This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2995408, IEEE
Transactions on Vehicular Technology

Trajectory Tracking Control of Autonomous Vehicle

with Random Network Delay
Zhongkai Luan, Jinning Zhang, Wanzhong Zhao* and Chunyan Wang
Abstract—Random network delay will introduce uncertainty
essential to improve the trajectory tracking system robustness
into trajectory tracking model of the autonomous vehicle, which and stability in autonomous vehicle [7].
seriously deteriorates the vehicle’s control system stability and In control theory, robust control is a key approach for
trajectory tracking accuracy. In this paper, considering steering controller design that explicitly deals with uncertainty,
angle oscillation caused by random network delay, trajectory
therefore robust control theory is widely applied in uncertain
tracking system robustness and stability is analyzed and a linear
uncertain time-delay system is established. Comprehensively time-delay systems [8]–[11]. Wang et al. [12] presented a
considering control system accuracy, robustness, and robust H∞ model reference tracking controller for networked
computational efficiency in the rolling optimization of Model control systems, considering communication constraints of
Predictive Control (MPC) , Adaptive Model Predictive Control both controller-to-actuators and sensor-to-controllers. Robust
for Uncertain model (UM-AMPC) algorithm is proposed to control truly performs better in time-delay systems and can
predict control variables for the next sampling time and alleviate provide sufficient stability. However, limited by its
the target angle discontinuity. This is achieved by operating conservative characteristics, response speed performance is
target angle signal and augmented state variables, which are
received by the lower nodes during the period from the current
inevitably compromised. In addition, when the robust control
sampling time to network delay upper bound. The hardware-in- strategy is applied to stability control of actuator model with
the-loop simulation results show that the proposed algorithm can noise and other uncertainties, the target signal sent by the
effectively guarantee system stability and tracking accuracy of controller to the actuator is generally considered to be accurate.
the autonomous vehicle under random network delay. But for the trajectory tracking problem with random network
delay, the target signal received by the lower nodes will be
Index Terms—Autonomous vehicle; Trajectory tracking; discontinuous which means the target signal tracked by
Random network delay; Model predictive control. actuator is inaccurate. For example, the target signal that the
actuator tracking at the current sampling time may be sent by
I. INTRODUCTION the decision-making layer several sampling times ago.
Compared with conventional vehicle steering system, Therefore, it is impossible to eliminate the influence of
steering-by-wire (SBW) system of autonomous vehicle has incorrect signal imposed on the actuator by means of
more sensors, actuators and control units [1], [2]. The optimizing the control strategy in execution layer. The impacts
trajectory tracking control is realized by planning ideal of network delay must be considered in the process of target
trajectory, producing target angle based on collected data from signal generation in decision layer to provide the accurate
sensors, and transmitting target angle to SBW system low-level target angle signal for the actuator. In view of the above,
controller through the CAN [3], [4]. However, due to vehicle predictive control algorithm is considered as a promising
network bandwidth limitations in practice, the increasing approach, whose core idea is to design predictive controller on
burden in signal transmission will induce random network the basis of uncertainty modeling, so as to compensate for the
delay in lower level nodes when receiving the target message impact of network delay [13]–[16]. Liu et al. [17] proposed a
produced by upper controller nodes. And random network dynamic predictive feedback linearization controller to
delay may give rise to unexpected step changes in target angle compensate for the system delay caused by network
signals. Consequently, transient performance and control transmission delay and data loss. Li et al. [18] presented an
system stability are highly likely to be degraded, because the adaptive discrete sliding mode control (k-adsmc) network
system may oscillate when tracking the continuously varying predictive control algorithm for robust tracking control of
target signal [5], [6]. Therefore, taking the influence of random networked control systems with communication constraints. In
network delay into account in the control strategy design is the research field of trajectory tracking control of autonomous
vehicles, model predictive control (MPC) is widely used [19]–
[21]. However, since the intrinsic property of MPC is to solve
 the constrained open-loop quadratic optimization problem in a
This research presented within this article were supported by the National
Key R&D Program of China (Grant No.2017YFB0103600), the Jiangsu key rolling way at each sampling time, the optimality of MPC
R&D Plan (Grant No. BE2018124) and Industrial technology innovation
project of Anhui new energy vehicle and intelligent network vehicle in 2019. cannot guarantee closed-loop system stability [22]. Therefore,
Copyright (c) 2015 IEEE. Personal use of this material is permitted. with high-level random network delay, the uncertainty would
However, permission to use this material for any other purposes must be change the original system model greatly, and the existing
obtained from the IEEE by sending a request to [email protected].
Zhongkai Luan (e-mail: [email protected] ), Wanzhong Zhao (e- model predictive control algorithm hardly handle the tradeoff
mail: [email protected] ) and Chunyan Wang (e-mail: between tracking accuracy, system stability and single task
[email protected]) are with the Department of Vehicle Engineering,
Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China; cycle time of the trajectory tracking system.
Jinning Zhang (e-mail: [email protected]) is with Cranfield
University, Cranfield, MK43 0AL, U.K.

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Edinburgh. Downloaded on June 14,2020 at 19:13:06 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2995408, IEEE
Transactions on Vehicular Technology

In this paper, from the perspective of steering angle In the fixed ground coordinate system oxy, the continuous
oscillation, the influence of random network delay on tracking-error model of autonomous vehicle can be described
trajectory tracking model robustness and stability is analyzed. as [21]. Due to the limitation of paper length, only derivation
Then, a linear model under time-delay uncertainty is results are given in here.
established. Considering system robustness and accuracy under X  At X  Bt u (1)
network delay, as well as computational efficiency in the
X   y  yr x  xr   r   r y  yr x  xr 
T
rolling optimization process of Model Predictive Control
(MPC), an uncertain-model adaptive model predictive control u  uf  ufr
(UM-AMPC) algorithm is proposed. The algorithm can predict  -2  C f  Cr  f y 2  lr Cr  l f C f  
 0 x  0 0
control variable at the next sampling time according to target  mx x mx 
angle and augmented state variables, which are received by the  2C f  f f x 2l f C f  f 
  0 y 0 0
lower nodes during the period from the current sampling time  mx x mx 
to the upper bound of the network delay. The proposed At   0 0 0 1 0 0
 
 2  lr C r  l f C f  f 2  l f 2 C f  lr 2 Cr  
algorithm can effectively guarantee the stability and tracking  0 0 0
 Iz x x Iz x 
accuracy of the system under random network message delay.  cos   sin   x cos    y sin   0 0 0
The contributions of this work are summarized as follows.  
  sin   cos    y cos    x sin   0 0 0 
First of all, an adaptive model predictive control (UM-AMPC) T
  y  lf  
algorithm for uncertain system with CAN network delay is  2C f  2 f   
Bt  
2C f  x  2l f C f
0 0
proposed, comprehensively considering the robustness, 0
 m m Iz 
accuracy and real-time computation capability of the upper
f y 2C f  yr  l f r   2Cr  y  lr 
layer controller. The algorithm can handle with the uncertainty  
caused by random network delay to a certain extent, so as to x mx 2
provide a more reasonable target signal for the lower layer f x 2C f  f  y  l f  

actuator. Furthermore, a trajectory tracking control method x mx 2
considering random network delay is proposed for autonomous f 2l f C f  y  l f    2lr Cr  y  lr 

vehicle on basis of UM-AMPC algorithm. The uncertainties in x I z x2
signal transmission are taken into account to effectively where uf is the control input front wheel angle, x and y
mitigate the incurred actuator oscillation, and further avoid the
are longitudinal and lateral displacement of vehicle
real trajectory deviation from the ideal trajectory. In addition,
respectively; x and y are vehicle longitudinal and lateral
this work improves system robustness in the presence of road
disturbance to promote driverless vehicle safety on full speeds respectively;  and  are vehicle yaw angle and
workload and long-distance driving. yaw rate respectively. X r   yr xr r r yr xr  is the
T

The rest of this paper is arranged as follows: Section II is

value of the status parameters at the reference point; ufr is the
dynamic modeling of trajectory tracking system. In Section III,
the impact of random network delay on trajectory tracking reference value of the control input. Cf, Cr are the rigidity of
system performance is analyzed by double lane change (DLC) front and rear wheels respectively. m is the sprung mass of
test. In Section IV, an uncertain-model adaptive model vehicle.  f is the front wheel angle. lf, lr are the distances
predictive control algorithm is proposed. Section V presents from the center of mass to the front and rear axles
the hardware-in-the-loop simulation results and analysis. respectively.
Section VI is the conclusion of this paper. Discretizing the continuous tracking-error model in
equation 1, the discrete tracking-error model is shown in
II. DYNAMIC MODEL OF TRACKING SYSTEM equation 2.
The network message transmission process of autonomous X  k  1  Ad X  k   Bd u  k  (2)
vehicle trajectory tracking system is shown in Figure 1. The  Ad  e AtTs  L1  sI  At 1
trajectory tracking controller receives the message sent by the 
 Ts
A T  
sensor node through the CAN network, and calculates the  Bd   e t s Bt d
target angle required for planning trajectory tracking based on  0

the plant model, following which the target angle is transmitted where Ts is sampling time; L1 is the sign of Laplace inverse
to the steering-by-wire (SBW) system node through the CAN transformation; s is the Laplace operator.
network to realize the trajectory tracking. In the tracking system control, smoother control effect can
be obtained by taking control increment as control variable.
Therefore, a new state variable including the original system
state variables and control variables is established, as shown in
Formula 3.
 X k  
 k     (3)
u  k  1 
Fig.1 The trajectory tracking dynamics model

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Edinburgh. Downloaded on June 14,2020 at 19:13:06 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2995408, IEEE
Transactions on Vehicular Technology

In conclusion, the state space expression of discrete ideal

trajectory tracking system without considering random network
delay is shown in equation 4.
  k  1  Ad   k   Bd u  k  (4)
where,
A Bd 
， Bd   Bd 
Fig.3 The influence of network delay on the target signal of lower level nodes
Ad   d
0 I  I  In order to quantitatively analyze the influence of random
network delay on trajectory tracking system performance, the
III. ANALYSIS OF RANDOM NETWORK DELAY trajectory tracking error (TTE) index and the yaw rate tracking
In this Section, the mechanism of model uncertainty caused error (YRTE) index are defined, as shown in Equation 5.
by network delay is analyzed. The CAN network message e y  x   yr  x   y  x 

(5)
transmission is shown in Figure 2. When the sampling time is k, 
the target angle message sent by the controller is recorded as e  t   r  t     t 

Δuk. In no network delay case, the SBW system node receives where yr(x) is the ideal lateral coordinate corresponding to the
the message at k-th sampling time and executes the message in current longitudinal coordinate of the vehicle. y(x) is the
the time interval [kTs, (k+1)Ts]. However, due to random current actual lateral coordinate of the vehicle.   t  is the
network delay, the target angle message that the SBW system actual yaw rate at time t. r  t  is the ideal yaw rate at time t.
node responds at the k-th sampling time may be sent by the
controller node before the k-th sampling time [23]. The ideal yaw rate is calculated as shown in Equation 6
Limited by CAN bus sampling mechanism, SBW node will [24].
Vx  t 
only collect the signals received at the beginning of each r  t   f t  (6)
sampling cycle. If a signal is received between two sampling mVx (t) 2  Cr lr  Cf lf 
lf  lr 
points in a sampling period due to network delay, the signal 2Cf C r  lf  lr 
will not actually work. For example, as shown in Figure 2,
where Vx  t  is the longitudinal speed of the vehicle at time t.
although all messages are sent out from the controller node to
the SBW node, due to network delay and the sampling  f  t  is the front wheel angle at time t.
mechanism limitation, the control signal u  k  4 occupies Then, the influence of network delay on the performance
three sampling periods in the steering system node, robustness and system stability is analyzed by trajectory
consequently the other three signals u  k  3 , u  k  1 and tracking double lane change (DLC) test. The simulation speed
u  k  cannot function well in the calculation process of the is 120km/h, and Gaussian random signals with peak value of 1,
2 and 3 times of sampling time are added to the output signals
actual angle at the steering system node. of the controller. The length and width settings of each
segment in the DLC scene are based on ISO3888-2-2002_BS
standards, as shown in Table 1. In the simulation process, the
center line of the DLC scene is taken as the target trajectory to
evaluate the tracking performance.
Table 1 Length and Width of Each Segment of DLC
Length Lane offset Width
Section
/(m) /(m) /(m)
1 12 \ 1.1* width+0.25
2 13.5 \ \
3 11 1 width+1
4 12.5 \ \
Fig.2 Signal transmission of automobile CAN network
max{1.3*width
5 12 \
Therefore, compared with the smooth target angle +0.25, 3}
transmitted by the controller, the target angle actually received The controller output under the random network delay with
by the SBW system node may have multiple step changes, different peaks is shown in Figure 4, which indicates that with
which may cause the steering angle to oscillate in the tracking the increase of the peak value of the random network delay, the
process. That is, compared with the trajectory tracking model fluctuation in the angle signal output accordingly increases.
under the ideal network, the random network delay causes
uncertainties in the model. As shown in Figure 3, the message
sent by the controller node is a standard sinusoidal signal, but
due to network delay, the signal received by the SBW node has
multiple step changes.

a. Target angle of controller output

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Edinburgh. Downloaded on June 14,2020 at 19:13:06 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2995408, IEEE
Transactions on Vehicular Technology

polyhedron, and an adaptive model predictive control

algorithm for the uncertain model is proposed. The algorithm
flow is shown in Figure 6.

b. Angle error of controller output

Fig. 4 Controller output in test
The comparison of system performance under different
peak values of random network delay is shown in Figure 5. In
terms of TTE index, the average tracking error caused by the
random network delay with a peak value of 3*TS is 250.76% of Fig.6 UM-AMPC Algorithm
that of TS. In terms of YRTE index, the yaw rate error caused
A. Adaptive prediction model
by the random network delay with a peak value of 3*TS is
174.23% of the peak value of TS. Moreover, in the segment 5, At each sampling time, prediction horizon and control
straight-line segment of DLC scene, the vehicle has been horizon parameters will change the dimension of prediction
already in an oscillating state. equation coefficient matrix, thus affecting the solution
direction of rolling optimization [25]. In this section, the DLC
scenario centerline tracking problem is also taken as an
example to simulate and analyze the system performance with
different predictive horizon parameters.
In the simulation process, the relationship between the
control horizon and the predictive horizon is as shown in
a. Trajectory tracking error (TTE) Equation 7.
 Np 
N c  round   (7)
 n 
where Np is the predictive horizon, Nc is the control horizon,
round() is the rounding function, n is a coefficient required to
be determined, n  1, 2 N p  .
b. Yaw rate tracking error (YRTE) The simulation speed is 120km/h, and a random delay with
Fig.5 Performance analysis under random network delay a peak value of sampling period is added to the output of the
In conclusion, the impact of random network delay on the controller. The result is shown in Figure 7. The vehicle
system performance is positively correlated with the peak value. performance with different predictive horizon is analyzed
Therefore, in the design of trajectory tracking controller, the according to the DLC scene segments. In Segment 1 and
uncertainties caused by random network delay should be Segment 2, Np=20 achieves the optimal control performance,
considered. At the same time, the controller needs to adjust but there is an overshoot in the tracking process after Segment
parameters adaptively according to the working conditions to 3.
obtain the optimal control performance. In addition, yaw rate fluctuates volatilely with Np=20.
When the prediction time domain increases, the tracking error
IV. ADAPTIVE MODEL PREDICTIVE CONTROL CONSIDERING of the whole process is smaller, and the fluctuation of yaw rate
RANDOM NETWORK DELAY is also mitigated.
It can be seen from the above analysis that the network Compared with conventional optimal control, model
delay and the sampling mechanism will cause inconsistencies predictive control strategy considers the trajectory tracking
between the target angle message received by the lower node errors of both current and future sampling times in predictive
and transmitted from the controller, which is the source of horizon, which contributes to future trajectory prediction with a
system oscillation. However, although partial target angle smooth control process. However, an unexpected high value is
message from the controller does not work properly during also inappropriate for predictive horizon. For example, when
SBW system node computation, the proposed controller can the predictive horizon is selected as Np=30 in Fig.7, the linear
predict control variables for the next sampling time to trajectory of Segment 5 in the ideal case will dominate the
compensate for the target signal discontinuity caused by the prediction process due to large predictive horizon, which leads
network delay. the vehicle to perform "take a short cut" in Segment 2-4, and
In this Section, an adaptive prediction model is proposed induces a larger tracking error.
based on the impact of algorithm parameters under random
network delay on trajectory tracking performance, system
stability and computational efficiency. Then, the formula of
time-delay uncertainties is derived based on the matrix

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Edinburgh. Downloaded on June 14,2020 at 19:13:06 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2995408, IEEE
Transactions on Vehicular Technology

a. Vehicle lateral position in DLC test

Fig.9 The flow of adaptive prediction model
where k1, k2 are weighting factors; s1, s2 are scale factors,
which transform the two objectives to the same order of
magnitude; Objective function fTime is the CPU computational
time required for single rolling optimization, and objective
function fError() is the average trajectory tracking error of the
b. Yaw rate in DLC test whole process. The physical meaning of the constraint is that
Fig.7 Prediction horizon analysis the CPU computational time required for a single rolling
In addition, the increase in predictive horizon will impose optimization should not exceed the sampling time.
more computational burden on the controller. For the trajectory After obtaining the undetermined parameters through off-
tracking model, the CPU computational time required for line optimization, the system prediction equation is expressed
single rolling optimization with different predictive horizon is in equation 10. When the predictive horizon NP changes
shown in Figure 8. CPU computational time is obtained by adaptively according to the working conditions, the coefficient
MATLAB. matrix of the state variable and the control variable in the
When the predictive horizon is greater than 80, the prediction equation also changes adaptively.
computational time required for single rolling optimization will Y  k     k   u  k  (10)
be larger than 50 ms, far greater than the sampling time of
single chip microcomputer, which obviously does not satisfy where Y  k  is the prediction output at time k;   k  is the
the real-time requirements of embedded system. state vector; u  k  is the control increment vector;  and
 are coefficient matrix of state variable and control variable
respectively, and the expression is as follows:
 Cd Bd 0 0   Cd Ad 
 
 Cd Ad Bd Cd Bd   
C A2
 0   d d 
   
 Cd Bd   
  Cd Ad 
Np

Fig.8 CPU time required for one cycle  

Cd Ad N p 1 Bd Cd Ad
N p 2
Bd Cd Ad p c Bd 
N  N 1
Therefore, the selection of prediction horizon and control
horizon need to comprehensively consider the environmental B. UM-AMPC Algorithm
parameters (road curvature), operating conditions (vehicle
speed), and the dynamic network characteristics. Based on the In order to achieve superior trajectory tracking control
above analysis, this paper proposes an adaptive predictive effect, it is necessary to establish an augmented model of
horizon function, as shown in Equation 8. And the algorithm uncertainty in the algorithm design. The generalized state
flow is shown in Figure 9. vector of the augmented model at k-th time is shown in
N p  f N Vxk , Ck , k   k1Vxk  b1  k2Ck  b2  k3 k  b3 (8) equation 11.
  k    X  k  u  k  1 u  k    1
p T
(11)
where Vxk , Ck ,  k are vehicle longitudinal speed, track k

curvature and network delay obtained from time stamp at k-th where u(k-i) is the value of the control variable at the i-
sampling time respectively. k i and bi are undetermined th sampling time before time k; k    1 is the upper
coefficients of off-line optimization function, , i=1,2,3. bound of delay of vehicle CAN network at k-th time.
The off-line parameter optimization model is shown in Compared with Formula 3, the control variable of k-i time
equation 9, to change the predictive and control horizon values until the upper bound of delay k    1 is taken as the state
by adjusting the undetermined coefficients in multiple variable in Formula 11, and the delay length of each message
iterations, so as to obtain a set of coefficients which can which is sent from the controller node to the SBW node can be
simultaneously minimize the computational time and tracking obtained by time stamp. The control variables before k-th time
error under the current working condition. may not have delays, and the delay lengths are random and
 k1 k2 may not be equal, which would bring uncertainties to the
min f fitness  s f Time  N P , N c   s f Error  N P , N c  (9) original system.
 1 2
 s.t. max  f  N , N   T
 Time P c s

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Edinburgh. Downloaded on June 14,2020 at 19:13:06 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2995408, IEEE
Transactions on Vehicular Technology

The state equation of discrete-time system considering the  x k ,0   0 0 0 

 x  
uncertainty of random network delay is established, as shown
  k ,1   x k 0 0 
in equation 12.  x k 2    x k x k , j 2 0  (15)
  k  1  Ad   k   Bd u  k    k U  k  (12)    
k k k k
   
where,  x k ,h 1   x x k 2 x k h 
   k
 Ad 1, k  1, k  , k   Bd 
0 0  , 
k
, min0,Ts  k  At Ts  
x k  vTs
 I 0
 I  Bd k   e Bt d
Ad k 0 0 I 0  Bd   0 
0
where v  [0,1) is the length of time that the control message
   k


 0
   u  k  i  from  k  i  Ts will work when calculating k-th time
0 0 I   0



 u  k  1 
control variables.
 1, k  2, k  , k 
 I ,   The establishment of the random network delay uncertainty
  U k   u  k  2  
k   I      coefficient matrix can be represented by polyhedron as shown
    in equation 15.
   u  k    
 I     0, k  0,0  k   0,1  k   0,h  k  
       
In equation 12,  i , k is the uncertainty coefficient caused by  1, k  1,0  k   ˆ  1,1  k   ˆ   1,h  k    ˆ (16)
random network delay, as shown in equation 13.     0
  1
  h
       
i , k  
k i
e
At Ts  
Bt d   , k   ,0  k    ,2  k    ,h  k 
(13) ˆ 

0
 x k ,0 T 
i  1, 2   
0
 T 
 ˆ 1   x k ,1  B
 k i   k i   k i 1
    Ak
  t

 k i  min max 0, k i  iTs  , Ts 
   
  ˆ  T
 h  x k , h 
where  i is the upper limit of integral, whose physical
meaning is the length of time that the signal u  k  i  sent by where i ,h  k  is the time-varying coefficient related to the
the controller node at k-i time extends to the corresponding network delay K at each vertex of the polyhedron, as shown in
equation 16, which represents all signals acting in one
sampling interval at k-th time due to network delay.  k  i is the
sampling period.
total time length of messages sent by the controller node at k-i Following that, based on the uncertainty modeling and the
time. adaptive prediction time domain function strategy proposed in
The linear expression of delay uncertainty coefficient of the previous section, an adaptive model predictive control
stochastic network is established. The linear expression of algorithm for uncertain systems is established. The essence of
delay uncertainty coefficient of random network is established model predictive control is to achieve the current control action
based on polyhedron. And the integral term in delay by solving the finite-time open-loop optimal control problem at
uncertainty coefficient  i , k of random network is separated each sampling instant. Therefore, the trajectory tracking
and expanded by Taylor series, as shown in equation 14. dynamic optimization index is established based on the
  x   A k X k (14) dynamic model, as shown in Equation 17.
J   k  , U  k  1 , u  k   
   12 0  1
3
 1
h+1

 A k  e At Ts  At At 1
At h 1
 Np

   k  i     k  i  (17)
2
  1! 2! h!  k k
Q
 T
i 1
 x k h 
 X  k   x k x k

2 Nc

 u  k  i  R   2
2

where, h is the expansion order selected according to the i 1

accuracy in Taylor expansion. where Np is the predictive horizon, Nc is the control

All possible situations of random network delay are horizon, Q and R are weight matrix,  is the relaxation
modelled by polyhedron in [6], [26]. In one sampling interval,
factor weight coefficient, and  is the relaxation factor.
the integral interval of the uncertainty coefficient  i , k
When solving the model, the centroid sideslip angle
corresponding to all control signals is located at [0, vTs]. constraint, tire sideslip angle constraint and vehicle attachment
Therefore, the vertices of the polyhedron model are defined, as condition constraint are also required to be considered [27].
shown in equation 15. Among them, the sideslip angle of the center of mass and the
sideslip angle of the tire are hard constraints to ensure the
vehicle stability in the trajectory tracking process. Taking the
lateral acceleration as soft constraint, the relaxation factor  is
increased.
By transforming the objective function into quadratic
standard form, the dynamic optimization model of the

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Edinburgh. Downloaded on June 14,2020 at 19:13:06 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2995408, IEEE
Transactions on Vehicular Technology

trajectory tracking controller of the autonomous vehicle with from the controller. The main vehicle parameters used for HIL
network delay in each control cycle can be derived in equation simulation are shown in Table 2.
18. Because the predictive horizon and the control horizon
would adjust adaptively at each sampling time, the prediction Table 2 Main parameters of HIL simulation and vehicle model
model also changes adaptively, that is, the  in equation 18 Parameter Value Description Unit
changes adaptively, so as to balance the efficiency and stability m 1274 Sprung mass kg
Ixx 606.1 Roll inertia kg*m2
when solving the quadratic programming problem at each Iyy 1523 Pitch inertia kg*m2
sampling time. Izz 1523 Yaw inertia kg*m2
T
min  u  t  ,   H t  u  t  ,    Gt  u  t  ,  
T T T rw 0.05 Pinion radius m
      J 0.01 Steering system inertia kg*m2
B 0.025 Column damping Nm*s/deg
 U min   U   U max  (18)
 y      So as to realize the random delay processing of the target
s.t.  hc min    yhc    yhc max  angle message, the random signal library is established. In
 ysc min     ysc   yhc max    addition to Gaussian random signals, this research also
  T Q  R 0  ， establishes random signals which are uniformly distributed,
Ht    Gt   2eT Q 0 
 0   exponentially distributed and Rayleigh distributed as random
where U is the control increment constraint, yhc is the hard network delay for fault injection. As shown in Figure 11.
output constraint, ysc is the soft output constraint, the soft
constraint range can be adjusted by the relaxation factor  , and
e is the tracking error in the predictive horizon.
V. RESULT AND ANALYSIS
In order to verify the proposed algorithm, the hardware in
the loop (HIL) test platform is established as shown in Figure
10. Firstly, a single pure CAN network is established to ensure
that messages sent from VCU to gateway and messages sent
from gateway to ECU will not be preempted by other high a. Random signal obeying uniform distribution
priority messages. Then, random delay is added to the target
angle message through the gateway established by CANoe [28].
In addition, in order to ensure the real-time performance of
HIL simulation process, the control algorithm model and road
environment model established on MATLAB/Simulink and
CarSim will be run on NI PXI platform.

b. Random signal with exponential distribution

c. Rayleigh distribution signal

Fig.11 Random signal library
Fig.10 Hardware in the loop test platform The undetermined parameters in the adaptive predictive
The DLC centerline tracking problem is still taken as the horizon function need to be obtained by off-line optimization.
test scenario to analyze the trajectory tracking performance and In this paper, particle swarm optimization (PSO) algorithm is
system stability under random network delay. The test vehicle applied for optimization calculation. The parameters of PSO
speed is 120km/h. During the test, any distribution in the algorithm are set as shown in Table 3. The framework design
random signal library is selected to delay the target angle signal and parameter setting of PSO algorithm have an important

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Edinburgh. Downloaded on June 14,2020 at 19:13:06 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2995408, IEEE
Transactions on Vehicular Technology

influence on the solution process. For the research and Compared with the other two algorithms, the fluctuation of
application of PSO algorithm, please refer to other works of the UM-AMPC control algorithm is gentle which also validates the
author [29], [30]. analysis of the reasons for the influence of random network
Table 3 PSO Algorithm Parameter delay in Section 3.
Parameter Value Next, the tracking performance and stability of the tests are
Iteration times / It 1000 analyzed separately.
Population size / N 100
Inertia factor / w 0.9 A. Trajectory tracking performance analysis
Individual learning factor/ c1 1.2
Social learning factor / c2 1.2 Firstly, the trajectory tracking performance is analyzed, and
As shown in Figure 12, the initial value of the normalized the vehicle lateral position and lateral tracking error in HIL
loss function is 1, and it converges to 0 approximately after simulation are shown in Figure 14. Compared with MPC
1000 iterations. algorithm, the average tracking error of UM-AMPC algorithm
is reduced by 54.57%, the peak tracking error is reduced by
20.8%, and there is no overshoot. However, the average
tracking error of AMPC algorithm is 39.06% higher than that
of MPC. From Figure 14a, it can be seen that the AMPC
algorithm makes the vehicle tend to track the straight-line track
of the second half of the DLC by means of "taking a shortcut",
Fig.12 Parameter optimization results of adaptive predictive horizon function
which results in a large tracking error of the vehicle in section
2-4 of the DLC working condition.
The optimal solution obtained from the optimization is
selected for the design of UM-AMPC, as shown in Table 4.
Table 4 Optimization results
Parameter Value Unit
k1 5.32 \
b1 8.21 \
k2 0.05 \
b2 0.45 \ a. Vehicle lateral position
k3 0.23 \
b3 1.12 \
HIL simulation compares MPC, AMPC and UM-AMPC.
Among them, the algorithm parameters of MPC algorithm are
fixed and the delay is not considered in the dynamic model.
Compared with the MPC algorithm, the AMPC algorithm uses
an adaptive model predictive control algorithm, but the random b. Trajectory tracking error (TTE)
network delay is also not considered. UM-AMPC can not only Fig.14 Trajectory tracking performance
adaptively adjust the control parameters according to the B. Vehicle stability analysis
working conditions, but also consider the uncertainty brought
by the random network delay in the prediction model. The Then, the stability of the vehicle is analyzed in terms of
DLC tests are performed by the above three methods yaw rate index and sideslip angle index in the DLC test process.
respectively, and the controller outputs are shown in Figure 13. The measured yaw rate and sideslip angle are shown in Figure
15 and Figure 16. Compared with MPC algorithm, UM-AMPC
algorithm can reduce 69.6804% of the average yaw rate error
and 79.5047% of the average sideslip angle error.

a. Target angle of controller output

a. Yaw rate

b. Target angle error of controller output

Fig. 13 Controller output in test
b. Yaw rate tracking error (YRTE)
Fig. 15 Analysis of yaw rate

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Edinburgh. Downloaded on June 14,2020 at 19:13:06 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2995408, IEEE
Transactions on Vehicular Technology

VI. CONCLUSIONS
In this paper, the trajectory tracking predictive control
method of autonomous vehicle considering random network
delay is studied. In order to improve the tracking accuracy,
robust stability and computational efficiency in the process of
rolling optimization, an adaptive model predictive control
a. Sideslip angle (UM-AMPC) algorithm to address the system uncertainty is
proposed. Firstly, the adaptive predictive horizon model is
established, which can adaptively adjust the predictive and
control horizon values according to the vehicle longitudinal
speed, trajectory curvature and the network delay Then, a
predictive model is established according to the target angle
messages received from the current time to the upper bound of
b. Sideslip angle error network delay, whose function is to alleviate the discontinuity
Fig. 16 Analysis of Sideslip angle of the target angle received by the lower node. Finally, HIL
In addition, it can be noticed that the results of the AMPC simulation proves that the average yaw rate error of UM-
algorithm provides the best performance on the sideslip angle AMPC algorithm is 69.68% lower than that of MPC, and the
index. Compared with MPC algorithm, the error of the mean average tracking error can be reduced by 54.57%.
sideslip angle in AMPC algorithm is 49.3771% lower, which The testing results show that the proposed method can
effectively presents the advantages of AMPC in the stability accurately suppress the adverse effects caused by the random
control of uncertain systems. However, if the time-delay network delay, and predict the target angle signal according to
uncertainties are not considered in the model, the tracking the current working condition. This control algorithm provides
performance of the system will be seriously degraded. protection for the safety of critical automatic driving system,
In addition, to validate the system robustness and stability and can be widely applied to chassis network control and road
of the proposed control algorithm, when the uncertainties exist network control of various ground vehicles.
in vehicle dynamics parameters, the ideal trajectory tracking
performance with road adhesion coefficient μ of 0.4 is REFERENCES
simulated. In the test, the random network delay is injected [1] L. Guo, S. Manglani, Y. Liu, and Y. Jia, “Automatic Sensor
through the CANoe gateway between the PXI upper computer Correction of Autonomous Vehicles by Human-Vehicle Teaching-
and-Learning,” IEEE Trans. Veh. Technol., vol. 67, no. 9, pp. 8085–
and the steering actuator controller. Figure 17 shows the 8099, 2018.
vehicle response on the low adhesion road. It can be seen that [2] S. Kato, E. Takeuchi, Y. Ishiguro, Y. Ninomiya, K. Takeda, and T.
under low road adhesion coefficient condition, the vehicle is Hamada, “An open approach to autonomous vehicles,” IEEE Micro,
understeer. Compared with MPC algorithm, the mean square vol. 35, no. 6, pp. 60–68, 2015.
[3] Y. Xia, F. Pu, S. Li, and Y. Gao, “Lateral Path Tracking Control of
value of trajectory tracking error using UM-AMPC is reduced Autonomous Land Vehicle Based on ADRC and Differential
by 23.72%. It also can be observed from figure 17 that the yaw Flatness,” IEEE Trans. Ind. Electron., vol. 63, no. 5, pp. 3091–3099,
rate with MPC control shows instability at the final stage. It 2016.
indicates that the front wheel cannot provide sufficient lateral [4] S. K. Ider, M. K. Özgören, and V. Ay, “Trajectory tracking control
of robots with flexible links,” Mech. Mach. Theory, vol. 37, no. 11,
force because of its entry-into nonlinear range. The yaw rate pp. 1377–1394, 2002.
saturation controlled by UM-AMPC algorithm is obviously [5] F. Zhang, G. Tan, C. Yu, N. DIng, C. Song, and M. Liu, “Fair
much smaller, and the actual yaw rate can largely track the Transmission Rate Adjustment in Cooperative Vehicle Safety
reference value. It indicates that UM-AMPC algorithm can Systems Based on Multi-Agent Model Predictive Control,” IEEE
Trans. Veh. Technol., vol. 66, no. 7, pp. 6115–6129, 2017.
definitely provide better trajectory tracking performance in [6] Z. Shuai, H. Zhang, J. Wang, J. Li, and M. Ouyang, “Combined
terms of stability and robustness. AFS and DYC control of four-wheel-independent-drive electric
vehicles over CAN Network with time-varying delays,” IEEE Trans.
Veh. Technol., vol. 63, no. 2, pp. 591–602, 2014.
[7] X. Zhu, H. Zhang, D. Cao, and Z. Fang, “Robust control of
integrated motor-transmission powertrain system over controller
area network for automotive applications,” Mech. Syst. Signal
Process., vol. 58, pp. 15–28, 2015.
[8] L. Chen, Z. Han, and Z. Ma, “Research on robust control and
(a) Track tracking performance exponential stabilization for large scale impulsive hybrid network
systems with time-delay,” J. Supercomput., vol. 70, no. 3, pp. 1023–
1035, 2014.
[9] J. Song, Y. Niu, and Y. Zou, “Robust finite-time bounded control
for discrete-time stochastic systems with communication
constraint,” IET Control Theory Appl., vol. 9, no. 13, pp. 2015–
2021, 2015.
[10] H. Yan, H. Shi, H. Zhang, and F. Yang, “Quantized H∞ control for
(b) Yaw rate performance networked systems with communication constraints,” Asian J.
Fig. 17 Vehicle stability analysis with uncertain vehicle dynamics parameters Control, vol. 15, no. 5, pp. 1468–1476, 2013.
(road adhesion coefficient is 0.4)

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Edinburgh. Downloaded on June 14,2020 at 19:13:06 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2995408, IEEE
Transactions on Vehicular Technology

[11] A. Elahi and A. Alfi, “Finite-time H∞ control of uncertain [27] K. Nam, H. Fujimoto, and Y. Hori, “Lateral stability control of in-
networked control systems with randomly varying communication wheel-motor-driven electric vehicles based on sideslip angle
delays,” ISA Trans., vol. 69, pp. 65–88, 2017. estimation using lateral tire force sensors,” IEEE Trans. Veh.
[12] Y. L. Wang and G. H. Yang, “Robust H∞ model reference tracking Technol., vol. 61, no. 5, pp. 1972–1985, 2012.
control for networked control systems with communication [28] G. I. Networks, G. Xie, G. Zeng, R. Kurachi, and H. Takada,
constraints,” Int. J. Control. Autom. Syst., vol. 7, no. 6, pp. 992– “WCRT Analysis of CAN Messages in Gateway-Integrated In-
1000, 2009. Vehicle Networks,” IEEE Trans. Veh. Technol., vol. 66, no. 11, pp.
[13] D. Wu, X. M. Sun, C. Wen, and W. Wang, “Redesigned Predictive 9623–9637, 2017.
Event-Triggered Controller for Networked Control System with [29] W. Zhao, Z. Luan, and C. Wang, “Parameter optimization design of
Delays,” IEEE Trans. Cybern., vol. 46, no. 10, pp. 2195–2206, 2016. vehicle E-HHPS system based on an improved MOPSO algorithm,”
[14] P. Ojaghi and M. Rahmani, “LMI-Based Robust Predictive Load Adv. Eng. Softw., vol. 123, no. March, pp. 51–61, 2018.
Frequency Control for Power Systems with Communication [30] W. Zhao, Z. Luan, and C. Wang, “Parametric optimization of novel
Delays,” IEEE Trans. Power Syst., vol. 32, no. 5, pp. 4091–4100, electric–hydraulic hybrid steering system based on a shuffled
2017. particle swarm optimization algorithm,” J. Clean. Prod., vol. 186,
[15] Z. H. Pang, G. P. Liu, and D. Zhou, “Design and Performance pp. 865–876, 2018.
Analysis of Incremental Networked Predictive Control Systems,”
IEEE Trans. Cybern., vol. 46, no. 6, pp. 1400–1410, 2016.
[16] T. Zhang and G. Liu, “Predictive tracking control of network-based Zhongkai Luan received the B.S. and M.S. degree
agents with communication delays,” IEEE/CAA J. Autom. Sin., vol. in vehicle engineering from Nanjing University of
5, no. 6, pp. 1150–1156, 2018. Aeronautics and Astronautics, Nanjing, China, in
[17] Y. M. Liu and K. I-Fong, “Robust predictive tracking control of 2016 and 2019, respectively. He is currently a Ph.D.
networked control systems with time-varying delays and data candidate with the Department of Vehicle
dropouts,” IET Control Theory Appl., vol. 7, no. 5, pp. 738–748, Engineering, Nanjing University of Aeronautics and
2013. Astronautics, Nanjing, China. His research interests
[18] M. Li and Y. Chen, “Robust Tracking Control of Networked include vehicle system dynamics and control
Control Systems with Communication Constraints and External systems.
Disturbance,” IEEE Trans. Ind. Electron., vol. 64, no. 5, pp. 4037–
4047, 2017. Jinning Zhang received the B.S. in vehicle
[19] J. Ji, A. Khajepour, W. W. Melek, and Y. Huang, “Path planning engineering from Nanjing University of Aeronautics
and tracking for vehicle collision avoidance based on model and Astronautics, Nanjing, China, in 2018. She is
predictive control with multiconstraints,” IEEE Trans. Veh. currently a master student of Thermal Power
Technol., vol. 66, no. 2, pp. 952–964, 2017. (Aerospace Propulsion) with Cranfield University,
[20] M. N. Zeilinger, M. Morari, and C. N. Jones, “Soft constrained Cranfield, U.K. Her research interests include
model predictive control with robust stability guarantees,” IEEE vehicle system dynamics and control system.
Trans. Automat. Contr., vol. 59, no. 5, pp. 1190–1202, 2014. Wanzhong Zhao received the B.S. and M.S.
[21] G. V. Raffo, G. K. Gomes, J. E. Normey-Rico, C. R. Kelber, and L. degrees in vehicle engineering from Jiangsu
B. Becker, “A predictive controller for autonomous vehicle path University, Zhenjiang, China, in 2004 and 2005,
tracking,” IEEE Trans. Intell. Transp. Syst., vol. 10, no. 1, pp. 92– respectively, and Ph.D. degree in vehicle
102, 2009. engineering from Beijing Institute of Technology,
[22] Xianzhong Chen, Mohsen Heidarinejad, Jinfeng Liu, and Panagiotis Beijing, China, in 2009. He is currently a professor
D. Christofides, “Composite Fast-Slow MPC Design for Nonlinear and the director in the Department of Vehicle
Singularly Perturbed Systems,” Process Syst. Eng., vol. 58, no. 6, Engineering, Nanjing University of Aeronautics and
pp. 1802–1811, 2012. Astronautics, Nanjing, China. His research interests
[23] E. Choi, S. Han, J. Lee, S. Lee, S. Kang, and J. W. Choi, include vehicle system dynamics.
“Compatibility analysis of the turbo controller area network
(TURBO CAN),” IEEE Trans. Veh. Technol., vol. 67, no. 6, pp. Chunyan Wang received the B.S. and Ph.D.
5146–5157, 2018. degrees in mechanical engineering from Jilin
[24] S. Ding, L. Liu, and W. X. Zheng, “Sliding Mode Direct Yaw- University, Changchun, China, in 2000 and 2008,
Moment Control Design for In-Wheel Electric Vehicles,” IEEE respectively. She is currently an Associate Professor
Trans. Ind. Electron., vol. 64, no. 8, pp. 6752–6762, 2017. with the Department of Vehicle Engineering,
[25] Y. Zheng, S. Li, and X. Wang, “Horizon-varying model predictive Nanjing University of Aeronautics and
control for accelerated and controlled cooling process,” IEEE Trans. Astronautics, Nanjing, China. Her research interests
Ind. Electron., vol. 58, no. 1, pp. 329–336, 2011. include vehicle system dynamics.
[26] X. Zhu, H. Zhang, and Z. Fang, “Speed synchronization control for
integrated automotive motor-transmission powertrain system with
random delays,” Mech. Syst. Signal Process., vol. 64–65, pp. 46–57,
2015.

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Edinburgh. Downloaded on June 14,2020 at 19:13:06 UTC from IEEE Xplore. Restrictions apply.