See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224241345

Integral Sliding Mode Control for Nonlinear Systems With Matched and
Unmatched Perturbations

Article in IEEE Transactions on Automatic Control · December 2011

DOI: 10.1109/TAC.2011.2159420 · Source: IEEE Xplore

CITATIONS READS
226 2,563

5 authors, including:

Matteo Rubagotti Antonio Estrada

Nazarbayev University Conacyt - Center for Engineering and Industrial Development
68 PUBLICATIONS 1,355 CITATIONS 42 PUBLICATIONS 775 CITATIONS

SEE PROFILE SEE PROFILE

Fernando Castanos Antonella Ferrara

Center for Research and Advanced Studies of the National Polytechnic Institute University of Pavia
70 PUBLICATIONS 1,651 CITATIONS 424 PUBLICATIONS 9,703 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Fault Tolerant Control for Switched Systems View project

HYCON2 View project

All content following this page was uploaded by Fernando Castanos on 27 May 2014.

The user has requested enhancement of the downloaded file.

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO. 11, NOVEMBER 2011 2699

Integral Sliding Mode Control for Nonlinear Systems bounded perturbations, it is natural to try to compensate such perturba-
With Matched and Unmatched Perturbations tions by means of an auxiliary control that retains the effect of the con-
troller designed for the unperturbed system. The sliding mode-based
Matteo Rubagotti, Member, IEEE, Antonio Estrada, auxiliary controller that compensates the perturbation from the very
Fernando Castaños, Member, IEEE, beginning of the control action, while retaining the order of the un-
Antonella Ferrara, Senior Member, IEEE, and compensated system, is the ISM controller. This technique has been
Leonid Fridman, Member, IEEE deeply studied in the last years (see, e.g., [5]–[12] and the references
therein).
Recently, the problem of analyzing how to minimize the disturbance
Abstract—We consider the problem of designing an integral sliding terms using ISM has been taken into account for systems with a non-
mode controller to reduce the disturbance terms that act on nonlinear linear drift term and a constant input matrix [13]. This result has been
systems with state-dependent drift and input matrix. The general case used also in connection with other control strategies, for example with
of both, matched and unmatched disturbances affecting the system is
model predictive control in [14], allowing the use of less conservative
addressed. It is proved that the definition of a suitable sliding manifold
and the generation of sliding modes upon it guarantees the minimization high level controllers. Nevertheless, a very important class of systems
of the effect of the disturbance terms, which takes place when the matched have a state-dependent input matrix, e.g., in mechanical systems the
disturbances are completely rejected and the unmatched ones are not am- control is premultiplied by the inverse of the inertia matrix.
plified. A simulation of the proposed technique, applied to a dynamically In this technical note, we consider the general class of nonlinear
feedback linearized unicycle, illustrates its effectiveness, even in presence
of nonholonomic constraints. control-affine systems with both, matched and unmatched perturbations,
and a state-dependent input matrix. Due to the appearance of partial
Index Terms—Disturbance reduction, integral sliding mode (ISM), derivatives in the state-dependent input matrix, the methodology
sliding mode control, uncertain systems.
of [13] cannot be directly applied. The contribution of this work
consists in the definition of an integral sliding manifold which leads
I. INTRODUCTION to the minimization of the effect of the disturbance terms also
in this case (provided that some integrability conditions are met).
Sliding mode control [1], [2] is a robust technique for the control Moreover, it is proved that in the particular case of systems in
of nonlinear systems. The most positive feature of sliding mode con- the so-called regular form, it is possible to use a constant matrix
trol consists in the complete compensation of the so-called matched to define the sliding manifold, thus simplifying the design phase.
disturbances (i.e., disturbances acting on the control input channel) Note that a preliminary version of the theoretical development of
when the system is in the sliding phase and a sliding mode is en- this technical note, where only the particular case of systems in
forced. This latter takes place when the state is on a suitable subspace regular form are considered, can be found in [15].
of the state space, called sliding manifold. The compensated dynamics The technical note is organized as follows: Section II introduces the
become insensitive to matched disturbances and uncertainties under considered class of systems and the control problem, while the pro-
sliding mode control. The price for this insensitivity is control chat- posed solution is analyzed in Section III. Simulation examples are re-
tering and a reaching phase, during which the system dynamics are ported in Section IV, and Section V draws the conclusions.
vulnerable to disturbances/uncertainties.
The integral sliding mode (ISM) technique was first proposed in
II. PROBLEM STATEMENT
[3], [4] as a solution to the reaching phase problem for systems with
matched disturbances only. The ISM control can also be regarded as a The system taken into consideration is of the form
way to combine the use of the sliding mode controller with that of an-
other controller (called high level controller in the following). The latter x_ (t) = f (x; t) + B (x)u(x; t) + (x; t) (1)
aims at stabilizing the nominal system. Systems compensated with this
type of controllers are of full order, i.e., of order equal to the order of where x 2 IRn is the state of the system with initial condition x(t0 ) =
the uncompensated system. When the system is subjected to external x0 , u 2 IRm is the control variable, (x; t) 2 IRn is an unknown
vector representing modeling uncertainties and external disturbances,
Manuscript received September 06, 2010; revised January 26, 2011 and May f 2 IRn is a known nonlinear function, and B 2 IRn2m is a known
23, 2011; accepted May 30, 2011. Date of publication June 13, 2011; date full rank state-dependent matrix.
of current version November 02, 2011. This work was supported in part by Assumption 1: The uncertain vector (x; t) is such that
Project FONCICYT (Fondo de Cooperatión International en Ciencia y Tec-
nología) 93302. Recommended by Associate Editor X. Chen.
M. Rubagotti is with the Department of Mechanical and Structural Engi- 1
(x; t) 2 8; 8 = fv 2 IRn s:t: kvk2  8sup g
neering, Università degli Studi di Trento, Trento 38100, Italy (e-mail: matteo.
[email protected]).
where 8sup is a constant scalar value.
A. Estrada is with the Department of Control and Robotics, Universidad Na-
cional Autónoma de México (UNAM), Mexico City 04510, Mexico (e-mail: The approach used in ISM control consists of splitting the control
[email protected]). variable into two parts
F. Castaños is with the McGill Center for Intelligent Machines, Montreal H3A
2A7, QC, Canada (e-mail: [email protected]).
A. Ferrara is with the Department of Computer Engineering and Systems u(x; t) = u0 (x; t) + u1 (x; t) (2)
Science, Università degli Studi di Pavia, Pavia 27100, Italy (e-mail: antonella.
[email protected]). where the term u0 (x; t) is generated by the high level controller
L. Fridman is with the Departamento de Control Automatico, CINVESTAV- (which can be designed according to any suitable design method),
IPN, Mexico City 07000, Mexico and also with the Department of Control and
Robotics, Universidad Nacional Autónoma de México (UNAM), Mexico City while u1 (x; t) is a discontinuous control action designed to reject the
04510, Mexico (e-mail: [email protected]). disturbance terms, forcing the system state on a suitably designed
Digital Object Identifier 10.1109/TAC.2011.2159420 sliding manifold s(x; t) = 0. In the following, the dependence of x

0018-9286/$26.00 © 2011 IEEE

2700 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO. 11, NOVEMBER 2011

and x_ on t is omitted in some cases, when it is obvious, for the sake of equal to zero, then determining the value of the equivalent control, and
simplicity. The proposed integral sliding manifold can be defined as finally substituting it into the state equations. In the present case, the
derivative of s(x; t) is
s(x; t) = g(x) 0 z (x; t) (3)
s_ (x; t) = G(x)x_ 0 z_ (x; t)
where g (x) : IR
n ! IR
m is a nonlinear function, the total derivative
= G(x) [f (x; t) + B (x) (u0 (x; t) + u1 (x; t)
of which is +
+B (x)(x; t)

g_ (x) = G(x)x_ (4) +U (x; t) 0 f (x; t) 0 B (x)u0 (x; t)]

+
= G(x)B (x) u1 (x; t) + B (x)(x; t)
with
+ G(x)U (x; t): (12)
@g(x) m2n
G(x) =
@x
2 IR (5) Then, the equivalent control, defined as the continuous control vari-
able such that s_ (x; t) = 0, results being
representing the Jacobian matrix, and the integral term z (x; t) is
t u1eq (x; t) = 0B + (x)(x; t) 0 (G(x)B (x))01 G(x)U (x; t): (13)
z (x; t) = g(x0 ) + G(x) (f (x;  ) + B (x)u0 (x;  )) d: (6) Substituting this value into the system (1), one has that the matched dis-
t turbance is eliminated, and the trajectories of the system at the sliding
manifold are given by
First of all, note that the system is in sliding mode at the initial time
instant, i.e. s(x0 ; t0 ) = 0. Also, note that the integral sliding manifold x_ eq = f (x; t) + B (x)u0 (x; t)
in (3) is analogous to those proposed in [6] and [13], with the main dif- + I 0 B (x) (G(x)B (x))
01 G(x) U (x; t):
ference that in our case the matrix G(x) is not imposed to be constant,
and is specifically designed to minimize the effect of the unmatched In conclusion, the action of the ISM control strategy has transformed
disturbances. The following assumption is required. the original uncertain term (x; t) into a new term
Assumption 2: G(x) is such that
1
eq (x; t) = I 0 B (x) (G(x)B (x))01 G(x) U (x; t): (14)
rank ( G(x)B (x)) = m; 8x 2 IR
n: (7)
An optimal choice of the state-dependent matrix G(x) would minimize
The control law is designed relying on the unit vector approach [2], this term. The goal of the reminder of this work is then to solve the
where one has following problem.
G(x)B (x))> s(x; t)
Problem 1: For system (1) fulfilling Assumptions 1 and 2, find a
u1 (x; t) = 0(x; t) function g 3 (x) such that
(
(8)
> s(x; t)
(G(x)B (x))
@g3 (x)
2
=
1
G3 (x) = arg min keq x; t)k2 :
( (15)
@x G(x)2IR
with  2 IR a gain that guarantees the enforcing of the state motion on
the sliding manifold, provided that Assumptions 1 and 2 hold.
In practice, the ideal aim of the ISM strategy would be to keep at III. PROPOSED SLIDING MANIFOLD
zero the difference between the nominal and perturbed evolutions of the
system. Taking into account that this could be done only when n = m, A. Case 1: System in General Form
in the general case m < n one has that only the projection of the
difference between the two state evolutions along the projection onto A general result is hereafter introduced for the minimization of
the span of the rows of G(x) can be kept equal to zero. the equivalent disturbance (14) for system (1), when the ISM control
The uncertain vector  for a system in form (1) can always be ex- strategy is applied. First of all, consider the distribution given by
pressed by separating the matched disturbance M from the unmatched
1( x) = span Bi? (x) ; i = 1; . . . ; n 0 m (16)
one U , as follows:

(x; t) = M (x; t) + U (x; t) (9) where Bi? stands for the i-th column of B ? . We introduce the fol-
1 lowing assumption.
M (x; t) = B (x)B+ (x)(x; t) (10) Assumption 3: 1(x) is involutive, i.e.
U (x; t) = B (x)B ?+ (x)(x; t)
1 ?
(11)
@Bj? (x) ?
Bi? (x); Bj? (x) = Bi (x)
where B ? (x) 2 IRn2(n0m) is a matrix with independent columns ?
@x
that span the null space of B (x), i.e. B ?> (x)B (x) = 0(n0m)2m , 0 @B@x
i x B? x 2
(
j
)
x); 8i; j ; ; n 0 m (17)
?
rank(B (x)) = n0m. Moreover, B (x) is the left pseudo-inverse of
+
( ) 1( = 1 ...

B (x), i.e., B (x) = (B (x)B (x)) 1 B > (x). Analogously, we have

+ 1 > 0 where [1; 1] is the Lie bracket of two vector fields.
B ?+ (x) = 1
(B
?> (x)B ? (x))01 B ?> (x). This separation principle Lemma 1: If Assumption 3 is fulfilled, there exists a function g~(x)
relies on Proposition 1 in [13], which ensures that In = B (x)B + (x) + such that
B ? (x)B ?+ (x) for any full rank B (x), being In 2 IRn2n an identity @ g~(x)
matrix. Thus, given the rank condition on B (x), the decomposition
1
= ~( G x) = M (x)B > (x) (18)
@x
(10), (11) is without loss of generality. To determine the state equations
when the state is confined to the sliding manifold, the equivalent control where M (x) 2 IRm2m is a full rank matrix. Note that (18) guarantees
method [16] is used. This consists in forcing the derivative of s(x; t) that Assumption 2 holds.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO. 11, NOVEMBER 2011 2701

Proof: According to Frobenius’ Theorem (see, e.g., [17]), the in- Assumption 4: System (1) is such that it can be written in the fol-
1( )
volutivity of x is equivalent to the existence of m independent func- lowing regular form:
~( )
tions gi x such that
@ g~ (x) ? x_ [1] (t) = f[1] (x; t) + [1] (x; t) (21)
i
B (x) = 0 81  i  m; 1  j  n 0 m x_ [2] (t) = f[2] (x; t) + B (x)u(x; t) + [2] (x; t)
@x j (22)
~(x)B? (x) = 0. Since the m columns of G~ > (x)
or, more compactly, G where x[1] 2 n0m , x[2] 2 m , f[1] f[2] >
IR IR [ f , B 2 m2m ] = IR
are independent, they span the orthogonal complement of 1(x). That is a full rank matrix, while [2] 2 m and [1] 2 n0m are the
IR IR
is matched and unmatched disturbances, respectively, clearly separable
?
span G~> (x) = span
i Bi? (x) : (19)
in this form.
Such a structure for the system is often found in the sliding mode
control literature, where it is widely used thanks to its nice properties
Recall that the double orthogonal complement of a closed subspace
(see, e.g., [1]). By virtue of Assumption 4, for the system that we are
is equal to the subspace itself [18, p. 118], so (19) is equivalent to
span ~ ( ) = span ( )
fG>i x g fBi x g. The columns of G> x and B x ~ () () considering, it is possible to simply get
are bases of the same subspace and the matrix M > x in (18) is simply
() M (x; t) = 0 ... 0 [2] . . . [2] >
the transformation matrix relating them.
Remark 1: The sufficiency part of the proof of Frobenius Theorem U (x; t) = [1] . . . [1] 0 ... 0 >:
is constructive, thus providing an explicit procedure for finding g x ~( ) In the following corollary it is shown that, for systems in regular form, it
[17, pp. 24–26].
The main result of the technical note is now formulated. is possible to use a simple linear sliding manifold in the ISM controller
Theorem 1: Consider system (1) fulfilling Assumptions 1 and 3. design.
~( )
Then, g x solves Problem 1. Moreover, the resulting equivalent dis- Corollary 1: For system (1) fulfilling Assumptions 1 and 4, Problem
1 can be solved by a linear function g x Gx. ~( ) = ~
turbance (14) is such that
Proof: If Assumption 4 holds, one has
kU (x; t)k2 = min keq (x; t)k2 :
B~ (x)
(20)
(x)2IR
B ? (x) =
G

Proof: First of all, note that Assumption 3 leads to the fulfillment

0m2(n0m)
of Lemma 1. Therefore, if G x ( ) = ~( ) = ( ) ( )
Gx M x B > x , it automat-
~ IR
()
ically follows that Assumption 2 is also fulfilled, because B x is full where B 2 (n0m)2(n0m) is a full rank matrix. If we consider the
rank. Then, if one sets g x( ) = ~( )
g x , it makes sense to face Problem 1 i-th and j -th columns of B ? x , it yields
()
@B ? (x) ? @B ? (x) ?
with the given assumptions. Analogously to [13], define
j
B (x) 0 B (x) = i

'(x; t) =
1
(G(x)B(x))01 G(x)U (x; t): @x i
@x j

Bi;?1
@B @B
@x
111 @x
Remarking that .. .. ..
..
. .
kU (x; t) 0 B (x)'(x; t)k2
. .
@B @B ? 0m
Bi;n
111
= I 0 B(x)(G(x)B(x))01 G(x))U (x; t) 2
@x
0 111
@x
0 0
.. .. ..
we reformulate Problem 1 as the problem of finding '3 x; t such that
( )
.. .
. . .
0 111 0 0
'3 (x; t) = arg min k (x; t) 0 B(x)'(x; t)k2 : Bj;?1
U @B @B
'(x) 2IR @x
111 @x 3
.. .. .. .. ..
According to the projection theorem in [18, p. 51], an explicit solution . . .
is '3 x; t
( )= ( ) ( ) ~( )
B + x U x; t . Substituting the value of G x , it yields 0
@B
.
111
@B
.
? 0m
Bj;n = 30 :
'(x; t) = (G(x)B (x))01 G(x) (x; t) U
@x
0 111
@x
0 0 ..
..
01 .. .. ..
= M (x)B> (x)B(x) M (x)B> (x) (x; t) . . .
0
. .
U
0 111 0 0
= B+ (x) (x; t) = '3 (x; t):
U
Since the result of any Lie bracket belongs to the span of the columns
According to Lemma 1, Assumption 3 implies the existence of g x ~( ) of B ? , the distribution 1( )
x is involutive, i.e., Assumption 3 is ful-
~( )
generating the Jacobian matrix G x . Finally, since given a matrix A 2 filled (and consequently Assumption 2, which again allows us to refer
n2m
IR , one has (see, e.g. [13]) kI 0 A A> A 01 A> k2
( ) =1 , using to Problem 1). Theorem 1 can thus be applied, which leads to the pos-
A B x one can see that kI 0 B x B > x B x 01 B > x k2
= () ( )( ( ) ( )) () = sibility of explicitly finding an integral sliding manifold which mini-
1 IR
; 8 x 2 n , which leads to (20). This implies that it is not possible to ~( )
mizes the equivalent disturbance. In particular, the function g x can be
obtain an equivalent disturbance with a 2-norm which is smaller than chosen such that its Jacobian matrix is G x ~( ) = ( ) ( )
M x B > x , choosing
the 2-norm of the unmatched disturbance. ()
M x as M x ( )=  ( )
N B 0> x , where N 2 m2m is any constant full
IR
rank matrix. It yields
B. Case 2: System in Regular Form
Hereafter, we focus on the task of finding a simple solution for G~ (x)= N B 0> (x)B > (x)= N B 0> [0 B > ]=[0 N ]= G:
~
()
choosing g x , when the system structure falls into a precise family,
as follows. Therefore, g x ~ = [0 N ]x.
~( ) = Gx
2702 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO. 11, NOVEMBER 2011

Remark 2: This result has a very intuitive meaning: like in the case with kp1 ; kd1 ; kp2 ; kd2 > 0. Note that this controller requires that the
analyzed in [13] for a constant value of B (x), it is clear that the sliding translational velocity of the robot never goes to zero (see [20] for a
manifold must be defined such that the ISM control action is not trying detailed analysis of this aspect).
to compensate the unmatched disturbance, because any attempt to do If the presence of disturbances is taken into account, the behavior of
it would increase the norm of the equivalent disturbance. In fact, the the unicycle can be quite different than expected, and ISM can be used
function g~(x) can be expressed as g~(x) = Nx[2] , which means that to reduce the disturbances. The following system is then considered:
the ISM control variable u1 (t) only acts on M (x; t). This result is
possible also in the state-dependent case because the matched distur-
bance and the control variable act on the same state components at any x_ (t) = (uv (t) +  (t)) cos x (t) 0  (t) sin x (t)
1 1 3 2 3

time instant: as a consequence, the unmatched uncertainties cannot, in x_ (t) = (uv (t) +  (t)) sin x (t) +  (t) cos x (t)
2 1 3 2 3
any past, present of future time instant, act on the same direction of the
matched one and this ensures a “separation” property which makes it
x_ (t) = u! (t) +  (t)
3 3

possible to use a simple sliding manifold to optimize the performance. where

 (t) cos x (t) 0  (t) sin x (t)
1 3 2 3
IV. A CASE STUDY: DYNAMIC FEEDBACK
(x; t) =  (t) sin x (t) +  (t) cos x (t)
1
1 3 2 3
LINEARIZATION OF THE UNICYCLE
 (t) 3

In this section the proposed method will be applied to the unicycle,

a very common example of nonholonomic system in mobile robotics is the disturbance vector. If we assume that each component of the
(the reader is referred to [19] for an overview on this kind of systems). vector  is in absolute value smaller than a constant (1 , 2 , 3 , re-
spectively), we obtain 8 = 1 + 2 + 3 (as required in Assump-
2 2 2
Without taking the disturbance terms into account, we introduce the
kinematic model of the nominal system as tion 1), while the control variables are defined as
x_ (t) = uv (t) cos x (t)
1 0 3 uv (t) = uv (t) + uv (t) 0 1

x_ (t) = uv (t) sin x (t)

2 0 3 u! (t) = u! (t) + u! (t) 0 1

x_ (t) = u! (t) uv 1 and u! 1 being the ISM contributions. Note that this system can
3 0

where (x1 ; x2 ) is the position of the robot in Cartesian coordinates in be written in form (1), with f (x; t) = 0. To apply the ISM control
the world reference frame, while x3 is its orientation with respect to the strategy, we must check if Assumption 3 is fulfilled. The distribution
1(x) is
x1 -axis; uv and u! represent the translational and rotational velocities,
respectively, which are regarded as inputs. Note that u0 = [uv0 u!0 ]> 0 sin x
1
3

is the high level control variable. Therefore, we are considering a third x

1( ) = span cos x
order system with state x = [x1 x2 x3 ]> and two inputs. A possible
3
1
0
high level controller for this system was introduced in [20] using dy-
namic feedback linearization. The structure of the controller is which is involutive since it is spanned by a single vector field. As a
consequence, all the assumptions are fulfilled, and the minimization of
_(t) =  (t) cos x (t) +  (t) sin x (t)
1 3 2 3 the disturbance terms can be performed. To define the sliding manifold
uv (t) = (t)
0
(3), take

u! (t) = 0 sin x (t)+(t) (t) cos x (t)

1 3 2 3

g(x) = x x 3

x x x
0

1 cos 3 + 2 sin 3

where  2 IR is the state of the dynamic compensator [20]. As for the

auxiliary control variables 1 and 2 , they can be defined according
with the corresponding

G(x) = x :
to the following considerations. It is possible to define a new set of 0 0 1
coordinates as cos x 3 sin x 0x3 1 sin x 3 + x
2 cos 3

z (t) = x (t)
1 1 Note that, as expected from Lemma 1, this latter can be written as
z (t) = x (t)
2 2

z (t) = x_ (t) = (t) cos x (t)

3 1 3
G(x) = 0x 1 sin x 3
1
x
+ 2 cos x 3
0
1 cos
0
x 3
0
sin x 3
1
0
z (t) = x_ (t) = (t) sin x (t)
4 2 3 Mx ( ) B x ( )

which leads to the possibility of representing the extended system with where M (x) is full rank for all x. The ISM control variable is
two chains of integrators z1 (t) = 1 (t), z2 (t) = 2 (t). If the objective then computed according to (8). As for the disturbance reduction,
is to follow a desired trajectory for x1 (t) and x2 (t), namely x1d (t) exploiting the definition of the unmatched disturbance in (11) (being
and x2d (t), it is possible to design a globally exponentially stabilizing B? = [0 sin x3 cos x3 0]> ), it can be computed as
0 sin x
feedback controller defining
2 3

 (t) = x d (t) + kp (x d (t) 0 x (t))

1 1 1 1 1
U =  cos x 2 3

+ kd (x_ d 0 x_ (t))
0
1 1 1 (23)
 (t) = x d (t) + kp (x d (t) 0 x (t))
2 2 2 2 2 leading to kU (x; t)k2 = 2 (t) and therefore kU k2  2 .
+ kd (x_ d 0 x_ (t))
2 2 2 (24) The 2-norm of the equivalent disturbance in (14) is obtained as
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO. 11, NOVEMBER 2011 2703

Fig. 1. Path of the unicycle in case of high level controller with no disturbances
(top), high level controller with disturbances (middle) and high level controller Fig. 2. Time evolution of (from top to bottom) of the control variables and the
plus ISM with disturbances (bottom). The reference trajectory is depicted as a components of the sliding manifold, for the simulation example regarding the
dashed line, while the actual ones are represented as solid lines. use of high level controller and ISM with disturbances.

keq (x)k2 = 2 (t). As expected, the maximum norm of the equiv- and u! , and the two components of the sliding manifold s, namely s1
alent disturbance coincides with that of the unmatched disturbance. and s2 . For the reader’s interest, a simulation example for systems in
In conclusion, the disturbance term is reduced to 8 = 2 , that is the regular form can be found in [15].
strongest disturbance reduction obtainable with ISM.
In the simulation example shown in the following, the distur-
bances are chosen as 1 (t) = 1:2 sin(5t), 2 (t) = 0:4 sin(20t), V. CONCLUSION
3 (t) = 0:8 sin(t), leading to 8 ' 1:5. The high level controller This technical note introduces the definition of an integral sliding
in (23), (24) is designed with kp1 = kp2 = 15, kd1 = kd2 = 1, manifold for control-affine nonlinear systems. Two cases are consid-
while the ISM control law (8) is defined with a constant gain value ered. For the general case it is shown that a solution for the minimiza-
2 2
 = 1 + 3 ' 1:45, in order to compensate the matched distur- tion of the disturbance terms (i.e., the matched disturbances are elim-
bance M = [1 cos x3 1 sin x3 3 ]> . Moreover, in order to reduce inated and the unmatched ones are not amplified) can be obtained if
the so-called chattering effect, the well known equivalent control some involutivity properties of the system are fulfilled. For systems in
method [1] is used, applying a linear low-pass filter to the obtained regular form, a linear sliding surface can be exploited, obtaining anal-
discontinuous control variable. First of all, we show (Fig. 1, top) the ogous results. The proposed ISM control law is finally tested on a sim-
path of the unicycle in the x1 0 x2 plane in case there is no disturbance ulation example of a simple nonholonomic system.
and the high level controller only is used. As expected, after a transient
(since the initial condition is taken on purpose different from the
reference), the unicycle trajectory (solid line) settles on the desired one REFERENCES
(dashed line). If the disturbances are added, the high level controller [1] V. I. Utkin, J. Guldner, and J. Shi, Sliding Mode Control in Electro-
has a poor performance (Fig. 1, middle), since it is not designed to mechanical Systems. London, U.K.: Taylor & Francis, 1999.
work in their presence. Using the proposed ISM strategy, the bound [2] C. Edwards and S. Spurgeon, Sliding Mode Control: Theory and Ap-
plications. London, U.K.: Taylor & Francis, 1998.
on the disturbances is reduced to 8 = 0:4, and the performance of [3] G. P. Matthews and R. A. DeCarlo, “Decentralized tracking for a class
the overall control law is improving (Fig. 1, bottom). In this last case, of interconnected nonlinear systems using variable structure control,”
we show also (Fig. 2) the time evolution of the control variables uv Automatica, vol. 24, pp. 187–193, 1988.
2704 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 56, NO. 11, NOVEMBER 2011

[4] V. Utkin and J. Shi, “Integral sliding mode in systems operating under Stochastic Receding Horizon Control With Bounded
uncertainty conditions,” in Proc. 35th IEEE Conf. Decision Control, Control Inputs: A Vector Space Approach
Kobe, Japan, Dec. 1996, pp. 4591–4596.
[5] A. Poznyak, Y. B. Shtessel, and C. Jiménez, “Min-max sliding mode
Debasish Chatterjee, Peter Hokayem, and John Lygeros
control for multimodel linear time varing systems,” IEEE Trans.
Autom. Control, vol. 48, no. 12, pp. 2141–2150, Dec. 2003.
[6] W. J. Cao and J. X. Xu, “Nonlinear integral-type sliding surface for
both matched and unmatched uncertain systems,” IEEE Trans. Autom.
Control, vol. 49, no. 8, pp. 1355–1360, Aug. 2004. Abstract—We design receding horizon control strategies for stochastic
[7] M. Basin, J. Rodriguez, and L. Fridman, “Optimal and robust control discrete-time linear systems with additive (possibly) unbounded distur-
for linear state-delay systems,” J. Franklin Inst., vol. 344, no. 6, pp. bances while satisfying hard bounds on the control actions. We pose the
801–928, 2007. problem of selecting an appropriate optimal controller on vector spaces
[8] J. X. Xu, Y. J. Pan, T. H. Lee, and L. Fridman, “On nonlinear H-infinity of functions and show that the resulting optimization problem has a
sliding mode control for a class of nonlinear cascade systems,” Int. J. tractable convex solution. Under marginal stability of the zero-control
Syst. Sci., vol. 36, no. 15, pp. 983–992, 2005. and zero-noise system we synthesize receding horizon polices that ensure
[9] Y. Niu, W. C. Ho, and J. Lam, “Robust integral sliding mode control bounded variance of the states while enforcing hard bounds on the con-
for uncertain stochastic systems with time-varying delay,” Automatica, trols. We provide examples that illustrate the effectiveness of our control
vol. 41, no. 5, pp. 873–880, 2005. strategies, and how quantities needed in the formulation of the resulting
[10] J. X. Xu and K. Abidi, “Discrete-time output integral sliding mode optimization problems can be calculated off-line.
control for a piezo-motor driven linear motion stage,” IEEE Trans. Ind.
Electron., vol. 55, no. 11, pp. 3917–3926, Nov. 2008. Index Terms—Constrained control, receding horizon control, stochastic
[11] Y. B. Shtessel, P. Kaveh, and A. Ashrafi, “Integral and second order control.
sliding mode control of harmonic oscillator,” in Proc. 44th IEEE Conf.
Decision Control Eur. Control Conf., Seville, Spain, Dec. 2005, pp.
3941–3946.
[12] M. Defoort, A. Kokosy, T. Floquet, and W. Perruquetti, “Integral I. INTRODUCTION
sliding mode control for trajectory tracking of a unicycle type mobile
robot,” Integr. Comp.-Aided Eng., Inform. Control, Autom. Robot., Receding horizon control is a popular paradigm for designing con-
vol. 13, no. 3, pp. 277–288, 2006. trol policies. In the context of deterministic systems it has received a
[13] F. Castaños and L. Fridman, “Analysis and design of integral sliding
manifolds for systems with unmatched perturbations,” IEEE Trans. considerable amount of attention over the last two decades, and signif-
Autom. Control, vol. 51, no. 5, pp. 853–858, May 2006. icant advancements have been made in terms of its theoretical founda-
[14] M. Rubagotti, D. M. Raimondo, A. Ferrara, and L. Magni, “Robust tions as well as industrial applications. The motivation comes primarily
model predictive control with integral sliding mode in continuous-time from the fact that receding horizon control yields tractable control laws
sampled-data nonlinear systems,” IEEE Trans. Autom. Control, vol. 56,
no. 3, pp. 556–570, Mar. 2011. for deterministic systems in the presence of constraints, and has conse-
[15] M. Rubagotti, A. Estrada, F. Castaños, A. Ferrara, and L. Fridman, quently become popular in the industry. The counterpart in the context
“Optimal disturbance rejection via integral sliding mode control for of stochastic systems, however, is a relatively recent development. In
uncertain systems in regular form,” in Proc. Int. Workshop Variable this article we solve the problem of stochastic receding horizon con-
Structure Syst., Mexico City, Mexico, Jun. 2010, pp. 78–82. trol for linear systems subject to additive (possibly) unbounded distur-
[16] V. I. Utkin, Sliding Mode in Control and Optimization. Berlin, Ger-
many: Springer-Verlag, 1992. bances and hard norm bounds on the controls, over a class of feedback
[17] A. Isidori, Nonlinear Control Systems. London, U.K.: Springer- policies. Methods for guaranteeing hard bounds on the controls, within
Verlag, 1996. our context, while ensuring tractability of the underlying optimization
[18] D. G. Luenberger, Optimization by Vector Space. New York: Wiley, problem are, to our knowledge, not available in the current literature.
1969.
[19] A. Bloch and J. Baillieul, Nonholonomic Mechanics and Control. Preliminary results in this direction were reported in [1].
New York: Springer-Verlag, 2003. In the deterministic setting, the receding horizon control scheme is
[20] A. De Luca, G. Oriolo, and M. Vendittelli, “Stabilization of the unicycle dominated by worst-case analysis relying on robust control and ro-
via dynamic feedback linearization,” in Proc. 6th IFAC Symp. Robot bust optimization methods, see, for example, [2]–[9] and the refer-
Control, Vienna, Austria, Sep. 2000, pp. 397–402. ences therein. The central idea is to synthesize a controller based on the
bounds of the noise such that a certain target set becomes invariant with
respect to the closed-loop dynamics. However, such an approach tends
to yield rather conservative controllers and large infeasibility regions.
Moreover, assigning an a priori bound to the noise seems to demand
considerable insight. A stochastic model of the noise is a natural al-
ternative approach to this problem: the conservativeness of worst-case
controllers may be reduced, and one may not need to impose any a

Manuscript received March 31, 2009; revised October 15, 2009; accepted
May 30, 2011. Date of publication June 13, 2011; date of current version
November 02, 2011. This work was supported in part by the Swiss Na-
tional Science Foundation under grant 200021-122072, and FeedNetBack
FP7-ICT-223866. Recommended by Associate Editor C. Szepesvari.
D. Chatterjee is with the Systems and Control Engineering Department, In-
dian Institute of Technology, Bombay, India (e-mail: [email protected]).
P. Hokayem and J. Lygeros are with the Automatic Control Laboratory,
ETH Zürich, Zürich 8092, Switzerland (e-mail: [email protected];
[email protected]).
Color versions of one or more of the figures in this technical note are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TAC.2011.2159422

0018-9286/$26.00 © 2011 IEEE

View publication stats