Received January 9, 2020, accepted March 2, 2020, date of publication March 11, 2020, date of current version March

24, 2020.
Digital Object Identifier 10.1109/ACCESS.2020.2980188

An Improved Kinematic Model Predictive Control

for High-Speed Path Tracking of
Autonomous Vehicles
LUQI TANG 1,2,3 , FUWU YAN 1,2,3 , BIN ZOU 1,2,3 , KEWEI WANG 1,2,3 ,

AND CHEN LV 4 , (Member, IEEE)

1 Hubei Key Laboratory of Advanced Technology for Automotive Components, Wuhan University of Technology, Wuhan 430070, China
2 Hubei Collaborative Innovation Center for Automotive Components Technology, Wuhan University of Technology, Wuhan 430070, China
3 Hubei Research Center for New Energy & Intelligent Connected Vehicle, Wuhan 430070, China
4 School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore 639798

Corresponding author: Bin Zou ([email protected])

This work was supported in part by the National Natural Science Foundation of China under Grant 51975434, in part by the Overseas
Expertise Introduction Project for Discipline Innovation under Grant B17034, in part by the Innovative Research Team in University of
Ministry of Education of China under Grant IRT_17R83, and in part by the Department of Science and Technology, Hubei Provincial
People’s Government under Grant 2019AEA169.

ABSTRACT Kinematic model predictive control (MPC) is well known for its simplicity and computational
efficiency for path tracking of autonomous vehicles, however, it merely works well at low speed. In addition,
earlier studies have demonstrated that tracking accuracy is improved by the feedback of yaw rate, as it
improves the system transients. With this in mind, it is expected that the performance of path tracking can
be improved by a cascaded controller that utilizes kinematic MPC to determine desired yaw rate rather
than steering angle, and uses proportional-integral-derivative (PID) control to follow the reference yaw rate.
However, directly combining MPC with PID feedback control of yaw rate results in a controller with poor
tracking accuracy. The simulation results show that the cascaded MPC-PID controller has relatively stable
but larger error compared to classic kinematic and dynamic MPC. Based on the analysis of vehicle sideslip
angle, a novel path tracking control method is proposed, which is designed using a kinematic MPC to
handle the disturbances on road curvature, a PID feedback control of yaw rate to reject uncertainties and
modeling errors, and a vehicle sideslip angle compensator to correct the kinematic model prediction. The
proposed controller performances involving steady-state and transient response, robustness, and computing
efficiency were evaluated on Carsim/Matlab joint simulation environment. Furthermore, field experiments
were conducted to validate the robustness against sensor disturbances and time lag. The results demonstrate
that the developed vehicle sideslip compensator is sufficient to capture steer dynamics, and the developed
controller significantly improves the performance of path tracking and follows the desired path very well,
ranging from low speed to high speed even at the limits of handling.

INDEX TERMS Autonomous vehicles, path tracking, lateral control, model predictive control.

I. INTRODUCTION of the vehicle dynamics, time lag, uncertainties, and road

In recent years, research on autonomous vehicles has curvature disturbances, ensuring tracking accuracy and vehi-
seen great achievement together with computer and sen- cle stability simultaneously is considered to be a great chal-
sors technology advances. As one of major components lenge [1], [2]. The ideal path tracking controller should take
in autonomous vehicles, path tracking aims at following a into account future road information and be capable of reject-
desired path or trajectory via controlling the vehicle in lateral ing disturbances and parameter uncertainties.
and longitudinal motion. In general, this reference path is To date, extensive research on path tracking has
generated by path planning module. Due to the nonlinearity been carried out and usually involves in feedforward-
feedback or optimization control. Early tracking controllers
The associate editor coordinating the review of this manuscript and are mostly developed based on geometric vehicle model and
approving it for publication was Jenny Mahoney. feedback control theory due to its simplicity and stability,

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
51400 VOLUME 8, 2020
L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

namely, the deviation inputs of the feedback controller are such as fuzzy control [20], sliding mode control [2]. These
obtained by the geometrical relationship between vehicle types of classical control deal with worst–case disturbances,
and road. For instance, in the studies [3]–[5], different which often lead to too conservative performance [21].
proportional-integral-derivative (PID) control architectures In addition, these feedback control methods rely on system
are proposed to follow the given path. Pure pursuit method, instantaneous states and usually are incapable of predict-
as a standard benchmark, has been widely used in several ing future behaviors. Consequently, this drawback results
DARPA Challenge vehicles [6]. These methods are simple in the lack of flexibility of road curvature disturbances.
but merely work well in a narrow operating window, since Recently, due to advances in hardware, sensors and artificial
these controllers always calculate errors at one or several pre- intelligence, large amounts of data can be collected. Data-
view points and are unable to capture complete steer dynam- driven methods attract increasing attention in the field of
ics. To improve these methods, many adaptive approaches autonomous driving. For instance, NVIDIA trained a con-
that automatically tune look-ahead distance depending on volutional neural network to map raw pixels from a single
curvature and speed have been proposed [7]–[9]. front-facing camera directly to steering commands, which is
More recently, with the advances of computer perfor- well-known as end-to-end approach [22]. Nitin et al. inves-
mance, model predictive control (MPC) has been shown tigated the path tracking of racecar via iterative learning
to be an attractive control algorithm for path tracking control in consideration of the nonlinear vehicle dynamics
problem [9]–[12]. It has the advantage of handling the con- and unmodelled road conditions during racing task [23].
straints on the state variables and control inputs and achieving Shida et al. proposed a data-driven method, model-free
multi-objective optimization, such as driver comfort, time adaptive control for the lateral motion of an autonomous
consumption, tracking accuracy. For instance, Wang et al. vehicle [24]. The major drawback of these approaches is
proposed an improved MPC controller based on fuzzy adap- the huge amount of training data set representing various
tive control to improve both tracking accuracy and ride com- driving situations, which makes data-driven methods have
fort which can adjust the weights of cost function adaptively not yet applied to the real world as successfully as MPC
based on lateral position error and heading error [13]. Aiming techniques.
at the tracking error representation, Sun et al. believed that In summary, the path tracking controller need to have the
path tracking accuracy and vehicle stability can hardly be ability to handle the road curvature disturbances with predic-
accomplished by one fixed control frame in various condi- tion, reject uncertainties using feedback control, and become
tions. Then, the authors presented a novel MPC controller high efficient in computation. With this in mind, in this paper,
with switched tracking error which mainly involves different a new path tracking control architecture is proposed, which is
treatments regarding sideslip angle in computing the heading designed using a MPC controller based on vehicle kinematics
deviation [14]. to handle the disturbances on road curvature, a PID feed-
Inspired by more precise modeling, actuator dynamics is back control of yaw rate to reject uncertainties and modeling
incorporated to capture the transient response of the vehicle errors, and a vehicle sideslip angle compensator to correct the
into collision avoidance constraints [15]. Cai et al. presented above prediction process.
a MPC controller using a 4-DOF vehicle model to reflect Although it is well known that kinematic vehicle model is
the characteristics of vehicle dynamics to avoid rollover unsuitable for high-speed path tracking as they are inaccurate
accidents of automobiles [16]. In addition, considering the in regions of tire force saturation [25], the proposed controller
noise in the localization and planning stage, a model-based based on kinematic model follows the desired path very well,
linear quadratic gaussian control with adaptive Q-matrix was ranging from low speed to high speed even at the limits
proposed to tracking controller design [17]. Although MPC of handling. This is mainly caused by the involvements of
method with prediction has the ability to forecast future these aforementioned two elements, i.e., the feedback control
dynamic behaviors and significantly improves path tracking, of yaw rate and vehicle sideslip compensation. It has been
it requires solving optimization problem repeatedly at each demonstrated in earlier studies that the tracking accuracy is
control step. This may lead to heavy computational burden improved by additional feedback of the yaw rate which can
and potential risks in real-time implementation. Moreover, be measured by a gyro, as it improves the system transients,
many vehicle parameters play an important role in vehicle by changing the eigenvalues displacement of the steering
dynamic, however, these parameters probably change over dynamics [2], [4], [26]. Consequently, the main contributions
time, such as vehicle mass and cornering stiffness [18]. of this paper are as follows.
Actually, it should be noted that it is very difficult to accu- 1) To handle the challenges of path tracking at high speed
rately characterize the nonlinearities by existing several semi- and sharp curves, a novel tracking control architecture is
empirical tire model [19]. Therefore, pure MPC method developed, consisting of three main components: kinematic
may be unsatisfactory in real applications when taking into model predictive control, feedback control of yaw rate, and
account computational efficiency and prediction accuracy at vehicle sideslip compensation.
the same time. 2) A vehicle sideslip angle compensator is utilized to cor-
To reject the aforementioned uncertainties and distur- rect the prediction process using kinematic model, which is
bances, many classical control theories are also explored, designed based on the relationship between sideslip and yaw

VOLUME 8, 2020 51401

 L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

FIGURE 1. Kinematic bicycle model where αf and αr denote the front and
rear wheel slip angles, respectively. r is yaw rate.

FIGURE 2. Absolute lateral tire force as a function of slip angle, with

rate, and formulized as an expression involving yaw rate and different vertical tire load FN .
current vehicle speed.
The remainder of this paper is organized as follows:
Section II presents the vehicle lateral kinematic and dynamic In this paper, the path tracking control aims at minimizing
model; and Section III introduces the lateral control design the lateral and heading deviation of the autonomous vehicle
for path tracking; Section IV evaluates the proposed control with respect to a given reference path at arbitrary safe speed
method and compares it with classic kinematic and dynamic ranging from low speed to extremely high speed at the limits
MPC; and the field test is presented in Section V. Section VI of handling. Figure 1 illustrates the schematic diagram of
concludes this paper. path tracking model which demonstrates the geometric rela-
tionships between autonomous vehicle and the desired lane.
II. MODELING ea denotes the heading deviation that is the orientation error
In this section, kinematic bicycle modeling and dynamic between the heading of vehicle and the tangential direction of
modeling are carried out respectively. The kinematic model the road centerline. ey denotes the lateral deviation that is the
is the basis of the proposed control design and used to pre- distance of the c.g. of the vehicle from the center line of the
dictive control, however, the dynamic model is explored to lane.
understand steer dynamics and contribute to vehicle sideslip Generally, the kinematic bicycle model described above
compensator design. is suitable for control law design at low speed. However,
at high-speed scenarios, this prediction model will become
A. KINEMATIC BICYCLE MODEL increasingly unreliable, due to the rise of tire sideslip angles.
The kinematic bicycle model is given by the following set of Therefore, it is necessary to investigate the vehicle dynamics
equations in an inertial frame according to the axes system and tire side-slip characteristics for improving the controller
with SAE standards [19] (see Figure 1), under the assump- performance.
tions: 1) The vehicle is assumed to have planar motion, and
the vertical, pitch and roll motions are ignored; 2) The slip B. TIRE and VEHICLE DYNAMIC MODEL
angles at both wheels are zero. It is well known that tire force plays a key role in the
analysis of vehicle motion, as in addition to aerodynamic
Ẋ = v cos(ψ + β) (1a)
forces and gravity, all the forces that affect vehicle motion
Ẏ = v sin(ψ + β) (1b) are produced by the tires. To some extent, due to the com-
v cos(β) plexity of tire model, obtaining vehicle models of sufficient
ψ̇ = tan δf − tan (δr )



(1c)
lf + lr accuracy is not available in real time. Moreover, it is very

!
lf tan (δr ) + lr tan δf difficult to precisely formulate the nonlinearities by a unified
β = tan −1
(1d) tire force model. Most of the existing tire models are pre-
lf + lr
dominantly ‘‘semi-empirical’’ in nature, such as Burckhardt
where X denotes global X axis coordinate, Y global axis model, Magic formula, and Dugoff model [19]. These semi-
coordinate, v the speed of the vehicle, ψ the heading angle empirical tire models involve severe nonlinearities as shown
of the vehicle, β vehicle sideslip angle, lf and lr represent the in Figure 2, resulting in many difficulties in stability analysis
distance from the center of the mass of the vehicle to the front and real-time controller design [20]. Nevertheless, the typical
and rear axles, respectively. δf and δr are the steering angles tire model shown in Figure 2 also indicates that in normal
for the front and rear wheels. we assume δr = 0, as in most driving situations where the slip angles are small, the relation
vehicles the rear wheels cannot be steered. between lateral force and slip angle is nearly linear. Under

51402 VOLUME 8, 2020

L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

such assumption, the tire model can be characterized by a TABLE 1. Vehicle parameters.
simplified linear model that the lateral tire forces are approx-
imately linear with respect to the tire slips and given as
Fyf = Cf αf (2)
Fyr = Cr αr (3)
where Fyf and Fyr are the lateral tire force of the front and
rear wheels, respectively, αf and αr denote the front and rear
wheel slip angles, respectively, Cf and Cr denote the front
and rear wheel cornering stiffness, respectively. The tire slip
angles αf and αr can be expressed as
lf ψ̇
αf = β + − δf (4)
v
lr ψ̇
αr = β − (5)
v
With the aforementioned linear tire model and certain FIGURE 3. Lateral control scheme where rdesired , rreal denote the
assumptions: 1) Ignoring the weight transfers and road desired and real yaw rate, respectively. er is the error between the
desired and real yaw rate.
bank angle, the left and right tire sideslip angles on the
same axle are identical; 2) The roll and pitch dynamics are
neglected, the vehicle lateral dynamic model can be expressed
speed are obtained from existing modules. Therefore, the
as [2], [27]
steering angles of front wheels δf is the only output of the
ẍ = ψ̇ ẏ + ax (6a) proposed controller.
2 In this section, a novel lateral control scheme is proposed,
−ψ̇ ẋ + Fyf cos δf + Fyr



ÿ = (6b)
m which is the main contribution of our work. The proposed
2 control scheme is illustrated in Figure 3. It is designed as a
ψ̈ =


lf Fyf − lr Fyr (6c)
Iz hybrid MPC-PID cascade control loop. The external control
Ẋ = ẋ cos(ψ) − ẏ sin(ψ) (6d) loop produces the yaw rate reference signal using a kine-
matic MPC controller with vehicle sideslip compensation.
Ẏ = ẋ sin(ψ) + ẏ cos(ψ) (6e)
Compared to the PID control in Marino’s work [4], the MPC
where ax is longitudinal acceleration, m denotes vehicle mass, control law has considerable advantage on rejecting the dis-
Iz is yaw moment of inertia. turbances on road curvature and velocity variation, with the
It can be seen from the comparison between kinematic ability to predict future behaviors of vehicle. The inner PID
and dynamic model that although the kinematic model also control loop is to track rapidly the yaw rate reference coming
involves vehicle sideslip angle β, it assumes that all tire slip from the external one. As mentioned in introduction, this
angles are deemed to be zero which will lead to significant design based on yaw rate is inspired by the existing studies
model mismatch as tire slip angles increase, such as at high that additional feedback of the yaw rate leads to a significant
speed scenarios. This drawback of kinematic model motivates reduction of tracking error in nearly all driving maneuvers,
the proposed vehicle sideslip compensator which is one of our as it improves the system transients [2], [4], [26].
main contributions.
The main vehicle parameters are summarized in Table 1 and A. CASCADED MPC-PID CONTROL
the tire cornering stiffness is determined by the tire model Model predictive control has been widely used in the field of
depicted in Figure 2 that the lateral tire forces are calculated path tracking, in general, which can be roughly classified into
as a function of vertical load, lateral tire slip angle. The source two methods: kinematic MPC and dynamic MPC, depending
data in Figure 2 is from Carsim software by setting the type on the vehicle model [30]. Each method has its own pros and
of tires as ‘‘225/60 R18’’. cons. kinematic MPC is simple, but only works well at low
speed. As speed increases, the kinematic model mismatch
III. CONTROLLER DESIGN will result in large tracking error. On the contrary, dynamic
As mentioned above, the objective of path tracking control MPC can overcome the impact of increasing speed, however,
is to keep the vehicle as close as possible to the given it has the drawback of poor computational efficiency and
path under the desired speed. In this paper, we decouple becomes singular at low vehicle speeds, no matter linear
the problems of path tracking into lateral control design and dynamic model or nonlinear model.
longitudinal control design, which is similar to many previous To deal with the above dilemma, we explore a cascaded
work [28], [29]. Additionally, we only focus on the lateral kinematic MPC-PID controller in this section, with the expec-
control under the assumptions that the given path and desired tation that PID feedback control of yaw rate is capable

VOLUME 8, 2020 51403

 L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

of rejecting uncertainties and modeling errors, meanwhile,

the controller retains superior computational efficiency. Con-
sidering the fact that the output of predicting model is yaw
rate rather than the steering angles of front wheels δf , the δf
needs to be eliminated from (1c), (1d). Substituting from (1c)
into (1d), the vehicle sideslip angle β can be rewritten as
 
−1 lr
β = sin ψ̇ (7)
v
Then, substituting from (7) into (1a), (1b), the kinematic
model used in MPC can be rewritten as
  
−1 lr
Ẋ = v cos ψ + sin ψ̇ (8a)
v
  
lr
Ẏ = v sin ψ + sin−1 ψ̇ (8b)
v
ψ̇ = r (8c)
FIGURE 4. Simulation system architecture.
where yaw rate r is the output of MPC controller and yaw
rate reference tracked by the PID inner loop via controlling
the steering angles of front wheels. Once the MPC control loop is designed, the remaining step
Based on the kinematic model (8), the desired yaw rate is to design the PID inner loop controller. The goal of the
is obtained by a typical MPC module. We used the publicly PID controller is to minimize the yaw rate error er between
available solver IPOPT [31] to solve the following optimiza- the measured yaw rate rreal and desired yaw rate rdesired .
tion problem. At each time, the following constrained finite The relationship between the error er and output δf can be
horizon optimal control problem is solved: formulated in the following standard PID control law,
XHp
T
er = rdesired − rreal (11)
min zi − zref ,i Q zi − zref ,i Z t
der (t)
u i=1
XHc −1 h i δf = Kp er (t) + Ki er (t)dt + Kd (12)
+ (ui − ui−1 )T M (ui − ui−1 ) + uTi Rui 0 dt
i=0
with proportional, integral and differential gain Kp , Ki , Kd .
s.t. z0 = z(t), u−1 = u (t − ts )
It should be noted that yaw rate would not change with
zi+1 = f (zi , ui ) , i = 0, . . . , Hp − 1 steering angles δf at a standstill, which can also be derived
rmin,i ≤ ui ≤ rmax,i , ∀i by equation (1c). Therefore, this cascaded MPC-PID control
1rmin,i ≤ ui − ui−1 ≤ 1rmax,i , ∀i (9) method is not appropriate for stop-and-go scenarios and auto-
matic parking, if Ki is not set to zero.
where, as in standard MPC notation, Q, M and R are weight-
ing matrices of appropriate dimensions. The reference sig- B. STEADY-STATE RESPONSE with MPC-PID CONTROL
nal zref represents the desired output, where z = [ψ, Y ]0 . In order to test the above MPC-PID control law, we imple-
Hp , Hc denote the prediction horizon and control horizon, mented the MPC controller in C++ and evaluated it on
respectively. Time ts is the sampling time of the path tracking Carsim/Matlab joint simulation environment. An overview of
controller. f (zi , ui ) denotes the state update with the kine- the simulation system architecture is outlined in Figure 4. The
matic model derived in (8) through forward Euler. In partic- simulation system is composed of two personal computers.
ular, if f (zi , ui ) update with the kinematic model (1) or the One is used for the path tracking controller based on Ubuntu
dynamic model (6), we will achieve classic kinematic MPC OS and Robot Operating System with an intel i5-4590 pro-
and dynamic MPC controllers, respectively [30]. The vari- cessor. The other one aims at providing a simulator involv-
ables rmin,i , rmax,i , 1rmin,i , 1rmax,i denote lower and upper ing path and autonomous vehicle based on Carsim/Matlab.
bounds of the yaw rate and the constraints on increment Carsim, as a high-fidelity vehicle simulator, utilizes detailed
of yaw rate, respectively. ut,i = ut , . . . , ut+Hc −1 is the
 
nonlinear tire models and vehicle models to simulate the
optimization vector at time t. Especially, u−1 represents the dynamic behaviors of different types of vehicle and is widely
control action at the previous sampling step and the first value used in automotive industry [20]. The TCP/IP networking
ut,0 of the optimization vector is used as the optimal control protocol is used to communicate between the above two
action: computers. Table 2 lists the main controller parameters which
rdesired = ut,0 (10) are the same for all controllers presented in this paper.
Figure 5 shows the steady-state responses of three types
where rdesired is the MPC controller output, namely, the yaw of controllers for following a circular path with a radius
rate reference tracked by PID controller of the inner loop. of 100m as a function of vehicle speed. These controllers

51404 VOLUME 8, 2020

L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

TABLE 2. Main controller parameters.

FIGURE 6. The performance of PID control for tracking the desired yaw
rate.

rate results in a controller with poor tracking accuracy. The

cascaded MPC-PID control law never obtains the lowest
tracking error no matter how fast the vehicle moves, com-
pared to the other two MPC control methods. However,
the cascaded MPC-PID controller still makes a difference in
FIGURE 5. Steady-state response for following a circular path with a
radius of 100m. Each simulation test indicated by dots on the path tracking that the tracking error is relatively stable and
corresponding curve was implemented at constant vehicle speed, ranging is limited into a range of 0.4 m without excessive deviation
from 10 to 100 in intervals of 10 km/h.
from the desired path, which demonstrates the contribution
of the feedback of yaw rate and also implies some systematic
are classic kinematic MPC, dynamic MPC and MPC-PID bias.
control law, based on kinematic model (1), dynamic model To further find out the reason for the poor accuracy of
(6) and the modified kinematic model (8), respectively. It is MPC-PID controller, the performance of PID control for
apparent that reaching a certain point, the lateral tracking tracking the desired yaw rate is illustrated in Figure 6 which
errors grow with the vehicle speed under the classic kine- is extracted from one of the above steady-state tests under
matic MPC control, on the contrary, if the speed is decreased the control of MPC-PID controller. The results show that the
to a certain point, the tracking errors will be increased for inner PID control loop follows the desired yaw rate very well
dynamic MPC. The point of intersection of the above two when the vehicle speed is held at 50 km/h. Therefore, there
curves is close to 35 km/h. This confirms the expected must be some significant model mismatch in the external
results from the consensus in existing literature that kine- MPC control loop and the desired yaw rate generated via the
matic model is unsuitable for high-speed path tracking, and kinematic model (8) needs to be corrected. With this in mind,
the dynamic MPC control becomes singular at low vehicle a compensation for vehicle sideslip is proposed to alleviate
speeds [25], [30]. Moreover, it is of interest to note that the above model mismatch and improve tracking accuracy
directly combining MPC with PID feedback control of yaw under MPC-PID control law.

VOLUME 8, 2020 51405

 L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

C. COMPENSATION FOR VEHICLE SIDESLIP

The classic kinematic model (1) and the derived kinematic
model (8) both involve the vehicle sideslip angle β, however,
it should be noted that obtaining the vehicle sideslip angle β
in both model are under the assumptions that the slip angles
at both wheels are zero. As it is shown in (4), (5), there
exists direct correlation among β and the tire slip angles
αf , αr . Therefore, in high-speed scenarios, these assumptions
will lead to severe mismatch between predicting model and
vehicle dynamics inevitably. In this section, a more accurate
form of vehicle sideslip angle β is derived from steady-state
cornering conditions with linear tire model. First, at steady-
state conditions (road curvature rate k̇ = 0, v̇ = 0), the rear
tire forces can be given by the following simplified equation
as described by Kapania et al. [1].
mlf 2
Fyr = v k (13)
lf + lr
FIGURE 7. Steady-state response with sideslip compensation.
Then, substituting from (13) and linear tire model (3) into (5),
yields the vehicle sideslip angle:
mlf lr ψ̇ A. STEADY-STATE RESPONSE
β= v2 k + (14) Figure 7 demonstrates the steady-state response of the
(lf + lr )Cr v
designed MPC-PID control law with sideslip compensation
Since road curvature k = R1 = ψ̇v , the proposed sideslip (K-MPC-PID-C). The simulation setting is consistent with
compensator can be obtained by rewritten (14) as that in Figure 5. The results show the excellent performance
mlf lr ψ̇ of the proposed controller that outperforms all the other
β= vψ̇ + (15) controllers in this steady-state test. First, at low speed both
(lf + lr )Cr v
classic kinematic MPC and our controller follow the desired
where yaw rate ψ̇ is the output of the proposed MPC control path well, however, as vehicle speed increases the classic
loop. Therefore, at each prediction stage of solving the MPC kinematic MPC becomes inacceptable, but the tracking error
problem, the predicted sideslip angle β will update with ψ̇. of the proposed control law is immune to increasing. Second,
Then substituting from (15) into (1a), (1b), yields the kine- the proposed control law is superior to both dynamic MPC
matic model: and MPC-PID without compensation throughout the whole
Ẋ = v cos(ψ + β) (16a) range of vehicle speed. Third, at both low speed and high
speed, even at the limits of handling, the maximum of lateral
Ẏ = v sin(ψ + β) (16b)
errors under sideslip compensation are less than 0.1 m.
ψ̇
β = K1 vψ̇ + K2 (16c)
v B. TRANSIENT RESPONSE
ψ̇ = r (16d)
Figure 8 shows the transient state responses for tracking a
ml
where K1 = (lf +lrf )Cr , K2 = lr . In practice, parameters sinusoidal path with an amplitude of 4m and a wavelength
K1 , K2 can be empirically tuned through simulation or field of 50 m, mimicking the lane-change maneuver. The results
test, and their effects are further explored in Section IV. indicate that all controllers have the similar tracking perfor-
Note that the sideslip compensator will become singular at a mance as the steady-state tests and the designed controller
standstill because of the denominator involving the velocityv. still follows the sinusoidal path well throughout the whole
To avoid this situation, the velocity v will be replaced by a range of vehicle speed, despite approaching the limits of
threshold when the vehicle starts at a standstill. Replacing handling with the maximum lateral acceleration of 8.4 m/s2 .
kinematic model (8) with kinematic model (16) under the This figure also illustrates that although the classic kinematic
framework of cascaded MPC-PID control, we obtain the MPC achieves the best tracking accuracy when speed is
proposed controller. lower than 30 km/h, the maximum tracking errors of the
proposed controller also have not exceed a limit of 0.1 m.
IV. CONTROLLER PERFORMANCE Then, as speed increases, the designed controller becomes
To better understand the performance of the proposed con- the only available control law in this test with maximum error
trol law, in this section we analyze system steady-state and of 0.16 m. In addition, one can observe that the dynamic MPC
transient responses, the effect of the parameters K1 , K2 , the is not capable of following the desired path at high lateral
robustness against measurement error and parameters uncer- acceleration conditions as the steady-state test in Figure 7,
tainties, and computing efficiency. despite the fact that the vehicle reaches to a similarly high

51406 VOLUME 8, 2020

L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

 
FIGURE 8. Transient state responses as a function of vehicle speed for following a sinusoidal trajectory Y = 2 sin 2π
l
(x + 4l ) − 2, l = 50. Each
simulation test indicated by dots on the corresponding curve was implemented at constant vehicle speed, ranging from 10 to 60 in intervals of 5 km/h.

level of speed (beyond 40 km/h). The reason for this is that trend of the lateral tracking error for the designed controller
the tire dynamics become so difficult to model at the limits at varying group of the parameters K1 , K2 by steady-state
of handling where the mismatch of linear dynamic model test. The results reveal a tradeoff in tuning K1 and K2 . When
is unable to be neglected anymore. It is apparent that this K1 is set to zero, the larger K2 is capable of providing the
drawback makes it difficult to combine the kinematic MPC better tracking accuracy at low speed, but the worse tracking
and dynamic MPC, that is, the controller cannot trigger a accuracy at high speed. Then, when K2 is set to a proper
switch from kinematic MPC to dynamic MPC just accord- constant, K1 has the ability to improve the tracking accuracy
ing to the increase of vehicle speed, because their tracking at high speed while ensure the tracking performance at low
performances depend on not only vehicle speed, but also speed. It can be see that the maximum of lateral error takes
road curvature and the resulting lateral acceleration. The place at the speed of around 80 km/h with the coupling effect
above weakness of classic MPC for path tracking highlights of K1 and K2 , rather than the top speed of 100 km/h at the
the contribution of the proposed control law that with the limits of handling. Consequently, this coupling effect could
derived sideslip compensation, the cascaded MPC-PID con- be used to tune the two parameters K1 and K2 for the best
trol achieves exact tracking accuracy from low speed to high tracking performance in accordance with the highest design
speed with different road curvature. speed. In addition, the designed sideslip compensator could
be also incorporated into those advanced online estimations
C. EFFECT of the PARAMETERS K1 , K2 on tire cornering stiffness [32] and vehicle mass [33] for
Parameters K1 , K2 derived from the sideslip compensator (15) further improvement in path tracking.
and (16c) have their own theoretical definition, but in practice
we may be unable to obtain the precise value of those related D. ROBUSTNESS
parameters such as vehicle mass and tire cornering stiffness. In essence, the designed control law is a cascaded MPC-PID
Therefore, for a better sense of the effect of the parameters control that the inner PID feedback control loop struggles
K1 , K2 in sideslip compensator, Figure 9 shows the changing to track rapidly the desired yaw rate produced by the MPC

VOLUME 8, 2020 51407

 L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

is successful to be deployed, the mean of measurement error

can be eliminated by control parameter tuning. Therefore,
the simulation results prove that the controller is capable of
following the path with the measurement error of yaw rate
sensors, which is further validated in field test presented in
next section.
Controller robustness against parameters uncertainties:
The robustness of the controller is evaluated over vehicle
parameters uncertainty with respect to vehicle mass and tire
cornering stiffness. Figure 11 shows the steady-state and
transient responses where payloads of four passengers with
different mass are set in Carsim to simulate the effect of
mass uncertainties. The steady-state tracking error increases
as payload increases, especially at high speed, however, all
tracking trajectory in transient responses nearly coincide with
each other even at the limits of handling where the result-
ing peak lateral acceleration is 8.6 m/s2 . This interesting
FIGURE 9. The effect of the parameters K1 , K2 in sideslip compensator. observation is consistent with the mathematical derivation
of sideslip compensator (15) and (16c) that only gain K1
loop. Therefore, the measurement of yaw rate plays a key role involves the term mass and the product of K1 and vehicle
in this approach compared to other path tracking methods. speed is used to compensate the sideslip dynamics. Thus, The
For this reason, the robustness of the proposed controller robustness against vehicle mass decreases as speed increases.
performance with respect to the measurement error of the yaw Nevertheless, the designed controller is still able to meet the
rate sensor is tested in steady-state and transient scenarios demand of path tracking with acceptable errors despite the
near the limits of handling. In addition, the robustness against variations of vehicle mass, as shown in Figure 11.
parameters uncertainties regarding the vehicle mass and tire As for tire cornering stiffness, there are more compli-
cornering stiffness are also validated with simulation, because cated responses than that of vehicle mass variations. First,
the derived kinematic model (16) with sideslip compensation Figure 12 shows that when the cornering stiffness drops
is related to both of them. 25 percent or more, the vehicle will slide off the road at the
Controller robustness against measurement error of the limits of handling because the tires are incapable of producing
yaw rate sensor: adequate lateral forces. Second, it can be observed that the
The measurement error of the yaw rate is modeled as controller has the ability to reject uncertainties that cornering
Gaussian distribution with a Matlab block called Random stiffness has increased by 25%, however, as it further climbs,
Number and injected to the original yaw rate calculated by the tracking error is beyond 0.2 m indicating that the con-
the Carsim vehicle model. In field test presented in Section V, troller fails to follow the desired path. Like the impact of vehi-
the yaw rate is measured by the inertial navigation system cle mass, the cornering stiffness variations impose stronger
RT3002 which measures the yaw rate at the error level of disturbances at high speed than low speed with similarly high
0.01 ◦ /s, 1σ . Then, the disturbance of measurement error is level of resulting lateral acceleration.
set to two orders of magnitude higher than that in the sensor
specification, i.e., 1 ◦ /s. Moreover, in the transient test, we add E. COMPUTING EFFICIENCY
the case that measurement error mean or variance is 10 ◦ /s Figure 13 shows the comparison for computational cost of
to explore where the tracking performance will get worse, three types of MPC controller for path tracking. we imple-
although the measurement accuracy of real sensors will not mented all these MPC controllers in C++ with solver IPOPT
be so inferior. As shown in Figure 10, the results of steady on Linux system with an intel i5-4590 processor. Moreover,
and transient tests show that if the measurement error is the same MPC parameter configurations are set to ensure a
within the range that both the mean and variance are less than fair comparison that both the prediction horizon and control
1 ◦ /s, the tracking performance of the proposed controller is horizon are set to 10 steps, and sampling time are set to
not significantly affected by the disturbance of measurement 0.1 s. It can be observed that the proposed controller reaches
error and controller achieves the equivalent tracking accuracy a comparable level of computational efficiency compared
compared to the cases of no noises. As the level of distur- to classic kinematic MPC controller. The reason for this
bance further increases, the actual tracking path has gradually is that the proposed controller is also based on kinematic
deviated from the desired path as shown in Figure 10 (b). vehicle model. Thus, both kinematic MPC have almost the
The tracking error is approximately 0.5m when mean of same computing efficiency in solving the problem of path
measurement error is 10 ◦ /s and variance is 1 ◦ /s. However, tracking. In addition, it is obvious that the dynamic MPC con-
the 0.01◦ /s angular rate accuracy with sensor RT3002 is far troller has the worst computational efficiency compared to
better than the above error of 1 ◦ /s and once control system others.

51408 VOLUME 8, 2020

L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

FIGURE 10. The robustness against the measurement error of the yaw rate sensor: a) steady-state, b) transient-state at the limits of handling (60 km/h).

FIGURE 11. Robustness against parameters uncertainties: vehicle mass uncertainties. a) steady-state, b) transient-state at the limits of handling
(60 km/h).

V. FIELD TEST adequate to validate the robustness against measurement error

A. VEHICLE PLATFORM of the yaw rate sensor and time lag, which are the main con-
The proposed cascaded MPC-PID with sideslip com- cern about validation on the premise of the aforementioned
pensator was implemented on an autonomous electric simulations.
vehicle—Dongfeng A60EV, as shown in Figure 14. A real-
time kinematic (RTK) positioning system and inertial mea- B. EXPERIMENTAL RESULTS
surement unit RT3002 were used to obtain global vehicle Figure 15 presents the experimental results for following a
states. The controller was implemented in C++ under circular path on concrete road surface. The vehicle speed
Ubuntu and Robot Operating System with an intel i7-6700k varies from 10 km/h to 32 km/h in consideration of safety.
processor and operates at 50 Hz. The target vehicle longi- Note that although the vehicle speed is not very high in
tudinal speed was followed by a PID controller. Because of this field test, the resulting maximum lateral acceleration
space limitations of experimental field, only a circular path could be aggressive (5.84 m/s2 ), due to the small road
with a radius of 14 m was used for experimental validation. radius. It can be found that the proposed controller restricts
Although it does not cover all test cases to verify the proposed the lateral tracking error into a range of 0.08 m, which
approach, especially lacking of high speed scenarios, it is denotes remarkable tracking accuracy and is identical with

VOLUME 8, 2020 51409

 L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

FIGURE 12. Robustness against parameters uncertainties: cornering stiffness uncertainties. a) steady-state test, Fy-R1 denotes the lateral forces of right
front tire. b) transient-state at the limits of handling (60 km/h).

FIGURE 14. Dongfeng A60EV used for field test.

by MPC loop is followed well by the real yaw rate and

both yaw rate change smoothly, which confirms the outstand-
ing prediction performance of MPC control with sideslip
FIGURE 13. Comparison of computing time with the same MPC compensation and the robustness under real inertial mea-
configurations that both the prediction horizon and control horizon are
set to 10 steps, and sampling time are set to 0.1 s. surement unit. In short, this experimental results indicate
that in the real world the proposed controller is capable
the simulation results in Figure 7. Moreover, the plot of yaw of providing comparably good tracking performance with
rate in Figure 15 shows that the desired yaw rate generated simulation.

51410 VOLUME 8, 2020

L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

FIGURE 15. Experimental results for following a circular path with a radius of 14 m.

VI. CONCLUSION low speed, with the improvement for path tracking that the
This paper describes the design of a cascaded kinematic tracking errors are guaranteed less than 0.16 m, ranging from
MPC-PID controller with vehicle sideslip compensation for low speed to high speed even at the limits of handling. In addi-
path tracking of autonomous vehicles. A kinematic MPC tion, due to the utilization of kinematic model, the proposed
based on yaw rate is derived to tackle the disturbances of control method reaches a comparable level of computational
the upcoming road curvature at various speeds. Subsequently, efficiency compared to classic kinematic MPC. Furthermore,
in consideration of the kinematic model mismatch at high simulation and field experiments conducted with the A60EV
speed, a novel vehicle sideslip compensator is proposed to autonomous vehicle validate the robustness against sensor
correct model prediction and is integrated into the kinematic disturbances and time lag. Lastly, this research has also con-
model. Note that the MPC control loop outputs desired yaw firmed that
rate rather than steering angle compared to classic MPC 1) the developed vehicle sideslip compensator is sufficient
controller for path tracking. Then, a PID control is designed to capture steer dynamics and mitigate the effect of vehicle
to follow the reference yaw rate, which takes full advantage of sideslip angle in the proposed control architecture. This find-
the feedback of yaw rate to reject uncertainties and modeling ing may be incorporated into other control law based on the
errors. feedback of yaw rate to improve path tracking.
The proposed controller performances involving steady- 2) As for classic MPC controller for path tracking, the con-
state and transient response, robustness, and computing effi- troller cannot trigger a switch from kinematic MPC to linear
ciency were evaluated on Carsim/Matlab joint simulation dynamic MPC just according to the increase of vehicle speed,
environment. The simulation results demonstrate that the because their tracking performances depend on not only vehi-
proposed controller is successful to resolve the dilemma that cle speed, but also road curvature and the resulting lateral
kinematic MPC only works well at low speed while dynamic acceleration, consequently, which makes it difficult to com-
MPC has poor computational efficiency and gets worse at bine the kinematic MPC and dynamic MPC for improving

VOLUME 8, 2020 51411

 L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

tracking performance. The above weakness of classic MPC [21] A. Carvalho, Y. Gao, S. Lefevre, and F. Borrelli, ‘‘Stochastic predictive
for path tracking highlights the contribution of the proposed control of autonomous vehicles in uncertain environments,’’ in Proc. 12th
Int. Symp. Adv. Vehicle Control, 2014, pp. 712–719.
control law. [22] M. Bojarski, D. Del Testa, D. Dworakowski, B. Firner, B. Flepp,
Future work will focus on the implementation of the P. Goyal, L. D. Jackel, M. Monfort, U. Müller, J. Zhang, X. Zhang,
proposed control method with embedded platform such as J. Zhao, and K. Zieba, ‘‘End to end learning for self-driving cars,’’ 2016,
arXiv:1604.07316. [Online]. Available: https://arxiv.org/abs/1604.07316
NVIDIA Nano Kit. [23] N. R. Kapania and J. C. Gerdes, ‘‘Path tracking of highly dynamic
autonomous vehicle trajectories via iterative learning control,’’ in Proc.
Amer. Control Conf. (ACC), Jul. 2015, pp. 2753–2758.
REFERENCES [24] S. Liu, Z. Hou, T. Tian, Z. Deng, and Z. Li, ‘‘A novel dual successive
[1] N. R. Kapania and J. C. Gerdes, ‘‘Design of a feedback-feedforward projection-based model-free adaptive control method and application to an
steering controller for accurate path tracking and stability at the limits of autonomous car,’’ IEEE Trans. Neural Netw. Learn. Syst., vol. 30, no. 11,
handling,’’ Vehicle Syst. Dyn., vol. 53, no. 12, pp. 1687–1704, Dec. 2015. pp. 3444–3457, Nov. 2019.
[2] G. Tagne, R. Talj, and A. Charara, ‘‘Design and comparison of robust [25] S. Dixit, S. Fallah, U. Montanaro, M. Dianati, A. Stevens, F. Mccullough,
nonlinear controllers for the lateral dynamics of intelligent vehicles,’’ IEEE and A. Mouzakitis, ‘‘Trajectory planning and tracking for autonomous
Trans. Intell. Transp. Syst., vol. 17, no. 3, pp. 796–809, Mar. 2016. overtaking: State-of-the-art and future prospects,’’ Annu. Rev. Control,
[3] M. T. Emirler, İ. M. C. Uygan, B. A. Güvenç, and L. Güvenç, ‘‘Robust vol. 45, pp. 76–86, Jan. 2018.
PID steering control in parameter space for highly automated driving,’’ Int. [26] J. Ackermann, J. Guldner, W. Sienel, R. Steinhauser, and V. I. Utkin,
J. Veh. Technol., vol. 2014, Nov. 2014, Art. no. 259465. ‘‘Linear and nonlinear controller design for robust automatic steering,’’
IEEE Trans. Control Syst. Technol., vol. 3, no. 1, pp. 132–143, Mar. 1995.
[4] R. Marino, S. Scalzi, G. Orlando, and M. Netto, ‘‘A nested PID steering
[27] J. Guo, Y. Luo, K. Li, and Y. Dai, ‘‘Coordinated path-following and
control for lane keeping in vision based autonomous vehicles,’’ in Proc.
direct yaw-moment control of autonomous electric vehicles with sideslip
Amer. Control Conf., 2009, pp. 2885–2890.
angle estimation,’’ Mech. Syst. Signal Process., vol. 105, pp. 183–199,
[5] B. Mashadi, M. Mahmoudi-Kaleybar, P. Ahmadizadeh, and A. Oveisi,
May 2018.
‘‘A path-following driver/vehicle model with optimized lateral dynamic
[28] R. Attia, R. Orjuela, and M. Basset, ‘‘Combined longitudinal and lateral
controller,’’ Latin Amer. J. Solids Struct., vol. 11, no. 4, pp. 613–630,
control for automated vehicle guidance,’’ Vehicle Syst. Dyn., vol. 52, no. 2,
Aug. 2014.
pp. 261–279, Feb. 2014.
[6] M. Buehler, K. Iagnemma, and S. Singh, The 2005 DARPA Grand Chal- [29] J. Zhao and A. E. Kamel, ‘‘Integrated longitudinal and lateral control
lenge: The Great Robot Race, vol. 36. Berlin, Germany: Springer, 2007. system design for autonomous vehicles,’’ IFAC Proc. Volumes, vol. 42,
[7] R. Liu and J. Duan, ‘‘A path tracking algorithm of intelligent vehicle by pre- no. 19, pp. 496–501, 2009.
view strategy,’’ in Proc. 32nd Chin. Control Conf., 2013, pp. 5630–5635. [30] J. Kong, M. Pfeiffer, G. Schildbach, and F. Borrelli, ‘‘Kinematic and
[8] Y. Shan, W. Yang, C. Chen, J. Zhou, L. Zheng, and B. Li, ‘‘CF-pursuit: dynamic vehicle models for autonomous driving control design,’’ in Proc.
A pursuit method with a clothoid fitting and a fuzzy controller for IEEE Intell. Vehicles Symp. (IV), Jun. 2015, pp. 1094–1099.
autonomous vehicles,’’ Int. J. Adv. Robot. Syst., vol. 12, no. 9, p. 134, [31] L. T. Biegler, A. M. Cervantes, and A. Wächter, ‘‘Advances in simultaneous
Sep. 2015. strategies for dynamic process optimization,’’ Chem. Eng. Sci., vol. 57,
[9] N. H. Amer, H. Zamzuri, K. Hudha, and Z. A. Kadir, ‘‘Modelling and no. 4, pp. 575–593, Feb. 2002.
control strategies in path tracking control for autonomous ground vehicles: [32] Z. Chu, N. Chen, N. Zhang, and G. Li, ‘‘Path-tracking for autonomous
A review of state of the art and challenges,’’ J. Intell. Robot. Syst., vol. 86, vehicles with on-line estimation of axle cornering stiffnesses,’’ in Proc.
no. 2, pp. 225–254, May 2017. Chin. Control Conf. (CCC), Jul. 2019, pp. 6651–6657.
[10] T. Ming, W. Deng, S. Zhang, and B. Zhu, ‘‘MPC-based trajectory [33] B. L. Boada, M. J. L. Boada, and H. Zhang, ‘‘Sensor fusion based on a dual
tracking control for intelligent vehicles,’’ SAE, Warrendale, PA, USA, Kalman filter for estimation of road irregularities and vehicle mass under
Tech. Rep. 0148-7191, 2016. static and dynamic conditions,’’ IEEE/ASME Trans. Mechatronics, vol. 24,
[11] B. Zhang, C. Zong, G. Chen, and B. Zhang, ‘‘Electrical vehicle path no. 3, pp. 1075–1086, Jun. 2019.
tracking based model predictive control with a Laguerre function and
exponential weight,’’ IEEE Access, vol. 7, pp. 17082–17097, 2019.
[12] C. Shen, H. Guo, F. Liu, and H. Chen, ‘‘MPC-based path tracking controller
design for autonomous ground vehicles,’’ in Proc. 36th Chin. Control Conf.
(CCC), 2017, pp. 9584–9589. LUQI TANG received the B.S. degree in ther-
[13] H. Wang, B. Liu, X. Ping, and Q. An, ‘‘Path tracking control for mal energy and power engineering and the M.S.
autonomous vehicles based on an improved MPC,’’ IEEE Access, vol. 7, degree in power machinery and engineering from
pp. 161064–161073, 2019.
the School of Automotive Engineering, Wuhan
[14] C. Sun, X. Zhang, Q. Zhou, and Y. Tian, ‘‘A model predictive controller University of Technology, China, in 2012 and
with switched tracking error for autonomous vehicle path tracking,’’ IEEE
2015, respectively, where he is currently pursuing
Access, vol. 7, pp. 53103–53114, 2019.
the Ph.D. degree in automotive engineering. His
[15] M. Babu, R. R. Theerthala, A. K. Singh, B. P. Baladhurgesh, B. Gopalakr-
research interests include decision making, vehicle
ishnan, K. M. Krishna, and S. Medasani, ‘‘Model predictive control for
dynamics, and control for autonomous vehicles.
autonomous driving considering actuator dynamics,’’ in Proc. Amer. Con-
trol Conf. (ACC), Jul. 2019, pp. 1983–1989.
[16] J. Cai, H. Jiang, L. Chen, J. Liu, Y. Cai, and J. Wang, ‘‘Implementation and
development of a trajectory tracking control system for intelligent vehicle,’’
J. Intell. Robot. Syst., vol. 94, no. 1, pp. 251–264, Apr. 2019.
[17] S. Xu and H. Peng, ‘‘Design, analysis, and experiments of preview path
tracking control for autonomous vehicles,’’ IEEE Trans. Intell. Transp. FUWU YAN is currently the Dean of the School
Syst., vol. 21, no. 1, pp. 48–58, Jan. 2020. of Automotive Engineering, Wuhan University of
[18] J. Bechtoff, L. Koenig, and R. Isermann, ‘‘Cornering stiffness and Technology, where he is also a Chief Professor of
sideslip angle estimation for integrated vehicle dynamics control,’’ IFAC- automotive engineering. He is also the Director
PapersOnLine, vol. 49, no. 11, pp. 297–304, 2016. of the Hubei Research Center for New Energy &
[19] H. Pacejka, Tire and Vehicle Dynamics. Amsterdam, The Netherlands: Intelligent Connected Vehicle. His research inter-
Elsevier, 2005. ests include electrification of vehicles, powertrain
[20] C. Zhang, J. Hu, J. Qiu, W. Yang, H. Sun, and Q. Chen, ‘‘A novel fuzzy control, and autonomous vehicles.
observer-based steering control approach for path tracking in autonomous
vehicles,’’ IEEE Trans. Fuzzy Syst., vol. 27, pp. 278–290, Jul. 2019.

51412 VOLUME 8, 2020

L. Tang et al.: Improved Kinematic MPC for High-Speed Path Tracking of Autonomous Vehicles

BIN ZOU received the Ph.D. degree in power CHEN LV (Member, IEEE) received the Ph.D.
machinery and engineering from the Wuhan Uni- degree from the Department of Automotive Engi-
versity of Technology, Wuhan, China, in 2013. neering, Tsinghua University, Beijing, China,
He is currently an Associate Professor with the in 2016. He was a Research Fellow with the
School of Automotive Engineering, Wuhan Uni- Advanced Vehicle Engineering Center, Cranfield
versity of Technology. His current research inter- University, Cranfield, U.K., from 2016 to 2018,
ests include automobile electronic control and and a joint Ph.D. Researcher with the Electrical
autonomous vehicles. Engineering and Computer Sciences, University
of California, Berkeley, CA, USA, from 2014 to
2015. He is currently an Assistant Professor with
the School of Mechanical and Aerospace Engineering, Nanyang Techno-
logical University, Singapore. His research interests include cyber-physical
systems and advanced vehicle control and intelligence, in which he has
contributed more than 60 articles and obtained 11 granted China patents.
Dr. Lv was a recipient of the Highly Commended Paper Award of IMechE,
KEWEI WANG received the B.S. degree in U.K., in 2012, the NSK Outstanding Mechanical Engineering Paper Award,
automotive engineering from the School of in 2014, the China SAE Outstanding Paper Award, in 2015, the First Class
Automotive Engineering, Wuhan University of Award of China Automotive Industry Scientific and Technological Invention,
Technology, China, in 2013, where he is currently in 2015, and the Tsinghua University Outstanding Doctoral Thesis Award,
pursuing the Ph.D. degree in automotive engineer- in 2016. He serves as a Guest Editor for the IEEE/ASME TRANSACTIONS ON
ing. His research interests include environment MECHATRONICS and the IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, and
perception and multisensor fusion in autonomous an Associate Editor for the International Journal of Electric and Hybrid
driving. Vehicles and the International Journal of Vehicle Systems Modelling and
Testing.

VOLUME 8, 2020 51413