Received September 29, 2018, accepted October 14, 2018, date of publication October 17, 2018, date of current

version November 9, 2018.

Digital Object Identifier 10.1109/ACCESS.2018.2876558

Finite-Time Backstepping Control for Quadrotors

With Disturbances and Input Constraints
TAO JIANG , DEFU LIN, AND TAO SONG
Beijing Key Laboratory of UAV Autonomous Control, Beijing Institute of Technology, Beijing 100081, China
Corresponding author: Tao Jiang ([email protected])
This work was supported by the National Science Foundation of China under Grant U1613225.

ABSTRACT This paper proposes a novel finite-time backstepping framework combined with an auxiliary
input-saturation compensator and applies it to tackle the trajectory tracking problem of quadrotors with
disturbances and input constraints. Finite-time convergence property is guaranteed by introducing the
fractional power functions of tracking errors and the finite-time filters of target commands for the next
step. To avoid the singularity of rotation motion, quaternion-based attitude control is adopted, in which the
target command is obtained from the finite-time filter of desired thrust direction. Additionally, finite-time
disturbance observers based on a multivariable super-twisting algorithm are employed to estimate the lumped
perturbations, and then, we directly counteract them. To handle the adverse effect of input saturation, a novel
auxiliary system is developed to provide fast desaturation when input saturation occurs. The singularity of the
auxiliary dynamics is avoided by the cubic representation of the auxiliary variables. A rigorous proof of the
finite-time stability of the closed-loop system is derived by the Lyapunov theory, despite the presence of input
saturation and disturbances. Finally, several comparative simulations and experimental results demonstrate
the effectiveness and superiority of the proposed method.

INDEX TERMS Finite-time backstepping, trajectory tracking, multivariable super-twisting, finite-time

filter, input saturation compensator.

I. INTRODUCTION backstepping (BS) [8]–[14], model predictive control [15],

Unmanned quadrotor helicopters have received wide interest singular perturbation theory [16] and sliding mode (SM) con-
among the research community, due to their capabilities- trol [17], etc. As a powerful Lyapunov-based tool, recursive
vertical taking off and landing (VTOL), broad envelope of BS is generally utilized as a baseline controller due to the cas-
flight ranging from hovering to cruising, potential to fly at low caded structure of quadrotor dynamics. BS framework mainly
altitude, and highly agile maneuvering in tightly constrained has two great benefits: 1) it enables to provide accommoda-
environments [1], [2]. In practical application scenes, heli- tion of nonlinearities and avoid wasteful cancelations [18];
copters are usually required to quickly and accurately track 2) the different flight modes, such as position, velocity, and
a prescribed trajectory. However, controlling unmanned heli- attitude control mode can be integrated into the helicopters’
copters is a challenging task due to their complex nonlinear control system due to the hierarchical framework of BS.
dynamics, strong couplings, significant parameter uncertain- However, there exist several problems in traditional BS
ties and external disturbances [3]. method. First, for high-order nonlinear systems BS design
Traditional linear control methods are advantageous to causes ‘‘explosion of complexity’’ by the repeated differ-
project realization, and have been widely used in helicopters entiation of certain nonlinear functions. Second, traditional
control system design, including PID [4], linear quadratic reg- BS doesn’t provide a robustness performance guarantee
ulation [5], and H∞ robust control [6]. However, these meth- against perturbations. To tackle the limitations, a robust neu-
ods suffer from performance degradation when the quadrotor ral adaptive BS control is designed for quadrotors in [13],
helicopters leave away from their designed trim points or exe- where a radial basis function neural approximator is intro-
cute aggressive maneuvers. To tackle these limitations, var- duced to estimate and compensate for the perturbations. This
ious nonlinear control approaches have been developed for work [14] develops a novel immersion and invariance based
quadrotors’ controller design, including dynamic inverse [7], adaptive controller for quadrotors with the uncertain inertial

2169-3536
2018 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 6, 2018 Personal use is also permitted, but republication/redistribution requires IEEE permission. 62037
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 T. Jiang et al.: Finite-Time Backstepping Control for Quadrotors With Disturbances and Input Constraints

parameters. Whereas disturbances can be finally eliminated feedback inputs [26], [28], [29]. Our work improves the
through these adaptive techniques, they are rejected in a input-saturation compensator in [28] and [29] by the cubic
relatively slow and indirect manner. Recently, active distur- representation of auxiliary variables, such that the singularity
bance rejection control (ADRC) has been proposed to achieve of the auxiliary dynamics is avoided. Moreover, the auxiliary
rapid compensation of disturbance influence [19], where the dynamics are developed to guarantee the finite-time stability
lumped unknown disturbances are estimated by a disturbance of the closed-loop system.
observer (DO) and compensated for online [20]. The BS con- The main contributions of our work are listed as follows.
troller proposed by [10] applies command filtering to com- 1) The multivariable composite finite-time BS framework
pute the derivatives of the virtual control signals, and DO to is proposed. The fractional powers of tracking errors are
estimate and attenuate the effect of the disturbing forces and introduced to improve convergence property near the trim
moments. To guarantees accurate tracking in the presence of point. The derivatives of the target commands are estimated
lumped disturbances, [11] combines dynamic surface control by the finite-time filters, which can avoid ‘‘differential explo-
(DSC) with extent state observer. Under the DO based control sion’’. Moreover, multivariable super-twisting algorithm is
framework, high-precision and fast tracking for quadrotors developed to design finite-time DO, which is applied to com-
can be achieved. pensate for the disturbances.
To further improve the performance of BS control, our 2) Trajectory tracking problem of quadrotors subjected to
work develops finite-time convergence property for BS con- disturbances and input constraints is tackled by applying
troller. Finite-time controllers enable to improve control per- the proposed finite-time BS. The control system guarantees
formance when states are near the equilibrium point, which finite-time convergence of tracking errors. To avoid the sin-
provide a faster transient response, a higher precision and a gularity of rotation motion, the control law is designed based
better disturbance rejection than the controllers with asymp- on the quaternion-based attitude dynamics. The target com-
totic stability property [21]. There are two main theory mand for the angle loop is computed by filtering the desired
frameworks to complete finite-time stability analysis, includ- direction of thrust.
ing homogenous method [21] and Lyapunov-based method 3) A novel auxiliary compensation system for input saturation
[22], [23]. In [24] and [25], finite-time stabilizing controllers is proposed, which makes a significant improvement on the
for quadrotor’s trajectory tracking are designed based on method [28], [29] by introducing the cubic representation
the homogeneous technique. In our work, Lyapunov-based of the auxiliary variables. It provides fast desaturation when
method is combined with BS technique to design and analyze input saturation occurs. Additionally, this compensator not
the control system. First, to achieve finite-time convergence only guarantees finite-time convergence of the closed-loop
the fraction powers of the tracking errors are applied. Sec- system, but also has no existence of singularity.
ond, the finite-time filters are employed to approximate the 4) Finite-time stability of the closed-loop system with our
time derivative of desired commands, such that ‘‘differential proposed controller is proven based on the finite-time Lya-
explosion’’ is avoided. Third, in order to guarantee robustness punov theory. Some comparative simulations have illustrated
performance, the lumped disturbances are estimated by finite- the effectiveness and superiority of our proposed control law.
time DO based on multi-variable super-twisting algorithm, Moreover, the effectiveness of this method is investigated by
and eliminated directly in the corresponding BS step. The practical flight experiments.
simulations and experiments are conducted to demonstrate The configuration of this paper is arranged as follows.
the effectiveness of the finite-time controller. To the best Section II presents the dynamic model of quadrotor heli-
of our knowledge, no published experimental results on the copters and states our control problem. Section III proposes
finite-time control for trajectory tracking of quadrotor are the composite finite-time control method, and provides finite-
available till now. time stability analysis of the closed-loop system. Section IV
Another innovation of our work is to propose a novel input- discusses the comparison simulations conducted in this work.
saturation compensator, which not only quickly desaturates The experiment results are illustrated in Section V. Section VI
the control inputs, but also guarantees finite-time stability draws the conclusions.
of the closed-loop system with input constraints. It is well Notations: Given a vector x = [x1 , · · · , xn ] ∈ Rn , kxk =
known that a satisfactory control system should guarantee (x x)1/2 denotes the Euclidean norm of vector. Define a
T

nominal tracking performance and robustness when input multivariable sign function sign(x) = x/kxk. The power of
saturation occurs [12], [26], [27]. To realize high tracking x is expressed as xm = [|x1 |m sign(x1 ), · · · , |xn |m sign(xn )]T .
performance with large initial errors or/and disturbances, For any vector l ∈ R3 , S(l) is a skew-symmetric matrix so that
the controller may generate control inputs which exceed the l × n = S(l)n, where × represents the vector cross product.
maximum forces and torques that the actuators can produce.
This saturation problem would potentially give rise to system II. MATHEMATICAL MODEL AND PROBLEM STATEMENT
performance degradation or even loss of stability. Tradition- This section presents the nonlinear dynamics of an unmanned
ally, the saturation problem can be handled by introducing quadrotor helicopter. The helicopter is considered a six-
an auxiliary system, which takes the differences between degree-of-freedom rigid body model with simplified force
the virtual control input and the actual actuator input as and moment generation process. First, two reference frames

62038 VOLUME 6, 2018

T. Jiang et al.: Finite-Time Backstepping Control for Quadrotors With Disturbances and Input Constraints

Quaternion multiplication has the property R(q1 )R(q2 ) =

R(q1 ⊗ q2 ). With the identity element qI = 1 03×1 T ,

T T
each q = qw qv T has an inverse q−1 = qw −qv T ,
   

so that q−1 ⊗ q = q ⊗ q−1 = qI . Considering the vector

l ∈ R3 , the rotation action using quaternion is achieved with
the double product,

V(R(q) · l) = q ⊗ V(l) ⊗ q−1 (5)

 T T
where V(l) = 0 l denotes the ‘‘lift’’ operator. If the
angular velocity ω in BRF is known, the derivative of q is
FIGURE 1. ‘‘X’’ type quadrotor. represent as follows:

−qv T
   
q̇w 1 1
q̇ = = q ⊗ V(ω) = ω (6)
are defined as follows: earth reference frame (ERF) I = q˙v 2 2 qw I3 + S(qv )
{Oxyz}, which is fixed to the earth; body reference frame
If the helicopter’s attitude 2 is given, its unit quaternion
(BRF) B = {Ob xb yb zb }, whose origin is located at the
representation can be derived as follows:
helicopter’s center of gravity [1]. Fig. 1 shows the direction of
the two reference frames. The dynamic model of helicopter q(2)
can be described as follows [3]: φ θ
      
ψ φ θ
     
ψ
cos cos cos −sin sin sin
Ṗ = V

 2 2 2 2 2 2  
1 φ θ ψ φ θ ψ
            
V̇ = ge3 +
 
R(2)F cos sin sin +sin cos cos 
m  2 2 2 2 2 2 
=      
 
Ṙ(2) = R(2)S(ω)
cos φ sin θ cos ψ +sin φ cos θ sin ψ 
     
 
J ω̇ = −ω × J ω + M (1) 
 2 2 2 2 2 2  
T T
φ θ ψ φ θ ψ 
              
where P , x y z and V , Vx Vy Vz refer to the 
helicopter’s position and velocity vector in the ERF, respec- cos cos sin −sin sin cos
2 2 2 2 2 2
tively; m is the mass, and g is the gravitational acceleration (7)
T
constant; e3 , 0 0 1 is a unitary vector; J represents the
approximate inertia matrix In (1), F and M are the external force and torque vector
  exerted on fuselage in BRF, respectively.
Jxx 0 0    
J ≈0 Jyy 0 (2) 0 Mφ
0 0 Jzz F =  0  + 1Fe , M =  Mθ  + 1Me (8)
−T Mψ
The rotation matrix R(2) from  BRF to ERF is dependent
on the three Eular angles 2 , φ θ ψ , which denote  roll,
where T , Mφ , Mθ , and Mψ are the generated thrust and
pitch, and yaw angles, respectively. ω , ωx ωy ωz denote moment from the actuator output. 1Fe and 1Me are the
the angular rates in the BRF. external force and moment disturbances, involving gust dis-
Here, unit quaternion representations are applied to turbances and the drag of fuselage. Aiming at quadrotor
describe the rotational dynamics, the total lift force and control torque are generated
 motion T to 4avoid singularity prob-
lems [30]. Let q = qw qv T ∈ R denote unit quaternion as follows:
which parameterizes the rotation matrix R, where qw ∈ R  
T
 
u1
 
ut
and qv ∈ R3 are commonly refer to as the ‘‘scalar’’ and  Mφ  u2   uφ 
‘‘vector’’ parts of unit quaternion q. Unit quaternion satisfies   = Ap Aa   = Ap  
 Mθ  u3   uθ  (9)
that q2w +qv T qv = 1. The transformation from unit quaternion Mψ u4 uψ
to rotation matrix is defined as
where ui ∈ (0,  1] (i = 1,2, 3, 4) is the throttle input of
R(q) = I3 + 2qw S(qv ) + 2S(qv )2 (3) T
the ith motor. ut uφ uθ uψ denote the inputs to control
each individual channel, including the altitude, roll, pitch,
where In denotes an n × n identity matrix.TMultiplication and yaw angle channels. Ap is the matrix that reflects the
between two quaternions, qi = qi,w qi,v T , i ∈ {1, 2} ,
quadrotor’s physical characteristics, including its size, servo
is defined as
mounting position, motor characteristics, etc. Aa is the control
q1,w q2,w − q1,v T q2,v
 
allocation matrix which depends on the type of helicopters.
q1 ⊗ q2 = (4)
q1,w q2,v + q2,w q1,v + S(q1,v )q2,v Practically, the real control inputs are subjected to saturation

VOLUME 6, 2018 62039

 T. Jiang et al.: Finite-Time Backstepping Control for Quadrotors With Disturbances and Input Constraints

nonlinearities due to the physical limitations of actuators. The Aiming at the quadrotors subjected to input con-
input constraints are expressed as: straints (10) and disturbances under Assumption 1, design
a robust saturated control law, such that quadrotors’ actual
ui = sat(ui,c , ui,min , ui,max ) trajectory tracks the desired position command Pd (t) in finite

ui,max , if ui,c > ui,max
 time.
= ui,c , if ui,max ≥ ui,c ≥ ui,min (10)


ui,min , if ui,c < ui,min III. PRELIMINARIES
Some preliminary lemmas which will be utilized latter are
where i = {t, φ, θ, ψ}; ui,c is our designed control command; stated as follows:
ui,min and ui, max are the minimum and maximum constraints Lemma 1 [23]: Let c, d be positive real numbers and
of the input commands, respectively. Considering ‘‘X’’ type γ (x, y) > 0 a real-valued function. Then
quadrotor equipped with four rotors (see Fig. 1), we have
    cγ (x, y) c+d d
kf 0 0 0 1 1 1 1 |x|c |y|d ≤ |x| + γ (x, y)−c/d |x|c+d
 0 k f l1 0 0 c+d c+d
, Aa = −1 1 1 −1

Ap ≈ 
Lemma 2 [23]: The inequality (|x1 | + · · · |xn |)p ≤ |x1 |p +

0 0 k f l2 0   1 1 −1 −1
0 0 0 kq 1 −1 1 −1 · · · + |xn |p hold for xi ∈ R, i = 1, · · · , n, where p is a real
(11) number satisfying 0 < p ≤ 1.
Lemma 3 [31]: Consider system ẋ = f (x), x ∈ Rn .
where kf and kq are the coefficients that reflects the relation- Suppose there exists a continuous function V (x) such that the
ship from motors’ throttle input (Electrical Speed Controller, following conditions hold:
ESC) to thrust and torque generated by propellers, respec- (i) V (x) is positive definite.
tively; l1 and l2 are the constant lengths related to the size of (ii) There exist real numbers p1 ∈ (0, 1), p2 ∈ (−∞, p1 ),
quadrotor helicopter; Combining (1) with (8), we obtain α > 0, and β > 0 such that
kf R(2)1Fe V (x) ≤ −αV (x)p1 + βV (x)p2 < 0 (15)
V̇ = ge3 − R(2)e3 · ut +
m m
= ge3 − R(2) · kt0 ut + 1F (12) Then, the system is finite-time uniformly ultimately bounded.
 −1
Jxx kf l1 · uφ
 The set of attraction D is given by D = {x : V (x)p1 −p2 <
ω̇ = −J −1 (ω × J ω) + J −1 1Me +  Jyy
−1 k l · u  β/θ}, where θ ∈ (0, α) is a positive constant. The settling
f 2 θ
time T that needs to reach the set {x ∈ D} from {x ∈ / D} can
Jzz−1 kq · uψ
  be bounded as T ≤ V (x0 )1−p1 /[(α − θ)(1 − p1 )].
kφ0 · uφ
=  kθ 0 · uθ  + 1M (13) A. MULTIVARIABLE FINITE-TIME DISTURBANCE
kψ0 · uψ OBSERVER
where kt0 , kφ0 , kθ0 , and kψ0 are the approximation of the Consider system
control coefficients, which can be approximated through one x˙1 = −k1 kx1 k1/2 sign(x1 ) + x2
crude identification experiment. 1F and 1M are the lumped
x˙2 = −k2 sign(x1 ) + 11 (t) (16)
disturbances caused by external perturbations, parameter
uncertainties, and unmodeled dynamics. The dynamics (12) where x1 , x2 ∈ are the state variables; 11 (t) ∈
Rn is the Rn
and (13) can be further simplified by introducing the interme- unknown perturbation term; k1 and k2 are the design positive
diate control signals, which is described as gains. The time-varying perturbation 11 (t) is assumed to
satisfy that its derivative is bounded, i.e., k11 (t)k ≤ cδ ,
V̇ = ge3 + UV + 1F
where cδ is the constant.
ω̇ = UM + 1M (14) Lemma 4 [32], [33]: Aiming at system (16) subjected to
the disturbances, if the gains satisfy
where UV = −R(2)e3 · kt0 ut , UM = Kω0 [uφ uθ uψ ]T ,
Kω0 = diag(kφ0 , kθ0 , kψ0 ).
3cδ
For completing the design of finite-time DO, the follow- k1 > √ , k2 > 2cδ (17)
ing assumptions are required for the lumped disturbances: k2 − cδ
Assumption 1: The bounded 1F and 1M satisfy that Then, all trajectories of the states converge in finite time to the
their first time derivatives are bounded, i.e., k1̇F k ≤ C11 , origin x1 D x2 D 0, upperbounded by b T which depends on
k1̇M k ≤ C12 . k1 , k2 , and cδ .
Assumption 2: The desired trajectory command Pd (t) ∈ Rn Now, consider a perturbed system as follows:
is smooth and has the bounded high-order derivatives.
ẋ = f (x) + Ku u + d(t) (18)
Assumption 1 about the boundedness of first derivative is
one of the common assumptions for DO design [20]. The where x ∈ Rn
is the state vector; f (x) is the known function;
control objective of our work is stated as follows: u ∈ Rn denotes the input; d(t) ∈ Rn expresses the unknown

62040 VOLUME 6, 2018

T. Jiang et al.: Finite-Time Backstepping Control for Quadrotors With Disturbances and Input Constraints

disturbance vector due to model uncertainties or external system in the following section. The design parameters of our
disturbances, which satisfies that its derivative is bounded, DO include K1d , K2d , K3d , and K4d . The larger the values
i.e., kḋ(t)k ≤ cd . Ku ∈ Rn×n is the control coefficient of these gains, the faster the estimation errors converge. But
matrix which is invertible. The finite-time DO is expressed too large values may cause violent chattering phenomenon.
as follows: Thus, the selection of the observer gains is a trade-off process.
Practically, the gains are adjusted based on the results of the
x̂˙ = f (x) + Ku u − K1d kx̂ − xk1/2 sign(x̂ − x) + d̂ simulation and experimentation.
d̂˙ = −K2d sign(x̂ − x) (19)
IV. QUADROTOR FINITE-TIME CONTROL DESIGN
where K1d and K2d are the design positive constants. The A. CONTROL LAW DESIGN
convergence property of DO (19) is given by the following
1) STEP 1: POSITION LOOP
lemma.
Define the position error as Pe = P − Pd . From (1), the posi-
Lemma 5: Consider the system (18) with disturbances.
tion error dynamics are given as follows:
Applying the observer (19) with proper parameters, the esti-
mation error ed = d̂ − d will converge to zero after finite Ṗ e = V − Ṗ d (24)
time. Proof: Define the estimation error as ex = x̂ − x.
The error dynamics are given as follows: Design the virtual control

e˙x = −K1d kex k1/2 sign(ex ) + ed Vd = Ṗ d − Kp kPe kα1 sign(Pe ) (25)

e˙d = −K2d sign(ex ) − ḋ (20) where Kp is the position constant. The positive constant 12 <
α1 < 1 represents the finite-time convergence property of
Following Lemma 4 and the boundness of the derivative of
the system. Let the desired velocity command vector pass
d(t), we conclude that the state of system (20) will converge
through the finite-time filter
to zero after finite time. Then, we have
V˙ t = −KVf kVt − Vd kα2 sign(Vt − Vd ) (26)
x̂ = x and d̂ = d, when t ≥ td
where the constant α2 = 2 − α11 > 0.
Thus, the proof of this lemma is completed.
Remark 3: The traditional BS methods which are used
Applying Lemma 5, the lumped disturbances 1F and 1M
to avoid ‘‘differential explosion’’ include dynamic surface
in (14) can be estimated by the following finite-time DO.
control [11] and command filter BS [10]. These methods
V̂˙ = UV + ge3 − K1d kV̂ − V k1/2 sign(V̂ − V ) + 1ˆF smooth the target signal and estimate its derivative by a low-
˙ = −K sign(V̂ − V ) pass filter, whereas they just provide asymptotic convergence
1̂ F 2d (21) of estimation errors. In our finite-time BS framework, finite-
˙
ω̂ = UM − K3d kω̂ − ωk1/2 sign(ω −
ˆ ω) + 1ˆM time filters are proposed to approximate the derivatives of the
˙ = −K sign(ω −
1̂ ˆ ω) (22) target signals, while the finite-time convergence property is
M 4d
guaranteed.
where 1̂F and 1̂M denote the estimation of the lumped
disturbances 1F and 1F , respectively. According to Assump- 2) STEP 2: VELOCITY LOOP
tion 1 and Lemma 5, we conclude that the estimation error Let velocity error Ve = V − Vt . Following (14), we have
1ˆF − 1F and 1ˆM − 1M will converge to zero after finite
time, i.e., V˙e = ge3 + UV + 1F − V˙ t (27)

k1̂F − 1F k = 0, k1̂M − 1M k = 0, when t > T1 (23) The desired control is designed as follows:

where T1 is the convergence time which is determined by Uv,d = −ge3 + V˙ t − 1ˆF − Kv kVe kα2 sign(Ve ) (28)
Lemma 4 [33]. According to the definition of Uv , the throttle control signal
Remark 1: It should be note that the multivariable version can be calculated as follows:
of finite-time DO is implemented to compensate for the
effect of disturbances. Although the scalar algorithm can be kUv,d k + ξt2α1 −1
ut,c = , R3d = −Uv,d /kUv,d k (29)
obtained based on the decoupled structure of the system (14), kt0
some improved performances can be derived by using the where ξt is an auxiliary variable to handle the saturation. Its
multivariable algorithm, such as the chattering reduction and dynamics are constructed as follows:
the applicability of the proposed method to a class of non-
α1 2
decoupled uncertain systems [32], [33]. χ̇t = −Kχt |ξt |α1 +2 sign(ξt ) − Kχt ,f fχt ξt + Ut,1 ξt (30)
Remark 2: Actually, there are lots of versions of
DO [34]–[36]. Here, the finite-time DO (19) is used to guar- where χt = 3 ξt .
1 3
Kχt and Kχt ,f are the positive constants.
2
α1 −2
antee finite-time convergence of the estimation error, which fχt = kVe k Ve T R3d Ut,1 + |Ut,1 |1+α1 , where Ut,1 =
is also applied to prove finite-time stability of the closed-loop kt0 (ut − ut,c ). Note that there exists a singularity in (29) if

VOLUME 6, 2018 62041

 T. Jiang et al.: Finite-Time Backstepping Control for Quadrotors With Disturbances and Input Constraints

α1 +1 2−αc2
kUv,d kâĂ− = 0. Actually, this case happens rarely, unless ≤ −Kγ ,1 kγe k αc2
+ kγe k αc2
kR˙3d k (36)
the drone sharply declines. Following [3], [10], and [11],
the singular case can be avoided by constraining the reference Considering the boundedness of Uv,d , it yields kR˙3d k ≤
trajectory and the control gains. Furthermore, to completely CU kU ˙v,d k, where CU is a constant.
avoided the singularity, the saturation is applied as
Uv,d,3 = sat(Uv,c,3 , −∞, −T ) (31) 3) STEP 3: ANGLE LOOP
Define the rotation error Re = RTt · R. Its unit quaternion
where Uv,c,3 is the third element of the control vector Uv,d representation is given qe = [qe,w qe,v T ]T = qt T ⊗q. Using
in (28). The vertical channel of Uv,d is constrained by the the property of unit quaternion, its time derivative is given as
bounded constant −T < 0. Then, it is concluded that follows:
kUv,d k ≥ T . These strategies are reliable and simply to
realize and apply in the practical flight controller. 1
q˙e = qe ⊗ V(ω − RTe ωt ) (37)
Define the error vector of the thrust direction γ = 2
[γ1 γ2 γ3 ]T = RTt R3d . Rt = R(qt ), where qt denotes the Design the virtual angular rate ωd as
target quaternion. The following quaternion-based finite-time
filter is proposed ωd = RTe ωt + ωe,d (38)
1
q˙t = qt ⊗ V(ωt ) (32) ωe,d = −Kq · sign(qe,w )kqe,v kα1 sign(qe,v ) (39)
2
The target angular rate ωt = [ω1t ω2t ω3t ]T is The finite-time filter is expressed as:
expressed as
ω˙tf = −Kωf kωtf − ωd kα2 sign(ωtf − ωd ) (40)
α α
ω1t = −Kγ γ2 3 , ω2t = Kγ γ1 3 , ω3t = ω3d (33)
where ωtf ∈ R3 is the state vector of the filter; Kωf > 0 is a
where the constant α3 = α1 +α c2 −1
αc2 > 0, αc2 = 2α1 − 1. constant design gain.
Considering that the length of the error vector γ is equal to 1, Remark 6: In the intermediate control law (38), the item
it requires another freedom to be specified. Hence, ω3d are sign(qe,w ) is used to avoid the unwinding phenomenon of
chosen to be obtained from outside, which is often produced the quaternion-based control [30]. Additionally, the fraction
by the real-time remote control or the prescribed trajectory. powers of attitude tracking error qe,v are introduced to guar-
Furthermore, from the expression of ωt it is obviously found antee finite-time convergence property of attitude tracking.
that the derivative of target attitude motion q˙t is bounded. Thus, it achieves global finite-time stability of attitude track-
Remark 4: According to the definition of ξt , the proposed ing.
auxiliary system (30) can be rewritten as:
Kχt ,f fχt α1 4) STEP 4: ANGULAR RATE LOOP
ξ̇t = −Kχt |ξt |α1 sign(ξt ) − + Ut,1 (34)
ξt Define the angular rate error ωe = [ωe,1 ωe,2 ωe,3 ]T =
In the above expression, there exists singularity when ξt = 0. ω − ωtf ∈ R3 , which dynamics are given as:
Applying our improved control form, this problem can be
avoided. Actually, our method is a significant improvement to ω˙e = UM + 1M − ω˙tf (41)
the input saturation auxiliary method in [28] and [29], which
applies the boundary layer approach to avoid singular prob- The control inputs are designed as follows:
lem. This causes the discontinuous switching and the inexact
UM,c = −Kω · kωe kα2 sign(ωe ) − 1ˆM + ω˙tf + ξM 2α1 −1
convergence. By introducing the cubic representation of the
−1
auxiliary state in (30), these problems have been overcome, [uφ,c uθ,c uψ,c ]T = Kω0 UM,c (42)
while finite-time convergence of auxiliary system has been
achieved by the proposed representation (30). where Kω is the positive control gain to be designed; ξM =
Remark 5: The time derivative of γ is described as [ξM ,1 ξM ,2 ξM ,3 ]T are the auxiliary variables for anti-
windup compensation, which are expressed as:
γ̇ = −S(ωt )γ + RTt R˙3d (35)
αc2
2 χ̇M ,i = −KχM |ξM ,i |α1 +2 sign(ξM ,i ) − KχM ,f fχM ,i ξM ,i
αc2
Define the positive Lyapunov function Vγ = 2 kγe k ≤ α1
,1,i ξM ,i
2
1 + UM (43)
(γ12 + γ22 ) αc2 , where γe = e3 − γ . Applying Lemma 2, its
derivative is expressed as where χM = [χM ,1 χM ,2 χM ,3 ]T and χM ,i = 31 ξM ,i .
2−αc2 KχM and KχM ,f are the positive constants. fχM ,i =
V̇γ = −kγe k αc2
γ̇3 2
−2
2−αc2 kωe k α1 ωe,i T UM ,1,i + |UM ,1,i |1+α1 , where UM ,1 =
1+α3 1+α3
≤ kγe k αc2
(−Kγ (|γ1 | + |γ2 | ) + kR˙3d k) [UM ,1,1 UM ,1,2 UM ,1,3 ]T = UM − UM,c .

62042 VOLUME 6, 2018

T. Jiang et al.: Finite-Time Backstepping Control for Quadrotors With Disturbances and Input Constraints

B. PROOF OF STABILITY where the positive constant C1 is dependent on the derivatives

The closed-loop error dynamics can be divided into the inner- of qt . In the light of (40) and Lemma 1, we have
loop part (attitude tracking) and the outer-loop part (position 
α1 2
0
α1
tracking), which are expressed as: kωef k
2
The outer-loop part:
2 h i
= kωef k α1 ωef T − Kωf kωef kα2 sign(ωef ) − ω˙d
−2
P˙e = −Kp kPe kα1 sign(Pe ) + Ve + Vef
V˙e = −KV kVe kα2 sign(Ve ) + 1F − 1ˆF 1+α1 2−α1
≤ (Kωf − γωf ,2 )kωef k α1
+γq,2 kqe,v k1+α1 +C1 kωef k α1
− R3d (Ut,1 + ξt2α1 −1 ) − (Rt − Rd + R − Rt )e3 kt0 ut
(44) (49)
where γωf ,2 and γq,2 are the positive constants. Considering
The inner-loop part:
the control law (43), it follows that
−qv T −Kq sign(qe,w )kqe,v kα1 sign(qe,v )
  
1 0
α1

q˙e = 2
α1
2 qw I3 +S(qv ) +ωef +ωe kωe k
α2
2
ω˙e = −Kω · kωe k sign(ωe ) + 1M − 1ˆM α2
T −Kω · kωe k sign(ωe ) + 1M − 1M
ˆ
 
2
α1 −2
+UM,1 + ξM 2α1 −1 (45) = kωe k ωe
+ UM,1 + ξM 2α1 −1
where Vef = Vt − Vd , ωef = ωtf − ωd ∈ R3 . 1+α1 2−α1
≤ − Kω − γωe ,2 kωe k α1 + kωe k α1 k1M − 1ˆM k


The proof process is divided into two steps. First, the atti- 2
−2
tude tracking error qe is proven to finite-time converge to a + γξ,1 kξ k1+α1 + kωe k α1 ωe T UM,1
small region near zeros. Then, we illustrate that the state vec-
tors of the outer loop will not escape to infinite in any finite where γωe ,2 and γξ,1 are the proper constants produced by
time. Finally, the conclusion that the quadrotor’s trajectory inequality transformation. From (43), we can get
reaches to a small region near the desired trajectory in finite  0
1 2
time, is drawn. kξM k
2KχM ,f
Theorem 6: Consider the inner-loop error dynamics (45)
under disturbances with Assumption 1. The control law is 3
1 X
given as the form in (42) with the finite-time filter (32), finite- = ξM ,i ξ̇M ,i
Kχ ,M
time DO (22) and auxiliary system (43). If the control gains i=1
are selected properly, the attitude tracking error qe can finite- 3
−KχM 1
ξM T UM,1 α1
X
time converge to a small region of origin, while all signals ≤ kξM k1+α1 − fχM ,i +
in the closed-loop system are bounded in any finite time. KχM ,f KχM ,f
i=1
Proof: 1). Assign the positive Lyapunov function KχM 2
−2
α1 2 ≤ −( − γξ,2 )kξM k1+α1 − kωe k α1 ωe T UM,1
L = (1 − |qe,w |)2 + kqe,v k2 + kωe k α1 KχM ,f
2
(50)
α1 2 kξM k2
+ kωef k α1 + (46)
2 2KχM ,f where γξ,2 is a positive constant. From the above analysis,
the derivative of the Lyapunov function can be simplified as
For obtaining the time derivative of L, some simplifications
follows:
are needed. Following Lemma 1, the intermediate control law
1+α1 1+α1
(38) and the properties of the unit quaternion, we have L̇ ≤ −Cq kqe,v k1+α1 − Cωf kωf k α1
− Cωe kωe k α1
h i0
(1 − |qe,w |)2 + kqe,v k2 2−α1 2−α1
−Cξ kξM k1+α1 +kωe k α1
k1M − 1ˆM k+C1 kωef k α1

= 2sign(qe,w )q̇e,w
(51)
≤ −Kq kqe,v k1+α1 + kqe,v kkωe k + kqe,v kkωef k
1+α1 where Cqv = Kq − γq,1 − γq,2 , Cωf = Kωf − γωf ,1 − γωf ,2 ,
≤ −(Kq − γq,1 )kqe,v k1+α1 + γωe ,1 kωe k α1
K
1+α1 Cωe = Kω − γωe ,1 − γωe ,2 and Cξ = KχχM,f − γξ,1 − γξ,2 .
M
+ γωf ,1 kωf k α1
(47) To guarantee stability of the closed-loop system, the control
gains should be selected such that the following conditions
where γq,1 , γωe ,1 , and γωf ,1 are the positive constant gener-
are satisfied:
ated by the transformation of inequalities. According to the
definition of the intermediate control (38) and the bounded- Cqv > 0, Cωf > 0, Cωe > 0, Cωe > 0 (52)
ness of the reference signals qt and ωt , it yields
From Lemma 5, it has been proven that the norm of
kω˙d k ≤ γ1 kqe,v k2α1 −1 + C1 (48) the estimation error k1M − 1ˆM k converges to zero in

VOLUME 6, 2018 62043

 T. Jiang et al.: Finite-Time Backstepping Control for Quadrotors With Disturbances and Input Constraints

finite time T1M . It follows that k1M − 1ˆM k must be control (28) and its auxiliary system (30), it yields
bounded. Hence, there exists a constant, i.e., CM 1 such that 0
α1

2
k1M − 1ˆM k ≤ CM 1 in any finite time. Obviously, kVe k α1

From (51) and Lemma 2 the following inequality is always 2

 α1 +1 2−α1 
satisfied −Kv kVe k α1 + kVe k α1 k(R(qe ) − I3 )e3 kkt0 ut
1+α1 2−α1
 2−α1 2−α1 
L ≤ −CL,1 L + BL,1 L (53) ≤  +kVe k α1 k1F − 1ˆF k + kVe k α1 kγe kkt0 ut 
2 2

2−2α1

where CL,1 > 0. The constant BL,1 > 0 depends on CM 1 −kVe k α1

Ve T R3d (Ut,1 + ξt2α1 −1 )
α1 +1 2−α1
and C1 . According to Lemma 3, the states are uniformly
 
−(Kv − γVe ,2 )kVe k α1 + kVe k α1 k1F − 1ˆF k
ultimately bounded. After a finite time T1 , the estimate error  2−2α1 
will converge to zero, i.e., k1M − 1ˆM k = 0. Then, we ≤ −kVe k α1 Ve T R3d Ut,1 + γξt ,1 |ξt |α1 +1

 
2−α1 α1 +1
obtain + 4kt0 ut kVe k α1
kqe,v k + γγe ,1 kγe k αc2
1+α1 2−α1
L ≤ −CL,1 L 2 + BL,2 L 2 (54) (58)
0
|ξt |2

where the constant BL,2 is only dependent on C1 . Applying 2Kχt ,f
Lemma 3, it is concluded that Lyapunov function will reach  
−( Kχχt,f − γξt ,2 )|ξt |α1 +1
K
the set of attraction DM after finite time TM . The set of t
2α1 −1 ≤ 2−2α1 (59)
attraction DM = L 2 < BL,2 /θM , where θM ∈ (0, CL,1 ).

+kVe k α1
Ve T R3d Ut,1
The tracking error can be arbitrarily small by choosing the
large enough control gains. Due to the positive definiteness where γVe ,2 , γξt ,1 , and γξt ,2 are the constants. Following
of L, we have that the attitude tracking error qe,v is bounded Assumption 1-2, eq. (26), (28), (36) and applying Lemma 1-
in any time and will converge to Dq = {qe,v : kqe,v k < Bq }, 2, it yields
when t > TM . 
αc2 2
0
αc2
Theorem 7: Consider the error dynamics (44) and (45) kγe k
2
subjected to disturbances with Assumption 1, which control 2−αc2
law [ut,c uφ,c uθ,c uψ,c ]T is given by (29) and (42) with ≤ kγe k αc2 (−Kγ (|γ1 |1+α1 + |γ2 |1+α1 ) + kR˙3d k)
the finite-time filters (26), (32) and (40), finite-time DOs (21)  α1 +1

−(K − γ )kγ k αc2
+ γ kP kα1 +1
(22) and auxiliary systems (30) (43). If the control gains are γ ,1 γe ,2 e Pe ,3 e
α1 +1 α1 +1
 
selected properly, the origin of the error dynamics is finite- ≤  +γV ,3 kVe k α1 + γV ,3 kVef k α1
 
 (60)
e ef
time uniformly ultimately bounded. Proof: Select the  2−α1

positive Lyapunov function +Cγe kγe k αc2

1 α1 2 α1 2 Thus, the derivative of Lyapunov function is described as

LV = kPe k2 + kVe k α1 + kVef k α1 α1 +1 α1 +1
2 2 2
L̇V ≤ −CPe kPe kα1 +1 − CVef kVef k α1
− CVe kVe k α1

|ξt |2 αc2 2 α1 +1 2−α1

+ + kγe k αc2 (55) − Cξt |ξt |α1 +1 − Cγe kγe k αc2
+ kVe k α1
k1F − 1ˆF k
2Kχt ,f 2
2−α1 2−α1 2−α1
Similarly, some inequality transformations are stated for the + 4kt0 ut kVe k α1
kqe,v k+CVf kVef k α1
+Cγe kγe k αc2
following proof process. Applying Lemma 1, from the imme- (61)
diate control (25) and the finite-time filter (26) we have
 0 where CPe = Kp − γP,1 − γP,2 − γP,3 . CVef = KVf − γVf ,1 −
1 K
γVf ,2 − γVf ,3 . CVe = Kv − γVe ,1 − γVe ,2 − γVe ,3 . Cξt = Kχχt,f −
 
kPe k2 = Pe T − Kp kPe kα1 sign(Pe ) + Ve + Vef t
2 γξt ,1 −γξt ,2 . Cγe = Kγ ,1 −γγe ,1 −γγe ,2 . The outer-loop control
α1 +1 gains are properly selected such that the following conditions
≤ −(Kp − γP,1 )kPe kα1 +1 +γVf ,1 kVef k α1
are satisfied:
α1 +1
+ γVe ,1 kVe k α1 (56) CPe > 0, CVef > 0, CVe > 0, Cξt > 0, Cγe > 0 (62)
α1 +1
The estimation error k1F − 1ˆF k is bounded and converges
 
0
α1  −(KVf − γVf ,2 )kVef k
α1

2
kVef k α1 ≤ 2−α1
 to zero when t > T1 . Moreover, from Theorem 6, the attitude
2 +γP,2 kPe kα1 +1 + CVf kVef k α1 tracking error kqe,v k is bounded and converge to the set Dq
(57) when t > TM . So it is inferred that Lyapunov function LV and
the states of the closed-loop system are bounded in any finite
where γP,1 , γVe ,1 , γVf ,1 , γVf ,2 , and γP,2 are the constants time. Furthermore, when t > max⥹{TM , T1 }, we have
produced by inequality transformations. CVf is the constant α1 +1 2−α1
2 2
which depends on the desired command. Then, following the L̇V ≤ −CLV ,1 LV + BLV ,1 LV (63)

62044 VOLUME 6, 2018

T. Jiang et al.: Finite-Time Backstepping Control for Quadrotors With Disturbances and Input Constraints

TABLE 1. Parameters of quadrotor helicopter. as follows:

(
[0 0 0]T , if t ≤ 8s
Vw = (65)
[5 5 0]T , if t > 8s
The force and moment perturbation caused by wind is
expressed as follows:
1Fw,i = −0.5ρAb,i BVhw,i
2
sign(BVhw,i ), i = {x, y, z}
1Mw = [0 0 0]T (66)
where BV = RT (V
− Vw ) denotes the helicopter’s
hw
relative airspeed in BRF; ρ is the air density; Ab =
2α1 −1 [Abx Aby Abz ]T = [0.02 0.02 0.04]T m2 is a vector
So we calculate the set of attraction DV = LV 2 < BLV ,1 /θV , including the quadrotor’s effective drag areas along each
where θV ∈ (0, CLV ,1 ). Due to positive definiteness of LV , axis of BRF. The position of CoM in BRF is PC =
the ultimate boundedness of all signals in the closed-loop [PCx PCy 0]T = [0.03 0.03 0]T . The force and torque
system can be guaranteed. In addition, by increasing the disturbances caused by CoM are described as
values of observer bandwidths and control gains and reducing T
1FM = 0 0 0

the values of time constants in filtering, the ultimate upper
bound for tracking errors can be made arbitrarily small. This
 
0 kf PCy 0 0
concludes the proof. 1Mw = 0 0 kf PCx 0 Aa U (67)
Remark 7: All of design parameters include KP , KV , Kq , 0 0 0 0
Kω for nominal tracking performance, KVf , Kγ , Kωf for filter
part, and Kχt , Kχt ,f , KχM , KχM ,f for input saturation auxiliary The offset of CoM has a significant impact on the roll and
compensation. The gains KP , KV , Kq , Kω are the most impor- pitch moments by altering the location of the rotors’ thrust.
tant parameters which are related to the convergence rate of Moreover, the servo dynamics of rotors are also consid-
system’s error, i.e., the larger their values chosen, the faster ered, which can be approximated as a first-order filter
convergence rate of tracking error. However, too large values 1
u̇i,a = (ui − ui,a ), i = 1, 2, 3, 4 (68)
may result in an excessive transient peaking. Thus, it’s a Tr
trade-off process to choose these parameters according to the where Tr = 0.05s is the time constant; ui,a represents the
specific applications. actual ith rotor’s response. The constraints of the control input
Remark 8: Note that traditional BS control (DSC) is a are given as follows:
special case of our finite-time framework, when α1 is set as
1. If we select α1 to satisfy that 21 < α1 < 1 and αi > 0, i = ut,min = 0.1, ut,max = 2
1, 2, 3, finite-time convergence property of the closed-loop
system can be guaranteed. The following comparison simu- ui,min = −1, ui,max = 1, (i = φ, θ, ψ)
lation proves that finite-time convergence property provides
better transient response and tracking performance. Before designing helicopter’s controller, control coeffi-
cients of the quadrotor model are approximated as kt0 = 10,
V. SIMULATION kφ0 = 400, kθ0 = 400, kψ0 = 300. These parameters are the
In this section, numerical simulation results are presented to only we need to know about the helicopter’s model, which
illustrate the effectiveness of our proposed composite control can be obtained from the crude identification experiments.
law for helicopters. Table 1 summarizes the corresponding The proposed finite-time controller is given in (29) and (42)
helicopter model parameters. Note that all of these parameters with the parameters for each part as follows: α1 = 9/10,
in Table 1 are unknown in our controller design process. KP = 1, KV = 4, Kq = 15, Kω = 40 for the nominal tracking
Following [37], the lumped force and moment disturbances performance of the closed-loop system; K1,d = K3,d = 10,
in (65) are introduced in simulation to test the robustness K2,d = K4,d = 8 for the DO part; Kχt = KχM = 1,
of the helicopter’s control system. These disturbances are Kχt ,f = KχM ,f = 0.1 for the auxiliary system; KVf =
caused by the external disturbances (wind gusts), and the Kγ = Kωf = 15 for the finite-time filter part. Two cases
position of the center of mass (CoM). are simulated to prove the superiority of our proposed control
law.
Case 1: The desired commands are Pd = [10 10 −5]T
1Fe = 1Fw + 1FM
(m), ω3d = 0 (rad/s). The simulation results of the controllers
1Me = 1Mw + 1MM (64) with and without finite-time DO are depicted in Fig. 2 and
titled as ‘‘Proposed’’ and ‘‘Without DO’’. It is demonstrated
Wind disturbance velocity Vw is assumed to possess wind that the finite-time DO provide good performance in estimat-
components along x- and y-direction in ERF. The wind is set ing and counteracting the disturbances. BS method (α1 = 1)

VOLUME 6, 2018 62045

 T. Jiang et al.: Finite-Time Backstepping Control for Quadrotors With Disturbances and Input Constraints

FIGURE 3. The control input with constraints in Case 1.

FIGURE 2. The position responses of the quadrotor in Case 1.

is also used to illustrate the superiority of finite-time control. the position tracking is: ‘‘Proposed’’ > ‘‘BS’’ >‘‘Without
The steady tracking errors of the proposed method in x- and y- DO’’. The framework which combines finite-time conver-
directions are smaller than BS method, while tracking errors gence and disturbance elimination achieves the best tracking
in z-direction are both less than 1 cm. The response results performance. Fig. 5 depicts the attitude response. The angu-
illustrate that finite-time control provides superior transient lar responses of our proposed method have a more intense
and steady tracking performance. Moreover, the control law oscillation to guarantee high tracking performance. Violent
without the input saturation auxiliary system labeled as chattering exists at t = 4s due to the switching of reference
‘‘Without AU’’, is applied to show the effectiveness of our trajectory. Fig. 6 illustrates the angular rate in z-direction can
proposed saturation treatment in Fig. 3. Due to large initial track the desired commands.
errors, the control in thrust channel emerges the saturation
phenomenon. Applying the auxiliary compensation of input VI. EXPERIMENT
saturation, the control command ut can be quickly desatu- Trajectory tracking experiment of quadrotor has been con-
rated. It is noted that the control commands chatter violently ducted to demonstrate the efficiency and control perfor-
to fast eliminate the effect of wind disturbances. mances of the developed algorithm. The experiment testbed
Case 2: The desired ‘‘8-shape’’ reference trajectory is consists of the ‘‘X-type’’ quadrotor (see Fig. 7), the VICON
described as: motion capture system, and Bluetooth pairs. The overall size
of the helicopter platform is 280mm × 320mm × 55mm with
Pd (t) = [0 0 − 4(1 − e−0.3 t )]T (m), for t ≤ 4 s a weight of 550 g. The VICON system is used to estimate
8 − 8 cos 2π(t−4)
 
12
the quadrotor’s position, velocity, yaw angle and yaw angular
Pd (t) =  4 sin 4π(t−4)  (m), for t > 4s velocity. The data of helicopter’s states is sent to quadrotor via
12
−4(1 − e −0.3 t ) Bluetooth at 100 Hz.
ω3d = 0 (rad/s) (69) Before designing our proposed quadrotor’s control sys-
tem, the control coefficients kt0 , kφ0 , kθ 0 , and kψ0 in (12)
Firstly, the quadrotor vertically climbs for 4s to simulate the and (13) are needed to be identified from one crude iden-
take-off stage. Then it follows an ‘‘8-shaped’’ path while tification experiment. The algorithm is based on the least
continuing to lift. In this stage, quadrotor’s lift, sideslip and squares regression. The results of identification are shown
forward performances are evaluated comprehensively. Sim- in Fig. 8. The control coefficients are obtained as kt0 = 40.3,
ulation results are illustrated in Figs. 4-6. Fig. 4 shows the kφ0 = 1273, kθ 0 = 881, and kψ0 = 276. The following
three dimensional position response curves of the closed- experiment of trajectory tracking is used to demonstrate the
loop system based on the above three methods. Accuracy of control performance of the composite finite-time controller.

62046 VOLUME 6, 2018

T. Jiang et al.: Finite-Time Backstepping Control for Quadrotors With Disturbances and Input Constraints

FIGURE 6. The angular rate ωz response of the quadrotor in Case 2.

FIGURE 7. The quadrotor experiment platform.

FIGURE 4. The position responses of the quadrotor in Case 2.

FIGURE 5. The angle responses of the quadrotor in Case 2.

FIGURE 8. Quadrotor identification results.

The desired circle trajectory is described as follows:

[0 0 − 0.2t ] ,
 2 T if t ≤ 2s

0 − 0.8] , T if 2 < t ≤ 6s Firstly, the quadrotor vertically climbs for 2s with a accelera-

[0


  tion of 0.4 m/s2 . Then it hovers for 4s. Finally, it tracks a circle
Pd (t) = 0.75(1 − cos 2π 5 (t − 6))
trajectory 4 times. Through this experiment, quadrotor’s lift,
2π
− 6)  if t > 6s

0.75 sin (t

  



 5 hover, and tracking performances are evaluated comprehen-
 0.8 sively. Additionally, to further show the performance of the
ω3d (t) = 0 rad/s (70) developed method, it is compared with the BS controller with

VOLUME 6, 2018 62047

 T. Jiang et al.: Finite-Time Backstepping Control for Quadrotors With Disturbances and Input Constraints

with an error less than 1.5 cm, which illustrates that our
controller enables to eliminate the adverse effect of dynamic
coupling. From the comparative results, it can be seen that
the finite-time backstepping control has better transient per-
formance, and faster convergence property. There exist large
parameter uncertainties caused by the crude identification.
Through the finite-time DO, these uncertainties can be elim-
inated and the nominal control performance is recovered.
Actually, without DO the proposed control system is difficult
to achieve the hover flight, not to mention the maneuver
flight. The attitude response is shown in Fig. 10. To achieve
high tracking performance, the desired attitude commands
have a bit chattering. Note that the desired yaw position
is unlocked, which depends on the finite-time filtering pro-
cess of quaternion, and the desired angular rate signal ω3d
obtained from the remote control or prescribed signals.

VII. CONCLUSION
In this paper, the problem of trajectory tracking is handled for
quadrotor helicopters with parameter uncertainties, external
disturbances, and actuator’s constraints. In the control sys-
FIGURE 9. The position responses when tracking the circle. tem, finite-time BS framework is applied to provide finite-
time convergence property of tracking errors. To improve
robustness, the disturbances are estimated by the finite-time
DOs based on multi-variable super-twisting algorithm, and
counteracted directly in the corresponding loop. Moreover,
a novel auxiliary system is proposed to compensate for
the adverse effect of actuators constraints. The compara-
tive simulation results have illustrated the advantages of
our proposed approach, including finite-time stability, high-
precision tracking, disturbances cancellation, and fast desat-
uration. The practical flight experiments further show the
effectiveness of the control system.

REFERENCES
[1] Q. Quan, Introduction to Multicopter Design and Control. Singapore:
Springer, 2017.
[2] R. Mahony, V. Kumar, and P. Corke, ‘‘Multirotor aerial vehicles: Modeling,
FIGURE 10. The angle responses when tracking the circle. estimation, and control of quadrotor,’’ IEEE Robot. Autom. Mag., vol. 19,
no. 3, pp. 20–32, Sep. 2012.
[3] S. Omari, M.-D. Hua, G. Ducard, and T. Hamel, ‘‘Hardware and
software architecture for nonlinear control of multirotor helicopters,’’
DO and the opensource flight controller APM.1 The APM IEEE/ASME Trans. Mechatronics, vol. 18, no. 6, pp. 1724–1736,
is one of the most popular open source flight controllers, Dec. 2013.
which adopts the quaternion-based nest saturation control [4] R. A. García, F. R. Rubio, and M. G. Ortega, ‘‘Robust PID control of the
quadrotor helicopter,’’ in Proc. 2nd IFAC Conf. Adv. PID Control, Brescia,
[3] and realizes a variety of flight modes. BS + DO is a Italy, Oct. 2012, pp. 229–234.
recent classical framework for flight control [10], [11], and [5] S. Bouabdallah, A. Noth, and R. Siegwart, ‘‘PID vs LQ control techniques
our proposed controller is the development and improvement applied to an indoor micro quadrotor,’’ in Proc. IEEE/RSJ Int. Conf. Intell.
Robots Syst. (IROS), Sendai, Japan, Sep./Oct. 2004, pp. 2451–2456.
of this approach. [6] G. V. Raffo, M. G. Ortega, and F. R. Rubio, ‘‘An underactuated H∞ control
The experiment results are depicted in Fig. 9-10. In the strategy for a quadrotor helicopter,’’ in Proc. Eur. Control Conf., Budapest,
climb stage, the quadrotor with the proposed controller takes Hungary, 2009, pp. 3845–3850.
[7] A. Das, K. Subbarao, and F. Lewis, ‘‘Dynamic inversion with zero-
less than 4s to climb to the desired height. Then it hovers at dynamics stabilisation for quadrotor control,’’ IET Control Theory Appl.,
the desired point with average root-mean square (RMS) errors vol. 3, no. 3, pp. 303–314, Mar. 2009.
of 0.7 cm in x-, 0.6 cm in y-, and 1.1 cm in z-direction. During [8] T. Madani and A. Benallegue, ‘‘Backstepping control for a quadrotor heli-
copter,’’ in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Beijing,
the tracking stage, the quadrotor can quickly and exactly track China, Oct. 2006, pp. 3255–3260.
the circle trajectory. Moreover, it keeps at the specified height [9] W. Jasim and D. Gu, ‘‘Integral backstepping controller for quadrotor
path tracking,’’ in Proc. Int. Conf. Adv. Robot., Istanbul, Turkey, 2015,
1 APM Autopilot Suite: http://ardupilot.com/ pp. 593–598.

62048 VOLUME 6, 2018

T. Jiang et al.: Finite-Time Backstepping Control for Quadrotors With Disturbances and Input Constraints

[10] A. Aboudonia, A. El-Badawy, and R. Rashad, ‘‘Active anti-disturbance [31] Q. Hu, B. Jiang, and Y. Zhang, ‘‘Observer-based output feedback attitude
control of a quadrotor unmanned aerial vehicle using the command- stabilization for spacecraft with finite-time convergence,’’ IEEE Trans.
filtering backstepping approach,’’ Nonlinear Dyn., vol. 90, no. 1, Control Syst. Technol., to be published.
pp. 581–597, Oct. 2017. [32] I. Nagesh and C. Edwards, ‘‘A multivariable super-twisting sliding mode
[11] X. Shao, J. Liu, H. Cao, C. Shen, and H. Wang, ‘‘Robust dynamic sur- approach,’’ Automatica, vol. 50, no. 3, pp. 984–988, 2014.
face trajectory tracking control for a quadrotor UAV via extended state [33] B. Tian, L. Liu, H. Lu, Z. Zuo, Q. Zong, and Y. Zhang, ‘‘Multivariable finite
observer,’’ Int. J. Robust Nonlinear Control, vol. 28, no. 7, pp. 2700–2719, time attitude control for quadrotor UAV: Theory and experimentation,’’
May 2018. IEEE Trans. Ind. Electron., vol. 65, no. 3, pp. 2567–2577, Mar. 2018.
[12] R. Wang and J. Liu, ‘‘Trajectory tracking control of a 6-DOF quadrotor [34] J. Han, ‘‘From PID to active disturbance rejection control,’’ IEEE Trans.
UAV with input saturation via backstepping,’’ J. Franklin Inst., vol. 355, Ind. Electron., vol. 56, no. 3, pp. 900–906, Mar. 2009.
no. 7, pp. 3288–3309, 2018. [35] Q.-C. Zhong, A. Kuperman, and R. K. Stobart, ‘‘Design of UDE-based
[13] M. A. M. Basri, ‘‘Trajectory tracking control of autonomous quadrotor controllers from their two-degree-of-freedom nature,’’ Int. J. Robust Non-
helicopter using robust neural adaptive backstepping approach,’’ J. Aerosp. linear Control, vol. 21, no. 17, pp. 1994–2008, 2011.
Eng., vol. 31, no. 2, p. 04017091, Mar. 2018. [36] L. B. Freidovich and H. K. Khalil, ‘‘Performance recovery of feedback-
[14] Y. Zou and Z. Meng, ‘‘Immersion and invariance-based adaptive controller linearization-based designs,’’ IEEE Trans. Autom. Control, vol. 53, no. 10,
for quadrotor systems,’’ IEEE Trans. Syst., Man, Cybern., Syst., to be pp. 2324–2334, Nov. 2008.
published. [37] G. Antonelli, E. Cataldi, F. Arrichiello, P. R. Giordano, S. Chiaverini, and
[15] K. Alexis, G. Nikolakopoulos, and A. Tzes, ‘‘Switching model predictive A. Franchi, ‘‘Adaptive trajectory tracking for quadrotor MAVs in presence
attitude control for a quadrotor helicopter subject to atmospheric distur- of parameter uncertainties and external disturbances,’’ IEEE Trans. Control
bances,’’ Control Eng. Pract., vol. 19, no. 10, pp. 1195–1207, Oct. 2011. Syst. Technol., vol. 26, no. 1, pp. 248–254, Jan. 2018.
[16] S. Bertrand, N. Guénard, T. Hamel, H. Piet-Lahanier, and L. Eck, ‘‘A hier-
archical controller for miniature VTOL UAVs: Design and stability analy-
sis using singular perturbation theory,’’ Control Eng. Pract., vol. 19, no. 10,
pp. 1099–1108, 2011.
[17] B. Zhao, B. Xian, Y. Zhang, and X. Zhang, ‘‘Nonlinear robust sliding TAO JIANG received the B.S. degree from the
mode control of a quadrotor unmanned aerial vehicle based on immersion Beijing Institute of Technology, Beijing, China, in
and invariance method,’’ Int. J. Robust Nonlinear Control, vol. 25, no. 18, 2014, where he is currently pursuing the Ph.D.
pp. 3714–3731, 2016.
degree with the Institute of UAV Autonomous
[18] H. K. Khalil, Nonlinear Systems. Englewood Cliffs, NJ, USA:
Prentice-Hall, 1996.
Control. His main research interests include robust
[19] Y. Huang and W. Xue, ‘‘Active disturbance rejection control: Methodology and nonlinear control of unmanned aircraft vehi-
and theoretical analysis,’’ ISA Trans., vol. 53, no. 4, pp. 963–976, 2014. cles, and development of robotic control systems.
[20] W.-H. Chen, J. Yang, L. Guo, and S. Li, ‘‘Disturbance-observer-based
control and related methods—An overview,’’ IEEE Trans. Ind. Electron.,
vol. 63, no. 2, pp. 1083–1095, Feb. 2016.
[21] S. P. Bhat and D. S. Bernstein, ‘‘Finite-time stability of continu-
ous autonomous systems,’’ SIAM J. Control Optim., vol. 38, no. 3,
pp. 751–766, Jan. 2000.
[22] S. P. Bhat and D. S. Bernstein, ‘‘Continuous finite-time stabilization of DEFU LIN received the Ph.D. degree from the
the translational and rotational double integrators,’’ IEEE Trans. Autom. School of Aerospace Engineering, Beijing Insti-
Control, vol. 43, no. 5, pp. 678–682, May 1998.
tute of Technology, in 2004. He is the Head of the
[23] C. Qian and W. Lin, ‘‘A continuous feedback approach to global strong
Institute of UAV Autonomous Control, and also
stabilization of nonlinear systems,’’ IEEE Trans. Autom. Control, vol. 46,
no. 7, pp. 1061–1079, Jul. 2001. the Director of the Beijing key laboratory of UAV
[24] X.-N. Shi, Y.-A. Zhang, and D. Zhou, ‘‘Almost-global finite-time trajectory Autonomous Control. He is currently a Professor
tracking control for quadrotors in the exponential coordinates,’’ IEEE of flight vehicle design with the Beijing Institute
Trans. Aerosp. Electron. Syst., vol. 53, no. 1, pp. 91–100, Feb. 2017. of Technology. He focuses on the overall design
[25] B. Tian, H. Lu, Z. Zuo, Q. Zong, and Y. Zhang, ‘‘Multivariable finite-time of fixed-wing and rotor-wing flight vehicles.
output feedback trajectory tracking control of quadrotor helicopters,’’ Int.
J. Robust Nonlinear Control, vol. 28, no. 1, pp. 281–295, 2018.
[26] N. Cao and A. F. Lynch, ‘‘Inner–outer loop control for quadrotor UAVs
with input and state constraints,’’ IEEE Trans. Control Syst. Technol.,
vol. 24, no. 5, pp. 1797–1804, Sep. 2016.
[27] S. Li, Y. Wang, and J. Tan, ‘‘Adaptive and robust control of quadro- TAO SONG received the Ph.D. degree from
tor aircrafts with input saturation,’’ Nonlinear Dyn., vol. 89, no. 1, the School of Aerospace Engineering, Beijing
pp. 255–265, 2017. Institute of Technology. He completed his Post-
[28] M. Chen, S. S. Ge, and B. Ren, ‘‘Adaptive tracking control of uncertain
Doctoral Research at the Institute of Flight Sys-
MIMO nonlinear systems with input constraints,’’ Automatica, vol. 47,
tem Dynamic, Technical University of Munich,
no. 3, pp. 452–465, Mar. 2011.
[29] J. Du, X. Hu, M. Krstić, and Y. Sun, ‘‘Robust dynamic positioning of in 2016. He is currently an Associate Professor
ships with disturbances under input saturation,’’ Automatica, vol. 73, with the Institute of UAV Autonomous Control,
pp. 207–214, Nov. 2016. Beijing Institute of Technology, Beijing, China.
[30] C. G. Mayhew, R. G. Sanfelice, and A. R. Teel, ‘‘Quaternion-based hybrid He focuses on flight system identification in the
control for robust global attitude tracking,’’ IEEE Trans. Autom. Control, frequency domain.
vol. 56, no. 11, pp. 2555–2566, Nov. 2011.

VOLUME 6, 2018 62049