Automatica 75 (2017) 144–153

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Continuous-time model predictive control of under-actuated

spacecraft with bounded control torques✩
Huiping Li a,1 , Weisheng Yan a , Yang Shi b
a
School of Marine Science and Technology, Northwestern Polytechnical University, Xi’an, 710072, China
b
Department of Mechanical Engineering, University of Victoria, Victoria, B.C., Canada, V8W 3P6

article info abstract

Article history: The stabilization problem of rigid spacecraft is essential for space explorations and operations. This paper
Received 21 January 2016 studies the model predictive stabilization problem of a class of underactuated rigid spacecrafts with
Received in revised form two bounded control torques. A novel model predictive control (MPC) algorithm is designed by making
8 August 2016
use of the homogeneity of the system dynamics. In addition, a local homogeneous Lyapunov function
Accepted 14 August 2016
Available online 1 November 2016
is constructed based on which the approach to designing the terminal set and other parameters are
developed. Finally, the conditions for ensuring algorithm feasibility and closed-loop stability are provided.
Keywords:
We show that under the given conditions, the designed MPC algorithm is feasible, and the closed-loop
Spacecraft system is asymptotically stable. Simulation and comparison studies verify that the developed results are
Model predictive control (MPC) effective and valid, and the designed controller fulfills the constraint satisfaction and achieves much faster
Stabilization convergence rate in comparison with conventional continuous time-varying controllers.
Nonholonomic systems © 2016 Elsevier Ltd. All rights reserved.
Control constraints

1. Introduction discontinuous controllers (Casagrandea, Astolfi, & Parisini, 2008),

The stabilization problem of underactuated spacecraft (i.e., nor-
mally with two control torques) is a very challenging issue due
to the nonholonomic constraints. The study can be traced back to
Byrnes and Isidori (1991), where it has been shown that there does
not exist a (local) static state feedback controller to stabilize the
spacecraft with two control torques, because it violates the Brock-
ett necessary condition (Brockett, 1982).
Due to the design challenges, much attention has been
directed to studying the stabilization problem of spacecraft of
two control torques. In the literature, several strategies have
been developed to solve this stabilization problem, including the
time-varying continuous feedback controllers (Morin & Samson,
1997; Morin, Samson, Pomet, & Jiang, 1995), the time-invariant
✩ This work was supported in part by National Natural Science Foundation of
China (NSFC) under Grants 61473225, 61502395, 51579210; the Fundamental
Research Funds for the Central Universities under grant 3102016AXXX012; the
Scientific Research Foundation for the Excellent Returned Overseas Scholars,
Shaanxi Province. The material in this paper was not presented at any conference.
This paper was recommended for publication in revised form by Associate Editor
Riccardo Scattolini under the direction of Editor Ian R. Petersen.
E-mail addresses: lihuiping@nwpu.edu.cn, peter.huiping@gmail.com (H. Li),
weisheng_yan@yahoo.cn (W. Yan), yshi@uvic.ca, yangshi.uvic@gmail.com (Y. Shi).
1 Fax: +86 029 88495278.
http://dx.doi.org/10.1016/j.automatica.2016.09.024
0005-1098/© 2016 Elsevier Ltd. All rights reserved.
the hybrid control strategy (Teel & Sanfelice, 2008) and the
‘‘Quadratic Immersion’’-based approach (Carravetta, 2014). In
some special cases such as spin-axis stabilization, when there
are no requirements on stabilizing the angular velocities, the
stabilization problem of full states can be reduced to that of partial
states and may be solved by linear feedback controllers (Tsiotras,
1997). However, the above mentioned results with few exceptions
(e.g., Tsiotras, 1996; Tsiotras & Luo, 2000), neither consider the
optimal control performance nor practical control constraints,
though these two factors are particularly important in practical
implementations.
The references in Gui and Vukovich (2015) and Tsiotras and
Luo (1997, 2000) investigate the control problem of underactu-
ated spacecraft with bounded control inputs. In Tsiotras and Luo
(1997), the stabilization and tracking problem of underactuated
axisymmetric spacecraft with two control inputs is studied, and
large control inputs are avoided by designing the controllers that
divide the state space into two regions, and drive the state away
from the region where the large control inputs are generated. The
idea is further extended into Tsiotras and Luo (2000), where a priori
specified control bound is given and needs to be satisfied. Recently,
an adaptive spin-axis stabilization strategy of a symmetric space-
craft with two bounded torques is proposed in Gui and Vukovich
(2015), where a saturated proportional–derivative controller with
an adaptive scheme is developed to handle saturation and uncer-
tainties simultaneously. Note that the aforementioned results are
 H. Li et al. / Automatica 75 (2017) 144–153
focused on the axisymmetric spacecraft and only consider kine-
matics, which is quite different from this paper in which we will
study the asymmetric underactuated spacecraft and consider com-
plete system model including kinematics and dynamics.
It is well known that the model predictive control (MPC), also
known as receding horizon control, can provide almost optimal
control performance and well respect system constraints, which
has been widely implemented in the industry. This motivates the
current study to solve the model predictive stabilization problem
of spacecraft with two bounded control torques. Note that the
model predictive stabilization problem for such underactuated
systems is nontrivial, as the well-developed MPC strategies (Chen
& Allgöwer, 1998; Li & Shi, 2014a,b; Michalska & Mayne, 1993) are
not valid and cannot be directly applied. This is because the design
of these MPCs (Chen & Allgöwer, 1998; Li & Shi, 2014a,b; Michalska
& Mayne, 1993) heavily relies on a hypothesis that there exists
a state feedback controller for corresponding linearized systems,
which does not hold for the underactuated spacecraft dynamics. A
general MPC framework is proposed in Fontes (2001) which mainly
depends on the stability condition SC5, however no way has been
provided to construct such a condition, and it is generally very hard
to find such a condition for underactuated spacecraft.
In the literature, most of the MPC strategies for spacecraft
control are focused on fully actuated models and/or linearized
ones. For example, a fixed-point MPC strategy with quadratic
cost functions is developed for tracking attitude set points of a
full-actuated spacecraft in Guiggiani, Kolmanovsky, Patrinos, and
Bemporad (2014); an explicit model predictive control algorithm
is proposed for a linearized spacecraft model in Hegrenæ, Gravdahl,
and Tøndel (2005); a nonlinear MPC algorithm is designed for
attitude control of full-actuated spacecraft on SO(3) in Kalabic,
Gupta, Di Cairano, Bloch, and Kolmanovsky (2014), where the
Lie group variational integrator based discrete-time model is
adopted. It is worth noting that for spacecraft with two control
torques, an MPC-based stabilization strategy has been designed
in Marchand and Alamir (2003), where a numeric approximation
of the system trajectory is utilized by Chebyshev’s polynomial
basis, but the theoretical properties of the designed algorithm
are not studied. The theoretical results of MPC strategies for
underactuated underwater vehicles are developed in Li, Xie,
and Yan (2016) and Li and Yan (2016), but they cannot be
directly applied for underactuated spacecraft. Therefore, it is
very necessary to investigate the model predictive stabilization
problem of underactuated spacecraft with practical constraints,
and further provide a rigorous analysis to uncover theoretical
properties of the designed MPC algorithm.
In this study, we will study the MPC-based stabilization
problem of rigid spacecraft with two control torques and practical
control input constraints, and develop a systematic parameter-
design approach to ensure the theoretical properties of the
designed MPC algorithm. The main contributions of this paper are
three-fold:
• A novel model predictive stabilization algorithm is designed
for rigid spacecraft with two bounded control torques. The
novelty of the designed algorithm lies in that the design
of cost function makes full use of the homogeneity of the
system dynamics. Different from the conventional MPCs (Chen
& Allgöwer, 1998; Li & Shi, 2014a,b; Li & Yan, 2015; Michalska
& Mayne, 1993), where the cost functions use the Euclidean
norm of the system states and control inputs, the stage cost
function is built on the homogeneous norm of partial system
states, and the terminal cost function is time-varying and
captures the properties of system dynamics. In addition, the
terminal constraint is imposed in terms of the time-varying cost
function unlike the conventional ones in which a terminal set is
constructed by using the time-invariant terminal cost function
(Chen & Allgöwer, 1998; Li & Shi, 2014a,b; Michalska & Mayne,
1993).
145
• An approach to designing parameters for the MPC algorithm
is developed. Due to the fact that the terminal cost is time-
varying and parameter-dependent, we develop conditions to
construct a local stable set for the closed-loop system by using
the time-varying control law in Morin and Samson (1997), and
show that the terminal cost function is qualified as a local
homogeneous Lyapunov function for such closed-loop system
with appropriate parameters. Furthermore, an optimization
problem which combines the control input constraints and
the local stable set is formulated for designing the terminal
constraint.
• The theoretical conditions on ensuring algorithm feasibility
and closed-loop stability are established. Under the way
of parameter design (i.e., Algorithm 2), we prove that the
designed MPC algorithm is iteratively feasible, and the closed-
loop system is asymptotically stable and the control input
constraints are satisfied. These theoretical results provide a
useful tool for algorithm design and analysis in practical
implementations.
Notations: We use the symbol R and R+ to denote the real space
and real space with positive numbers, respectively. We adopt
‘‘T’’ to represent the matrix transposition, and col{x1 , x2 , . . . , xn }
to denote the column operation as [xT1 , xT2 , . . . , xTn ]T for column
vectors x1 , x2 , . . . , xn . For a column vector x ∈ Rn , its Euclidean
norm (and its absolute value for n = 1) is denoted by |x|.
For a vector x ∈ R3 , its skew operation is written as S (x) =
[0, x3 , −x2 ; −x3 , 0, x1 ; x2 , −x1 , 0]. Given a matrix P ∈ Rn×n , we
use P > 0 (P > 0) to denote its positive definiteness (semi-positive
definiteness). For a symmetric matrix, we use ∗ to represent its
symmetric elements with respect to its main diagonal. For two sets
U ⊆ Rn and W ⊆ Rn with U ⊆ W , define W \ U = {x ∈ Rn |x ∈
W , ̸∈ U}. We say that a function f belongs to the class C k , if its
derivatives up to the order of k exist and are continuous.
2. Problem formulation and preliminaries
2.1. Problem statement
Consider a rigid spacecraft with two control torques as:
Ṙ = S (ω)R,
J ω̇ = S (ω)J ω + Bτ , (1)
where R ∈ SO(3) is the rotation matrix from the body frame B to
the inertial frame I, ω ∈ R3 is the angular velocity in the frame
I, J = diag(j1 , j2 , j3 ) is the principal moments of inertia, τ =
col(τ1 , τ2 , 0) with τi , i = 1, 2, are torques driving the spacecraft,
and B = I3 . Note that there are only two control torques, making
the system in (1) nonholonomic. The control torques are bounded
by
τi ∈ Γi , i = 1, 2, (2)
where Γi ⊆ R, contains zero as its interior point.
Taking x = col(x1 , x2 , x3 ) = −k tan θ as in Morin and Samson
(1997) and Morin et al. (1995), where ⃗ k is the unit vector around
which R is rotated, θ ∈ (−π , π ) is the rotated angle, and k is
coordinate of ⃗
k in the frame I, the system in (1) can be transformed
into the following form (Morin & Samson, 1997; Morin et al., 1995)
1
ẋ = (ω + S (ω)x + (xT ω)x),
2
ω̇1 = c1 ω2 ω3 + u1 ,
ω̇2 = c2 ω1 ω3 + u2 ,
ω̇3 = c3 ω1 ω2 , (3)
146 H. Li et al. / Automatica 75 (2017) 144–153

where c1 = 2 j 3 , c2 = 3 j 1 , c3 = 1 j 2 , u1 = τ1 /j1 and u2 = τ2 /j2 . An intermediate system of (5) by ignoring g1 (z ) and g2 (z ) is

j −j j −j j −j
1 2 3
The control input constraints become given as

ui ∈ Γ̃i , i = 1, 2, (4) ż = f (z , u). (6)

with Γ̃i = Γi /ji . It is assumed that c3 ̸= 0, which is to ensure In addition, a reduced system for (6) can be obtained when defining
the nonlinear system in (3) is controllable. Otherwise, the state ω3 v1 = ω1 and v2 = ω2 as the control inputs:
cannot be controlled. Defining z , col(x, ω), the system in (3) can
also be written as ż = h(z , u), where h(z , u) = col( 12 (ω + S (ω)x + 1
η̇ = col(v1 , v2 , ω3 + v2 x1 − v1 x2 , 2c3 v1 v2 ), (7)
(xT ω)x), c1 ω2 ω3 + u1 , c2 ω1 ω3 + u2 , c3 ω1 ω2 ). 2
The objective of this study is to design an MPC algorithm to where η , col(x, ω3 ).
stabilize the system in (3) and respect the control bounds in (4).
By making using of the homogeneity of (5) and the property of
(7), a continuous time-varying feedback control can be designed as
2.2. Preliminaries of homogeneous systems in the following lemma.

The definitions and properties of homogeneous systems are

Lemma 1 (Morin and Samson (1997)). For the system in (3), the
recalled in this subsection (MCloskey & Murray, 1997; Pomet &
Samson, 1993). feedback control law

u1 (t ) = −µ1 (ω1 − v1 (x, ω3 , t )),

Definition 1. A dilation δλr : Rn → Rn , is defined by
u2 (t ) = −µ2 (ω2 − v2 (x, ω3 , t )), (8)
δλr x = (λr1 x1 , λr2 x2 , . . . , λrn xn ),
exponentially stabilizes the closed-loop system with respect to the
for any λ > 0 and ri > 0, i = 1, . . . , n. A homogeneous norm homogeneous norm ρ(x, ω), for large enough µ1 > 0 and µ2 > 0,
with respect to the dilation operator δλr is defined by ρpr (x) = with
1/p
|xi |p/ri , p > 0.
n
i=1 v1 (x, ω3 , t ) , − k1 x1 − ρ(x, ω3 ) sin(t /ϵ),
In this study, the homogeneous norm ρpr (x) will be written as (x3 + ω3 ) sin(t /ϵ)
ρ(x) in case of not causing any confusions. v2 (x, ω3 , t ) , − k2 x2 + ,
ρ(x, ω3 )
Definition 2. For a continuous function f : Rn × R → R and a where k1 > 0, k2 > 0, ϵ > 0 is a small enough constant,
dilation δλr , if f (δλr x, t ) = λm f (x, t ), ∀λ > 0, then it is said to be a ρ(x, ω3 ) = (x41 + x42 + x23 + ω32 )1/4 , ρ(x, ω) = (x41 + x42 +
homogeneous function and of homogeneous degree m associated x23 + ω34 + ω34 + ω32 )1/4 and the dilation is defined by δλr (x, ω) =
with δλr . (λx1 , λx2 , λ2 x3 , λω1 , λω2 , λ2 ω3 ).
Definition 3. For a continuous vector field g : Rn × R → Rn × R When applying v1 = v1 (x, ω3 , t ) and v2 = v1 (x, ω3 , t ) to the
and a dilation δλr , if gi (δλr x, t ) = λm+ri gi (x, t ), ∀i = 1, . . . , n, for reduced system in (7), the closed-loop system becomes
any λ > 0, then the vector field is a homogeneous one and of
homogeneous degree m with respect to the dilation δλr . 1
η̇ = f0 (η) + f1 (η) sin(t /ϵ) + f2 (η) cos(2t /ϵ), (9)
2
Definition 4. For the system ẋ = f (x, t ), if there are c1 > 0 and
where f0 (η) = 21 col(−k1 x1 , −k2 x2 , ω3 + (k1 − k2 )x1 x2 ,
c2 > 0 such that ρ(x(t )) 6 c1 ρ(x(0))e−c2 t , ∀t > 0, x(0) ∈ N , x (x +ω )
2c3 k1 k2 x1 x2 − (ω3 + x3 )), f1 (η) = 12 col(ρ(x, ω3 ), ρ(3x,ω 3) , 1ρ(x3,ω )3
x +ω
where N ∈ Rn is a neighborhood of the origin, then it is locally 3 3
exponentially stable with respect to the homogeneous norm ρ(x). + ρ(x, ω3 )x2 , 2c3 k2 x2 ρ(x, ω3 ) − 2c3 k1 x1 ρ(x3 +ω3
x,ω3 )
) , and f 2 (η) =
col(0, 0, 0, x3 + ω3 ). The system in (9) is a time-varying system
T
2.3. A time-varying stabilizable controller with period T0 = 2π ϵ . Applying the operation 1/T0 0 0 (.)dt to
both sides of (9), the averaged system of (9) can be obtained as
For the system in (3), let us recall its homogeneous properties
and the continuous time-varying feedback control law from Morin η̇ = f0 (η). (10)
and Samson (1997). Note that the system in (3) can be written as
The asymptotic stability of the system in (10) can be established
ż = f (z , u) + g1 (z ) + g2 (z ), (5) by inspecting that the linearized system is stable as in Morin and
where Samson (1997).

1
f ( z , u) = col(ω1 , ω2 , ω3 + ω2 x1
2 3. Model predictive control algorithm
− ω1 x2 , 2u1 , 2u2 , 2c3 ω1 ω2 ),
1 This section first develops the result on how to design
g1 (z ) = col(ω3 x2 − ω2 x3 + ω1 x21 + ω2 x2 x1 , homogeneous Lyapunov functions for the averaged dynamics in
2
(10), and then the MPC algorithm is designed.
− ω3 x1 + ω1 x3 + ω1 x1 x2 + ω2 x22 ,
ω1 x1 x3 + ω2 x2 x3 , 2c1 ω2 ω3 , 2c2 ω1 ω3 , 0),
3.1. Homogeneous Lyapunov function design
1
g2 (z ) = (ω3 x3 x1 , ω3 x3 x2 , ω 2
3 x2 x3 , 0, 0, 0).
2 To facilitate the design of the cost function in MPC, we develop
It can be verified that, for a dilation δλr z = (λx1 , λx2 , λ2 x3 , λω1 , a class of homogeneous Lyapunov functions with appropriate
λω2 , λ2 ω3 ), g1 (z ) is of degree l1 = 2, and g2 (z ) is of degree l2 = 4. bounds for the system in (10) in the following lemma.
 H. Li et al. / Automatica 75 (2017) 144–153 147

Lemma 2. For the system in (10), define the function V (x, ω3 ) , For the system in (3), the stage cost function at each time instant
ax41 + bx42 + cx23 + dω32 + ex3 ω3 . For large enough a > 0 and b > 0, if tk is constructed as

c > 0, d > 0, 2c3 d − e > 0, ec3 > 0, L(z (s; tk ), u(s; tk ))

β β

ec3 (2c3 − e) > |c − 0.5ec3 − c3 d|, (11)
= Aρ ρ β (x(s; tk ), ω3 (s; tk )) + Bu1 u1 (s; tk ) + Bu2 u2 (s; tk ),

then V (x, ω3 ) is a Lyapunov function for (10). In addition, there exist where ρ(x, ω3 ) = (x41 + x42 + x23 + ω32 )1/4 , β = 4, Aρ > 0, Bu1 > 0,
constants K1 > 0 and K2 > 0 such that V (x, ω3 ) 6 K1 ρ 4 (x, ω3 ) and and Bu2 > 0 are scalar constants, which will be specified later. The
∂ V (x,ω3 ) terminal cost function is designed by
∂η
f0 (η) 6 −K2 ρ 4 (x, ω3 ).
F (z (tk + T ; tk ), tk + T )
Proof. Taking derivative of V (x, ω3 ) along the trajectory of (10),
one has V̇ (x, ω3 ) = −2ak1 x41 − 2bk2 x42 + [c (k1 − k2 ) + = W (η(tk + T ; tk ), tk ) + Φ (z (tk + T ; tk ), tk ),
c3 ek1 k2 ]x1 x2 x3 + [2dc3 k1 k2 + 1/2e(k1 − k2 )]x1 x2 ω3 − 1/2ec3 x23 − where
(c3 d − 1/2e)ω32 + (c − 1/2ec3 − c3 d)x3 ω3 . Due to the condition in ∂ V (x, ω3 )
(11), there exist ξ1 > 0, ξ2 > 0, β1 > 0 and β2 > 0 such that W (η, t ) , V (x, ω3 ) − ϵ(1 − cos(t /ϵ)) f1 (η)
∂η
V̇ (x, ω3 ) 6 −2ak1 x41 − 2bk2 x42 + [c (k1 − k2 ) + c3 ek1 k2 ]x1 x2 x3 1 ∂ V (x, ω3 )
ϵ sin(2t /ϵ)
− f2 (η), (14)
+ [2dc3 k1 k2 + 1/2e(k1 − k2 )]x1 x2 ω3 4 ∂η
− (ξ12 + ξ2 )x23 − (β12 + β2 )ω32 .
 
 2
1 ωir +1 + r vir +1 (x, ω3 , t )
Φ (z , t ) , √ − vi (x, ω3 , t )ωi ,
r

By taking a = (ã2 + b̃2 )/k1 and b = (ã2 + b̃2 )/k2 , with ã = i =1

µi r +1
c (k1 −k2 )+c3 ek1 k2 2dc3 k1 k2 +1/2e(k1 −k2 )
2ξ
and b̃ = 2β
, we obtain V̇ (x, ω3 ) 6 (15)
1 1
−(ã + b̃ )
2 2
x41 − (ã + b̃ ) − ξ
2 2
x42 2
2 x3 − β2 ω 6 −K2 ρ (x, ω3 ), where
2
3
4
r = 3, ϵ > 0 and V (x, ω3 ) is a function designed as in Lemma 2.
K2 can be taken as min{(ã + b̃ ), ξ2 , β2 }. 2 2

On the other hand, there exist constants ξ3 > 0 and ξ4 > 0 such Remark 5. It is worth noting that for some small ϵ > 0, W (η, t )
that V (x, ω3 ) > ax41 + bx42 + (c + ξ3 )x23 + (d + ξ4 )ω32 > K1 ρ 4 (x, ω3 ), will be positive and equal to zero only when η = 0, and the
where K1 can be taken as min{a, b, (c + ξ3 ), (d + ξ4 )}. The proof is detailed design of ϵ will be given later. Φ (z , t ) is always positive
ωr +1 +r v r +1 (x,ω3 ,t )  ωi
completed. by noting that i i
r +1
− vir (x, ω3 , t )ωi = vi (x,ω3 ,t ) sr −
vi (x, ωi , t )ds > 0, i = 1, 2.
r
Remark 3. It can be verified that the Lyapunov function V (x, ω3 )
is a homogeneous function of degree 4 with respect to the Finally, the cost function is constructed by
dilation δλr (x, ω3 ) = (λx1 , λx2 , λ2 x3 , λ2 ω3 ), which differs from
J (z (s; tk ), u(s; tk ), s)
conventional Lyapunov functions. This homogeneity will be  tk + T
utilized to construct Lyapunov functions for the time-varying and
, L(z (s; tk ), u(s; tk ), s)ds + F (z (tk + T ; tk ), tk + T ),
periodic system in (9). tk

Remark 4. It is worthwhile to note that in the proof of Lemma 2, where T > τ is the prediction horizon.
the way of designing K1 and K2 might be conservative. A less Based on the cost function, an optimization problem solved at
conservative way to design K1 and K2 can be found by solving the each time tk for the system in (3) is formulated as follows:
following matrix inequalities:
  Problem P (tk , z (tk ), T ). minu(s;tk ) J (z (s; tk ), u(s; tk ), s), subject to
a − K1 0 0 0
∗ b − K1 0 0  ż (s; tk ) = h(z (s; tk ), u(s; tk )),
> 0, (12)

 ∗ ∗ c − K1 e/2  u1 (s; tk ) ∈ Γ̃1 , u2 (s; tk ) ∈ Γ̃2 , s ∈ [tk , tk + T ], (16)
∗ ∗ ∗ d − K1
F (z (tk + T ; tk ), tk + T ) 6 ε, (17)
c (k1 − k2 ) + c3 ek1 k2 2dc3 k1 k2 + e(k1 − k2 )/2 
z (tk ; tk ) = z (tk ),

ā
2 2
where ε > 0 is called the terminal set level.
 
 c − ec3 /2 − c3 d  6 0,

∗ −(0.5ec3 − K2 )

2

Unlike the conventional MPC strategies (Chen & Allgöwer,
∗ ∗ −(c3 d − e/2 − K2 ) 1998; Li & Shi, 2014a,b; Li, Shi, & Yan, 2016; Michalska & Mayne,
(13) 1993), where the cost functions are in quadratic forms, the
√ cost function in this study is time-varying and related to the
where ā = −2 (2ak1 − K2 )(2bk2 − K2 ). homogeneous norm of the system under study. In addition, the
terminal constraint in (17) is imposed by a level set associated with
3.2. Design of MPC algorithm the periodic time-varying terminal cost function, which differs
from the conventional one in Chen and Allgöwer (1998), Li and Shi
For the system in (3), define the sampling time instants as tk , (2014a,b), Li, Shi et al. (2016) and Michalska and Mayne (1993).
k = 0, 1, . . . . The sampling period τ > 0 is a constant, and Using the optimization Problem P , the model predictive
the MPC controller samples the system state col(x(tk ), w(tk )) at stabilization algorithm is designed as in Algorithm 1.
each sampling time instant. For the ease of presentation, denote Due to the nonholonomic property of the system in (3), the
the following control and state trajectories: u(s; tk )—a functional design of the parameters ε , Aρ , Bu1 and Bu2 will be nontrivial, and
to characterize the control input trajectory; u∗ (s; tk )—the optimal the detailed design approach is postponed to the next section.
control input trajectory; ū(s; tk )—a functional to characterize In practice, the detailed steps for implementing the MPC
the feasible control input trajectory. The corresponding state Algorithm 1 numerically are as follows: (1) Take an appropriate
trajectories are defined accordingly. prediction horizon T as in the continuous-time case; (2) Discretize
148 H. Li et al. / Automatica 75 (2017) 144–153

Algorithm 1 Model Predictive Stabilization Algorithm
Require: Sampling period τ ; prediction horizon T ; other parame-
ters ϵ , ε , Aρ , Bu1 and Bu2 ; initial state z0 ; index k = 0.
Ensure: z (t0 ) = z0 .
1: while The control action is not stopped do
2: Sample system state z (tk );
3: Solve Problem P to generate the optimal control trajectory
u∗ (s; tk ), s ∈ [tk , tk + T ];
4: Apply control u(s; tk ) = u∗ (s; tk ), s ∈ [tk , tk + τ );
5: k = k + 1.
6: end while
the continuous-time system model by the sampling period τ >
0; (generally set the constant τ being small and ensure N =
T /τ be an integer.) (3) Set N as the prediction horizon in the
discrete-time implementation; (4) Discretize the continuous-time
control trajectory u(s; tk ), s ∈ [tk , tk + T ] as a control sequence
{u(k|k), u(k + 1|k), . . . , u(k + N − 1|k)}; (5) Discretize the
cost function J (z (s; tk ), u(s; tk ), s) and obtain the cost function in
discrete-time format. Through the above steps, the MPC algorithm
1 can be executed numerically.
4. Terminal set and parameter design
This section first provides a way to design a set within which
F (z , t ) is a local Lyapunov function for the system in (3) using
the control law in Lemma 1. Then the parameter design algorithm
including the terminal constraint set design is developed.
4.1. A local stable set
Before proceeding, let us recall a result from Morin and Samson
(1997), which states that a class of time-varying and periodic
systems are exponentially stable if their averaged systems are
asymptotically stable and certain conditions are satisfied. Consider
the system
p
 homogeneous norm is denoted by ρ(x, y1 , y2 ). By taking l as an odd
ẋ = f0 (x) + gi (t /ϵ)fi (x), (18)
i=1
where fi : Rn → Rn , i = 0, . . . , p, belonging to C 1 on Rn \
{0}, are continuous and homogeneous vector fields, which have
homogeneous degree zero with respect to a dilation δλr x, gi , i =
1, . . . , p, are continuous scalar periodic functions with period T0
T
and satisfy 0 0 gi (s)ds = 0. The averaged system for (18) is defined
as ẋ = f0 (x).
 ∂ V (x)
 
δi = sup  fi (x) ,


ρλ (x)=1
r ∂ x
 ∂ ∂ V (x)
   
γi,j = sup  fi (x) fj (x) .

ρλr (x)=1 ∂ x ∂x
In the next lemma, we extend Proposition 5 in Morin and
Samson (1997) to the case that there are two integrators added to
a class of homogeneous time-periodic systems. Consider a system
ẋ = f (x, v1 (x(1) , t ), v2 (x(2) , t ), t ), (20)
where f (x, y1 , y2 , t ) : Rn × R × R × R → Rn , is a continu-
ous function with period T0 , x(1) = col(x1 , . . . , xm1 ) and x(2) =
col(x1 , . . . , xm2 ), m1 6 n, m2 6 n, and the functions vi (x(i) , t ) :
Rmi × R → R, i = 1, 2, are continuous with period T0 , differen-
tiable with respect to t, homogeneous of degree q with respect to
a dilation δλr x and belong to class C 1 on (Rmi \ {0}) × R.
Lemma 7. For the system in (20), suppose that it is homogeneous
of degree zero with respect to the dilation δλr x, and is asymptotically
stable at zero. Then, for large enough µ1 > 0 and µ2 > 0, the
following system
ẋ = f (x, y1 , y2 , t ),
ẏ1 = −µ1 (y1 − v1 (x(1) , t )),
ẏ2 = −µ2 (y2 − v2 (x(2) , t )), (21)
is asymptotically stable at zero.
The proof of Lemma 7 is obtained by following the similar line
of Proposition 5 in Morin and Samson (1997). Since the parameter
design of the MPC strategy will need some intermediate results
which have not provided in Morin and Samson (1997), the proof
is presented in below.
Proof. Define δλr x = (λr1 x, . . . , λrn xn ) as the dilation with respect
to which the system in (20) is of homogeneous degree zero, and
the homogeneous norm is denoted by ρ(x). Define δ̃λr (x, y1 , y2 ) =
(λr1 x, . . . , λrn xn , λq y1 , λq y2 ) as the dilation associated with which
the system in (21) is of homogeneous degree zero, and the
integer such that l > max{ri /q, i = 1, . . . , max{m1 , m2 }}, as in
Morin and Samson (1997), it can be shown that vil (., .) are at least
of class C 1 on Rmi × R.
By applying Theorem 2 in Rosier (1992), there exists a Lyapunov
function V (x, t ) for the system in (20) which is with period T0 ,
and homogeneous of degree β = (l + 1)q. Similar as in Morin
and Samson (1997), define the following function W (x, y1 , y2 , t ) =
V (x, t ) + i=1 √1µ φi (yi , x(i) , t ), where φi (yi , x(i) , t ) = l+11 (yli+1 +
2
i

lvil+1 ) − vil yi . Since φi (yi , x(i) , t ) can be equivalently written as

φi (yi , x(i) , t ) = v i(x(i) ,t ) sl − vil (x(i) , t )ds, it can be verified that
y
Lemma 6 (Morin and Samson (1997)). For the system in (18),
i
suppose that its averaged system is asymptotically stable at x = 0 W (x, y1 , y2 , t ) is positive, equal to zero only at (x = 0, y1 =
and the associated Lyapunov function V (x) is in class C 2 and of 0, y2 = 0) and also proper (Morin & Samson, 1997). In addition,
homogeneous degree β with respect to the dilation δλr x, and satisfies due to that vil (., .) are at least of class C 1 on Rmi × R, φi (., ., .) are
V (x) > α1 (ρλr (x))β and ∂∂Vx f0 (x) 6 −α2 (ρλr (x))β for some constants also at least of class C 1 .
α1 > 0 and α2 > 0. Then for any ϵ ∈ (0, ϵ0 ), the system in (18) is In what follows, we show that W (x, y1 , y2 , t ) is a Lya-
exponentially stable, and the corresponding Lyapunov function is punov function for the system in (21). Taking the derivative of
p  t /ϵ
V1 (x, t ) = V (x) − ϵ i=1 0 gi (s)ds ∂∂Vx fi (x), with V̇1 (x, t ) 6 W (x, y1 , y2 , t ) with respect to t yields
β ∂V ∂V
p p
− α2 − i=1 Ii (γi,0 + j=1 Cj γi,j ) ϵ (ρλ (x)) , where
   r
Ẇ = f (x, v1 (x(1) , t ), v2 (x(2) , t ), t ) +
  ∂x ∂t
∂V 
f (x, y1 , y2 , t ) − f (x, v1 (x , t ), v2 (x(2) , t ), t )
(1)

 
 

 α
 +
α2 ∂x
 

1
ϵ0 = min p ,  , (19) 2
∂φ ∂φ (i) ∂φ
  
 I δ  p p  1
( , , , ) ,
 
Ii γi,0 + Cj γi,j  + ẏi + f x y y t +
 
 i i i
 √ 1 2
µi ∂ yi ∂ x(i) ∂t
 

i=1 j=1 i =1

where f (i) are the vector functions with their components are the
 t 
Ci = sup |gi (t )|, gi (s)ds ,
 
Ii = sup 
first mi component of f .
t ∈R t ∈R 0
 H. Li et al. / Automatica 75 (2017) 144–153 149

Due to the fact that V is a Lyapunov function for the system in Proof. According to Lemma 2, V (x, ω3 ) is a Lyapunov function for
(20) and homogeneous of degree β , there exists a constant K1 > 0 the system in (10) and satisfies the conditions of Lemma 6. As a
such that ∂∂Vx f (x, v1 (x(1) , t ), v2 (x(2) , t ), t ) + ∂∂Vt 6 −2K1 (ρ(x))β . result, by applying p = 2, g1 (t /ϵ) = sin(t /ϵ) and g2 (t /ϵ) =
By following the similar idea as in Morin and Samson (1997), it 1
2
cos(t /ϵ) in Lemma 6, for all ϵ ∈ (0, ϵ0 ), the function W (η, t )
can be verified that there exists a constant K2 > 0 such that defined in (14) is a Lyapunov function for the system in (9),
∂V
∂x
(f (x, y1 , y2 , t ) − f (x, v1 (x(1) , t ), v2 (x(2) , t ), t )) 6 K1 (ρ(x))β +  ϵ0 is defined
where  as in (19). In addition, Ẇ (η, t ) 6 −(α2 −
K2 [(y1 − v1 (x(1) , t ))r +1 + (y2 − v2 (x(2) , t ))r +1 ], where K2 is the ( pi=1 Ii (γi,0 + pj=1 Cj γi,j ))ϵ)ρ 4 (x, ω3 ).
constant such that Next consider the closed-loop system in (6) using the control
law in (8). By substituting y1 = ω1 , y2 = ω2 , x =
∂V

max (f (x, y1 , y2 , t ) col(x, ω3 ), δλr x = (λx1 , λx2 , λ2 x3 , λ2 ω3 ), δ̃λr (x, y1 , y2 ) = (λx1 , λx2 ,
ρ(x,y1 ,y2 )=1 ∂x λ2 x3 , λω1 , λω2 , λ2 ω3 ), q = 1, and l = 3 into Lemma 7, it follows
that the function F (z , t ) = W (η, t ) + Φ (z , t ) is a Lyapunov

(1) (2)
− f (x, v1 (x , t ), v2 (x , t ), t ) K1 (ρ(x))β
 
function for the system in (6) using the control law in (8), where
Φ (z , t ) is defined in (15).
+ K2 (y1 − v1 (x(1) , t ))r +1 + (y2 − v2 (x(2) , t ))r +1 6 1.
 
(22) Now consider the closed-loop system in (5) with the control
law given in (8). Taking the derivative of F (z , t ) with respect to
Furthermore, using the same way as in Morin and   Samson (1997),  it
∂φ t along the trajectory of the closed-loop system in (5), one has
can be shown that ∂ y ẏi = −µi yli − vil (x(i) , t ) yi − vil (x(i) , t ) 6 ∂ F (z ,t ) ∂ F (z ,t ) ∂ F (z ,t )
i Ḟ (z , t ) = ∂ z f (z , u) + ∂ t + ∂ z (g1 (z ) + g2 (z )). According
− 2l1−1 µi (yi − vi (x(i) , t ))l+1 . Finally, since the following function to Lemma 7, we have
∂φi (i) ∂φ (y ,x(i) ,t )
∂ x(i)
f (x,y1 ,y2 ,t )+ i i∂ t ∂ F (z , t ) ∂ F (z , t )
is homogeneous function of degree zero f ( z , u) +
(
(ρ(x))β + 2i=1 yi −vi (x(i) ,t ) )
l+1
∂z ∂t


and is well defined outside (x = 0, y1 = 0, y2 = 0), there exist 6 −A0 (µ1 , µ2 )ρ 4 (x, ω3 ) − C01 (µ1 , µ2 )(µ1 − v1 (x, ω3 , t ))4
∂φ
constants K3 > 0 and K4 > 0 such that (11) f (1) (x, y1 , y2 , t ) +
∂x − C02 (µ1 , µ2 )(ω2 − v2 (x, ω3 , t ))4 ,
∂φ1 (y1 ,x(1) ,t ) β
)) ( v1 (x(1) , t ))l+1 ], and
2
∂t
6 K 3 [(ρ( x + i =1 y1 − √
where A0 (µ1 , µ2 ) = K1 − √µ3 − √µ4 , C01 (µ1 , µ2 ) = 41 µ1 −
K K
∂φ2 (2) ( )
∂φ (y ,x ,t )
2
f (x, y1 , y2 , t ) + 2 2∂ t 6 K4 [(ρ(x))β + i=1 (y2 −
2 1 2
∂ x(2) √
K2 − √µ3 − √µ4 , C02 (µ1 , µ2 ) = 14 µ2 − K2 − √µ3 − √µ4 , with
K K K K
(2)
v2 (x , t )) l+1
], where K3 and K4 can be calculated as follows: 1 2 1 2

∂φ (y ,x(1) ,t ) 
     
 ∂φ1 (1) p p
 ∂ x(1) f (x, y1 , y2 , t ) + 1 1∂ t 1  
K1 = α2 − Ii (γi,0 + Cj γi,j ) ϵ , (26)

K3 6 max , (23) 2
ρ(x,y1 ,y2 )=1 2  l+1 i =1 j=1
(ρ(x))β + yi − vi (x(i) , t )

i =1 and K2 , K3 and K4 being calculated as in (22)–(24) by replacing
∂φ (y ,x(2) ,t )  y1 = ω1 , y2 = ω2 , x = η, f (i) (x, y1 , y2 , t ) as (7), and vi (x(i) , t )
 
 ∂φ2 (2)
 ∂ x(2) f (x, y1 , y2 , t ) + 2 2∂ t  as vi (x, ω3 , t ).
K4 6 max . (24)
Furthermore, since g1 (z ) is homogeneous of degree l1 = 2 with
ρ(x,y1 ,y2 )=1 2  l+1
(ρ(x))β + yi − vi (x(i) , t ) respect to δλr (x, ω) (which is defined in Theorem 8), there exists a

i =1 ∂ F (z ,t )
constant M1 > 0 such that ∂z
g1 (z ) 6 M1 ρ 6 (x, ω), where
By collectively considering the above results, one has Ẇ 6
 ∂ F (z , t )
l+1 l+1  
−β1 (ρ(x))β − β2 y1 − v1 (x(1) , t ) − β3 y2 − v2 (x(2) , t )
 
g1 (z ) .

, M1 = max  (27)
K K √
where β1 = (K1 − √µ3 − √µ4 ), β2 = ( 2l1−1 µ1 − K2 − √µ3 − √µ4 )
K K ρ(x,ω)=1 ∂z
1 2 1 2
√
and β3 = ( 2l1−1 µ2 − K2 − √µ3 − √µ4 ). The proof is completed by
K K Note that g2 (z ) is homogeneous of degree l2 = 4. By using the same
∂ F (z ,t )
1 2
reasoning, there exists a constant M2 > 0 such that ∂ z g2 (z ) 6
designing µ1 and µ2 large enough such that β1 > 0, β2 > 0 and
and β3 > 0. M2 ρ (x, ω), where
8

 ∂ F (z , t )
 
Based on Lemmas 6 and 7, in the next theorem, we provide a g2 (z ) .

M2 = max  (28)
result on how to design an invariant set for the system in (3) with ρ(x,ω)=1 ∂z
the control law in (8).
As a result, one obtains that Ḟ (z , t ) 6 −A0 (µ1 , µ2 )ρ 4 (x, ω3 ) −
C01 (µ1 , µ2 )(ω1 −v1 (x, ω3 , t ))4 − C02 (µ1 , µ2 )(ω2 −v2 (x, ω3 , t ))4 +
Theorem 8. For the system in (3) with the control law in (8), define M1 ρ 6 (x, ω) + M2 ρ 8 (x, ω).
a dilation as δλr (x, ω) = (λx1 , λx2 , λ2 x3 , λω1 , λω2 , λ2 ω3 ), and the Note again that the function A0 (µ1 , µ2 )ρ 4 (x, ω3 ) + C01 (µ1 , µ2 )
homogeneous norm as ρ(x, ω) = (x41 + x42 + x23 + ω14 + ω24 + ω32 )1/4 . (ω1 − v1 (x, ω3 , t ))4 + C02 (µ1 , µ2 )(ω2 − v2 (x, ω3 , t ))4 is homo-
Then for large enough µ1 > 0 and µ2 > 0, there exist constants geneous of degree 4. As a result, there exists a constant K̄1 >
ϵ0 > 0 and ε0 > 0, such that for any ϵ ∈ (0, ϵ0 ) and ε ∈ (0, ε0 ], 0 such that M1 ρ 6 (x, ω) 6 K̄1 ρ 2 (x, ω)(A0 (µ1 , µ2 )ρ 4 (x, ω3 ) +
the set inclusion col(x, ω) ∈ Ωε (x, ω) implies Ḟ (z , t ) 6 0 along the C01 (µ1 , µ2 )(ω1 −v1 (x, ω3 , t ))4 + C02 (µ1 , µ2 )(ω2 −v2 (x, ω3 , t ))4 ),
trajectory of the closed-loop system in (3) using control in Lemma 1, where the constant K̄1 is calculated as
where the set Ωε (x, ω) , {(x, ω) ∈ R6 |ρ(x, ω) 6 ε}, the value of ϵ0
K̄1 = max M1 / A0 (µ1 , µ2 )ρ 4 (x, ω3 )

is calculated as in (19) and
ρ(x,ω)=1

+ C01 (µ1 , µ2 )(ω1 − v1 (x, ω3 , t ))4



 K̄ 2 + 4K̄ − K̄
 1 2 1
+ C02 (µ1 , µ2 )(ω2 − v2 (x, ω3 , t ))4 .

ε0 = (25) (29)
2K̄2
Similarly, there exists a constant K̄2 > 0 such that M2 ρ (x, ω) 6 8

with K̄1 in (29) and K̄2 in (30). K̄1 ρ 4 (x, ω)(A0 (µ1 , µ2 )ρ 4 (x, ω3 ) + C01 (µ1 , µ2 )(ω1 − v1 (x, ω3 , t ))4
150
+ C02 (µ1 , µ2 )(ω2 − v2 (x, ω3 , t ))4 ), where the constant K̄2 is
calculated as
K̄2 = max M2 / A0 (µ1 , µ2 )ρ (x, ω3 ) 4

ρ(x,ω)=1
+ C01 (µ1 , µ2 )(ω1 − v1 (x, ω3 , t )) 4
+ C02 (µ1 , µ2 )(ω2 − v2 (x, ω3 , t ))4 .

Therefore, we arrive at
Ḟ (z , t ) 6 −Ξ (ρ(x, ω)) A0 (µ1 , µ2 )ρ 4 (x, ω3 )

+ C01 (µ1 , µ2 )(ω1 − v1 (x, ω3 , t ))4
+ C02 (µ1 , µ2 )(ω2 − v2 (x, ω3 , t ))4 ,

where Ξ (ρ(x, ω)) = 1 − K̄1 ρ 2 (x, ω) − K̄2 ρ 4 (x, ω). Since ε0 is
designed as in (25) in Theorem 8, it follows that Ξ (ρ(x, ω)) > 0
when 0 < ρ(x, ω) 6 ε0 . As a result, if µ1 and µ2 are designed such
that
A0 (µ1 , µ2 ) > 0, C01 (µ1 , µ2 ) > 0, and
C02 (µ1 , µ2 ) > 0,
and 0 < ρ(x, ω) 6 ε0 , then Ḟ (z , t ) 6 0. That is, col(x, ω) ∈
Ωε0 (x, ω) implies Ḟ (z , t ) 6 0. This completes the proof.
Remark 9. From the proof of Theorem 8, it can be seen that the
function F (z , t ) is a local Lyapunov function for the system in
(3) given that the parameters µ1 and µ2 are designed to satisfy
the condition in (32), ϵ ∈ (0, ϵ0 ) and ε ∈ (0, ε0 ]. And the
corresponding local stable set is Ωε0 (x, ω).
4.2. Parameter design
Using the result in Theorem 8 and considering the control input
constraints in (2), an upper bound of the terminal set level ε > 0
can be designed by the following optimization problem.
Problem Q.
εT = sup F (z , t ), subject to u1 (t ) ∈ Γ1 , u2 (t ) ∈ Γ2 ,
z ∈Ωε0 (x,ω),t ∈R
where u1 (t ) and u2 (t ) are given as in (8).
In the optimization problem P (tk , z (tk ), T ), the terminal set
level ε is designed as ε ∈ (0, εT ]. The constants Aρ , Bu1 and Bu2
are designed by
Aρ ∈ (0, Ξ (ε)A0 ], Bu1 ∈ (0, Ξ (ε)C01 ]
and Bu2 ∈ (0, Ξ (ε)C02 ].
In summary, the steps for designing parameters in the opti-
mization problem P (tk , z (tk ), T ) are detailed as in Algorithm 2.
5. Feasibility and stability analysis
This section first investigates the iterative feasibility problem
of the optimization problem P (tk , z (tk ), T ), and then studies the
closed-loop stability problem of (3) using the optimal control law
by solving P (tk , z (tk ), T ). Before proceeding, we define all the
initial states as Z0 under which Problem P (t0 , z (t0 ), T ) has a
solution.
Theorem 10. For any z (t0 ) ∈ Z0 , if the parameters are designed as
in Algorithm 2, then the optimization Problem P (tk , z (tk ), T ) has at
least a solution for all tk > t0 , and the designed Algorithm 1 is feasible.
H. Li et al. / Automatica 75 (2017) 144–153
Algorithm 2 Parameter Design
1: Design V (x, ω3 ) according to the condition in (11);
2: Design K1 and K2 by solving (12) and (13);
3: Apply α1 = K1 and α2 = K2 , calculate δi , Ci , Ii and γi,j as in
Lemma 6, and calculate ϵ0 according to (19);
4: Choose ϵ ∈ (0, ϵ0 );
(30) 5: Design K1 , K2 , K3 and K4 as in (26), (22), (23) and (24),
respectively;
6: Design µ1 and µ2 according to (32), and calculate A0 (µ1 , µ2 ),
C01 (µ1 , µ2 ), and C02 (µ1 , µ2 );
7: Calculate M1 and M2 according to (27) and (28);
8: Calculate K̄1 and K̄2 using (29) and (30);
Solve ε0 according to (25);
(31) 9:
10: Solve the optimization problem Q to calculate εT , and design
ε ∈ (0, εT ];
11: Design Aρ , Bu1 and Bu2 according to (33).
Proof. The proof can be obtained by following the similar way as
in Li and Shi (2014a,b) and Michalska and Mayne (1993) by noting
(32) that, at time tk+1 , ūi (s; tk+1 ), s ∈ [tk+1 , tk+1 + T ] is a feasible
solution, where
ū(s; tk+1 ) = u∗ (s; tk ), s ∈ [tk+1 , tk + T ],
ū(s; tk+1 ) = col(u1 (s; tk+1 ), u2 (s; tk+1 )), s ∈ (tk + T , tk+1 + T ],
with u∗ (s; tk ) being the optimal control input at time instant tk and
u1 (s; tk ) and u2 (s; tk ) defining as in (8).
In next theorem, we show that the closed-loop system in (3)
is stabilized by the designed control Algorithm 1. Before that, we
recall a lemma which extends the Barbalat’s lemma (Slotine & Li,
1991).
Lemma 11 (Jiang and Nijmeijer (1999)). For any differentiable
function f : R+ → R, if limt →∞ f (t ) = 0, and its derivative can
be written as f˙ (t ) = f0 (t ) + η(t ) with f0 being uniformly continuous
and limt →∞ η(t ) = 0, then limt →∞ f˙ (t ) = 0 and limt →∞ f0 (t ) = 0.
Theorem 12. For the system in (3), if the control law is designed
by using Algorithm 1 in which the parameters are designed as in
Algorithm 2, then the closed-loop system is locally asymptotically
stable at the origin.
Proof. Using the idea as in Fontes (2001), denote the optimal value
of the optimization problem P (t , z (t ), T −(t − tk )) by Wtk (t , z (t )),
and define U (t , z (t )) , Wδ (t , z (t )), with δ = sup{tk 6 t }.
According to the principle of Dynamic programming, one can
obtain that
t

U (t , z (t )) = Wtk (tk , z (tk )) −
∗ ∗
L(z ∗ (s), u∗ (s), s)ds. (34)
(33) tk
Note that Wtk (tk , z ∗ (tk )) = J (z ∗ (s; tk ), u∗ (s; tk ), s) and Wtk+1 (tk+1 ,
z ∗ (tk+1 )) 6 J (z̄ (s; tk+1 ), ū(s; tk+1 )). Thus,
Wtk+1 (tk+1 , z ∗ (tk+1 )) − Wtk (tk , z ∗ (tk ))
6 J (z̄ (s; tk+1 ), ū(s; tk+1 ), s) − J (z ∗ (s; tk ), u∗ (s; tk ), s).
Furthermore, by plugging the detailed terms into J (z̄ (s; tk+1 ),
ū(s; tk+1 ), s) − J (z ∗ (s; tk ), u∗ (s; tk ), s), we can get
J (z̄ (s; tk+1 ), ū(s; tk+1 ), s) − J (z ∗ (s; tk ), u∗ (s; tk ), s)
tk+1

=− L(zi∗ (s; tk ), u∗i (s; tk ), s)ds
tk
 tk+1 +T
+ L(z̄ (s; tk ), ū(s; tk ), s)ds
tk +T
+ F (z̄ (tk+1 + T ; tk+1 ), tk+1 + T ) − F (z̄ (tk + T ; tk ), tk + T ).
 H. Li et al. / Automatica 75 (2017) 144–153 151

Because Aρ , Bu1 and Bu2 are designed as in (33), one has

L(z̄ (s; tk ), ū(s; tk ), s) 6 Ξ (ε)[A0 (µ1 , µ2 )ρ 4 (x̄(s; tk ), ω̄3 (s; tk )) +
C01 (µ1 , µ2 )(ω̄1 (s; tk ) − v1 (x̄(s; tk ), ω̄3 (s; tk ), s))4 + C02 (µ1 , µ2 )
(ω̄tk2+(1s+;Ttk ) − v2 (x̄(s; tk ), ω̄3 (s; tk ), s)) ]. Using (31), it follows that
4

tk + T
L(z̄ (s; tk ), ū(s; tk ), s)ds + F (z̄ (tk+1 + T ; tk+1 ), tk+1 + T ) −
F (z̄ (tk + T ; tk ), tk + T ) 6 0, and J (z̄ (s; tk+1 ), ū(s; tk+1 ), s) −
t
J (z ∗ (s; tk ), u∗ (s; tk ), s) = − t k+1 L(z ∗ (s; tk ), u∗ (s; tk ), s)ds. As a
k
result,
 t
U (t , z ∗ (t )) − U (tk , z ∗ (tk )) 6 − L(z ∗ (s), u∗ (s), s)ds, (35)
tk

which implies that the closed-loop system is Lyapunov stable.

Next we need to prove that the states converge to the origin.
Using (35) further yields
 t
U (t , z ∗ (t )) − U (t0 , z ∗ (t0 )) 6 − L(z ∗ (s), u∗ (s), s)ds
t0
t Fig. 1. Comparison of state x1 (t ).

6− ρ 4 (x∗ (s), ω3∗ (s))ds. (36)
t0

result, for all t > tk , U (t , z (t )) is bounded, and this implies

As a  ∗
t
that t ρ 4 (x∗ (s), ω3∗ (s))ds is bounded for all t > t0 . Because the
0
system in (3) is continuous and the control input u∗ (s) is bounded,
dρ 4 (x∗ (s),ω3∗ (s))
ds
is bounded. By using the Barbalat’s Lemma (Slotine
& Li, 1991), we can conclude that limt →∞ ρ 4 (x∗ (s), ω3∗ (s)) = 0,
which leads to limt →∞ x∗ (t ) = 0 and limt →∞ ω3∗ (t ) = 0.
Now reconsider the system in (3), we have
 
3
1 
∗
ẋ1 = ω1 + ω3 x2 − ω2 x3 +
∗ ∗ ∗ ∗ ∗
ωi xi x1 ,
∗ ∗ ∗
2 i=1
 
3
1 
∗
ẋ2 = ω2 − ω3 x1 + ω1 x3 +
∗ ∗ ∗ ∗ ∗
ωi xi x2 .
∗ ∗ ∗
2 i=1

Since limt →∞ x∗ (t ) = 0 and limt →∞ ω3∗ (t ) = 0, limt →∞ [ω3∗ (t )

x∗2 (t ) − ω2∗ (t )x∗3 (t ) + i=1 ωi∗ (t )x∗i (t )x∗1 (t )] = 0. Furthermore, due
3
to the continuity of (3) and the bounds of control input u∗1 (t ), ω̇1∗ (t )
is bounded for all t > t0 . Therefore, the result in Lemma 11 can be
used, which implies that limt →∞ ω1∗ (t ) = 0. For the same way,
limt →∞ ω2∗ (t ) = 0. The proof is completed.
6. Simulation and comparison studies
This section presents the simulation and comparison studies for
the designed MPC Algorithm 1.
For the system in (3), by setting j1 = 500, j2 = 400 and
j3 = 100, U1 = [−1000, 1000] and [−800, 800], the parameters
c1 = 0.6, c2 = −1 and c3 = 1, and the control bound sets are
Γ1 = [−2, 2] and Γ2 = [−2, 2]. Design V (x, ω3 ) = x41 + x42 + x23 +
ω14 + ω24 + ω32 , which satisfies the condition in (11).
In Lemma 6, α1 = 1/2, α2 = 1/8 and ϵ0 is calculated as
0.0029. According to Lemma 7, K1 = 5.82e − 2, K2 = 2e + 5,
K3 = 8.26e + 3 and K4 = 9.69e + 3. µ1 and µ2 are designed as
µ1 = 8e + 11, and µ2 = 8e + 11, which satisfies the condition
in (32). Finally, εT is estimated to be 1e − 12, Aρ = 0.039, and
Bu1 = Bu2 = 5.76e − 4e. In the simulation, the prediction
horizon is taken as T = 1.6 s, and the sampling period is τ =
0.1 s. To implement Algorithm 1 numerically, the system in (3) is
discretized according to the chosen sampling period τ , the control
input trajectory over the prediction horizon T is approximately
represented by T /τ = 16 control input variables at each step, and
the cost functions in Problem P becomes a summation from the
integral. The simulation is conducted by using the MATLAB package
and the optimization algorithm is executed by fmincon function.
Fig. 2. Comparison of state x2 (t ).
For the same conditions, we simulate the closed-loop systems
using Algorithm 1, and the time-varying control law (Morin &
Samson, 1997), respectively. The simulated results are reported
as follows: The comparisons for the system states are displayed
in Figs. 1–6. The comparisons for the control inputs are shown in
Figs. 7–8.
From Figs. 1–6, it can be seen that (1) the designed MPC
algorithm stabilizes the system in (3); (2) it takes about 5 s
for the MPC algorithm to stabilize the closed-loop system, but
it needs more than 20 s for time-varying control law in Morin
and Samson (1997), implying that the designed MPC algorithm
can provide a much faster convergence speed than that of the
conventional control algorithm in Morin and Samson (1997). From
Figs. 7–8, it can be observed that the designed MPC algorithm
fulfills the control input constraints. However, the conventional
time-varying control violates the control input constraints. In
summary, these comparison results on states and control inputs
verify the effectiveness and advantages of the designed MPC
algorithm.
7. Conclusions
In this paper, we have studied the MPC-based stabilization
problem of a class of nonholonomic spacecrafts with only two
control torques and subject to control input constraints. By making
use of the homogeneity of the system dynamics, a novel MPC
152 H. Li et al. / Automatica 75 (2017) 144–153

Fig. 3. Comparison of state x3 (t ). Fig. 6. Comparison of state ω3 (t ).

Fig. 4. Comparison of state ω1 (t ).

Fig. 7. Comparison of control u1 (t ).

Fig. 5. Comparison of state ω2 (t ).

Fig. 8. Comparison of control u2 (t ).

algorithm has been designed. In particular, we have developed a
systematic approach to design parameters for the MPC algorithm.
We have proven that under the way of designing these parameters,
the designed MPC algorithm is iteratively feasible, the closed-loop effectiveness and superiority of the designed MPC algorithm. The
system is asymptotically stable and the control input constraints future study would consider the MPC for nonholonomic spacecraft
are fulfilled. Simulation and comparison studies have verified the with disturbances and unmodeled dynamics.
 H. Li et al. / Automatica 75 (2017) 144–153
References
Brockett, R. W. (1982). Asymptotic stability and feedback stabilization. In Differen-
tial geometry control theory (pp. 181–191).
Byrnes, C. I., & Isidori, A. (1991). On the attitude stabilization of rigid spacecraft.
Automatica, 27(1), 87–95.
Carravetta, F. (2014). Almost-global exponential state-feedback stabilization of an
underactuated rigid-body in 3D. In Proceedings of the IEEE 53rd annual conference
on decision and control. CDC (pp. 5612–5617).
Casagrandea, D., Astolfi, A., & Parisini, T. (2008). Global asymptotic stabilization of
the attitude and the angular rates of an underactuated non-symmetric rigid
body. Automatica, 44(7), 1781–1789.
Chen, H., & Allgöwer, F. (1998). A quasi-infinite horizon nonlinear model predictive
control scheme with guaranteed stability. Automatica, 34(10), 1205–1217.
Fontes, F. A. C. C. (2001). A general framework to design stabilizing nonlinear model
predictive controllers. Systems & Control Letters, 42(2), 127–143.
Gui, H., & Vukovich, G. (2015). Robust adaptive spin-axis stabilization of a
symmetric spacecraft using two bounded torques. Advances in Space Research,
56(11), 2495–2507.
Guiggiani, A., Kolmanovsky, I., Patrinos, P., & Bemporad, A. (2014). Fixed-point
constrained model predictive control of spacecraft attitude. arXiv:1411.0479.
Hegrenæ, Ø., Gravdahl, J. T., & Tøndel, P. (2005). Spacecraft attitude control using
explicit model predictive control. Automatica, 41(12), 2107–2114.
Jiang, Z.-P., & Nijmeijer, H. (1999). A recursive technique for tracking control of
nonholonomic systems in chained form. IEEE Transactions on Automatic Control,
44(2), 265–279.
Kalabic, U., Gupta, R., Di Cairano, S., Bloch, A., & Kolmanovsky, I. (2014). Constrained
spacecraft attitude control on so(3) using reference governors and nonlinear
model predictive control. In Proceedings of the 2014 American control conference.
ACC (pp. 5586–5593).
Li, H., & Shi, Y. (2014a). Distributed receding horizon control of large-scale nonlinear
systems: Handling communication delays and disturbances. Automatica, 50(4),
1264–1271.
Li, H., & Shi, Y. (2014b). Robust distributed model predictive control of constrained
continuous-time nonlinear systems: A robustness constraint approach. IEEE
Transactions on Automatic Control, 59(6), 1673–1678.
Li, H., Shi, Y., & Yan, W. (2016). Distributed receding horizon control of constrained
nonlinear vehicle formations with guaranteed γ -gain stability. Automatica, 68,
148–154.
Li, H., Xie, P., & Yan, W. (2016). Receding horizon formation tracking control of
constrained underactuated autonomous underwater vehicles. IEEE Transactions
on Industrial Electronics, http://dx.doi.org/10.1109/TIE.2016.2589921.
Li, H., & Yan, W. (2015). Receding horizon control based consensus scheme in
general linear multi-agent systems. Automatica, 49(7), 1031–1036.
Li, H., & Yan, W. (2016). Model predictive stabilization of constrained underactuated
autonomous underwater vehicles with guaranteed feasibility and stability.
IEEE/ASME Transactions on Mechatronics, http://dx.doi.org/10.1109/TMECH.
2016.2587288.
Marchand, N., & Alamir, M. (2003). Receding horizon stabilization of a rigid
spacecraft with two actuators. Journal of Dynamic Systems, Measurements, and
Control, 125(3), 489–491.
MCloskey, R. T., & Murray, R. M. (1997). Exponential stabilization of driftless
nonlinear control systems using homogeneous feedback. IEEE Transactions on
Automatic Control, 42(5), 614–628.
Michalska, H., & Mayne, D. Q. (1993). Robust receding horizon control of
constrained nonlinear systems. IEEE Transactions on Automatic Control, 38(11),
1623–1633.
Morin, P., & Samson, C. (1997). Time-varying exponential stabilization of a rigid
spacecraft with two control torques. IEEE Transactions on Automatic Control,
42(4), 528–534.
Morin, P., Samson, C., Pomet, J.-B., & Jiang, Z.-P. (1995). Time-varying feedback
stabilization of the attitude of a rigid spacecraft with two controls. Systems &
Control Letters, 25(5), 375–385.
Pomet, J.-B., & Samson, C. (1993). Time-varying exponential stabilization of
nonholonomic systems in power form. Rr-2126, INRIA.
Rosier, L. (1992). Homogeneous lyapunov function for homogeneous continuous
vector field. Systems & Control Letters, 19, 467–473.
Slotine, J.-J. E., & Li, W. (1991). Applied nonlinear control. Englewood Cliffs: Prentice
Hall.
153
Teel, A. R., & Sanfelice, R. G. (2008). On robust, global stabilization of the attitude
of an underactuated rigid body using hybrid feedback. In Proceedings of 2008
American control conference (pp. 2909–2914).
Tsiotras, P. (1996). Stabilization and optimality results for the attitude control
problem. Journal of Guidance, Control, and Dynamics, 19(4), 772–779.
Tsiotras, P. (1997). Optimal regulation and passivity results for axisymmetric rigid
bodies using two controls. Journal of Guidance, Control, and Dynamics, 20(3),
457–464.
Tsiotras, P., & Luo, J. (1997). Reduced-effort control laws for underactuated rigid
spacecraft. Journal of Guidance, Control, and Dynamics, 20(6), 1089–1095.
Tsiotras, P., & Luo, J. (2000). Control of underactuated spacecraft with bounded
inputs. Automatica, 36(8), 1153–1169.
Huiping Li, received his B.Sc. and M.Sc. degrees in mechan-
ical engineering and automatic control from Northwestern
Polytechnical University, Xi’an, China, in 2006 and 2009,
respectively, his Ph.D. degree in mechanical engineering
from University of Victoria, BC, Canada, in 2013. From May
to October 2013, he was a Postdoctoral Research Fellow
with the Department of Mechanical Engineering, Univer-
sity of Victoria, BC. Since November 2013, he has been with
the School of Marine Science and Technology, Northwest-
ern Polytechnical University, Xi’an, where he is currently
a Full Professor. His research interests include model pre-
dictive control, networked control systems, multi-agent systems, and distributed
and cooperative control of underwater vehicles. He was a Guest Technical Editor
for a Special Section in IEEE/ASME Transactions on Mechatronics.
Weisheng Yan, received the B.Sc., M.Sc. and Ph.D. degrees
in mechanical engineering and automatic control from
Northwestern Polytechnical University (NWPU), Xi’an,
China, in 1990, 1993 and 1999, respectively. Since
1993, he has been with the Department of Mechanical
Engineering and Automation, School of Marine Science
and Technology, NWPU, Xi’an, China, as Lecturer (1993),
Associate Professor (1999) and Full Professor (2003). From
2003 to 2013, he served as the Dean of the Department of
Mechanical Engineering and Automation. Since 2013, he
has been serving as the Associate Dean of the School. His
research interests are mainly focused on advanced control theory and control of
underwater vehicles.
Yang Shi received the Ph.D. degree in electrical and
computer engineering from the University of Alberta,
Edmonton, AB, Canada, in 2005. From 2005 to 2009, he
was an Assistant Professor and Associate Professor in
the Department of Mechanical Engineering, University of
Saskatchewan, Saskatoon, Saskatchewan, Canada. In 2009,
he joined the University of Victoria, and now he is a
Professor in the Department of Mechanical Engineering,
University of Victoria, Victoria, British Columbia, Canada.
His current research interests include networked and
distributed systems, model predictive control, system
identification, mechatronics, navigation and control of autonomous systems (AUV
and UAV), and energy system applications.
Dr. Shi received the University of Saskatchewan Student Union Teaching
Excellence Award in 2007, the Faculty of Engineering Teaching Excellence Award at
the University of Victoria in 2012. He received the JSPS Invitation Fellowship (short-
term), and was a Visiting Professor at the University of Tokyo during Nov–Dec
2013. He serves as Associate Editor for IEEE Trans. Control Systems Technology,
IEEE/ASME Trans. Mechatronics, IEEE Trans. Fuzzy Systems, IET Control Theory and
Applications, ASME Journal of Dynamic Systems, Measurement, and Control. He is
currently a senior member of IEEE, a member of ASME, and a registered Professional
Engineer in British Columbia, Canada.