IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 62, NO.

6, JUNE 2017 2965

Sliding Mode Control of Constrained Nonlinear Systems

Gian Paolo Incremona, Matteo Rubagotti, and Antonella Ferrara

Abstract—This technical note introduces the design of sliding finite-time interval if the control action is large enough to coun-
mode control algorithms for nonlinear systems in the presence teract the effect of the uncertain terms. After reaching the sliding
of hard inequality constraints on both control and state variables.
Relying on general results on minimum-time higher-order sliding manifold, the evolution of the state variables is insensitive to the
mode for unconstrained systems, a general order control law is for- so-called matched disturbances, i.e., those acting in the same chan-
mulated to robustly steer the state to the origin, while satisfying all nel as the control variable [4], [5]. The main drawback of SMC is
the imposed constraints. Results on minimum-time convergence to the so-called chattering [6], [7], which is the high-frequency oscilla-
the sliding manifold, as well as on the maximization of the domain
of attraction, are analytically proved for the first-order and second-
tory motion around the sliding manifold due to the discontinuity of
order sliding mode cases. A general result is presented regarding the control law. Higher order sliding mode (HOSM) is a possible
the domain of attraction in the general order case, while numer- solution for chattering reduction (see, e.g., [8]–[17], and the ref-
ical results on the estimation of the domain of attraction and on erences therein included). In particular, [16] proposes an algorithm that
minimum-time convergence are discussed for the third-order case, guarantees a time-optimal reaching of the sliding manifold for arbitrary
following a procedure applicable to a sliding mode of any order.
order (i.e., dimension of the sliding manifold, see, e.g., [18]). Note that
Index Terms—Constrained control, higher order sliding mode chattering reduction is not the only possible purpose of HOSM, since
(HOSM), second-order sliding mode, sliding mode control, uncer- this latter allows the use of a relative degree greater than one between
tain systems.
the discontinuous control input and the sliding variable.
In the first-order SMC formulation, input saturations are immediately
satisfied if the control variable switches between values that are inside
I. INTRODUCTION
the imposed boundaries. When HOSM is used for chattering reduction,
Input constraints are present in all practical control implementations, specific solutions have been proposed to achieve the satisfaction of
mainly in the form of saturations. In order to guarantee an acceptable saturation constraints (see, e.g., [19], [20] for the second-order case).
performance of the controlled system in their presence, different On the other hand, the possible presence of state constraints is usually
solutions have been proposed, mainly in the field of anti-windup not taken into account in SMC formulations. Recently, few solutions
control (see, e.g., [1], [2]). In addition, in order to avoid failures or have been proposed in order to merge SMC and MPC, and combine
critical conditions of the controlled system, certain regions of the the constraints satisfaction property of MPC with the robustness of
state space need to be avoided during the execution of many tasks. SMC [21]–[25]. As an alternative, in order to avoid the additional
In some cases, a conservative tuning of the control laws can lead computational burden of MPC, the presence of state constraints has
to the avoidance of these regions. On the other hand, control laws been directly inserted in the SMC law in [26]–[28] for the first-order
that directly consider the presence of state constraints can reduce sliding mode case, in [29]–[32] for the second-order sliding mode case,
conservativity and improve the overall system performance. The most and in [33] for third-order sliding mode with box constraints.
well-known control methodology able to manage both input and state In this technical note, an HOSM control law of general order r is
constraints is model predictive control (MPC), for which the reader is proposed, aimed at guaranteeing the minimum-time convergence of the
referred to [3] and the references therein. state onto the sliding manifold, and, at the same time, satisfying the im-
A control technique that naturally handles the presence of some posed hard inequality constraints on input and states, in the presence of
classes of uncertainties and disturbances is sliding mode control matched disturbances. More specifically, the main contributions of the
(SMC), in which a discontinuous control law steers the state onto a present work are the following: first of all, the formulation of a control
suitably-defined hyper-surface (the so-called sliding manifold), and, law capable of solving an r-th order SMC problem for uncertain non-
under suitable design conditions, makes the origin of the state space linear affine systems with inequality constraints on both input and state
an asymptotically stable equilibrium point for the closed-loop sys- variables (note that this was an open problem, to the best of the authors’
tem. The convergence to the sliding manifold is guaranteed in a knowledge); second, a procedure to select the sliding variable in order
to satisfy the constraints, including two different sufficient conditions
which provide guidance in the choice of the sliding variable; third, the
proof of the minimum-time convergence with constraint satisfaction
Manuscript received March 7, 2016; revised June 20, 2016; accepted
and maximization of the domain of attraction for the first and second-
August 22, 2016. Date of publication August 31, 2016; date of current
version May 25, 2017. Recommended by Associate Editor X. Yu. order cases (a preliminary result on this aspect, limited to the second-
G. P. Incremona is with the Dipartimento di Ingegneria Civile ed Ar- order case, was presented in [30], without proof of the minimum-time
chitettura, University of Pavia, 27100, Pavia, Italy (e-mail: gianpaolo. convergence); fourth, a general result on the domain of attraction for the
[email protected]). r-th order case, along with the specific numerical study of the domain
M. Rubagotti is with the Department of Engineering, University of Le-
icester, Leicester LE1 7RH, United Kingdom (e-mail: [email protected]). of attraction and of the minimum-time convergence for the third-order
A. Ferrara is with the Dipartimento di Ingegneria Industriale e case. Note that preliminary results on the numerical evaluation of the
dell’Informazione, University of Pavia, 27100, Pavia, Italy (e-mail: domain of attraction for a third-order sliding mode controller (only
[email protected]). for the case of box state constraints) were described in [33], and
Color versions of one or more of the figures in this paper are available
can be considered now as particular cases of the proposed general
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TAC.2016.2605043 formulation.

0018-9286 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Nazarbayev University. Downloaded on February 18,2025 at 09:33:07 UTC from IEEE Xplore. Restrictions apply.
2966 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 62, NO. 6, JUNE 2017

II. PROBLEM FORMULATION on its value, for each specific application, to the designer of the
control law. 
In this technical note, we consider a class of uncertain nonlinear
Remark 2: The analysis of chattering is outside the scope of this
dynamical systems, defined by
paper. Yet, it is an interesting topic that the reader may deepen making
ẋ(t) = φ(x(t), t) + γ(x(t), t)u(t) (1) reference to [6], [34]–[37]. Indeed, in systems with chains of integrators
and certain dynamics of lag type in series, chattering may exist or not
where x ∈ Rn is the state vector, u ∈ R is the control input, φ : depending on the type of the discontinuous control law (ideal relay or
Rn + 1 → Rn and γ : Rn + 1 → Rn are uncertain smooth vector fields, relay with hysteresis) and the effect of parasitic dynamics. 
and the whole state vector is assumed to be available for feedback. The In the particular case when u is directly defined as the discontinuous
control objective is the regulation of the state x to the origin. Differ- control variable (i.e., m = 0 and therefore w = u) the constraints in
ently from classical SMC formulations, a set of (possibly mixed) hard (2) are required to be formulated as disjoint input and state constraints,
input and state inequality constraints is introduced. More specifically, and precisely
it is required that
  X̄ = {(x, u) ∈ Rn × R : x ∈ X , u ∈ [−α, α]} (7)
x̄ ∈ X̄ , x̄  x u ∈ Rn + 1 (2)
where X is, in general, a closed set, while α > 0 is the same fixed
where  denotes the transposition operator. X̄ ⊆ Rn + 1 is a closed (pos- parameter as in (4), which defines the maximum amplitude of the
sibly unbounded) set that includes the origin in its interior, and is control variable. In order to consider (5) as the general formulation, the
formulated as a system of (possibly nonlinear) algebraic equations. case m = 0 can be formulated as a particular case of (5), with xa = x,
Φ(·, ·) = φ(·, ·), Γ(·, ·) = γ(·, ·), and Xa = X .
III. THE PROPOSED HOSM CONTROL LAW
B. Definition of the Sliding Manifold
A. Definition of the Augmented System
After defining the augmented system (5) and the related constraint
The control law can either be defined as a discontinuous control sets in (6), the next step consists of defining a suitable output variable
law (classical SMC), or as the result of an m-fold time integration of σ1 (t) ∈ R, as a (in general, nonlinear) function of xa . Following the
a discontinuous signal w(t) (HOSM aimed at chattering reduction). standard design procedure of HOSM control [10], a vector of time
The case m > 0, m being an integer number, is considered first. An derivatives of σ1 is defined as
integrator-chain dynamics is added to the system, starting from xn + 1 
   
u, as σ = σ1 σ2 . . . σr  σ1 σ̇1 . . . σ1(r −1 ) ∈ Rr (8)
⎧
⎪
⎪ ẋn + 1 (t) = xn + 2 (t) where r ∈ [0 , n + m] is the well defined, uniform and time-invariant
⎪
⎪
⎪
⎪ relative degree of the system (assuming w as input and σ1 as output).
⎪
⎨ẋn + 2 (t) = xn + 3 (t)
(3) In the (n + m)-dimensional space defined by the components of xa ,
⎪
⎪
. the manifold Σ  {xa : σ(xa ) = 0} is referred to as sliding manifold.
⎪ ..
⎪
⎪
⎪ The control variable will be defined in order to ensure the finite-time
⎪
⎩ẋ
n+m (t) = w(t). convergence of xa on the sliding manifold, which (assuming a correct
definition of σ) will in turn imply the asymptotic convergence of xa
Note that the chain of integrators will be an element of the closed loop to zero. With reference to [38, Theorem 13.1], a diffeomorphism Ω :
control system for which the stability results proved in Sections IV and Rn + m → Rn + m −r × Rr is defined, such that
V hold.
In order to provide bounds on the derivatives of the actual control (ζ, σ) = Ω(xa ) (9)
variable u, and recalling that the constraints on u = xn + 1 are already n + m −r
where ζ ∈ R is the internal state vector. The diffeomorphism
imposed in (2), a set of box constraints is defined, as
allows one to transform system (5) into the normal form
w ∈ [−α, α], xn + i ∈ [−βi , βi ], i = 2, . . . , m (4) ζ̇(t) = ψ(ζ(t), σ(t), t) (10a)
α, βi > 0 being fixed parameters. The overall augmented system dy- σ̇i (t) = σi + 1 (t), i = 1, . . . , r − 1 (10b)

namics, with state xa  [ x1 . . . xn + m ] ∈ Rn + m is now repre-
sented by σ̇r (t) = f (ζ(t), σ(t), t) + g(ζ(t), σ(t), t)w(t) (10c)

ẋa (t) = Φ(xa (t), t) + Γ(xa (t), t)w(t) (5) with ψ : Rn + m + 1 → Rn + m −r , f : Rn + m + 1 → R, g : Rn + m + 1

→ R. The dynamics of system (5) includes uncertain terms, as specified
where Φ : Rn + m + 1 → Rn + m and Γ : Rn + m + 1 → Rn + m are when defining the vector fields φ and γ in (1). As a consequence, ψ, f
smooth vector fields immediately obtainable from (1) and (3). For and g in (10) also contain uncertain terms, but some information about
the sake of compactness, the sets of disjoint constraints in (2) and (4) them is available, as stated in the following assumption.
can be merged as Assumption 1: There exist positive constants F, G1 , G2 , such that

xa ∈ Xa , w ∈ [−α, α], (6) |f (ζ(t), σ(t), t)| ≤ F, (11)

with implicit definition of Xa . 0 < G1 ≤ g(ζ(t), σ(t), t) ≤ G2 , (12)

Remark 1: As m increases, the smoothness of the control signal u
αr  G1 α − F > 0. (13)
also increases [12], together with the number of states of system (5),
with a consequent increase in the complexity of the controller. On the Moreover, the following two hypotheses are valid for the internal dy-
other hand, the chattering amplitude does not necessarily decrease as namics (10a). Firstly, for all initial conditions (ζ(0), σ(0)) ∈ Rn + m ,
m increases [34]. Thus, there is not a “best value” for m in general: we for all realizations of the uncertain terms satisfying (11), (12), the in-
propose an approach valid for any value of m, and leave the decision ternal dynamics (10a) does not present finite time escape phenomena.

Authorized licensed use limited to: Nazarbayev University. Downloaded on February 18,2025 at 09:33:07 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 62, NO. 6, JUNE 2017 2967

Secondly, the zero dynamics with

x1 − x2 σ1 + sin−1 (σ2 )
ζ̇(t) = ψ(ζ(t), 0, t) (14) σ = Ω(xa ) = xa = Ω−1 (σ) = .
sin(x2 (t)) sin−1 (σ2 )
is globally asymptotically stable. 
These are typical assumptions in SMC: (11), (12) require the bound- The constraint set is defined as S = {(σ1 , σ2 ) : |σ1 + sin−1
edness of the uncertain terms; (13) ensures that the control amplitude (σ2 )| ≤ 1, |σ2 | ≤ sin(1)}, which is a function of σ only. 
is large enough to counteract their effect; the assumptions on the in- Lemma 2: Assume that the state vector xa ∈ Rn + m of the aug-
ternal dynamics and on the zero dynamics (14) imply that, during the mented system is composed of two subvectors, as xa = [ xb xc ] ,
reaching phase the internal states remain bounded, while, once σ has with xb ∈ Rn + m −p , xc ∈ Rp , p ∈ {1, 2, . . . , n − m}. As a conse-
been steered to zero, ζ will also converge to zero asymptotically. By quence, (5) takes the form
means of the diffeomorphism Ω(xa ) in (9), the set Xa can be in general
ẋb (t) = Φb (xb (t), xc (t), t) (16)
mapped into a new set S, i.e.,
ẋc (t) = Φc (xc (t), t) + Γc (xc (t), t)w(t) (17)
xa ∈ Xa ⇐⇒ (ζ, σ) ∈ S (15)
with implicit definition of Φb , Φc , and Γc . In addition, assume that
which will be used to enforce the constraints in the coordinate system the constraint set Xa is a function of xc only. Then, if σ = σ(xc ) with
defined by (10). The presence of these constraints requires an additional r = p, Assumption 2 is satisfied. 
hypothesis. Proof: Since r = p, the same development of the proof of Lemma
Assumption 2: The diffeomorphism Ω(xa ) is defined such that S = 1 can be followed, for subsystem (17). Ω−1 (σ) will map σ to xc , and
S(σ) ⊂ Rr , i.e., the expression of the constraint set does not depend the evolution of the internal state variables ζ will not influence the
on the internal state ζ.  dynamics of σ. Again, by substituting the expression of σ into the
Assumption 2 is required by the fact that, in SMC, the con- expression of Xa , the set S will be obtained as a function of σ only. 
trol law is defined as a function of σ. As a consequence, it Example 2: Consider as augmented system
would not be possible to enforce constraint satisfaction on the
ẋ1 (t) = −x1 (t)e−x 1 (t ) + x52 (t)
internal state ζ during the reaching phase. For this reason, the
constraints have to be expressed in the r-dimensional space defined ẋ2 (t) = −x32 (t) + w(t)
by σ.
An explicit expression for Ω(xa ) is not available in gen- with constraint set Xa = {(x1 , x2 ) : x2 ∈ [−1, 3]}. In this case,
eral. However, in the following we provide two different suffi- the state vector can be divided into two scalar components
cient conditions, for which Ω(xa ) can be immediately defined. xb = x1 and xc = x2 . Choosing ζ = x1 and σ1 = x2 , we obtain a
We also show an example in which neither of these conditions is satis- scalar σ ∈ R, and the dynamics of the system in form (10) is described
fied, but nonetheless Ω(xa ) can be easily found. For the general case, by
the solution to this problem has to be considered by the designer of the
ζ̇(t) = −ζ(t)e−ζ (t ) + σ15 (t)
control system as a specific step of the design process.
σ̇1 (t) = −σ13 (t) + w(t)
C. Design of the Sliding Manifold with
The definition of Ω(xa ) (or, equivalently, of σ1 ) such that σ satisfies ζ x1 ζ
Assumption 2 requires, in general, some attention. Guidelines are pro- = Ω(xa ) = , xa = Ω−1 (ζ, σ) = .
vided in the following. Let us start by describing two particular cases σ1 x2 σ1
in which Assumption 2 can be easily satisfied.
The constraint set is defined as S = {σ1 : σ1 ∈ [−1, 3]}, which is a
Lemma 1: Consider a given augmented dynamics (5) with associ-
function of σ only. 
ated constraints (6). If σ is defined such that r = n + m, then Assump-
In the general case, Assumption 2 is not straightforwardly verified
tion 2 is satisfied. 
so that a careful choice of σ1 must be made for each case. A simple
Proof: Since r = n + m, Ω(xa ) defines a mapping from xa to
example is shown next.
σ, and no internal state ζ exists. Therefore, each element of xa can
Example 3: The augmented system dynamics is given by
be expressed as a function of the elements of σ. Substituting the cor-
responding expressions of σ in the formulation of Xa , the set S is ẋ1 (t) = −x1 (t) + x2 (t) + w(t)
obtained, which is a function of σ only. 
Example 1: Consider the augmented system given by ẋ2 (t) = −x2 (t) + w(t)

with associated constraint set Xa = {(x1 , x2 ) : |x1 + x2 | ≤ 1}. Since

ẋ1 (t) = sin(x2 (t)) + u(t)
Xa depends on both state variables, it is not possible to use the result
ẋ2 (t) = u(t) in Lemma 2. The constraint set S has to be a function of σ only, so we
choose σ1 = x1 + x2 . The internal state can be chosen as ζ = x1 − x2 .
with constraint set Xa = {(x1 , x2 ) : |x1 | ≤ 1, |x2 | ≤ 1}. Choosing As in the previous example, we obtain a scalar σ ∈ R, and the dynamics
σ1 = x1 − x2 , we obtain a vector σ ∈ R2 , the dynamics of which is of the system in form (10) is
defined by
3 1
ζ̇(t) = − ζ(t) + σ1 (t)
2 2
σ̇1 (t) = σ2 (t)
1 1
σ̇2 (t) = cos(x2 (t))w(t) σ̇1 (t) = ζ(t) + σ1 (t) + 2w(t)
2 2

Authorized licensed use limited to: Nazarbayev University. Downloaded on February 18,2025 at 09:33:07 UTC from IEEE Xplore. Restrictions apply.
2968 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 62, NO. 6, JUNE 2017

with following. The first is the determination of the domain of attraction SI

ζ x1 − x2 of the origin for the closed-loop system, i.e., all the initial conditions
1 ζ + σ1
= Ω(xa ) = , xa = Ω−1 (ζ, σ) = . σ(0) for which σ(t) converges to the origin in a finite time without
σ1 x1 + x2 2 ζ − σ1 violating the constraints. The second is the problem of determining if
The constraint set is defined as S = {σ1 : |σ1 | ≤ 1}, which is a func- the obtained domain of attraction can be enlarged by using a different
tion of σ only.  control law w(t), with the same system dynamics, the same bounds on
the uncertain terms and on the control amplitude, and the same state
D. Definition of the Control Law constraints. The third issue is related to investigating if w(σ) in (22)
still leads to minimum-time convergence to σ = 0 for the constrained
The control law w(t) has to be defined as a discontinuous function case, as it was in [16] for the unconstrained case. In the next section, the
of σ. In [16, Sec. III.C], an HOSM control law for a system in form cases r = 1 and r = 2 with box constraints on σ will be considered.
(10b),(10c) was proposed (i.e., for arbitrary value of r), which guar- For such cases, formal results will be presented. The general case will
anteed the reaching of the sliding manifold in minimum time, for the be tackled in Section V.
worst-case realization of the uncertain terms (i.e., f (ζ(t), σ(t), t) ≡ Remark 3: In many practical applications, SMC laws are im-
−F · sgn(w(t)) and g(ζ(t), σ(t), t) ≡ G1 ). In this worst-case formu- plemented using a finite sampling time, or one of the so-called
lation, αr in (13) can be interpreted as a reduced control amplitude pseudo-sliding techniques (typically, employing a saturation function
(however, notice that αr and α have different units of measure). We instead of the sign function, or low-pass filtering the discontinuous con-
denote the control law proposed in [16], which did not take into ac- trol signal). All of these techniques typically lead to the convergence to
count the presence of constraints, as wi n (σ)  (−1)r + 1 · α · c(σ), a boundary layer of the sliding manifold, rather than on the sliding man-
where function c(σ) : R→ {−1, +1} is determined as reported in [16]. ifold itself [41]. If such techniques are directly applied to approximate
In general, one obtains an (r − 1)-dimensional manifold in the r- the proposed control law, the imposed constraints can be slightly vio-
dimensional space generated by vector σ, described by the nonlinear lated, as the state evolves on a boundary layer around them. The typical
equation s(σ) = 0. This manifold, which includes the origin, divides solution consists in defining more conservative constraint set S̃ ⊂ S,
the σ-space into two half-spaces. By imposing a constant value of c so as to prevent the original constraints from being violated. 
(either −1 or +1) in each half-space, the control law can be written as
wi n (σ) = −α · sgn(s). (18) IV. ANALYTICAL RESULTS ON FIRST AND SECOND-ORDER SLIDING
MODES WITH BOX CONSTRAINTS
As mentioned in [16], the complexity of the expression of s(σ) grows
very fast with the order r, and the derivation of efficient methods In this section, the convergence in a finite (minimum) time of the slid-
(numerical or exact-algebraic) in order to define it, is still an open ing variable associated with the proposed closed-loop control system is
problem. A method for determining the expression of s(σ) for generic discussed. Note that this result directly implies asymptotic stability of
order is proposed in [39]. For relatively low-dimensional cases, the the origin of the closed-loop system since, by assumption, the zero dy-
analytical expression of s(σ) is given by namics (14) of system (1) transformed via the diffeomorphism Ω(xa )
is globally asymptotically stable.
σ ∈ R1 ⇒ s(σ) = σ1 (19)
σ2 |σ2 | A. First-Order Sliding Mode Control
σ ∈ R2 ⇒ s(σ) = σ1 + (20)
2αr
The formulation of the control law (22) for r = 1, using (19), be-
σ3 σ 2 sgn(σ3 ) comes
σ ∈ R ⇒ s(σ) = σ1 + 32 + sgn σ2 + 3
3
3αr 2αr
w(σ) = −α sgn(σ1 ) (23)
3
1 σ 2 sgn(σ3 ) σ2 2 σ2 σ3
× √ sgn σ2 + 3 σ2 + 3 + . which coincides with the “unconstrained” control law wi n in (18). The
αr 2αr 2αr αr
constraint set in this simple case can only take the form
(21)
S = {σ1 : σ1 ∈ [σ 1 , σ 1 ]}. (24)
To be precise, in the problem defined in [16], the expression of wi n is
found also in the different sub-cases when σ = 0, which would make with σ 1 < 0 and σ 1 > 0. Then, the following result can be proved.
the formulation of the control law more complicated. However, as Theorem 1: Given system (10) with r = 1, assuming that Assump-
pointed out also in [16], being s(σ) a null-measure set, the definition tion 1 holds, if the control law (23) is applied, then σ1 is steered to
of the control law for σ = 0 has no practical relevance. For this reason, the origin with domain of attraction SI coinciding with S in (24).
the value of wi n (σ) when s(σ) = 0 will not be explicitly defined, and, The set SI is the maximum obtainable domain of attraction for the
when needed in the proofs, will be determined as the Filippov solution given set of constraints, and the convergence takes place in minimum
of the given discontinuous vector field at σ (see [40]). time for the worst possible realization of the disturbance terms (i.e.,
In order to take the presence of constraints into account, the control f (ζ(t), σ1 (t), t) ≡ −F · sgn(w(t)) and g(ζ(t), σ1 (t), t) ≡ G1 ). 
law proposed in this technical note coincides with wi n (σ) when σ ∈ S, Proof: See Appendix A. 
while for σ ∈/ S is defined as wo u t (σ)  −α · sgn(σr ). Overall
⎧
⎨−α · sgn(s(σ)), σ ∈ S B. Second-Order Sliding Mode Control
w(σ) = (22)
⎩−α · sgn(σr ), σ∈ /S In the case r = 2, the second-order SMC law, according to (20) and
(22), is defined as
which is an extremely simple law, yet capable of solving a rather ⎧  
complex nonlinear constrained control problem for uncertain systems. ⎨−α sgn σ1 + σ 2 |σ 2 | if (σ1 , σ2 ) ∈ S
2α r
After the introduction of the proposed control law (22), three in- w(σ) = (25)
⎩−α sgn(σ ) if (σ1 , σ2 ) ∈
/S
2
teresting issues regarding the closed-loop system are addressed in the
Authorized licensed use limited to: Nazarbayev University. Downloaded on February 18,2025 at 09:33:07 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 62, NO. 6, JUNE 2017 2969

while the box constraints can be expressed as

S = {σ ∈ R2 : σ1 ∈ [σ 1 , σ 1 ], σ2 ∈ [σ 2 , σ 2 ]} (26)
with σ 1 , σ 2 < 0 and σ 1 , σ 2 > 0.
Theorem 2: Given system (10) with r = 2, assuming that Assump-
tion 1 holds, if the control law (25) is applied, then σ is steered to the
origin with domain of attraction
SI  S \ {M0 ∪ M1 } (27)
where
 
σ2 |σ2 |
M0  (σ1 , σ2 ) : σ1 > − + σ 1 , σ2 > 0 , (28)
2αr
 
σ2 |σ2 |
M1  (σ1 , σ2 ) : σ1 < − + σ 1 , σ2 < 0 . (29)
2αr
The set SI is the maximum obtainable domain of attraction for the
given set of constraints, and the convergence takes place in minimum
time for the worst possible realization of the disturbance terms (i.e.,
f (ζ(t), σ(t), t) ≡ −F · sgn(w(t)) and g(ζ(t), σ(t), t) ≡ G1 ). 
Proof: See Appendix B. 
Fig. 1. Numerical estimate of the domain of attraction SI for Example 4:
V. GENERAL CASE: HIGHER ORDER SLIDING MODES WITH (18) with α = 1 (a), (25) with α = 1 (b), (18) with α = 5 (c), and (25) with
ARBITRARY CONSTRAINTS α = 5 (d).

In Section IV, the properties of the proposed scheme have been

analyzed for constrained first-order and second-order sliding modes corresponding domains of attraction are shown in Figs. 1(c) and (d),
with box constraints. These represent very common and relevant cases again for (18) and (22), respectively. The corresponding domains of
(for instance, a mechanical system with constraints on position and attraction amount at 3% and 86% of the sphere volume, respectively.
velocity). Yet, sometimes one has to design higher-order sliding mode Two observations are in place:
controllers and/or to deal with more general constraints. We provide 1) The domain of attraction using the constrained control law (22) is
now general considerations and results for the higher-order case with not reduced (actually, it is enlarged) with respect to that obtained
arbitrary constraints on σ. by using (18), as expected from Lemma 3.
In the cases analyzed in Section IV, it has been proved that the ob- 2) The difference between the two domains of attraction seems to
tained domain of attraction is maximized for the given set of constraints. widen as α increases: this is due to the ability of the proposed
The following result can be proved for the general case. control law (22) to impose a state trajectory that, when possible,
Lemma 3: Given system (10) with generic r, assuming that As- moves on the boundary of S before moving to its interior. 
sumption 1 holds, if the control law (22) is applied, then the domain of The result regarding minimum-time convergence also in the presence
attraction SI , for the given set of constraints, includes the domain of of constraints, formally proved in the previous section for lower-order
attraction SI obtained by applying the unconstrained control law (18). cases, becomes analytically intractable for higher sliding mode orders.
 However, in order to show that the proposed method might be achieving
Proof: The result immediately follows by observing that, if σ(0) ∈ the minimum-time convergence for the general case, we show the
SI , then w(0) = wi n (0). Being SI by definition a robust positively following example.
invariant (RPI) set [42, Def. 4.3] when applying (18), the time evolution Example 5: Consider again the system of Example 4, this time
of σ will coincide with that obtained by applying (18). This implies with constraints defined as σ 2 ≤ 0.1. In order to test the minimum-
that SI ⊆ SI .  time converge properties of the controller, we set α = αr = 0.5.
Example 4: In order to show the domain of attraction in a rather Considering an initial condition σ = [0.05 0.05 0] , the control
general case, we consider the third-order SMC, for which the dynamics law (22) is simulated with MATLAB using the Ode4 solver with
of the sliding variable is defined as a fixed discretization interval of 10−3 s. With the same initial
σ̇1 (t) = σ2 (t) condition and the same constraints, a minimum-time constrained
optimization problem is solved numerically by using CVX [43], by
σ̇2 (t) = σ3 (t) implementing a bisection routine which runs a feasibility problem
σ̇3 (t) = w(t). at every step, for a fixed time interval. In Figure 2, it is possible to
see that the state trajectories obtained with the two control laws are
Note that, in order to be able to make comparisons, the uncertain terms indistinguishable 
have been fixed as f (ζ(t), σ(t), t) ≡ 0 and g(ζ(t), σ(t), t) ≡ 1, which A similar procedure has been repeated for different systems and
implies α = αr . The set of constraints is defined as σ 2 ≤ 1, which different sets of constraints, always obtaining indistinguishable state
is a sphere centered at the origin. By fixing α = 1, we numerically trajectories, and the same time interval for convergence to the ori-
obtain the domains of attraction shown in Figs. 1(a) (63% of the gin. This can lead us to the conjecture that the proposed con-
sphere volume) and 1(b) (76% of the sphere volume) by applying trol law achieves minimum-time convergence in the general r-order
the unconstrained control law (18) and the new proposed control case. The formal proof of such a result can be a topic for further
law (22), respectively. Increasing the control amplitude to α = 5, the research.

Authorized licensed use limited to: Nazarbayev University. Downloaded on February 18,2025 at 09:33:07 UTC from IEEE Xplore. Restrictions apply.
2970 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 62, NO. 6, JUNE 2017

APPENDIX A
PROOF OF THEOREM 1
Assume that σ1 (0) ∈ S, i.e., σ1 (0) ∈ [σ 1 , σ 1 ]. Specifically, from
Assumption 1 it follows that

0 < σ1 (t) ≤ σ 1 ⇒ σ̇1 (t) ≤ F − G1 α = −αr < 0, (30)

σ 1 ≤ σ1 (t) < 0 ⇒ σ̇1 (t) ≥ −F + G1 α = αr > 0. (31)

As a consequence, σ1 (t) is steered to zero in finite time without leaving

S for all σ1 (0) ∈ SI ≡ S. Being S the set of state constraints, a larger
set SI cannot be obtained, which proves the maximization of the domain
of attraction. Also, being signals w(σ(t)) and wi n (σ(t)) coinciding for
all t, the controller solves a minimum-time problem for the worst-case
disturbance, as proved in [16].

APPENDIX B
PROOF OF THEOREM 2
Define the external perimeter ∂SI of SI as the union of the segments
Fig. 2. Comparison between the state trajectories using the pro-
AB, CD, EF , GH, BC, DE, F G, and HA (see Fig. 3). The extreme
posed control law (22) and the numerical minimum-time control law, for points (e.g., A and B in AB) are not considered as part of the segment
Example 5. for the first four ones, while they are taken into account for the other
segments. Fig. 3 shows a realization of the invariant set SI in the simple
case σ 1 = σ 2 = 1, σ 1 = σ 2 = −1, αr = 1. The white region is SI ,
while the black regions represent the set {M0 ∪ M1 }. The so-called
‘switching line’ of equation

σ2 |σ2 |
σ1 = − (32)
2αr
is also shown as a solid blue line.
A. Positive Invariance of SI : As a preliminary result, it
will be shown that SI is a robust positively invariant (RPI) set [42, Def.
4.3] for the closed-loop system, which is proved by checking that for
each σ ∈ ∂SI , the vector field σ̇ = [σ̇1 , σ̇2 ] never points outside SI
[42, Theorem 4.10].
Case 1 (σ ∈ HA ∨ σ ∈ DE): Assume σ ∈ DE so that σ̇ =
[σ2 , f − gα] . Notice that, from this point and for all the remain-
der of this proof, the dependency of f and g from their arguments
will be omitted for the sake of readability. Consider that σ ∈ DE
implies σ2 < 0, while −F − G2 α ≤ f − gα ≤ F − G1 α < 0. Then,
the vector field is pointing down-left, that is towards the interior of SI .
Analogous considerations can be done if σ ∈ HA, where the vector
field is always pointing up-right.
Case 2 (σ ∈ BC ∨ σ ∈ F G): Assume σ ∈ F G \ {G} so that
σ̇ = [σ2 , f + gα] , with σ2 = σ 2 < 0 and 0 < −F + G1 α ≤ f +
Fig. 3. Graphical representation of a possible invariant region SI in the gα ≤ F + G2 α. Then, the vector field σ̇ is always pointing up-left,
second-order case. which means towards the interior of SI . Analogous considerations
can be done if σ ∈ BC, where the vector field is always pointing
down-right.
Case 3 (σ ∈ CD ∨ σ ∈ GH): Assume σ ∈ GH so that σ̇ =
VI. CONCLUSIONS
[σ2 , f + gα] . One has always that σ2 < 0 is on this segment, while
This technical note has presented a new approach to design HOSM 0 < −F + G1 α = αr ≤ f + gα ≤ F + G2 α. It is easy to notice that,
control laws for nonlinear uncertain systems with arbitrary relative since all the points on this segment verify σ1 = − σ22 α|σr2 | + σ 1 , σ̇ can
degree subject to input and state constraints. The presence of con- be at most tangent to the line but never points outside. The same con-
straints, while being a topic of paramount importance in practical siderations can be stated for σ ∈ CD.
applications, has not been dealt with extensively in the SMC liter- Case 4 (σ ∈ AB ∨ σ ∈ EF ): Assume σ ∈ AB. The control law
ature up to now. The proposed control laws, apart from maintain- is discontinuous on AB. The vector field is therefore generated
ing the system state and the control variable always within the ad- as the Filippov solution (see, e.g., [4], [5]) of the state space
missible domain, are aimed at achieving the minimum-time conver- equations (10b)–(10c) for the second-order case. More precisely,
gence on the sliding manifold, and the maximization of the domain of σ̇ belongs to the convex hull of σ̇ + = [σ2 , f − gα] and σ̇ − =
attraction. [σ2 , f + gα] . The solution is obtained as σ̇ = μσ̇ + + (1 − μ)σ̇ − .

Authorized licensed use limited to: Nazarbayev University. Downloaded on February 18,2025 at 09:33:07 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 62, NO. 6, JUNE 2017 2971

Finding μ from condition ∇σ2 · σ̇ = 0, with ∇σ2 = [0, 1] , integrator plant
− +
one has that σ̇ = ∇σ ∇σ 2 ·σ̇
·( σ̇ − −σ̇ + )
σ̇ + − ∇σ ∇σ 2 ·σ̇
·( σ̇ − −σ̇ + )
σ̇ − = f2+ggαα σ̇ + −
f −g α
2 2
σ̇1 (t) = σ2 (t)
2g α
σ̇ − = [σ2 , 0] which is always tangent to the segment AB and
pointing left. An analogous procedure has been used to analyze the σ̇2 (t) = wr (t) (34)
case σ ∈ EF .
We can therefore conclude that SI is an RPI set for the considered in which wr = (αr /α)w implicitly takes into account the effect of the
closed-loop system. disturbance terms. Let [0, 0] = [σ1 (tc ), σ2 (tc )] , where tc is the time
B. Finite-Time Convergence to the Origin: The conver- needed to reach the origin from the given initial condition for a given
gence property will be proved in three parts, showing that, for any control law. Following the approach discussed in [45], given a linear
initial condition σ(0) ∈ SI , σ(t) reaches the origin in a finite time. system subject to convex state constraints and strongly convex control
Case 1: Assume that one has σ(0) ∈ SI ∩ {σ : σ1 > constraints, if a covariant function ψ(t) and functions η(t) and φ(t)
−σ2 |σ2 |/2αr , σ2 > σ 2 }. The vector field in this case is σ̇(0) = can be found such that
[σ2 (0), f − gα] , being w(σ(0)) = −α, and it is possible to state
that f − gα < 0. As a consequence, for all the considered σ(0), φ̇(t) = ψ(t) + ξ(t)η(t) (35)
σ̇(0) has a vertical component which is always strictly negative.
for some time-optimal control law in the absence of state constraints
When f − gα < −αr , σ(t) will reach in finite time either EF , or
wn c (σ), and for any 0 ≤ t ≤ tc , then the obtained evolution of the
the switching line σ1 = − σ22 α|σr2 | . An analogous proof is obtained for
sliding variable σ(t) is the one and only minimal time path and
σ(0) ∈ SI ∩ {σ : σ1 < −σ2 |σ2 |/2αr , σ2 < σ 2 }: in that case, the
sliding variable will reach in a finite time the segment AB, or the 
wn c (σ) if σ 2 < σ2 < σ 2
switching line. w(σ) = (36)
Case 2: Assume that σ(0) ∈ EF . As proved in Case 4 in 0 if σ2 = σ 2 or σ2 = σ 2
Appendix B-A, the trajectory of the system is kept on the line σ2 = σ 2 .
Since σ̇ has a strictly negative horizontal component during this time is the one and the only minimal time control connecting points
interval, point F (which is on the switching line σ1 = − σ22 α|σr2 | ) is (σ1 (0), σ2 (0)) and (0, 0). In our case, the defined set of state con-
reached in finite time. Analogous considerations hold for σ(0) ∈ AB. straints in (26) is convex, while the input constraint set [−αr , αr ] is
Case 3: Assume that σ(0) ∈ {σ : σ1 = −σ2 |σ2 |/2αr , σ2 < 0}. strongly convex. Also, relying on the linear system (34), it is possible to
Since the control law is discontinuous on the switching line, we need to define ψ(t) ≡ 0, ξ(t) = |σ2 (t)| + σ2 (t)w(t)/αr , η(t) = sgn(σ2 (t))
σ 22 if σ(t) ∈/ S, and η(t) ≡ 0 if σ(t) ∈ S. Moreover, it is known that the
use the Filippov solution, with ∇(σ1 − ) = [1, − ασ 2r ] and σ2 < 0,
 σ2
 2α r  σ2
 time-optimal control law in the absence of state constraints would be
∇ σ 1 − 2 α2
r
·σ̇ − ∇ σ 1 − 2 α2
r
·σ̇ + the bang-bang control law
obtaining σ̇ =  σ2
 σ̇ + −  σ2
 σ̇ − =
∇ σ 1 − 2 α2 ·( σ̇ − −σ̇ + ) ∇ σ 1 − 2 α2 ·( σ̇ − −σ̇ + ) σ2 |σ2 |
r r
wn c (σ) = −αr sgn σ1 + . (37)
f + g α −α r 2αr
2g α
σ̇ + − f −g2 αg α−α r σ̇ − = [σ2 , αr ] which is always tangent to
the switching curve so that the state moves towards the origin, and
Condition (35) is therefore satisfied, as φ̇(t) = ψ(t) + ξ(t)η(t) =
converges to it in a finite time. The same holds for σ(0) ∈ {σ : σ1 =
−σ2 |σ2 |/2αr , σ2 > 0, for which σ̇ = [σ2 , −αr ] . Combining the two sgn(σ2 (t))(|σ2 (t)| + σ2 (t) w n cα(σr (t )) ), from which it follows that
obtained vector fields, we can also obtain that σ̇ = [0, 0] for σ = 0. φ(t) = σ1 (t) + σ 2 (t2)|σ
αr
2 (t )|
. Given the definition of S in (26), the effect
We have therefore proved that SI is a domain of attraction for the of the application of (25) to the uncertain system (10b)–(10c) would
origin, with finite-time convergence. be the same of applying (36) to system (34), with wn c defined in (37).
C. Maximal Region of Attraction: Assume that σ(0) ∈ This proves that the proposed control law (25) drives σ to the origin in
M0 . Considering that σ1 (0), σ2 (0) > 0, then σ1 will continue to in- minimum time, for the worst-case realization of the disturbance terms.
crease in time until σ2 > 0. The quickest way to make σ2 decrease is to
use the control variable w(σ) = −α. In this case, the system will move REFERENCES
on a parabolic arc, the equation of which, in the worst case, taking into
[1] M. V. Kothare, P. J. Campo, M. Morari, and C. N. Nett, “A unified frame-
account the uncertain terms, is work for the study of anti-windup designs,” Automatica, vol. 30, no. 12,
pp. 1869–1883, Dec. 1994.
[2] S. Galeani, S. Tarbouriech, M. Turner, and L. Zaccarian, “A tutorial on
σ2 |σ2 |
σ1 = − + σ1 + ε (33) modern anti-windup design,” in Proc. European Control Conf., Budapest,
2αr Hungary, Jul. 2009, pp. 306–323.
[3] J. Rawlings and D. Mayne, Model Predictive Control: Theory and Design.
Madison, WI: Nob Hill Pub., 2009.
with ε > 0. It is immediate to see that this arc intersects the σ1 -axis [4] V. I. Utkin, Sliding Modes in Optimization and Control Problems. New
outside S. One can easily see that any other realization of the control York: Springer-Verlag, 1992.
variable will also lead to the same outcome. As a consequence, M0 [5] C. Edwards and S. K. Spurgeon, Sliding Mode Control: Theory and Ap-
cannot be part of the region of attraction for any realization of the plications. London, U.K.: Taylor and Francis, 1998.
[6] I. Boiko, L. Fridman, A. Pisano, and E. Usai, “Analysis of chattering in
control variable, given the constraints imposed on w, and on σ. Analo-
systems with second-order sliding modes,” IEEE Trans. Autom. Control,
gous considerations hold if σ(0) ∈ M1 . In conclusion SI is the largest vol. 52, no. 11, pp. 2085–2102, Nov. 2007.
achievable region of attraction. [7] I. Boiko, “Analysis of chattering in sliding mode control systems with
D. Minimum-Time Convergence: The proof of the continuous boundary layer approximation of discontinuous control,” in
minimum-time convergence to the origin of the space {σ1 , σ2 } fol- Proc. American Control Conf., San Francisco, CA, Jul. 2011, pp. 757–
762.
lows from [44, Ch. 8] and [45]. Considering the worst-case realization [8] G. Bartolini, A. Ferrara, and E. Usai, “Output tracking control of uncertain
of the disturbance terms, it is possible to express the system dynam- nonlinear second-order systems,” Automatica, vol. 33, no. 12, pp. 2203–
ics as that obtained by applying the control input (25) to the double 2212, Dec. 1997.

Authorized licensed use limited to: Nazarbayev University. Downloaded on February 18,2025 at 09:33:07 UTC from IEEE Xplore. Restrictions apply.
2972 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 62, NO. 6, JUNE 2017

[9] G. Bartolini, A. Ferrara, and E. Usai, “Chattering avoidance by second- [26] M. Innocenti and M. Falorni, “State constrained sliding mode controllers,”
order sliding mode control,” IEEE Trans. Autom. Control, vol. 43, no. 2, in Proc. American Control Conf., Philadelphia, PA, Jun. 1998, pp. 104–
pp. 241–246, Feb. 1998. 108.
[10] G. Bartolini, A. Ferrara, A. Levant, and E. Usai, “On second-order sliding [27] H. Tanizawa and Y. Ohta, “Sliding mode control under state and control
mode controllers,” in Variable Structure Systems, Sliding Mode and Non- constraints,” in Proc. IEEE Conf. on Control Applications, Singapore,
linear Control, ser. Lecture Notes in Control and Information, K. Young Oct. 2007, pp. 1173–1178.
and Ü. Özgüner, Eds. London, U.K.: Springer-Verlag, 1999, pp. 329–350. [28] J. Fu, Q.-X. Wu, and Z.-H. Mao, “Chattering-free SMC with unidirectional
[11] G. Bartolini, A. Ferrara, E. Usai, and V. Utkin, “On multi-input chattering- auxiliary surfaces for nonlinear system with state constraints,” Int. J. Innov.
free second-order sliding mode control,” IEEE Trans. Autom. Control, Comput., Inf. Control, vol. 9, no. 12, pp. 4793–4809, 2013.
vol. 45, no. 9, pp. 1711–1717, Sept. 2000. [29] F. Dinuzzo, “A second order sliding mode controller with polygonal con-
[12] A. Levant, “Higher-order sliding modes, differentiation and output- straints,” in Proc. 48th IEEE Conf. Decision Control, Shanghai, China,
feedback control,” Int. J. Control, vol. 76, no. 9–10, pp. 924–941, Dec. 2009, pp. 6715–6719.
Jan. 2003. [30] M. Rubagotti and A. Ferrara, “Second order sliding mode control of a
[13] T. Floquet, J. P. Barbot, and W. Perruquetti, “Higher-order sliding mode perturbed double integrator with state constraints,” in Proc. American
stabilization for a class of nonholonomic perturbed systems,” Automatica, Control Conf., Baltimore, MD, Jun. 2010, pp. 985–990.
vol. 39, no. 6, pp. 1077–1083, Jun. 2003. [31] M. Tanelli and A. Ferrara, “Switched second-order sliding mode con-
[14] A. Levant, “Quasi-continuous high-order sliding-mode controllers,” IEEE trol with partial information: Theory and application,” Asian J. Control,
Trans. Autom. Control, vol. 50, no. 11, pp. 1812–1816, 2005. vol. 15, no. 1, pp. 20–30, Jun. 2013.
[15] F. Plestan, A. Glumineau, and S. Laghrouche, “A new algorithm for high- [32] M. Tanelli and A. Ferrara, “Enhancing robustness and performance via
order sliding mode control,” Int. J. Robust Nonlin. Control, vol. 18, no. 4–5, switched second order sliding mode control,” IEEE Trans. Autom. Control,
pp. 441–453, 2008. vol. 58, no. 4, pp. 962–974, Apr. 2013.
[16] F. Dinuzzo and A. Ferrara, “Higher order sliding mode controllers with [33] A. Ferrara, G. P. Incremona, and M. Rubagotti, “Third order sliding mode
optimal reaching,” IEEE Trans. Autom. Control, vol. 54, no. 9, pp. 2126– control with box state constraints,” in Proc. 52th IEEE Conf. Decision
2136, Sep. 2009. Control, Los Angeles, CA, Dec. 2014, pp. 4727–4732.
[17] Y. Shtessel, C. Edwards, L. Fridman, and A. Levant, “Higher-order slid- [34] A. Swikir and V. I. Utkin, “Chattering analysis of conventional and su-
ing mode controllers and differentiators,” in Sliding Mode Control and per twisting sliding mode control algorithm,” in Proc. Int. Workshop on
Observation, ser. Control Eng.. Springer, New York, 2014, pp. 213–249. Variable Structure Systems, Nanjing, China, Jun. 2016, pp. 98–102.
[18] A. Levant, “Sliding order and sliding accuracy in sliding mode control,” [35] A. Levant, “Chattering analysis,” IEEE Trans. Autom. Control, vol. 55,
Int. J. Control, vol. 58, no. 6, pp. 1247–1263, Dec. 1993. no. 6, pp. 1380–1389, 2010.
[19] A. Ferrara and M. Rubagotti, “A sub-optimal second order sliding mode [36] I. Boiko and L. Fridman, “Analysis of chattering in continuous sliding-
controller for systems with saturating actuators,” IEEE Trans. Autom. mode controllers,” IEEE Trans. Autom. Control, vol. 50, no. 9, pp. 1442–
Control, vol. 54, no. 5, pp. 1082–1087, Jun. 2009. 1446, Sep. 2005.
[20] I. Castillo, M. Steinberger, L. Fridman, J. A. Moreno, and M. Horn, [37] I. Boiko, “Oscillations and transfer properties of relay servo systems—The
“Saturated super-twisting algorithm: Lyapunov based approach,” in Proc. locus of a perturbed relay system approach,” Automatica, vol. 41, no. 4,
Int. Workshop on Variable Structure Systems, Nanjing, China, Jun. 2016, pp. 677–683, 2005.
pp. 269–273. [38] H. K. Khalil, Nonlinear Systems, 3rd ed. Upper Saddle River, NJ: Prentice-
[21] M. Rubagotti, D. Raimondo, A. Ferrara, and L. Magni, “Robust model Hall, 1996.
predictive control with integral sliding mode in continuous-time sampled- [39] U. Walther, T. T. Georgiou, and A. Tannenbaum, “On the computation of
data nonlinear systems,” IEEE Trans. Autom. Control, vol. 56, no. 3, switching surfaces in optimal control: A Gröbner basis approach,” IEEE
pp. 556–570, Mar. 2011. Trans. Autom. Control, vol. 46, no. 4, pp. 534–540, 2001.
[22] A. Ferrara, G. P. Incremona, and L. Magni, “A robust MPC/ISM [40] A. F. Filippov, Differential Equations with Discontinuous Right-Hand
hierarchical multi-loop control scheme for robot manipulators,” in Sides, ser. Mathematics and its Applications, F. Arscott, Ed. Dordrecht,
Proc. 52th IEEE Conf. Decision Control, Florence, Italy, Dec. 2013, The Netherlands: Springer, 1988, vol. 18.
pp. 3560–3565. [41] K. D. Young, V. I. Utkin, and U. Ozguner, “A control engineer’s guide to
[23] A. Chakrabarty, V. Dinh, G. T. Buzzard, S. H. Zak, and A. E. Rundell, sliding mode control,” IEEE Trans. Control Syst. Technol., vol. 7, no. 3,
“Robust explicit nonlinear model predictive control with integral slid- pp. 328–342, May 1999.
ing mode,” in Proc. American Control Conf., Portland, OR, Jul. 2014, [42] F. Blanchini and S. Miani, Set-Theoretic Methods in Control. Springer
pp. 2851–2856. Science & Business Media, 2007.
[24] D. M. Raimondo, M. Rubagotti, C. N. Jones, L. Magni, A. Ferrara, and [43] M. Grant, S. Boyd, and Y. Ye, “CVX: Matlab software for disciplined
M. Morari, “Multirate sliding mode disturbance compensation for model convex programming,” 2008.
predictive control,” Int. J. Robust Nonlin. Control, vol. 25, no. 16, [44] M. Athans and P. L. Falb, Optimal Control. New York: McGraw-Hill,
pp. 2984–3003, 2015. 1966.
[25] A. Ferrara, G. P. Incremona, and L. Magni, “Model-based event-triggered [45] S. S. L. Chang, “Optimal control in bounded phase space,” Automatica,
robust MPC/ISM,” in Proc. Eur. Control Conf., Strasbourg, France, vol. 1, no. 1, pp. 55–67, Jan. 1963.
June 2014, pp. 2931–2936.

Authorized licensed use limited to: Nazarbayev University. Downloaded on February 18,2025 at 09:33:07 UTC from IEEE Xplore. Restrictions apply.