This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS 1

Constrained Adaptive Robust Trajectory Tracking

for WIP Vehicles Using Model Predictive
Control and Extended State Observer
Ming Yue, Cong An, and Zhijun Li, Senior Member, IEEE

Abstract—This paper is concerned with a model predictive I. I NTRODUCTION

control (MPC) technique together with an adaptive robust scheme
VER the past few decades, wheeled inverted pendu-
for the trajectory tracking of a wheeled inverted pendulum vehi-
cle in the absence of platform velocity information, in addition to
taking dynamic uncertainty and external disturbance into con-
O lum (WIP) vehicles have become widely available and
continuously attracted much attentions owing to its high
sideration. Specifically, to deal with velocity information loss and mobility, simple structure and low energy characteristics
dynamic uncertainty, an extended state observer is introduced to (see [1]–[8] and the reference herein). For practical imple-
evaluate the velocity vectors and model dynamics, where the uni-
mentation, the control objective for a WIP vehicle yields two
formly ultimately bounded property of observer system can be
guaranteed by using the Lyapunov stability theorem. With these aspects: 1) the trajectory tracking for a reference path given
observations, MPC is employed for the underactuated longitudi- by earth-fixed frame which is closely related to the vehicle
nal subsystem to achieve longitudinal velocity tracking, as well posture and 2) the stability for the unstable pendulum-like
as holding the pendulum-like vehicle body stability, and in par- vehicle body. In particular, the vehicle body stability is the
ticular taking the state and input saturation into account; at the prerequisite of the trajectory tracking which needs to be first
same time, an adaptive robust controller is constructed for the
rotational subsystem to realize the rotational velocity tracking, ensured when the vehicle moves. However, unfortunately, for
in which the adaptive laws can enhance the vehicle adaptabil- the sake of simplification of structure and cost reduction, the
ity in diverse environment. In addition, a saturated trajectory WIP vehicle possesses only two drive motors installed on the
generator with closed-loop characteristic is introduced so as to left and right wheels; consequently, there has no other force or
properly handle the velocity limitation and nonholonomic con- torque which can be directly utilized to stabilize the inverted
straint simultaneously. The simulation results validate that the
control system is robust against the external disturbance and pendulum-like vehicle body. But there exists coupled nonlinear
model uncertainty, thereby demonstrating the effectiveness and dynamics between the vehicle motions and inverted pendulum
robustness of the proposed control strategy. tilt, and then just this sophisticated coupled dynamics turns
Index Terms—Adaptive robust control, extended state observer, out to be the sole control effect to make the unstable vehicle
input saturation, model predictive control (MPC), wheeled body hold upright position. Therefore, the WIP vehicle is a
inverted pendulum (WIP) vehicle. classical underactuated mechanical system, characterized by
controlling more variables with fewer actuators, which gives
rise to the difficulty for WIP vehicle control system synthesis.
The underactuated mechanical systems in recent years have
been widely used in diverse practical fields, such as surface
ships [9], [10], underwater robots [11], [12], mobile manip-
Manuscript received June 19, 2016; revised August 19, 2016; accepted ulators [13], [14], acrobat or pendubot linkages [15], [16],
October 19, 2016. This work was supported in part by the National Natural tower cranes [17], [18], humanoid robots [19], [20], and so
Science Foundation of China under Grant 61573078, Grant 61573147, and on. Correspondingly, as for the control system synthesis of
Grant 91520201, in part by the State Key Laboratory of Robotics and
System under Grant SKLRS2015ZD06, in part by the Guangzhou Research these underactuated plants, numerous efforts have been made
Collaborative Innovation Projects under Grant 2014Y2-00507, in part by for their related problems to various degrees. However, unlike
the Guangdong Science and Technology Research Collaborative Innovation other underactuated systems, the equilibrium of WIP vehi-
Projects under Grant 2014B090901056, and in part by the Guangdong Science
and Technology Plan Project (Application Technology Research Foundation) cle, i.e., the upright position of the inverted pendulum-like
under Grant 2015B020233006. This paper was recommended by Associate vehicle body, is unstable in nature. Once the vehicle body
Editor S. Tong. (Corresponding author: Ming Yue.) is likely beyond the neighborhood of equilibrium point, the
M. Yue is with the School of Automotive Engineering, Dalian University of
Technology, Dalian 116024, China, and also with the State Key Laboratory of WIP vehicle system will be inevitably out of control, because
Robotics and System, Harbin Institute of Technology, Harbin 150001, China the vehicle actuator (drive motor) has a rigorous rated power
(e-mail: yueming@dlut.edu.cn). and only supplies the limited capabilities on account of the
C. An is with the School of Automotive Engineering, Dalian University of
Technology, Dalian 116024, China. physical feature so that this saturated actuator cannot pro-
Z. Li is with Colllege of Automation Science and Engineering, South China vide enough power to stabilize the inverted vehicle body
University of Technology, Guangzhou 510641, China. again. Therefore, the tilt angle of vehicle body, in practice,
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. must be constricted to a relatively small range with respect
Digital Object Identifier 10.1109/TSMC.2016.2621181 to the equilibrium in consideration of the system stability.
2168-2216 c 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

Consequently, these comprehensive issues associated with var-

ious constraints and underactuated problems, must be handled
simultaneously, which is still much more challenging for the
control community up to now.
Model predictive control (MPC) can readily handle with
these problems by using optimization method, where the afore-
mentioned bounded constraint of tilt angle for vehicle body
and actuators saturations can be taken into account as state
and input constraints. Besides, compared with the other tradi-
tional optimization methods, such as linear quadratic regulator,
MPC technique adopts the unique rolling optimization which
can correct the vehicle current dynamics based on the feed-
back and predictive behaviors. To the best of our knowledge,
considerable studies have been developed to apply MPC to
control varied mechanical systems, but only a few focused
on exploiting MPC for the underactuated plants. For instance, Fig. 1. Sketch of a WIP vehicle.
Li et al. [21] presented an MPC method based on a primal-
dual neural network to deal with the tracking problems of arbitrary trajectory given by earth-fixed frame, a trajectory
nonholonomic wheeled mobile robots in which the velocity generator is constructed so as to solve the so-called kinematic
limitation and constrained actuator capability were both con- nonholonomic constraint problem.
cerned. Likewise, as for tracking of underactuated vessels, an To sum up, the main contributions of this paper lie in the
MPC scheme was proposed in [10], [22], and [23], where a following.
single-layer recurrent neural network called the general pro- 1) An extended state observer is constructed for WIP vehi-
jection network was employed to achieve trajectory tracking of cle so as to deal with the unmeasurable velocity problem,
the vessel under arbitrary external disturbance. These works, in which not only the unavailable states but also the
formulating the control objective of developed plants as a uncertain disturbance can be estimated at the same time.
standard quadratic programming (QP), have significant com- The presented observer can reduce the conservation and
putational benefits and light computational burden. Then, the improve the applicability of the presented control system
presented MPC, as an optimization method in essence, covers to the great extent.
a nonlinear optimization problem including constraint condi- 2) A constrained MPC approach is proposed for an under-
tions for a cost function, and a corresponding optimization actuated WIP vehicle with various physical limits, such
method. Motivated by these investigations, this paper attempts as state limitation and input saturation. This approach
to transform the control problem of an underactuated WIP employs receding horizon optimization based on the
vehicle system into a QP problem, and then the optimization transformed QP problem instead of traditional control
method is employed to realize the trajectory tracking of this methods, by which various constraints can be read-
kind of unstable and underactuated mobile platform. ily handled and the robustness will be significantly
Furthermore, the velocity information of vehicle system is enhanced.
usually obtained indirectly by differentiating the posture state The rest of this paper is organized as follows. Section II
or integrating of the acceleration message because the veloc- briefly describes the WIP vehicle model and summarizes var-
ity sensors for a mobile platform are commonly hard to set ious physical limits. Section III presents the extended state
up restricted to the mechanical structure and economic fac- observer design and in the following Section IV shows the
tors. Regarding of this limitation, numerous studies explore trajectory generator and dynamics controller design, includ-
alternative methods to obtain velocity information without ing the adaptive robust control and MPC in details. Section V
velocity sensors. Among these methods, the state observer presents simulation results and Section VI finally concludes
technique is the most represented approach which can esti- this paper.
mate the velocity information of vehicle in realtime, thereby
enhancing the feasibility of control algorithm (see [24]–[31]).
II. F ORMULATION AND C ONTROL O BJECTIVE
However, the observed values are always contaminated with
the unavoidable disturbance caused by varied factors, such as A. WIP Vehicle Formulation
the internal modeling errors, wind force and unpredictable col- A WIP vehicle with an inverted pendulum-like vehicle body,
lision, and so on, which will degrade the system performances as shown in Fig. 1, is driven by two independent motors
seriously and even destroy the system stability if the distur- installed on the wheels, whose movement principle is similar
bance is ignored. To overcome this drawback, an extended to an acrobat that can balance a plate by a connected bam-
state observer is addressed in [32]–[34], by which the inacces- boo on the head. The vehicle motion can be described by four
sible disturbance can be estimated simultaneously. Inspirited generalized coordinates {x, y, ϕ, θ}, where x and y are coordi-
by these contributions, this paper intends to propose an adap- nate position, ϕ denotes the rotation angle of vehicle center,
tive robust control scheme combined with the merits of MPC and θ stands for the relative angle of pendulum-like body with
and extended state observer. To make the vehicle track an respect to vertical direction. Furthermore, for dealing with the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

YUE et al.: CONSTRAINED ADAPTIVE ROBUST TRAJECTORY TRACKING FOR WIP VEHICLES 3

nonholonomic problem, the dynamic model of vehicle can Assumption 1: It should hold that |ω| ≤ ωmax and |υ| ≤
be transformed into a reduced order formula described by υmax , with ωmax and υmax being the maximum positive scalars,
{υ ω θ̇}, where υ and ω represent the longitudinal and rota- for preventing the drive motor from exceeding its maximum
tional velocities, respectively, and θ̇ is the angular velocity of rated power and guaranteeing the vehicle to track the reference
pendulum-like body. Then the entire dynamic model of the trajectory securely. The pendulum tilt angle must strictly be
WIP vehicle can be formulated as restricted within a small scale, e.g., |θ | ≤ θmax , where θmax is
a positive scalar; otherwise, the pendulum-like vehicle body
ω̇ = fω1 + g1 τω + τd1 (1a)
will fall down and then the saturated actuator cannot make the
υ̇ = fυ2 + g2 τυ + τd2 (1b) vehicle return to work again.
θ̈ = fθ3 + g3 τυ + τd3 (1c) Assumption 2: Given that the actuator saturation, the con-
trol inputs have the boundedness, namely, |τl | ≤ τmax and
where τω = τl − τr , τυ = τl + τr with τl and τr being the driv-
|τr | ≤ τmax , where τmax is a positive scalar. Then the corre-
ing torques of left and right wheels, respectively; and τd1 , τd2 ,
sponding control input τυ and τω should satisfy |τυ | ≤ 2τmax
and τd3 represent the disturbances forced on the correspond-
and |τω | ≤ 2τmax .
ing subsystem, respectively. Likewise, the corresponding terms
Assumption 3: The disturbances, i.e., τd1 , τd2 , and τd3 , are
can be given by
also necessary to meet the norm-bounded condition, namely
mp l2 ωθ̇ sin 2θ [τd1 τd2 τd3 ]T  ≤ τd max with τd max being a positive scalar.
fω1 = −
Jϕ + mp l2 sin2 θ The trajectory tracking purpose of this paper is to make
fυ2 the WIP vehicle track the reference trajectory given by earth-
  fixed frame as well as hold the inverted pendulum-like vehicle
1 1 22
= Jθ mp lθ̇ sin θ − mp l ω sin θ cos θ − mp l g sin 2θ
2 2 3 2 2
within a required boundary. Therefore, define a time-varying
Ω 2
fθ3 reference trajectory as {xr , yr , ϕr }, and then the corresponding

  reference velocity can be computed by υr = ẋr2 + ẏ2r and
1 1 1 22 2
= mv mp gl sin θ + mv mp l ω sin 2θ − mp l θ̇ sin 2θ
2 2
ωr = ϕ̇r afterwards. Besides, considering the desired upright
Ω 2 2 position of the vehicle body, the reference tilt angle can be
d Jθ mp l cos θ set as θr = 0. In terms of the reference trajectory, the persis-
g1 =   ; g2 = ; g3 = −
r Jϕ + mp l sin θ
2 2 Ωr Ωr tent excitation condition is able to be satisfied such that the
where mp and mυ represent the mass of the vehicle body and following assumption needs to be hold.
chassis, respectively; Jυ is the moment of inertia of the chas- Assumption 4: The time-varying reference trajectory should
sis and pendulum about vertical direction; Jω and Jc are the be sufficiently smooth and bounded such that second derivative
moment of inertia of the wheel and the chassis with respect of the trajectory exists and holds that [ẋr ẏr ϕ̇r ]T  ≤ q̇max and
to the axle direction; l and r denote the height of center of [ẍr ÿr ϕ̈r ]T  ≤ q̈max with q̇max , q̈max being positive scalars.
mass and the radius of the wheel, respectively; d represents With the aforementioned assumptions, in order to achieve
the distance between the two wheels along the axle. Besides, the reference trajectory tracking purpose, this paper attempts
Jϕ = Jυ + d2 (mυ + Jω /r2 )/2 and Ω = mv Jθ − m2p l2 cos2 θ are to propose the constrained MPC and extended state observer
related variables introduced for simplicity. to enhance the tracking performance of the vehicle, and the
The first subequation in (1) concerns the rotation velocity ω developed control system can be shown in Fig. 2, which will
which is an actuated one because an independent control effort be described in details in the following.
τω can be utilized for controlling this state, and then it can be
defined as a rotation subsystem. In contrast, there has the only III. E XTENDED S TATE O BSERVER
one available control input τυ for the longitudinal velocity υ The WIP vehicle, similar to the other mechanical systems,
and pendulum angular velocity θ̇ , which validates the under- is difficult to achieve the velocity information to construct con-
actuated characteristic of WIP vehicle system, thus it can be troller. In this regard, it is intended to employ extended state
defined as longitudinal subsystem consisting of two velocity observer for the control system to enhance the feasibility of
vectors. It is noteworthy that determining how to control this the presented controllers without velocity measurement. Let xv
underactuated subsystem is the key issue and the importance be the longitudinal displacement, and the necessary position
of this paper. More details about the dynamic model can refer messages, i.e., xv , ϕ, and θ can be generally measured by the
to [35], and it is omitted here for simplicity. vehicle-mounted position sensors and GPS system. Whereas,
B. Various Physical Limits the velocity informations, including ω, υ, and θ̇ in addition to
For practical implementation, the WIP vehicle suffers from the disturbance, need to be observed by the position messages
various physical limits, such as the state variable limitation, and control inputs.
control input saturation, and so on. These constrained physi- Consider the rotational subsystem as
cal conditions must be considered when designing the system ϕ̈ = f1 + g1 τω (2)
controller, otherwise the control system cannot be applied to
application. Exploiting the vehicle reveals that the follow- where f1 = fω1 + τd1 can be treated as a comprehensive dis-
ing assumptions should be introduced before synthesizing the turbance, satisfying |ḟ1 | ≤ ḟ1 max with ḟ1 max being the bounded
control system. scalar. Define yϕ as the output of this subsystem, then the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

Fig. 2. Control system of the WIP vehicle.

system function can be rewritten as Assume that N is an arbitrary symmetric positive defi-
 nite matrix, and M is the solution of the Lyapunov equation
ẋϕ = Axϕ + B(f1 + g1 τω ) T
(3) Ã M + MÃ + N = 0. By taking the derivative of V1 along the
yϕ = Cxϕ
trajectory (5) with respect to time, it holds that
where xϕ = [ϕ ϕ̇]T = [x1 x2 ]T , A = [0 1; 0 0], B = [0 1]T , T
V̇1 = ε1 δ̇ ϕ Mδ ϕ + ε1 δ Tϕ Mδ̇ ϕ
and C = [1 0].  T  
For system (3), the proposed observer can update the current = Ãδ ϕ + ε1 B̃ḟ1 Mδ ϕ + δ Tϕ M Ãδ ϕ + ε1 B̃ḟ1
state variables by  T
⎧ α1   T
⎪ x̂˙ 1 = x̂2 + yϕ − x̂1 = δ Tϕ Ã Mδ ϕ + ε1 B̃ḟ Mδ ϕ + δ Tϕ MÃδ ϕ + ε1 δ Tϕ MB̃ḟ1
⎪
⎪  T 
⎪
⎪ ε1
⎨˙ α2   = δ Tϕ Ã M + MÃ δ ϕ + 2ε1 δ Tϕ MB̃ḟ1
x̂2 = g1 τω + f̂1 + 2 yϕ − x̂1
⎪ ε1 (4)      
⎪  
⎪
⎪ ˙ α 3   ≤ −δ Tϕ Nδ ϕ + 2ε1 MB̃ · δ ϕ  ·  ḟ1 
⎪
⎩f̂1 = 3 yϕ − x̂1
ε1  2   

  

≤ −λmin (N)δ ϕ  + 2ε1 ḟ1 max MB̃ · δ ϕ  (7)
where x̂1 , x̂2 , and f̂1 are the corresponding estimated variables;
ε1 > 0 is a design parameter; α1 , α2 , and α3 are all positive where λmin (N) is the minimum eigenvalue of
design parameters, with s3 + α1 s2 + α2 s + α3 satisfying the N. When ε1 is small enough, it follows that
Hurwitz condition. δ ϕ  ≥ 2ε1 |ḟ1 |max MB̃/λmin (N), thereby indicating
In addition, define the estimated error vector as δ ϕ = that V̇1 ≤ 0.
[δ1 δ2 δ3 ]T , where δ1 = (x1 − x̂1 )/ε12 , δ2 = (x2 − x̂2 )/ε1 , δ3 = It is worthy to point out that the convergence rate of esti-
f1 − f̂1 . Then, a corresponding theorem can be obtained as mated error δ ϕ is determined by the design parameter ε1 ,
follows. which means that the smaller ε1 generally makes δ ϕ with
Theorem 1: Consider the rotational subsystem (2). If all a faster convergence; however, if ε1 is too small, the initial
signals in the subsystem are bounded, it is said that the error [x1 (0) − x̂1 (0)]/ε12 will induce the peaking phenomenon.
extended state observer (4) can guarantee the estimated error Therefore, an appropriate value needs to be chosen by trial-
vector δ ϕ converges to zero, e.g., limt→∞ δ ϕ  → 0 as and-error when applying the presented observer (see [36] for
t → ∞. more details).
The asymptotic convergence property of the estimated error Likewise, define f2 = fυ2 + τd2 and f3 = fθ3 + τd3 , and
can be verified in the sense of Lyapunov stability theorem. the extended state observer for longitudinal subsystem can be
From (3) and (4), an estimated error model can be formu- obtained as follows:
⎧ α4  
lated as ⎪
⎪
⎪ x̂˙ 3 = x̂4 + yxυ − x̂3
⎪
⎪ ε2
ε1 δ̇ ϕ = Ãδ ϕ + ε1 B̃ḟ1 (5) ⎪
⎪ α  
⎪x̂˙ 4 = g2 τυ + f̂2 + 5 yxυ − x̂3
⎪
⎪
⎪ ε 2
where ⎪
⎪
⎡ ⎤ ⎪
⎪ ˙ α6   2
−α1 1 0 ⎪
⎨f̂2 = 3 yxυ − x̂3
ε2
à = ⎣−α2 0 1⎦ and B̃ = [0 0 1]T . If s3 + α1 s2 + α2 s + α3 α7   (8)
⎪
⎪ ˙
x̂5 = x̂6 + yθ − x̂5
−α3 0 0 ⎪
⎪ ε3
⎪
⎪ α8  
⎪
⎪ ˙
satisfies the Hurwitz condition, then à is a Hurwitz matrix. ⎪x̂6 = g3 τυ + f̂3 + ε2 yθ − x̂5
⎪
⎪
⎪
⎪
As a result, a positive Lyapunov candidate can be chosen as ⎪˙
⎪ α9   3
⎪
⎩3f̂ = y θ − x̂5
V1 = ε1 δ Tϕ Mδ ϕ . (6) ε33
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

YUE et al.: CONSTRAINED ADAPTIVE ROBUST TRAJECTORY TRACKING FOR WIP VEHICLES 5

where xxυ = [xυ ẋυ ]T = [x3 x4 ]T , yxυ = xυ = x3 ,

xθ = [θ θ̇ ]T = [x5 x6 ]T , yxθ = θ = x5 , with x̂3 , x̂4 , f̂2 , x̂5 , x̂6 ,
and f̂3 being the corresponding estimated variables, respec-
tively; ε2 > 0 and ε3 > 0 are the positive design values;
α4 , α5 , α6 , α7 , α8 , and α9 are all the positive design
parameters, following that both s3 + α4 s2 + α5 s + α6 and
s3 + α7 s2 + α8 s + α9 satisfy Hurwitz condition. Given the
previous facts, a corresponding theorem can also be obtained
as follows. Fig. 3. Adaptive robust controller for rotational subsystem.
Theorem 2: If all signals in the longitudinal subsystem
are bounded, the extended state observer (8) can make the
estimated values globally asymptotically approach to their where f1 can be obtained online by the aforementioned
actual values, e.g., limt→∞ x̂i (t) → xi (t) (i = 3, . . . , 6) and extended state observer. In general, if accurate knowledge
limt→∞ f̂j (t) → fj (t) (j = 2, 3) as t → ∞. of g1 is obtained, the traditional computed torque method
It is worth mentioning that owing to the unavoidable enables the control system with high-performance; however,
observed errors, the estimated values are considered as an this method is almost impossible for application due to the
approximation for the true ones; in order to achieve good con- unavailable accurate model, thus an adaptive law needs to be
trol performance, the design gains for observer must be chosen employed to update the true value of g1 , thereby enhancing
with faster convergence than the settling time in controllers the feasibility of the presented controller. The control structure
so as to guarantee the observer effectiveness for application. of the developed rotational subsystem is shown in Fig. 3.
Moreover, the observe errors can be deemed as a disturbance- Define α = 1/g1 , and the rotational subsystem can be
like signal and they can be rejected by the inherent robustness rewritten as
of MPC. α ω̇ = αf1 + τω . (11)
Let the rotational velocity tracking error be eω = ω − ωd ,
IV. C ONTROL S YSTEM S YNTHESIS
and then the control laws can be constructed by an adap-
A. Trajectory Generator tive compensation scheme and an update law, which are
As shown in Fig. 2, a saturated trajectory generator needs formulated as
to be introduced to provide the desired velocities, by which
τω = α̂ ω̇d − α̂ f̂1 − kω eω (12)
the nonholonomic constraint can be resolved and the satu-  
rated property for the desired velocities can be guaranteed α̂˙ = −γ eω ω̇d − f̂1 (13)
simultaneously.
where kω and γ are positive design parameters, α̂ is the
Defining xe = xr − x, ye = yr − y, and ϕe = ϕr − ϕ,
estimated value by the adaptive law.
and assuming ωd and υd as the desired velocities, referring
Theorem 3: Consider the rotational subsystem (10). If all
to [37]–[39], a trajectory generator based on the so-called
signals in the subsystem are bounded, then the velocity track-
tracking error dynamics can be given by
ing error eω will converge to zero as t → ∞, with the control
⎡ ⎤
  λ1 υr ye sin ϕe scheme (12) and adaptive law (13).
ωd ω + + λ tanh ϕ
= ⎣ r 1 + xe 2 + ye 2 ϕe 2 e⎦
(9) Proof: Define a positive Lyapunov candidate as
υd
υr cos ϕe + λ3 tanh xe 1 2 1 2
V3 = αeω + α̃ (14)
where λ1 , λ2 , and λ3 are all positive design parameters. It can 2 2γ
be further concluded that |ωd | ≤ |ωr |+|λ1 υr |+|λ2 | and |υd | ≤ where α̃ = α̂ − α is the evaluate error, and it holds that α̃˙ =
|υr | + |λ3 |, which imply that ωd and υd are both bounded. α̂˙ because α approximates a constant value compared with
The stability of the generator can be proven by choosing a computer processing speed.
Lyapunov function candidate as V2 = λ1 ln(1 + xe2 + y2e )/2 + Differentiating (14) with respect to time along the trajec-
ϕe2 /2, and the proof procedures are omitted for simplicity. tory of (11) and substituting (12) and (13) into the results, it
follows that:
B. Adaptive Robust Control 1
V̇3 = αeω ėω + α̃ α̃˙
With the estimated values for the unavailable states and γ
1
= eω (α ω̇ − α ω̇d ) + α̃ α̂˙
uncertain disturbance obtained by presented observer, this
paper aims to apply adaptive robust control to actuated rota- γ
   
tional subsystem and utilize MPC for the underactuated longi- = eω αf1 + α̂ ω̇d − α̂ f̂1 − kω eω − α ω̇d − α̃eω ω̇d − f̂1
tudinal subsystem, both of which will provide the closed-loop  
system with adaptive and robust features. = −kω e2ω + αeω f1 − f̂1 . (15)
Consider the rotational subsystem as
Then, if kω is a relative large value, it is straightforward
ω̇ = f1 + g1 τω (10) that V̇3 ≤ 0, considering the fact that f̂1 ≈ f1 . Consequently,
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

⎡ ⎤
the tracking error eω will globally asymptotically converge to g(x(k|k − 1))τυ (k − 1)
⎢ g(x(k + 1|k − 1))τυ (k − 1) ⎥
zero in the sense of Lyapunov theorem of stability [36]. ⎢ ⎥
g̃ = ⎢ .. ⎥ ∈ R3Np .
It is worthy mentioning that for practical application, to ⎣ . ⎦
protect the system from peaking phenomenon, the saturated   
g x k + Np − 1|k − 1 τυ (k − 1)
constraint can be added directly to control input by τω =
sat(α̂ ω̇d − α̂ f̂1 − kω eω ), where sat(·) is a saturation func- Besides, the aforementioned various constraints, such as
tion, which is introduced to simultaneously satisfy saturated velocity constraints, limited tilt angle, actuator saturations, and
requirement for the actuator (refer to [40]–[43]). actuator increment limitation can be described as
τυ min ≤ τυ (k) ≤ τυ max
C. Model Predictive Control
τ υ min ≤ τ υ (k) ≤ τ υ max
For the longitudinal subsystem, the velocity tracking and
xmin ≤ x(k) ≤ xmax (19)
inverted pendulum-like vehicle body stability must be achieved
by the only control input at the same time, as well as taking where τυ max = 2τmax , xmax = [υmax θmax θ̇max ]T ; τυ min =
various physical limits into account. In this section, MPC tech- −τυ max , τυ min = − τυ max , xmin = −xmax .
nique is employed to handle such a challenging issue by using Furthermore, the mathematical formulation of the control
rolling optimization approach instead of the traditional control objective consisting of longitudinal velocity tracking, inverted
algorithm. To begin with, recall x4 = υ, x5 = θ , x6 = θ̇ and vehicle body stability, and control input requirement for the
the underactuated longitudinal subsystem can be written as lower energy can be comprehensively assessed in discrete time
⎧ by a cost function as follows:
⎪
⎨ẋ4 = f2 + g2 τυ  2
⎪
ẋ5 = x6 (16) J(x(k), τ̄ υ (k)) = ȳr (k) − ȳ(k)Q +  τ̄ υ (k)2R (20)
⎩
ẋ6 = f3 + g3 τυ .
where ȳr (k) is the reference value; ȳ(k) = x̄(k) = G̃ τ̄ υ (k) +
With the aid of presented observer, the related state variables f̃ + g̃ is the predictive output; Q and R are the symmetric
and uncertain disturbance in (16) are all supposed to be known definite-positive weight matrices.
in this case. By using the Euler-discrete method, a predictive Based on the above preparation, the WIP vehicle optimiza-
model can be obtained with a fixed sampling time ε as follows: tion problem of MPC scheme can be expressed at sampling

x(k + 1) = f (x(k)) + g(x(k))τυ (k) instant k as follows:
(17)
y(k) = x(k) min J(x(k), τ̄ υ (k))
τ̄ υ (k)
where x(k) = [x4 (k) x5 (k) x6 (k)]T ∈ R3 is the state vector, subject to τ̄ υ min ≤ 1Nc ⊗ τ υ (k − 1) + Ĩ τ̄ υ (k) ≤ τ̄ υ max
τυ (k) ∈ R is the input vector, y(k) ∈ R3 is the output vector,
τ̄ υ min ≤ τ̄ υ (k) ≤ τ̄ υ max
and the other entries are given by
⎡ ⎤ x̄min ≤ G̃τ̄ υ (k) + f̃ + g̃ ≤ x̄max (21)
x4 (k) + εf2 (k)
f (x(k)) = ⎣x5 (k) + εx6 (k)⎦ ∈ R3×1 where 1Nc ∈ RNc ×1 with all elements being 1, ⊗ stands for
x6 (k) + εf3 (k) Kronecker product, and
⎡ ⎤ ⎡ ⎤
g2 (k) I 0 ... 0
g(x(k)) = ε⎣ 0 ⎦ ∈ R3×1 . ⎢I I ... 0⎥
g3 (k) ⎢ ⎥ N ×N
Ĩ = ⎢ . . .. .. ⎥ ∈ R c c .
⎣ .. .. . .⎦
Based on this predictive model, a future dynamic behavior
can be forecasted at sampling instant k as I I ... I

x̄(k) = G̃ τ̄ υ (k) + f̃ + g̃ (18) Notice that every element in vectors, such as τ̄ υ min , τ̄ υ max ,
τ̄ υ min , τ̄ υ max , x̄min , and x̄max , corresponds to the lower
where Np is the prediction horizon, Nc is the control horizon; and upper bounds of (19) at control horizon and prediction
x̄(k) = [x(k + 1|k) · · · x(k + Np |k)]T ∈ R3NP , τ̄ υ (k) = horizon.
[ τυ (k|k) . . . τυ (k + Nc − 1|k)]T ∈ RNc Then an optimization method can be exploited to obtain
G̃ ∈⎡R3Np ×Nc ⎤
the solution, but this issue is rather complicated to calculate;
g(x(k|k − 1)) ... 0 fortunately, it can be further converted into a standard QP
⎢ g(x(k + 1|k − 1)) ... 0 ⎥ problem by a series of translations. Substituting (21) with (17)
⎢ ⎥
=⎢ .. . . ⎥ and (18), it results in
⎣ . . . .
. ⎦
     
g x k + Np − 1|k − 1 . . . g x k + Np − 1|k − 1 1
⎡ ⎤ min τ̄ Tυ (k)W τ̄ υ (k) + cT τ̄ υ (k)
f (x(k|k − 1)) 2
⎢ f (x(k + 1|k − 1)) ⎥ s.t. L ≤ E τ̄ υ (k) ≤ H (22)
⎢ ⎥
f̃ = ⎢ .. ⎥ ∈ R3Np T T
⎣ . ⎦ where W = 2(G̃ QG̃ + R) ∈ RNc ×Nc , c = −2G̃ Q(yr (k) −
  
f x k + Np − 1|k − 1 g̃ − f̃ ) ∈ RNc , E = [−Ĩ Ĩ − G̃ G̃ I]T ∈ R(3Nc +6Np )×Nc , and
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

YUE et al.: CONSTRAINED ADAPTIVE ROBUST TRAJECTORY TRACKING FOR WIP VEHICLES 7

the other entries are given by

   
−∞ b
L= ∈ R3Nc +6Np , H = ∈ R3Nc +6Np
τ̄ υ min τ̄ υ max
⎡ ⎤
−τ̄ υ min + 1Nc ⊗ τ υ (k − 1)
⎢ τ̄ υ max − 1Nc ⊗ τ υ (k − 1) ⎥
b=⎢ ⎣
⎥ ∈ R2Nc +6Np .
⎦
−x̄min + f̃ + g̃
x̄max − f̃ − g̃
By this stage, the QP optimization methods are readily
available by considerable developed theory, such as active
set method and interior point approach. In the following, a
sequence of control input increments can be obtained over a
prediction horizon and only the first one is applied to the vehi-
cle at the next horizon, hence at sampling instant k + 1, the
control input can be obtained by
τ υ (k) = τ υ (k − 1) + [1 0 · · · 0] τ̄ υ (k). (23) Fig. 4. Tracking curve of a trifolium trajectory.

Furthermore, noticing that the real inputs of WIP vehicle are

the right and left motor torques, the practical control inputs τl
and τr can be computed by τυ and τω with the formulation as
follows:
τω + τυ τυ − τω
τl = , τr = . (24)
2 2
V. S IMULATION VALIDATION
To illustrate the effectiveness and robustness of the
presented control approaches, numerical simulations are per-
formed for a WIP vehicle in MATLAB environment. A
trifolium curve is selected as reference trajectory, which is
governed by
xr (t) = κ cos(3ωr t) cos(ωr t), pyr (t) = κ cos(3ωr t) sin(ωr t)
where κ = 3 denotes the length of each  petal,
and it holds ωr = π/45 and υr (t) = ẋr2 + ẏ2r .
Fig. 5. Posture tracking errors.
Besides, the initial error is supposed as follows:
[xe (0) ye (0) ϕe (0)]T = [0.5 0.5 0.1]T , [ϕ(0) xv (0) θ (0)]T =
[0.5 0.1 0.05]T , [ω(0) υ(0) θ̇ (0)]T = [0 0 0]T , and
˙ ˙ ˙
˙
[ϕ̂(0) ϕ̂(0) f̂1 (0) x̂˙ υ (0) υ̂(0)
˙ f̂2 (0) θ̂(0) θ̂˙ (0) f̂3 (0)]T = As for the adaptive robust controller, kω = 20, γ = 50; for the
T
[0.5 0 0 0.1 0 0 0.05 0 0] . As for the uncertain disturbance, MPC, ε = 0.1, Q = diag[1 50 1 0], R = diag[1 1], Np = 20,
a sinusoidal waveform disturbances with amplitude 1, e.g., and Nc = 6.
τd1 = sin(0.5t), is imposed to the rotational subsystem, and a A practical trajectory tracking process is depicted in Fig. 4,
different pulse disturbance with amplitude 10 is introduced to whose results suggest that the large initial errors are rejected
longitudinal subsystem at 30 s, which can simulate a scenario by a self-regulation within 8 s and the external disturbance
that the pendulum-like vehicle body suffers from a collision. is successfully suppressed by the adaptive robust control
Furthermore, to demonstrate the capability of handling con- and MPC strategy. Moreover, the detailed tracking errors are
straint, the bounds of state variables, control input and its shown in Fig. 5, in which the tracking errors fluctuate within
increment are hypothesized as follows: ωmax = 0.3, υmax = a small range rather than have a phenomenon of gradual-
1.5; θmax = 0.1 and θ̇max = ∞; τmax = 15, τυ max = 30; and approach convergence to zero. This is a tradeoff between the
τυ min = −10, τυ max = 10. Consider the small rotational tracking performance and vehicle body stability, determined
angle and then assume τω max = 0.8. In addition, for trajectory by selecting different corresponding design parameters Q, R,
generator (9), λ1 = 1, λ2 = 0.1 and λ3 = 1; while, for the ε, Np , and Nc .
extended state observer, α1 = α4 = α7 = 6, α2 = α5 = α8 = As illustrated in Fig. 6, the tilt angle of inverted pendulum-
11, α3 = α6 = α9 = 6. It is noteworthy that in case of the like vehicle body has some periodic oscillation rather than
peaking phenomenon, the design parameters εi (i = 1, 2, 3) are converge to zero. However, this periodic oscillation is just
chosen as around the equilibrium point and within acceptable limits,
 whose boundedness can be suppressed as much as possible
3, 0<t≤1
1
ε1 = ε2 = ε3 = 100t but cannot be eliminated because of the complicated coupling
100 , t > 1.
1
relationships and underactuated characteristic. Also, the results
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

Fig. 6. Response of tilt angle and angular velocity. Fig. 9. Observation for longitudinal subsystem.

Fig. 10. Parameter adaptive results.

Fig. 7. Responses of saturated velocity tracking processes.
rotational one can be rapidly followed within a short time due
to its relatively simple dynamics. More specially, the desired
velocities are both governed within the constraint condition,
thanks to the saturated generator property.
To verify the effectiveness of the extended state observer, the
estimated results of state and disturbance for the rotational and
longitudinal subsystems are shown in Figs. 8 and 9, respec-
tively. As shown in Fig. 8, provided that there has a 20% error
of system parameter g1 , there exists small estimated error for
the disturbance, but the above results imply that the adaptive
robust control can overcome this estimated error and the whole
control performance is almost unaffected. From Fig. 9, it tells
that if there has no assumed error, the estimated disturbances
f2 and f3 are nearly identical with practical ones. Besides, this
fact means that all state errors between the estimations and
the practice are so small that the difference can be negligible,
Fig. 8. Observation for rotational subsystem.
even when the disturbances appear in the system. Moreover, it
can be seen that, the peak phenomenon is successfully avoided
by means of the parameter design and input saturation.
indicate the robustness of vehicle body stabilizing system to Fig. 10 depicts the estimated and true values for g1 to
reject both initial errors and uncertain disturbance. verify the adaptability of the proposed controller. It can be
Besides, the time responses for velocity tracking are plotted observed that although the estimated values does not match the
in Fig. 7. Compared with the longitudinal velocity, the desired true ones, the multiple solution characteristic of the presented
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

YUE et al.: CONSTRAINED ADAPTIVE ROBUST TRAJECTORY TRACKING FOR WIP VEHICLES 9

system with physical limits, it is difficult to find an appropriate

Lyapunov candidate to guarantee the stability and boundedness
of all the variables and inputs, but MPC can deal with these
problems with a relatively easy approach which was proved in
the developed theory and was validated by the numerical sim-
ulation. The developed results can be applied to other similar
constrained or underactuated systems.

R EFERENCES
[1] Y.-S. Ha and S. Yuta, “Trajectory tracking control for navigation of the
inverse pendulum type self-contained mobile robot,” Robot. Auton. Syst.,
vol. 17, nos. 1–2, pp. 65–80, 1996.
[2] F. Grasser, A. D’Arrigo, S. Colombi, and A. C. Rufer, “JOE: A
mobile, inverted pendulum,” IEEE Trans. Ind. Electron., vol. 49, no. 1,
pp. 107–114, Feb. 2002.
[3] K. Pathak, J. Franch, and S. K. Agrawal, “Velocity and position control
of a wheeled inverted pendulum by partial feedback linearization,” IEEE
Fig. 11. System control inputs. Trans. Robot., vol. 21, no. 3, pp. 505–513, Jun. 2005.
[4] C.-H. Chiu and C.-C. Chang, “Wheeled human transportation vehi-
cle implementation using output recurrent fuzzy control strategy,” IET
Control Theory Appl., vol. 8, no. 17, pp. 1886–1895, Nov. 2014.
[5] Z. Li and C. Yang, “Neural-adaptive output feedback control of a class
of transportation vehicles based on wheeled inverted pendulum mod-
els,” IEEE Trans. Control Syst. Technol., vol. 20, no. 6, pp. 1583–1591,
Nov. 2012.
[6] Z.-Q. Guo, J.-X. Xu, and T. H. Lee, “Design and implementation of
a new sliding mode controller on an underactuated wheeled inverted
pendulum,” J. Frankl. Inst., vol. 351, no. 4, pp. 2261–2282, 2014.
[7] C. Aguilar-Avelar and J. Moreno-Valenzuela, “A composite controller
for trajectory tracking applied to the Furuta pendulum,” ISA Trans.,
vol. 57, pp. 286–294, Jul. 2015.
[8] A. Ghaffari, A. Shariati, and A. H. Shamekhi, “A modified dynamical
formulation for two-wheeled self-balancing robots,” Nonlin. Dyn.,
vol. 83, nos. 1–2, pp. 217–230, 2016.
[9] Z. H. Peng, D. Wang, Z. Y. Chen, X. Hu, and W. Y. Lan, “Adaptive
dynamic surface control for formations of autonomous surface vehicles
with uncertain dynamics,” IEEE Trans. Control Syst. Technol., vol. 21,
no. 2, pp. 513–520, Mar. 2013.
[10] Z. Yan and J. Wang, “Model predictive control for tracking of under-
actuated vessels based on recurrent neural networks,” IEEE J. Ocean.
Eng., vol. 37, no. 4, pp. 717–726, Oct. 2012.
Fig. 12. Torques for the right and left wheels. [11] Z.-P. Jiang, “Global tracking control of underactuated ships by
Lyapunov’s direct method,” Automatica, vol. 38, no. 2, pp. 301–309,
2002.
[12] K. D. Do, “Robust adaptive tracking control of underactuated ODINs
controller causes this phenomena; however, the boundedness under stochastic sea loads,” Robot. Auton. Syst., vol. 72, pp. 152–163,
of the estimated errors can also illustrate the stability and Oct. 2015.
adaptability of the closed-loop system. [13] W. He, Y. T. Dong, and C. Y. Sun, “Adaptive neural impedance control
of a robotic manipulator with input saturation,” IEEE Trans. Syst., Man,
Moreover, the control inputs are shown in Figs. 11 and 12, Cybern., Syst., vol. 46, no. 3, pp. 334–344, Mar. 2016.
both of which reveal that all the related control inputs satisfy [14] Y. M. Li, S. C. Tong, and T. S. Li, “Adaptive fuzzy output feedback
the saturation constraints. The continuity of the control signals control for a single-link flexible robot manipulator driven DC motor
via backstepping,” Nonlin. Anal. Real World Appl., vol. 14, no. 1,
guarantees the feasibility of the proposed controllers. pp. 483–494, 2013.
[15] B. T. Gao, Y. Q. Bao, J. H. Xie, and L. J. Jia, “Passivity-based con-
trol of two-dimensional translational oscillator with rotational actuator,”
VI. C ONCLUSION Trans. Inst. Meas. Control, vol. 36, no. 1, pp. 111–118, 2014.
This paper studied the WIP vehicle trajectory tracking with- [16] X.-Z. Lai, J.-H. She, W.-H. Cao, and S. X. Yang, “Stabilization of
underactuated planar acrobot based on motion-state constraints,” Int. J.
out velocity measurement, in presence of model uncertainty, Non-Lin. Mech., vol. 77, pp. 342–347, Dec. 2015.
external disturbances, and various constraints. In particular, an [17] Y. C. Fang, P. C. Wang, N. Sun, and Y. C. Zhang, “Dynamics analysis
extended state observer was employed to estimate the velocity and nonlinear control of an offshore boom crane,” IEEE Trans. Ind.
Electron., vol. 61, no. 1, pp. 414–427, Jan. 2014.
vectors and uncertain disturbance by using the only position [18] N. Sun, Y. C. Fang, and H. Chen, “A new antiswing control method for
measurements, which was an effective and efficient way to underactuated cranes with unmodeled uncertainties: Theoretical design
replace the costly sensors and even can achieve the approx- and hardware experiments,” IEEE Trans. Ind. Electron., vol. 62, no. 1,
pp. 453–465, Jan. 2015.
imation of inaccessible disturbance. On the other hand, the [19] W. He et al., “Model identification and control design for a humanoid
presented MPC scheme, was demonstrated to be a power- robot,” IEEE Trans. Syst., Man, Cybern., Syst., to be published,
ful solution for underactuated control system synthesis by doi: 10.1109/TSMC.2016.2557227.
[20] Y. J. Liu and S. C. Tong, “Barrier Lyapunov functions for Nussbaum gain
using an optimization method, especially when the system adaptive control of full state constrained nonlinear systems,” Automatica,
suffers from various physical constraints. For the sophisticated to be published, doi: 10.1016/j.automatica.2016.10.011.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

[21] Z. J. Li, H. Z. Xiao, C. G. Yang, and Y. W. Zhao, “Model predictive con- [42] Y.-J. Liu, Y. Gao, S. C. Tong, and Y. M. Li, “Fuzzy approximation-based
trol of nonholonomic chained systems using general projection neural adaptive backstepping optimal control for a class of nonlinear discrete-
networks optimization,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 45, time systems with dead-zone,” IEEE Trans. Fuzzy Syst., vol. 24, no. 1,
no. 10, pp. 1313–1321, Oct. 2015. pp. 16–28, Feb. 2016.
[22] S.-R. Oh and J. Sun, “Path following of underactuated marine surface [43] Y. M. Li, S. C. Tong, and T. S. Li, “Hybrid fuzzy adaptive output
vessels using line-of-sight based model predictive control,” Ocean Eng., feedback control design for uncertain MIMO nonlinear systems with
vol. 37, nos. 2–3, pp. 289–295, 2010. time-varying delays and input saturation,” IEEE Trans. Fuzzy Syst.,
[23] Z. Yan and J. Wang, “Robust model predictive control of nonlinear vol. 24, no. 4, pp. 841–853, Aug. 2016.
systems with unmodeled dynamics and bounded uncertainties based on
neural networks,” IEEE Trans. Neural Netw. Learn. Syst., vol. 25, no. 3,
pp. 457–469, Mar. 2014.
[24] Y. M. Li, S. Sui, and S. C. Tong, “Adaptive fuzzy control design Ming Yue received the B.S. degree in mechanical
for stochastic nonlinear switched systems with arbitrary switchings engineering from Zhengzhou University, Zhengzhou,
and unmodeled dynamics,” IEEE Trans. Cybern., to be published, China, in 1998, and the M.S. and Ph.D. degrees in
doi: 10.1109/TCYB.2016.2518300. mechatronics engineering from the Harbin Institute
[25] R. X. Cui, J. Guo, and Z. Y. Mao, “Adaptive backstepping control of Technology, Harbin, China, in 2004 and 2008,
of wheeled inverted pendulums models,” Nonlin. Dyn., vol. 79, no. 1, respectively.
pp. 501–511, 2015. He was a Visiting Scholar with Aerospace
[26] Y.-J. Liu, S. C. Tong, D.-J. Li, and Y. Gao, “Fuzzy adaptive control with Engineering, University of Michigan, Ann Arbor,
state observer for a class of nonlinear discrete-time systems with input MI, USA, from 2013 to 2014. Since 2008, he has
constraint,” IEEE Trans. Fuzzy Syst., vol. 24, no. 5, pp. 1147–1158, been with the Dalian University of Technology,
Oct. 2016. Dalian, China, where he is currently an Associate
[27] Y. M. Li and S. C. Tong, “Adaptive fuzzy output-feedback Professor with the Department of Automotive Engineering. His research
stabilization control for a class of switched nonstrict-feedback interests include control approaches, such as adaptive control, sliding mode
nonlinear systems,” IEEE Trans. Cybern., to be published, control, model predictive control, with application to wheeled mobile robots
doi: 10.1109/TCYB.2016.2536628. and intelligent vehicles.
[28] H. Y. Li, H. J. Gao, P. Shi, and X. D. Zhao, “Fault-tolerant control
of Markovian jump stochastic systems via the augmented sliding mode
observer approach,” Automatica, vol. 50, no. 7, pp. 1825–1834, 2014. Cong An received the B.S. degree in mecha-
[29] Y.-J. Liu and S. C. Tong, “Barrier Lyapunov functions-based adaptive tronics and automobile engineering from Yantai
control for a class of nonlinear pure-feedback systems with full state University, Yantai, China, in 2012. He is currently
constraints,” Automatica, vol. 64, pp. 70–75, Feb. 2016. pursuing the M.S. degree with the Department
[30] Q. L. Hu and J. Zhang, “Relative position finite-time coordinated track- of Automotive Engineering, Dalian University of
ing control of spacecraft formation without velocity measurements,” Technology, Dalian, China.
ISA Trans., vol. 54, pp. 60–74, Jan. 2015. His research interests include nonlinear dynamics
[31] H. Y. Li, P. Shi, D. Y. Yao, and L. G. Wu, “Observer-based adaptive slid- and control with application to the wheeled inverted
ing mode control for nonlinear Markovian jump systems,” Automatica, pendulum vehicle.
vol. 64, pp. 133–142, Feb. 2016.
[32] Q. L. Hu, B. Li, and J. T. Qi, “Disturbance observer based finite-time
attitude control for rigid spacecraft under input saturation,” Aerosp. Sci.
Technol., vol. 39, pp. 13–21, Dec. 2014.
[33] Y.-J. Liu and S. C. Tong, “Adaptive fuzzy identification and control
for a class of nonlinear pure-feedback MIMO systems with unknown Zhijun Li (M’07–SM’09) received the Ph.D.
dead zones,” IEEE Trans. Fuzzy Syst., vol. 23, no. 5, pp. 1387–1398, degree in mechatronics from Shanghai Jiao Tong
Oct. 2015. University, Shanghai, China, in 2002.
[34] Y. M. Li, S. C. Tong, and T. S. Li, “Observer-based adaptive fuzzy From 2003 to 2005, he was a Post-Doctoral
tracking control of MIMO stochastic nonlinear systems with unknown Fellow with the Department of Mechanical
control directions and unknown dead zones,” IEEE Trans. Fuzzy Syst., Engineering and Intelligent Systems, University
vol. 23, no. 4, pp. 1228–1241, Aug. 2015. of Electro-Communications, Tokyo, Japan. From
[35] M. Yue, C. An, Y. Du, and J. Sun, “Indirect adaptive fuzzy control 2005 to 2006, he was a Research Fellow with
for a nonholonomic/underactuated wheeled inverted pendulum vehicle the Department of Electrical and Computer
based on a data-driven trajectory planner,” Fuzzy Sets Syst., vol. 290, Engineering, National University of Singapore,
pp. 158–177, May 2016. Singapore, and Nanyang Technological University,
[36] H. K. Khalil and J. W. Grizzle, Nonlinear Systems, vol. 3. Upper Saddle Singapore. From 2007 to 2011, he was an Associate Professor with the
River, NJ, USA: Prentice-Hall, 1996. Department of Automation, Shanghai Jiao Tong University. In 2008, he
[37] Z.-P. Jiang, E. Lefeber, and H. Nijmeijer, “Saturated stabilization and was a Visiting Scholar with Microsoft Research Asia, Beijing, China. Since
tracking of a nonholonomic mobile robot,” Syst. Control Lett., vol. 42, 2012, he has been a Professor with the College of Automation Science and
no. 5, pp. 327–332, 2001. Engineering, South China University of Technology, Guangzhou, China. In
[38] K. D. Do and G. Seet, “Motion control of a two-wheeled mobile vehicle 2015, he was a Visiting Professor with the Faculty Science and Technology,
with an inverted pendulum,” J. Intell. Robot. Syst., vol. 60, nos. 3–4, University of Macau, Macau, China, and the Department of Advanced
pp. 577–605, 2010. Robotics, Italian Institute of Technology, Genoa, Italy. His current research
[39] J. S. Huang, C. Y. Wen, W. Wang, and Z.-P. Jiang, “Adaptive output interests include service robotics, tele-operation systems, nonlinear control,
feedback tracking control of a nonholonomic mobile robot,” Automatica, and neural network optimization.
vol. 50, no. 3, pp. 821–831, 2014. Dr. Li is serving as an Editor-at-Large for the Journal of Intelligent and
[40] Y.-J. Liu, Y. Gao, S. Tong, and C. L. P. Chen, “A unified approach to Robotic Systems, an Associate Editor of the IEEE T RANSACTIONS ON
adaptive neural control for nonlinear discrete-time systems with nonlin- N EURAL N ETWORKS AND L EARNING S YSTEMS, the IEEE T RANSACTIONS
ear dead-zone input,” IEEE Trans. Neural Netw. Learn. Syst., vol. 27, ON S YSTEMS , M AN , AND C YBERNETICS : S YSTEMS , and the IEEE
no. 1, pp. 139–150, Jan. 2016. T RANSACTIONS ON AUTOMATION S CIENCE AND E NGINEERING. He has
[41] Y. M. Li, S. C. Tong, and T. S. Li, “Composite adaptive fuzzy out- been the General Chair of 2016 IEEE Conference on Advanced Robotics
put feedback control design for uncertain nonlinear strict-feedback and Mechatronics, Macau. Since 2016, he has been the Chair of Technical
systems with input saturation,” IEEE Trans. Cybern., vol. 45, no. 10, Committee on Biomechatronics and Biorobotics Systems and IEEE Systems,
pp. 2299–2308, Oct. 2015. Man, and Cybernetics Society.