to the controls in [9] is quite
(50) different from the one presented in the present
1~4 J -B Pomet / 7tm,,-t ar~tng stabtltztng control laws

paper, and (52) is only a way of r e w r m n g the
same controls. However, V given by (53) may be
checked to be the unique solution of
i~l'
- - + o~IL~,,V + ce2Lq,eV = 0,
expression (58) because it turns out that taking
the norm cancels out the parts of 0 whmh cannot
be expressed simply
(57a) 3.3. A n extenston o f the present method

V(0, ~ ) = ½ ( z , +k(0, z 2, ~ ) ) e In the design proposed in Section 2.1, one may

use, instead of only one function a, m functions
+_~(z5+ 3+sc~) • (57b) % . . . . a,n, each with the properties (6)-(9), where
It is therefore the function V described in Section (9) becomes
3.3 below, Le. It satisfies (60)-(61) (changing the I o~,(t, x)f II f , ( x ) II _<K(1 + I[x Jl),
letter f into g) and it is given by (10) and (59).
Since (52) is exactly (62), the m e t h o d of Section and define the Lyapunov function V by (10) with
3.3 with a 1 and a 2 given by (54) does lead to the the flow of
same control laws as [10]. It may be checked that
(63) is met for instance if .f = f i a , ( t , x ) f , ( x )
1=]
k ( t , z e, so3) = z , sm t + ~ 3 cos t.
instead of 2 = a(t, x ) f l ( x ) , i.e. with
Note. (a) Actually, to get V exactly as in (53),
another function than the squared n o r m has to be --(t,, re, x) = ~,(t~, ~(t,, t2, x))
used in (10) ff k(0, z 2, ~:3) is not zero (see also
remark 2); for instance, if .f,(g~(t,, t z, x ) ) , (59a)
k ( t , z~, ~3) = z , sin tq-~: 3 COS t, ~0(t, t, x ) = x , (59b)
then one has to chose in (10) the n o r m according instead of (5). Proposition 1 may easily be ex-
to the matrix tended. Relation (14) is modified: defining, as m
the proof of T h e o r e m 1, a vector field F 0 on
1 0 00 × 0~" (for the sake of clarity, we keep ~ instead
of S l here and use t instead of the 'time modulo
0 2
0 0 2rr' 0) by
m
m the coordinates (z 1, z 2, s~a, s¢4) since
Fo = - - + ~ a k F k (60)
at k=l
v(o. : l [i .., + + + +
with the Fk'S defined as in the p r o o f of T h e o r e m
(b) The expression of V obtained from Sec- 1 (below equation (72)), (14) is replaced by
tion 3 3 involves the solution (tO) of
LFoV= 0 (61)
zl = a l ( c ° s ~4 + z e sin ~4), (58a)
5,_ = - a l z I sm ~:4, (58b) (which gwes (14) if a I = oz and a 2 . . . . . o~,,, =
0). The control is then defined by
~, = a, sm ~4, (58c)
ut, = c~/,(t, x)
(~ = c~,. (58d)
-L&V(t,x), k = l . . . . m. (62)
For c~l and 0% given by (54) and (55), and a
general function k(t, z 2, ~3), like The same p r o o f as for T h e o r e m 1 shows that, if
the a,'s have the above properties and
k ( t , z e, so3) = z 2 sin t + ~ : 3 COS t,

the solutions of (58) cannot be completely writ-

Ldd~'jY(t'x)=O' I ~ x=0, (63)
ten. Nevertheless, V given by (10) has the explicit k=l ..... m;j>O 1
 J -B Pomet / Ttme-varymg stabthzmg control laws
then the control law (62) yields global asymptotic
stability of the origin.
This extension is interesting for two different
reasons:
(a) In the context of this paper (systems with-
out drift meeting Assumption 1), it gives some
more degrees of freedom. As an aside conse-
quence, it allows one to interpret the control laws
proposed by Samson [10] in the framework of this
paper, see Section 3.2.
(b) It may also extend to the more general
case of controllable systems without drift, not
necessarily meeting Assumption 1. This exten-
sion, using results from [3], is developed and
discussed in the forthcoming [4]. Let us briefly
outline it. It is not easy to deduce from controlla-
bility (full rank Control Lie Algebra) an explicit
expression of some a,'s meeting (63), whereas
Assumption 1 allowed us to give the explicit
expression (19) for a. The central difficulty solved
by Coron [3] is to establish existence of some a,'s
such that (his notations are not the same as here):
Rank{ad~0Fk(-32ar, x ),
k=l ..... m;j>O}=n Vx~O.
From (13), this implies property (63), at least for
3
t = 7rr, and this is actually enough to conclude,
via an extension of the proof of Theorem 1.
4. The proofs
The proof of Proposition 2 and Theorem 2 are
left to the reader. All the properties in Proposi-
tion 2 can be checked by hand from (21), (22) and
(25). Theorem 2 is actually a particular case of
Theorem 1 since it is easy to verify that a given
by (25) satisfies all the properties required in
Theorem 1 except maybe (7) and (9) but these
were only required to ensure that V be well
defined and periodic, which is here obvious from
(22) and (23).
4.1. Proof of Proposition 1
First of all, (9) implies that qJ, the 'general
solution' of (4) is defined for any tl, t z and x
because the right-hand side of (4) is sub-linear. V
is therefore defined all over ~ × ~ ' .
155
To prove the other properties, let us explain
where formula (10) comes from, and for which
reasons it makes V a solution of the PDE (14).
This is just an application of the method of
characteristics for solving the linear PDE (14):
(14) exactly means that the function V is constant
along the vector field O/at + af~ in ~ x Rn. This
vector field is complete from (9) and its flow at
time r is the diffeomorphism of ~ × ~" given by
(t, x) ~ q,(t + ~, t, x ) .
Formula (10) therefore states that V(t, x) is equal
to V(0, ~) where (0, ~) is the image of (t, x)
under the flow of a / a t + a f ! at time - t , i.e.
(0, ~) is the only point on the same integral curve
of O/at + af~ as (t, x) with a zero t-component.
V is therefore obviously constant on the integral
curves of O/Ot + o f 1, and this proves (14). Note
that (14) may also be checked by hand: by differ-
entiating the identity
t/,(~r, ~', ~(~', or, ~)) = ~: (64)
with respect to ~-, one gets
~2(~, ,, q,(,, ~, ~))
+O--x
(65)
From (5), one may replace (Otb/atl)(z, o', ~) in
(65) with
a ( z , $(~-, or, ~))fl(~b(z, iT, ~:)).
Applying the resulting identity for or = 0, ~"= t
and ~b(r, tr, ~) = x (i.e. ~: = ~O(0, t, x)), one gets
0q,
a~b(O, t, x) + (0, t x ) a ( t , x ) f l ( x ) =0.
'
(66)
Together with the expression (10) of W, this
clearly implies that V satisfies the PDE (14).
The fact that, for any tj and t2, tO(t~, t 2, 0) is
zero (from (8)) and x ~ q,(tl, t2, x) is a diffeo-
morphism of ~" implies that V defined by (10)
satisfies (12), (13), (15) and (16); (16) actually uses
(11) established below.
The only remaining property to be proved is
(11). Consider a solution y(t) of (4); (7) implies
156 J -B Pomet 1 Ttme-t an'rag stabthzmg control law~

that y ( - t ) = y ( t ) , and in particular that y ( - w ) can then be considered as a tlme-mvariant system

= y(w); this and (7) imply that y is 2w-penodlc. on S l × ~n, gwen by
q, is therefore 2"rr-periodm w~th respect to its first
entry. Identity (64) applied twice, for ~ and Cr + ~/= l, (69)
2v, ymlds
= u (O, x ) L ( x )
k=l

where 0 hves in S 1. We may re-write (69), with

= ~0(~r + 2w, r, ~ ( r , ~r + 2w, ~)) = ~ (67) X = (0), into:
and, since ~ is 2w-periodic with respect to its X=F(X) (70)
first entry,
w~th

= ~0(~, r, ~ ( r , o'+ 2rr, so)), F = E =,uk(O, X ) f k ( X )

which lmphes where the functions u k are given by (17). In other

words,
g,(r, ~r, ~:) = tO(r, ~r + 2~r, ~:) 0 m
F=--+ Y'. UkFk (72)
because x ~ @(o-, r, x) is one-to-one. ~0 is there- a0 k= 1
fore 2w-periodm with respect to its second entry.
where F k stands for the vector field on S ~ x ~"
This implies (11). []
naturally induced by the original fk on ~" ( ) ( =
F~(X) means 0 = 0, i =f~(x)).
4.2. Proof of Theorem 1 The fact that V decreases along the solutions
of the closed-loop system is straightforward since
Remark 5. The stability proof given below is very the time-derivative of V(x(t), O(t)) along the so-
much in the spirit of the one given in [6]. Actu- lutions of the system is, from (69), (17) and (14),
ally, once the time-varying closed-loop system on m
~" has been made a time invariant one on S l × V(O, x ) = - E (L&V(O, x))'. (73)
~", and once property (81) (Lemma 1) has been /~=1

established, one may notice that (81) is very simi-
lar to the 'ad-condition' used in [6], and this
allows to make almost the same application of
LaSalle's invariance principle as in [6]. However,
since it is somehow modified (one extra-dimen-
sion, lnvariant set S 1 × {0} instead of equilibrium
point, a vanishes ...), we give the complete
proof rather than referring the reader to [6].
Part (b) of the theorem is obvious: one may
check by hand that a given by (19) meets all the
requirements of part (a). Let us proceed with the
proof of (b).
Since the functions a, V, u~ ..... u m from R ×
R" to R are 2w-periodic with respect to time,
they induce naturally some functions from S x ×
~" to ~, the circle S 1 being the quotient
S 1= ~/2w~.
The 2"rr-periodic time-variant closed-loop system
To prove that 0 is a unformly asymptotically
stable equilibrium point of the time-variant
closed-loop system, or equivalently that the in-
variant set S l × {0} (or the periodic orbit 0 = 1,
x = 0) of system (69) is globally asymptotically
attractive, we will use LaSalle's invariance princi-
ple [8].
Consider a solution of this closed loop system,
with initial condition (00, %). Since V decreases
along this solution, it stays in the set
V-I(V(Oo, Xo)), which is compact from (16). Any
solution is therefore bounded.
LaSalle's invariance principle states [8, Theo-
rem 2] that, since V decreases along any solution,
any bounded solution converges to the largest
invanant set included in
A = {(0, x), v(0, x) = o}
(68) :{X~S'X~", L F y ( X )=O, k = l . . . . . m}
(74)
 J.-B Pomet / Ttme-varymgstablhzmg control laws
(where the second equality comes from (73)),
which is also the reunion of all the integral curves
of the vector field F which stay in A.
From (17) and (74), if (0, x) e A then
Ul(O, x) = a(O, x),
U2(0 , X) = 0 , . . . , Um(O, X) = 0,
so that the vector field F coincides on A with the
vector field
3 at j (0, x ) = 0 ,
Fo = -~ + aF 1.
We therefore have to prove that the only integral
curve of F 0 (i.e. solution of X'= Fo(X)) which
stays in A i s d = l , x = 0 .
From (74), the functions LFV are identically
zero on these trajectories, and so are their deriva-
tives: at any point X = (0, x) of such a trajectory,
LJFoLFkV(X)=O, k = l . . . . . m;j>__O.
Now, from (14), V is a first integral of F0:
LFoV(X ) = 0 V X ~ - S 1X ~n,
and it is not difficult to prove by induction on k
that this implies
L ~FoLFy= LadkovkV.
From (79) and (77), one can deduce that, at any
point (0, x) of a trajectory which stays in A,
Lad~o&V(O,x)=O, k = l , . . . , m ; j > O .
We shall now make use of the following lemma:
Lemma 1. Under Assumption 1, we have, at any
(0, x) such that a(O, x ) ~ O,
Rank{ad~oFk(0, x ) , k = 1 , . . . , m ; j > 0} = n .
This lemma is proved below. To complete the
proof of the theorem, let us first prove that a is
identically zero on an integral curve of F 0 which
stays in A. In order to prove this by contradic-
tion, let us suppose that there is a point (0, x) on
such a curve where a(O, x) ~ O. From l.emma 1,
the vectors
ad~0Fk(0, x ) , k=l ..... re;j>0,
157
span the subspace {0} X TxRn of T(o,x)(Sl X ~"),
and from (80) they are in the kernel of the linear
form dV(0, x); this implies that dV is colinear to
dO at this point, which is impossible from (13).
This actually proves that, on an integral curve
(75a)
of the closed-loop system which stays in A, F 0
(75b) coincides with O/a0. Therefore, from (74) and the
fact that a is identically zero on this trajectory,
we have, at any point (0, x) of this trajectory,
~)Jot
j>0, (82a)
(76)
L/V(O, x) = 0 , k = 1 . . . . . m, (82b)
which is equivalent to x = 0 from (18). This proves
that any solution of the closed-loop system which
stays in A statisfies x = 0, and completes the
proof from LaSalle's invariance principle. []
Proof of Lemma 1. A simple inductive computa-
(77)
tion from (76) shows that
ad~oFk = a~adlFF~
(78) + linear combination of
ad~ Fk,, O<_j'<j-1;l<k'<m,
(83)
(79) and this proves the lemma. []
5. Conclusion
(80)
We have derived some stabilizing time-varying
control laws, for a class of controllable nonlinear
systems, characterized by Assumption 1. Exis-
tence has already been established independently
in [3] in a more general situation. The interest of
our approach is to provide the control laws them-
selves. Actually, it turns out that, for these sys-
(81) tems with a zero drift vector field, there is a very
wide range of possible stabilizing control laws.
We have kept as many degrees of freedom as we
possibly could w~thout complicating too much the
expressions.
Acknowledgements
The author is very grateful to Claude Samson,
from INRIA Sophia-Antipolis, France, for point-
15~ J -B Pomet / Ttme-t'arg'tng stabthzmg control la~ s
lng out to him the power of time-varying feed-
back laws for stabilization, demonstrated on some
mobde robots. The author wishes also to thank
Laurent Baratchart, from I N R I A Sophia-Antipo-
hs, France, for some frmtful discussions on the
subject and Jean-Michel Coron from Universlt6
Paris-Sud, Orsay, for suggesting condition (7) to
ensure periodicity of all the solutions of (4).
References
[1] R.B Brockett, Asymptotic stabdtty and feedback stabl-
hzatlon, m R W Brockett, R S. Mdlman and H J Suss-
mann, Eds., Dtfferenttal Geometrtc Control Theory (Birk-
hauser, Basel-Boston, 19831 181-191.
[2] G Campion, B. d'Andr~a-Novel and G Bastm, Control-
lablhty and State-Feedback Stabdlzatlon of Non Holo-
nom~c Mechamcal Systems, lnternat. Workshop tn Adap-
twe and Nonhnear Control Issues m Robottcs, Grenoble
(Sprmger-Verlag, Berhn-New York, 1990)
[3] J -M Coron, Global asymptohc stabdizatlon for control-
lable systems without drift, Preprmt (1991), to appear m
Math Control Stgn Systems
[4] J -M Coron and J.-B Pomet, A remark on the design of
brae-varying control law,, tor controllable systems wLthout
drift, Preprmt (19911
[5] L R Hunt, R Su and G Meyer, Design for multHnput
nonhnear systems m R W Brockett, R S Mdlman and
H J Sussmann, Eds. Dtfferenttal Geometrw ( 'ontrol The~
or':' (Blrkhauser, Basel-Boston, 1983) 268-2ct8
[6] V Jurdjevlc and J P Qumn, Controllabdlty and stabdlty,
J Dtf]erenttal Equatton~ 28 (1978) 381-389
[7] B Jakubczyk and B Respondek, On Ilneanzatlon of
control systems, Bull Acad Polonat~e Sol Ser Su Math
28 (19801 517-522
[8] J-P LaSalle, Stabdlty theory for ordinary differential
equahons, J Dtfferenttal Equattons 4 (1968) 57-65
[9] C Samson, Velocity and torque feedback control a non-
holonomlc cart, Internat Workshop m Adapttt'e and Non-
hnear Control Issue~ m Robottcs, Grenoble (Sprmger-
Verlag, Berhn-New York, 199(I)
[10] C. Samson, Time-varying stabdlzatlon of a nonholonomlc
car-hke mobde robot, Report INRIA Sophla-Antlpohs,
(19901
[11] R S6pulchre, Preprmt, Umv Cathohque de Louvam,
Belgmm (1991)
[12] E D Sontag, Feedback stabdlzatlon of nonhnear systems,
tn M A Kaashoek et al, Eds., Robust Control of Lmear
Systems and Nonhnear Control Proceedmgs of the Interna-
tional Symposium MTNS-8q (Blrkhauser, Basel-Boston,
19901 61-81