This article has been accepted for publication in IEEE Access.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI

Nonlinear Tire Model Approximation

Using Machine Learning for Efficient
Model Predictive Control
LUCAS CASTRO SOUSA1 , HELON VICENTE HULTMANN AYALA1
1
Department of Mechanical Engineering, Pontifical Catholic University of Rio de Janeiro, Rio de Janeiro 22451-900, Brazil
Corresponding author: Lucas Castro Sousa ([email protected])
This work was supported by the National Council for Scientific and Technological Development of Brazil under Grant 141360/2019-4

ABSTRACT Model Predictive Controller (MPC) is widely used as a technique for path tracking control
since it allows for dealing with system constraints and future forecasts. However, the performance of MPC
is directly affected by the adopted model. A complex dynamic model can guarantee accuracy in path
tracking but may not be suitable in computational terms. On the other hand, a simplified model may not
capture essential nonlinear aspects. Thus, to cope with these problems, this paper deals with data-driven
tire modeling to improve autonomous ground vehicle path tracking control. The main contribution of the
present work is to show that neural tires can capture the nonlinearities present in the interaction between
lateral and longitudinal vehicle dynamics, with a reduced computational cost for predictive controllers.
Simulated and experimental tire data are approximate to design data-driven tire models using radial basis
function and multilayer perceptron neural networks. Then, based on ground vehicles with neural tires, model
predictive controllers are designed to regulate wheel torque and steering angle inputs. Comparative tests
were conducted to compare the proposed data-driven MPC approach with the classical nonlinear MPC
controller. The results show that the neural tires approximate nonlinear tire models and experimental data
with arbitrary precision in terms of accuracy and error-based metrics. The proposed methodology was
successfully applied to perform trajectory and velocity tracking of ground vehicles. In addition, MPC
with a neural tire model as prediction inference reduces the computational effort compared to traditional
approaches.

INDEX TERMS Data-driven models, trajectory tracking, model predictive control, neural networks.

I. INTRODUCTION for urban routines where uncertainties are common. Thus,

UTONOMOUS ground vehicles (AGVs) are the high- dynamic-based controllers rise as a potential solution for
A est achievement for computational intelligence applied
to vehicles since they are considered a key to reducing traffic
accurate path tracking control. Common approaches include
Sliding Model Control [6], [7], H∞ Control [8], [9], Game-
accidents and improving existing transportation systems [1]. based Control [10], [11], and Model-based Controllers [12]–
However, due to different non-trivial issues, including tra- [14].
jectory tracking control, driving fully autonomous vehicles Model-Based Predictive Control (MPC) has become an
remains challenging [2]. In addition, path tracking is arduous attractive control method since it allows for handling system
for ground vehicles since these vehicles are subject to high constraints and future forecasts. This method applies optimal
slip conditions due to curvy trajectories, obstacles, and envi- control actions to the system within a limited horizon using
ronmental conditions, which may provide trajectory error. optimization mechanisms [15]. Besides, the advantages of
Many works have dealt with the problem of AGV trajec- MPC control include its ability to handle different variables
tory tracking control [3], [4]. Geometric-based controllers in- and constraints. In addition, the imposition of constraints
clude Stanley and Pure Pursuit controllers and are useful for on the vehicle states can ensure comfort and safety for the
applications at low speed, becoming popular due to reduced passengers [16]. However, a setback for the MPC approach
computational cost [5]. However, these controllers usually do is its dependence on an accurate vehicle model. Although a
not consider dynamic aspects and therefore are unsuitable complex and reliable dynamic vehicle model can guarantee

VOLUME 4, 2016 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

satisfactory path tracking in terms of accuracy, the computa- laps, while a second algorithm alters the vehicle trajectory.
tional effort required by the MPC solver may not be suitable The latter reference proposed a neural network architecture
for real-time implementation. using past states and inputs from the physical model. The
Linear MPC is generally designed by applying simplified proposed method results achieved satisfactory performance
vehicle dynamics through linearization processes at specific on an experimental ground vehicle.
operating points [17]–[19]. However, this approach may lose
essential nonlinear information for high slip situations. Thus, A. CONTRIBUTIONS
Nonlinear MPC designed through a nonlinear vehicle dynam- To the best of our knowledge, although researchers have
ics model tends to enhance path tracking, capturing some proposed data-driven models to approximate vehicle dynam-
nonlinear properties [14], [15], [20]. In particular, friction ics, such approaches have not been applied so far for tire
coefficient aspects in longitudinal and lateral directions are modeling for efficient predictive control. In addition, vehicle
essential for good performance of steering commands during dynamics simulations are time-consuming and challenging
path tracking, traction, and braking control applications [21], for real-time simulations. In summary, the points, accuracy
since the loss of adhesion leads to vehicle instability. There- of data-driven models in tire curves prediction, application
fore, several studies are available regarding tire modeling of MPC with neural tire to provide control laws, and compu-
approximation resulting in traditional models such as the tational burden of the predictive control need to be addressed.
Tmeasy tire model [22], Pacejka’s tire model (Magic For- The present paper contributes with:
mula), and Brush model [23]. However, it is a complex task to • Artificial neural networks can be applied to approximate
predict the parameter ranges without prior information about nonlinear tire models with arbitrary precision;
the system, which results in a time-consuming computational Radial basis function (RBF) and multilayer perceptron
procedure [24]. (MLP) neural networks are designed to approximate
Alternatively, data-driven modeling can provide a a suit- Pacejka’s tire model and experimental data. High com-
able trade-off between uncertainties throughout the system putational efforts are commonly required to accurately
and accuracy in representing measured data [5], [24]. More- predict tire curves from traditional tire models. More-
over, no prior information about the system is required result- over, vehicle control depends on the tire-road interac-
ing in an essential advantage for model approximation [25]. tion, demanding an accurate tire model. These issues
In [26], longitudinal model identification and velocity control can be relieved by using data-driven models. Therefore,
of an AGV vehicle are designed for applications at low speed. in this paper, data-driven neural tires are built using
An adaptive ARX model as a function of the operating point machine learning from simulated/measured data result-
is used to identify throttle level as input signal and vehicle ing in models with optimized architectures. It is worth
velocity as output. The velocity control is designed using noting that neural approaches to predict tire curves
a PI controller. In [24] and [27], data-driven techniques are have been used before [32]–[34], but that techniques
applied to derive longitudinal and the combination of lateral only consider lateral curves or parameter estimation of
and longitudinal dynamics of a vehicle, respectively. In both traditional tire modeling, unlike the approach adopted in
works, linear system identification is compared to nonlinear this paper;
physical modeling resulting in satisfactory results for normal • Neural tire models can be used effectively with MPC to
driving conditions. Alternatively, artificial neural networks provide nonlinear control laws;
(ANNs) rise as a potential solution since they have learning Predictive controllers based on AGVs with data-driven
characteristics and adaptation to different complex problems tire models (MPC-Neural) are designed to regulate vir-
with precision [21]. In [5], an MPC controller is designed tual plants’ torque and steering angle inputs consider-
with learned vehicle dynamics employing experimental data. ing simulated and experimental tire data. The present
The metric results demonstrated that the proposed technique method considers a reduced number of data in that it
could successfully represent vehicle behavior and be suitable does not depend upon numerous datasets present in
for real-time operation. Reference [28] proposed a data- system identification and data-driven control, which is a
driven model based on deep neural networks to represent standard procedure used in the literature [5], [24], [27].
the longitudinal characteristics of a ground vehicle. The • MPC with neural tire model as prediction inference
proposed approach predicts the distance and velocity of is computationally more efficient than traditional ap-
the vehicle in real time with accuracy. In [29], the authors proaches.
proposed a data-driven identification using neural networks The predictive control results show that it is possible
to learn vehicle operation data to implement MPC control to improve computational time by 25%, in some cases,
of a racing car. Reference [30] developed a data-driven which indicates that using a data-driven model moti-
identification of an AGV based on an LPV framework using vates the application in real-time. The gain in computa-
machine learning techniques. To improve racing performance tional efforts is relevant if compared to recent references
and capture vehicle dynamics, [31] and [21] use data-driven [5], [35] considering neural networks to predict control
methods. The former applies an iterative learning control laws.
to improve lateral and longitudinal tracking over multiple
2 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

between steering wheel angle δsw and steering transmission

ratio isw [37], as follows

δsw
δ= . (6)
isw
The vehicle motion over the global frame can be derived
from the kinematic model as

Ẋ = ẋcosΨ − ẏsinΨ , (7)

Ẏ = ẋsinΨ + ẏcosΨ , (8)
Ψ̇ = ψ̇ . (9)
FIGURE 1. Single-track model used as the virtual plant.

An accurate model representing tire behavior is essential

for vehicle simulation and control. In this work, the Pacejka
tire model [23] is applied to 1-generate simulated tire data for
B. PAPER ORGANIZATION
ANN tire models; 2-approximate experimental tire data to
The remainder of this paper is organized as follows. We ex-
compare with the proposed ANN tire models. In particular,
pose concepts regarding the vehicle dynamics model, Pacejka
the dataset is composed of longitudinal and lateral friction
Formula, Neural networks, and MPC methodology in Section
curves at the contact patch. The general equation of the
II. In Section III, the proposed approach is detailed. Section
Pacejka model is as follows:
IV and Section V expose the results of the proposed method-
ology. Section VI presents an overall discussion about the F = Dsin(Catan(Bϕ)) + Sv , (10)
results. Finally, conclusions and suggested future research
directions are highlighted in Section VII.
ϕ = (1 − E)(λ + Sh ) + (E/B)arctan(B(λ + Sh )), (11)
II. METHODS where λ parameter is replaced by longitudinal slip (s) or
A. VEHICLE DYNAMIC MODEL AND PACEJKA TIRE lateral slip angle (α) for longitudinal and lateral efforts on
MODEL the tire (F ), respectively. Besides, lateral slip angles for the
The single-track model (Fig. 1) is a vehicle dynamic model front (αf ) and rear (αr ) tires are defined from Eq. (12) and
widely used for control applications [36]. The vehicle is Eq. (13) as
assumed to travel on a rigid path with suspension and
aerodynamic aspects neglected. The vehicle motion can be ẏ + la ψ̇
described in the vehicle (x and y) and global frames (X and αf = δ − , (12)
ẋ
Y ). Moreover, ψ and Ψ are the heading angle in the vehicle
and global frames, respectively. Based on Newton’s theorem,
ẏ − lb ψ̇
the governing equations of motion can be expressed as αr = − . (13)
ẋ
mẍ = mẏ ψ̇ + 2(Fxf cosδ − Fyf sinδ) + 2Fxr , (1) On the other hand, the longitudinal slip can be derived as
mÿ = −mẋψ̇ + 2(Fxf sinδ + Fyf cosδ) + 2Fyr , (2)
ẋ − ωrd
I ψ̈ = 2la (Fxf sinδ + Fyf cosδ) − 2lb Fyr , (3) s= . (14)
max[ẋ, ωrd ]
Iωf ω̇f = −2Fxf rd + Tf , (4)
Iωr ω̇r = −2Fxr rd + Tr , (5) The four coefficients B, C, D, and E are the stiffness
factor, shape factor, peak value, and the curvature factor,
where I and m are the moment of inertia about the yaw axis respectively. Finally, Sh and Sv coefficients are horizontal
and vehicle mass, respectively; la and lb are the front and rear and vertical shifts [23].
axle distance from the center of gravity. Besides, Fxf , Fxr , The friction coefficients in longitudinal (µx ) and lateral
Fyf , and Fyr are the longitudinal and lateral forces applied (µy ) directions represent the relation between tire efforts and
to the front and rear axles, respectively. The parameters Iωf the vertical efforts on the tire as [38]
and Iωr are the mass moment of inertia of the wheels, rd
is wheel radius, Tf and Tr are the torque applied to the Fx
driven wheels when considering a front- and rear-wheel drive µx (s) := , (15)
Fz
vehicle, respectively. The wheel rotational speeds for front
and rear wheels are given by ωf and ωr , respectively. Finally, Fy
δ is the front wheel angle which can be defined as the relation µy (α) := . (16)
Fz
VOLUME 4, 2016 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

B. ARTIFICIAL NEURAL NETWORK (ANN) predictive controller is the prediction model. In this work, the
Artificial neural networks are a representation inspired in prediction model is derived from the vehicle dynamic model
the brain to perform tasks. ANNs are composed of neurons (Eq. 1-9) in continuous time t, and it is abbreviated as
that form a complex architecture that considers inputs to
generate approximations as a mapping [39]. From data, the ẋ = f (x, u) , (20)
ANN architectures can adapt themselves by adjusting their where we have the states x = [ẋ, ẏ, ψ̇, ωf , ωr , X, Y, Ψ], for
connections and coefficients [40]. In this paper, we consider inputs u = [δ, Tf /Tr ]. Besides, as output we have the states
radial basis functions and multilayer perceptron networks. y = [ẋ, Y ]. Considering a reduced sampling interval T, the
The RBF network is a network that uses radial basis vehicle dynamic model can be discretized as follows
functions as an activation function. RBF networks are built
considering the input, hidden, and output layers. In the input x(i + 1 | t) = x(i | t) + T f (x(i | t), u(i | t)) . (21)
layer, input data are connected to source nodes. In contrast, in
the hidden layer, different neurons of the activation functions Then, the trajectory error can be determined:
have their output data weighted and then summed towards
e(i | t) = y(i | t) − yref (i | t)
the output layer. The output of the neuron model is expressed
as: ..
.
M
X e(N | t) = y(N | t) − yref (N | t) , (22)
ŷ(t) = ωi ϕ(r(t), ci , σi ) , (17)
i=1
where yref and y are the output reference and predicted,
respectively, and N is the length of the horizon.
where ŷ(t) is the neuron output, M is the quantity of neurons For trajectory tracking of AGVs, the main optimization
inside the hidden layer, ωi is the output weights, r(t) is the task is to reduce the trajectory error, and therefore the first
input vector, σi and ci are the width and the center of the i-th part of the optimization function is:
hidden node, respectively.
In this paper, we have used the multi-quadratic function as N
X
the activation function. The multi-quadratic equation can be J1 = ∥ e(i | t) ∥2Q , (23)
expressed as i=1
q where Q is a 2 x 2 weighting matrix for outputs. Trajectory
ϕi (l) = l2 + σi2 , (18) tracking control also needs to guarantee that the vehicle
where l represent the norm between the ANN input r to a handles smoothly to ensure passenger comfort. Then, the
given center c, i.e. l = ∥r − c∥. second term of the optimization function is as follows:
MLP architectures were proposed to solve nonlinearly sep- N
X
arable problems and can include one or more hidden layers. J2 = ∥ ∆ u(i | t) ∥2S , (24)
Mathematically, MLP networks are complex and became i=1
viable when researchers started to use the backpropagation where
algorithm [40]. The output of the MLP networks can be
defined as ∆u(i | t) = u(i | t) − (u(i − 1 | t)) . (25)
Z
!
X Moreover, S is a 2 x 2 weighting matrix for control inputs.
ŷj (t) = f ωij xij + ωj , (19)
i=1
Finally, we have the completed optimization function:

where ŷj (t) is the MLP output, Z is the number of hidden J = J1 + J2 , (26)
layers, ωij are the weights between the i-th neuron, in the
prior layer, and the j-th neuron in the actual layer, ωj is the Therefore, the MPC problem can be determined by the opti-
bias weight, i is the number of neurons connected to the j-th mization problem:
neuron. Moreover, xij is the input data from the i-th neuron
to the j-th neuron. PN PN
min i=1 ∥ e(i | t) ∥2Q + i=1 ∥ ∆ u(i | t) ∥2S ,
In this paper, we have proposed MLP networks with nodes s.t. ∆u ∈ (∆umin , ∆umax ) ,
connected by a feed-forward approach, activated by the sig- x ∈ (xmin , xmax ) .
moid function that allows training through the backpropaga- (27)
tion algorithm [41]. Here, the weights Q and S provide an essential effect on
the controller performance. The matrix Q is responsible for
C. MODEL PREDICTIVE CONTROLLER weighing the importance of lateral positions and velocity to
Model predictive control (MPC) is a control procedure that is be followed by the vehicle. In addition, the weight S ponders
used to predict the future states of a dynamic system during the importance of the control actions to smooth both vehicle
a finite-time window (horizon) [42]. The basic part of the trajectory and steering control actions.
4 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

III. PROPOSED APPROACH

Now we shall focus on stating the proposed approach based
on data-driven models for tire modeling and MPC with ANN
tire model to perform path tracking of ground vehicles.

A. DATA-DRIVEN TIRE MODELING AND OPTIMIZATION

PROBLEM
The accuracy of the data-driven approach is relevant since the
experimental data may be non-viable to measure. Moreover,
experimental tests are costly to prepare and may demand
work days. On the other hand, simulated tests do not require
significant computational time and are adaptable for different
case studies. This paper requires that simulated and exper-
imental data present the relation between friction (output)
and slip (input) in longitudinal and lateral directions. Then, FIGURE 2. Neural networks architectures are designed to approximate
we decompose the data-driven approximation problem into friction coefficients curves using slip (input) and friction (output). Each
architecture has a vector H representing model parameters to be determined
parameter estimation to obtain data-driven models. Being so by optimization problems. In particular, for MLP (weights and biases) and RBF
after setting the number of neurons and hidden layers (when (centers, widths, and weights) networks.
required), each architecture is designed to have a vector H
representing MLP (weights and biases) and RBF (centers,
widths, and weights) model parameters. These parameters virtual plant with experimental tire data (Section V). In the
are determined by the optimization algorithm presented in latter, look-up tables represent the experimental tire data.
Eq. 28: These proposed control methodology aims to improve
computational efficiency and verify the application of neural
1 PNl tire models with MPC to derive nonlinear control laws.
min Jp = ξn ξn , (28) In this paper, the optimization problem (Eq. 27) is solved
Nl n=1
using “Ipopt" software package [43] with the open-source
where ξ(t) = ŷref (t) − ŷ(t) i.e. the prediction error. More- tool “CasAdi” [44]. On the other hand, the ODE system is
over, Nl indicates the amount of data used. Besides, one solved with CVodes from the SUNDIALS Suite [45], which
has to specify the limits of search of the network parameter provides an accurate and fast solution for nonlinear systems.
according to the problem. Moreover, the weights matrices are defined by trial and error.
Experimental data can also be used to determine the pa- The proposed approach is summarized in Fig. 3.
rameters present in the Pacejka equation. The procedure is In this paper, the vehicle parameters are assumed to be
based on the same optimization problem presented in Eq. 28, known. However, the parameters needed to represent the
but with the view to obtain coefficients of the well-known vehicle for the proposed MPC can be measured or estimated
equation. with accuracy based on tests with simulated and real data.
The optimization problems for RBF and MLP models, Recent works ( [24], [27], [46]) have explored this topic by
as well as the Pacejka curve with experimental data, are employing identification approaches to solve this issue in as-
solved using the “Ipopt” software package [43] with the pects of vehicle suspension systems and lateral-longitudinal
open-source software “CasAdi”. CasADi is a software that dynamics. Finally, neural tire model creation and path track-
uses a symbolic framework to enhance numeric optimization ing predictive control are implemented in MATLAB under a
[44]. laptop with an Intel i5-7300HQ CPU and 16 GB RAM.
Firstly, data-driven models are designed based on the
number of neurons and the dataset through an optimization IV. TRAJECTORY TRACKING RESULTS WITH
problem (Fig. 2). Then, for path tracking, at each time step, SIMULATED TIRE DATA
longitudinal and lateral slippage (input) are calculated in the This section presents the data-driven modeling results from
dynamic model (Eq. 1-9) and sent to the ANN tire models to simulated tire data acquired from the Pacejka tire model.
generate the value of friction coefficient (output). Moreover, we designed an MPC controller based on a vehicle
with neural tires to control a virtual plant with Pacejka’s tires.
B. MPC WITH ANN AS TIRE MODEL PREDICTION The results are presented in Subsections IV-A and IV-B.
Predictive controllers are herein designed to perform path and
velocity tracking. Dynamic vehicle models with data-driven A. NEURAL TIRE MODEL CREATION
tire modeling are used for MPC prediction of optimized In this section, neural tires are compared with the Pacejka
control actions. The optimized control actions, steering, and friction curves. Simulated data from the Pacejka formula
torque on the driven wheels are firstly simulated on a virtual were generated containing 200 data points considering a
plant with simulated tire data (Section IV) and then on a vehicle (with weight distribution 50/50) and tire parameters
VOLUME 4, 2016 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

FIGURE 3. General overview of the proposed approach: Once the neural network architectures are obtained, the neural tire model is linked to the dynamic vehicle
model to derive optimized control actions in the MPC controller. Finally, the control actions are sent to the virtual plant with simulated/experimental tires.

listed in Table 1 and Table 2, respectively. Besides, for RBF metrics are considered in this paper, Root Mean Squared
and MLP neural networks, the limits of search of the network Error - RMSE and the Multiple Correlation Coefficient (R2 ),
coefficients are summarized in Table 3. to compare the numerical results.
Table 4 depicts the metric results for longitudinal and
TABLE 1. Vehicle Parameters
lateral friction coefficients. The number of hidden layers
Symbol Definition Value inside the neural network is set from 1 ({X}) to 3 ({X X X})
m Vehicle mass 2500 kg
I Inertia moment about yaw axis 2200 kg.m2
and the number of neurons inside these layers is set from 1
Iωf /ωr Inertial moment of the wheel 2.5 kg.m2 to 5 (X=1,2,...5). These quantities of the number of neurons
lt Wheelbase 2.7 m and layers were selected aiming a reduced computational cost
rd Radius of the tire 0.42 m which leads to a less complex architecture to be implemented
isw Steering transmission ratio 30
as a data-driven tire model.

TABLE 4. Results in terms of R2 and RMSE metrics, and solver average

TABLE 2. Coefficients from Pacejka Formula computational time for all the architectures tested considering simulated tire
data
µx µy
Model µx µy Average
Symbol Value Symbol Value
{Architecture} R2 RMSE R2 RMSE time (s)
B 0.208 B 0.154
C 1.650 C 1.300 RBF {1} 0.2248 0.7055 0.2546 0.6922 0.0190
D 6213.4 D 5367.9 RBF {2} 0.9741 0.1289 0.9977 0.0382 0.1780
E 0.604 E -1.464 RBF {3} 0.9978 0.0374 1.000 0.0041 0.7405
Sh 0.0 Sh 0.0 RBF {4} 0.9998 0.0117 1.000 0.0022 0.6585
Sv 0.0 Sv 0.0 RBF {5} 0.9784 0.1179 0.9954 0.0687 0.8540
MLP {1 1} 0.9587 0.1628 0.9921 0.0711 0.0250
MLP {2 2} 0.9997 0.0155 1.000 0.0043 0.2530
MLP {3 3} 1.00 0.0011 1.000 0.0024 0.3305
MLP {1 1 1} 0.9832 0.1040 0.9988 0.0276 0.1655
TABLE 3. Limits of search spaces of the neural networks coefficients MLP {2 2 2} 1.000 4.10e-4 1.000 0.0028 0.2865
Network Symbol Definition Lower limit Upper limit MLP {3 3 3} 1.000 3.88e-4 1.000 0.0029 0.6260
c center -1 1
RBF d delta 0.01 500
W weights -80 80 The results achieved by the neural networks were satisfac-
b bias parameters -4 4 tory, specifically, above one neuron, with R2 metric ranging
MLP
W weights -4 4
from 0.9587 to 1.00 and 0.9921 to 1.00 for longitudinal
and lateral friction coefficients, respectively. On the other
We simulated different RBF and MLP networks, modify- hand, the RMSE metric ranges from 0.1628 to 3.88e-4 and
ing the number of neurons for both approaches and the num- 0.0711 to 0.0022 for longitudinal and lateral friction coef-
ber of hidden layers for the latter. In addition, two error-based ficients. Considering RBFNs, RBF with four neurons (RBF
6 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

trial and error.

1
 
1.5e3 0 0
Fricion coefficient

0.5
 
1e3 0
Longitudinal

Q= 0 2e2 0  ;S = . (29)

0 0 1e3
0 0 2.5e3
Pacejka
-0.5
RBF
MLP
Finally, constraints of -0.9 < δ < 0.9 [rad] and -2500
-1 < T < 2500 [N.m] are applied on the steering angle and
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Longitudinal slip the front-driven wheels representing typical vehicle operation
(a)
zones. Moreover, an initial velocity of 19 m/s is given to the
1 vehicle.
Fricion coefficient

0.5
1) Double lane change
Lateral

0
In the simulation test, the vehicle is motivated to perform
-0.5
Pacejka a double lane change trajectory with a constant referenced
RBF
MLP speed of 20 m/s from an initial velocity of 19 m/s. The
-1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 tracking performance and the computational effort can be
Lateral slip
(b)
seen in Table 5.

TABLE 5. RMS and maximum error under different lengths of horizon and
FIGURE 4. (a) Longitudinal friction approximation using MLP and RBF
control strategies for double lane change simulation
networks. (b) Lateral friction approximation using MLP and RBF networks.
MPC Horizon YRM SE |Ymax | Average solver
approach length (m) (m) time (s)
5 0.1907 0.3686 0.612
{4}) presented the best result for both friction coefficients Pacejka
10 0.1986 0.3834 0.951
with an average computational time (measured considering 15 0.2018 0.3819 1.221
20 0.2006 0.3855 1.523
longitudinal and lateral friction components) of 0.6585 s. For 5 0.1898 0.3680 0.512
MLP, the architecture composed of three hidden layers and Neural
10 0.1929 0.3726 0.840
three neurons in each layer (MLP {2 2 2}) presented the best 15 0.1972 0.3804 1.085
20 0.1982 0.3824 1.342
results. However, this implies a more complex optimization
calculation of network parameters, and therefore, the compu-
tational time increases. Considering both the simulation time The results indicated that both MPC approaches achieved
and the accuracy of the error-based metrics, the MLP network good path tracking performance for all cases. However, the
with two hidden layers with two neurons in each layer (MLP MPC-Neural case presented the smaller lateral offset and the
{2 2}) presented the best result. smaller RMSE metric regarding the lateral position, respec-
Looking at Fig. 4(a) and Fig. 4(b) we can see the longi- tively, 0.3680 and 0.1898 m. It is interesting to note that as
tudinal and lateral friction coefficient curves obtained from the horizon length increases, the lateral error increases. Also,
the best RBF and MLP neural network architectures. In as expected, the simulation time increases with the horizon,
particular, RBF {4} and MLP {2 2} neural networks. From but the MPC-Neural could reduce the computational time
the figures, we can observe that the curves of MLP and RBF compared to the MPC-Pacejka. Overall, the maximum lateral
architectures are close to the reference tire model. error, RMS error, and simulation time go to MPC approaches
with a horizon of 20. From this point of view, a horizon length
B. PATH TRACKING CONTROLLER of 5 has a considered advantage in terms of precision and
This section compares the results of the MPC controller, simulation time.
based on a front-wheel-drive vehicle with neural tires, with In particular, from Fig. 5(a), we can see that both control
those performed by an MPC controller, based on a vehicle methods (MPC-Neural and MPC-Pacejka with Horizon set
with Pacejka’s tire model. Here, the vehicle model with to 5) conducted the vehicle to track the double lane change
neural tires (MLP {2 2}) is used to predict future states of the trajectory satisfactorily. Fig. 5(c) demonstrates that the offset
vehicle over a finite horizon to perform path tracking using error, in the lateral direction, achieved a maximum of 0.370
optimized control actions. The effectiveness of the proposed m for both cases, approximately. From Fig. 5(b), the vehicle
control approach is verified by considering simulated tests on starts with a initial velocity of 19 m/s until reach the reference
a plant with simulated tire data (Table 2) under double lane velocity of 20 m/s. Moreover, the velocity state remains close
change and consecutive lane changes. to the reference through the simulation, as we can see in Fig.
For the predictive controllers, the horizon N is set from 5(d).
5 to 20, and the simulation sampling interval is set to 0.1 From Fig. 6(a) and Fig. 6(b), we can observe that the
seconds. The weighting matrices are shown in Eq. 29. Both MPC controllers provided similar responses, with smooth
weights are set equal for both MPC approaches after some control inputs to the steering wheel and driven wheels, giving
VOLUME 4, 2016 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

Reference MPC-Neural MPC-Pacejka

2) Lane changes
8 20.5
6.48
Here, the vehicle is motivated to perform a sequence of lane
6
6.46
6.44
changes with a constant and referenced speed of 20 m/s from

Velocity (m/s)
20
6.42 an initial velocity of 19 m/s. The tracking performance, as
Y (m)

102
4
well as the computational efforts, can be seen in Table 6.
19.5 20
2 19.98
TABLE 6. RMS and Maximum Error under different length of horizon and
7 8 9
0 19 control strategies for multiple lane changes simulation
0 50 100 150 200 0 5 10
X (m) Time (s)
(a) (b)
MPC Horizon YRM SE |Ymax | Average solver
MPC-Neural MPC-Pacejka approach length (m) (m) time (s)
0.5 1 5 0.1584 0.3686 0.623
0.02
0 0.8 10 0.1645 0.3834 0.982
Pacejka
-0.02 15 0.1663 0.3819 1.251
Offset (m/s)

7.9 8 0.6 0.022

Offset (m)

0.02
20 0.1658 0.1637 1.621
0 0.4 0.018 5 0.1575 0.3680 0.561
7.9 7.95 8
0.2 10 0.1599 0.3726 0.905
Neural
15 0.1632 0.3804 1.132
0
20 0.1637 0.3824 1.496
-0.5 -0.2
0 5 10 0 5 10
Time (s) Time (s)
(c) (d)
MPC results from both approaches achieved good path
FIGURE 5. Results for double lane change simulation considering simulated tracking with maximum error obtained during the double
data. (a) Trajectory. (b) Velocity. (c) Trajectory offset. (d) Velocity offset. lane change maneuver. The RMS error decreased during the
simulation, and a better convergence between reference and
predicted data was performed. However, the MPC-Neural
MPC-Neural MPC-Pacejka case also presented a smaller lateral offset and smaller RMSE
100 2500
metric regarding the lateral position. Besides, the simulation
2000
50 time increased, and when the horizon was set 20, MPC-
Torque (Nm)
angle (deg)

Pacejka performed better than MPC-Neural even with a more

Steering

1500
110
0
1000
100 significant error. Therefore, when the predictive horizon is
7.4 7.5 7.6
-50
500
set to 5, both approaches converge to have precision and
computational effort.
-100 0
0 5 10 0 5 10 Considering the application of the proposed MPC-Neural
Time (s) Time (s)
(a) (b) and the traditional MPC on the controllable plant, we can see
Front - MPC-Neural Front - MPC-Pacejka Rear - MPC-Neural Rear - MPC-Pacejka
1 0.5 in Fig. 7(a) and Fig. 7(b) that both MPC-Neural and MPC-
0.4 Pacejka achieved similar results, with an offset error of 0.370
0.5 10-3
m, approximately. From Fig. 7(c) and Fig. 7(d), we can see
Longitudinal

0.3 4
coefficient
coefficient
Lateral
friction

friction

2
0 0.2 0 that the vehicle velocity achieved the reference velocity with
0.1
8 8.5 9
a minimum error.
-0.5
0 The control inputs for the lane changes path can be seen
-1 -0.1 in Fig. 8(a) and Fig. 8(b). Both controllers achieved sim-
0 5 10 0 5 10
Time (s) Time (s) ilar input curves. Moreover, Fig. 8(c) and Fig. 8(d) show
(c) (d)
the lateral and longitudinal friction coefficients, respectively.
High levels of longitudinal slip occurred on front-driven tires
FIGURE 6. Results for double lane change simulation considering simulated
data. (a) Steering wheel angle input. (b) Torque input. (c) Lateral friction while rear tires remained with low levels of slip. On the other
coefficient. (d) Longitudinal friction coefficient. hand, lateral slips increase under both tires during the ma-
neuvers. Table 6 shows the average solver time considering
lane changes maneuver. MPC-Neural also performed faster
better control and stability on curves. Besides, Fig. 6(c) than the traditional MPC-Pacejka with a computational time
and Fig. 6(d) demonstrate the evolution of the lateral and reduction of 10%.
longitudinal friction coefficients, respectively. We can see
that both controllers achieved low values for friction during V. TRAJECTORY TRACKING RESULTS WITH
the simulation in the straight direction. However, lateral EXPERIMENTAL TIRE DATA
slips increase during the maneuver. From Table 5, we also This section presents the results obtained from the experi-
can see the computational effort for the double lane change mental data-based tire model. Besides, we designed a MPC
maneuver. MPC-Neural performed faster than the traditional controller based on a vehicle with neural tires to control a vir-
MPC-Pacejka with a computational time reduction of 10%- tual plant with experimental tires. The results are presented
17%. in Subsections V-A and V-B.
8 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

Reference MPC-Neural MPC-Pacejka

particular, for longitudinal and lateral friction curves to be
8 20.5 fit, we consider the results for a vertical load of 660 N,
2.93

6
2.925 approximately. The longitudinal and lateral friction curves
2.92

Velocity (m/s)
2.915
247 248
20 contain 322, and 1268 data points, respectively.
Y (m)

4
20.005
For this case, the limits of search of the network coef-
2
19.5 20 ficients are the same presented in Table 3 for longitudinal
19.995

16 18 20
and lateral friction curves. Besides, we also used the same
0
0 100 200 300 400
19
0 5 10 15 20
configuration regarding activation functions for the neural
X (m)
(a)
Time (s)
(b)
networks. For the Pacejka tire model, the following limits of
MPC-Neural MPC-Pacejka search of the coefficients are used:
0.5

0.8 0.01

0
TABLE 7. Limits of search spaces of the Magic Formula coefficients
0.6
Offset (m/s)
Offset (m)

-0.01
0.4 Parameter Definition Lower limit Upper limit
0 19 20
0.2 B stiffness factor -10 10
0 C shape factor -10 10
-0.2
D peak value -10 10
E curvature factor -10 10
-0.5 -0.4
0 5 10 15 20 0 5 10 15 20 Sh horizontal shift -0.2 0.2
Time (s) Time (s) Sv vertical shift -1 1
(c) (d)

FIGURE 7. Results for lane changes simulation considering simulated data.

(a) Trajectory. (b) Velocity. (c) Trajectory offset. (d) Velocity offset. Table 8 indicates the optimized parameters to fit the exper-
imental tire curves through Pacejka tire model.

MPC-Neural MPC-Pacejka TABLE 8. Coefficients from Pacejka Formula

100 2500
µx µy
2000
50 Symbol Value Symbol Value
Torque (Nm)
angle (deg)

B 10.3075 B -10.8138
Steering

1500 120
100
0
80
C 1.9157 C -1.6192
1000 60 D 2.6268 D 2.7166
18.5 19 19.5
-50 E 0.5182 E 0.4118
500
Sh 0.0322 Sh -0.0036
-100 0 Sv -0.2819 Sv 0.0694
0 5 10 15 20 0 5 10 15 20
Time (s) Time (s)
(a) (b)
Front - MPC-Neural Front - MPC-Pacejka Rear - MPC-Neural Rear - MPC-Pacejka
1 0.5 Friction curves from the Neural and Pacejka tire models
0.4 are compared based on error metrics and average compu-
0.5
10-3 tational time. From the metric results presented in Table
Longitudinal

0.3
coefficient
coefficient

20
Lateral
friction

friction

0 0.2 10 9, we can see that the increase of neurons leads to more

0
0.1 19 19.5 accurate results, demanding, however, an increase in compu-
-0.5
0 tational effort. Considering both average computational time
-1 -0.1 (to obtain longitudinal and lateral curves) and accuracy of the
0 5 10 15 20 0 5 10 15 20
Time (s) Time (s) error-based metrics, MLP {1 1} and RBF {2} achieved the
(c) (d)
best results. However, we selected MLP {1 1} due to lower
computational effort.
FIGURE 8. Results for lane changes simulation considering simulated data.
(a) Steering input. (b) Torque input. (c) Lateral friction coefficient.(d) Fig. 9 demonstrates the longitudinal and lateral friction
Longitudinal friction coefficient. coefficient curves obtained from the best RBF and MLP
architectures as well as the Pacejka model. One can see that
the proposed method can capture the nonlinearities presented
A. NEURAL TIRE MODEL CREATION in tire data.
Here, we show that experimental data can be used to derive a
data-based tire model as an approximate model for tire mod- B. PATH TRACKING CONTROLLER
eling. To further validate the proposed Neural tire approach, In this part, the predictive control of a rear-wheel-drive vehi-
experimental data [47] were obtained from tests performed cle with experimental data tire is implemented to corroborate
by the Calspan Tire Research Facility (TIRF) as part of the the effectiveness of the neural tire approach proposed in this
Formula SAE Tire Test Consortium (FSAE TTC). For details paper. Two different trajectories are considered: double lane
regarding the tire tests and procedures, please refer to [48]. change and lane changes. Besides, the prediction is solved
We apply RBF and MLP architectures and a fit by the by the vehicle model with the neural tire, and the simulation
Pacejka model to approximate the experimental data. In is done considering the vehicle using experimental tire data.
VOLUME 4, 2016 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

1) Double lane change

4
The plant with experimental tire data is motivated to per-
form a double lane change maneuver with a constant and
Fricion coefficient

2
Longitudinal

referenced speed of 25 m/s from an initial velocity of 24

0
Real
m/s. The metric results in terms of lateral tracking error and
-2
RBF computational efforts can be seen in Table 11.
MLP
Pacejka Metric results for a horizon of 5 presented the best per-
-4
-0.25 -0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2 formance compared to the horizon set to 10. It is interesting
Longitudinal slip
(a)
to note that MPC-Pacejka presented a maximum offset of
1% minor compared to the MPC-Neural for horizon equals
4
5. However, a discrepancy of almost 20% can be achieved
for a horizon equal to 10. In terms of computational time,
Fricion coefficient

MPC-Neural decreased the effort by approximately 22% for

Lateral

0
Real horizons equal to 5 and 10, respectively.
RBF
-2
MLP From Fig. 10(a), we can see the desired trajectory and the
-4
Pacejka
result performed by the proposed MPC-Neural. We also can
-0.25 -0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25 note from Fig. 10(b) and Fig. 10(c), the followed velocity and
Lateral slip
(b) the lateral offset during the vehicle tracking. In particular, the
maximum lateral offset achieved is 0.52 m.
FIGURE 9. Comparison between experimental data and tire models Fig. 11(a) and Fig. 11(b) illustrate the control inputs to
considering (a) longitudinal and (b) lateral friction curves.
the steering wheels and rear-driven wheels, respectively. The
evolution of the lateral and longitudinal friction coefficients
are shown in Fig. 11(c) and Fig. 11(d), respectively. We
The latter uses look-up tables with slip and friction data as can see that the rear longitudinal friction is high at the
input and output. beginning due to the torque acting on the rear-driven wheels.
For the predictive controllers, the horizon is set to 5 and However, as the vehicle moves, the longitudinal friction tends
10, and the simulation sampling interval is set to 0.1 seconds. to decrease. On the other hand, lateral friction increase only
Also, the weighting matrices are set equal for both MPC when a steering angle is given to the wheels, as expected.
approaches after some trial and error.
  2) Lane changes
2.5e3 0 0   In this scenario, the vehicle is supposed to track a sequence of
1e3 0
Q=  0 2e2 0  ; S = . (30) lane changes maneuver and the velocity of 25 m/s. The metric
0 1e3
0 0 1.5e3 results in terms of lateral tracking error and computational
efforts can be seen in Table 12.
The vehicle (with weight distribution 50/50) parameters are
The metric results show that RMS error and lateral offset
listed in Table 10. In addition, constraints of -0.9 < δ <
are similar for both MPC approaches, approximately 0.65 m.
0.9 [rad] and -1200 < T < 1200 [N.m] are applied on the
The main difference is that using MPC-Neural produced a
steering angle and the rear-driven wheels. Finally, an initial
computational effort reduction of 25% and 22%, considering
velocity of 24 m/s is given to the vehicle.
horizon lengths of 5 and 10, respectively. In Fig. 12 and Fig.

TABLE 10. Vehicle Parameters

TABLE 9. Results in terms of R2 and RMSE metrics, and solver average Symbol Definition Value
computational time for all the architectures tested considering experimental
m Vehicle mass 270 kg
tire data
I Inertia moment about yaw axis 120 kg.m2
Model µx µy Average Iωf /ωr Inertial moment of the wheel 0.3 kg.m2
{Architecture} R2 RMSE R2 RMSE time (s) lt Wheelbase 1.6 m
RBF {1} 0.4403 1.6741 0.2978 1.9688 0.0255 rd Radius of the tire 0.20 m
RBF {2} 0.9953 0.1512 0.9988 0.0829 0.2310
RBF {3} 0.9968 0.1273 0.9992 0.0679 0.9600
RBF {4} 0.9968 0.1261 0.9992 0.0647 1.1300 TABLE 11. RMS and maximum error under different lengths of horizon and
RBF {5} 0.9919 0.2013 0.9992 0.0655 4.2050 control strategies for double lane change considering experimental tire data
MLP {1 1} 0.9957 0.1472 0.9991 0.0713 0.0415
MLP {2 2} 0.9968 0.1259 0.9993 0.0639 0.7920 MPC Horizon YRM SE |Ymax | Average solver
MLP {3 3} 0.9969 0.1238 0.9993 0.0634 1.7550 approach length (m) (m) time (s)
MLP {1 1 1} 0.9957 0.1474 0.9991 0.0695 0.2670 5 0.2803 0.5287 0.7429
Pacejka
MLP {2 2 2} 0.9968 0.1249 0.9993 0.0635 1.9750 10 0.2940 0.5509 1.4619
MLP {3 3 3} 0.9970 0.1246 0.9993 0.0632 3.2150 5 0.2836 0.5289 0.5788
Neural
Pacejka 0.9966 0.1313 0.9992 0.0671 0.2550 10 0.3272 0.6785 1.1583

10 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

Reference MPC-Neural MPC-Pacejka

10 25.5
0.1
0.095
5

Velocity (m/s)
Reference MPC-Neural MPC-Pacejka
0.09 25
492 496 500

Y (m)
10 25.5
8.046 0
8.044 25.02
8 24.5
8.042 -5

Velocity (m/s)
25 25
6 104.6
Y (m)

24 26 28
-10 24
4 25.02 0 200 400 600 0 10 20 30
24.5 25 X (m) Time (s)
2 24.98 (a) (b)
MPC-Neural MPC-Pacejka
5 6 7
0 24 1
0 50 100 150 200 0 2 4 6 8 0.6
X (m) Time (s) 0.8
0.4
(a) (b) 0.01

Offset (m/s)
0.6

Offset (m)
MPC-Neural MPC-Pacejka 0.2
0
1 -0.01
0 0.4
-0.02
0.4 0.8 -0.2
0.02 0.2 17 18 19
-0.4
Offset (m/s)

0.2 0.6
Offset (m)

0.01 0
0 0.4 -0.6
-0.2
6 6.5 7 7.5 0 10 20 30 0 10 20 30
-0.2 0.2 Time (s) Time (s)
(c) (d)
-0.4 0

-0.6 -0.2
0 2 4 6 8 0 2 4 6 8 FIGURE 12. Results for lane changes simulation considering experimental
Time (s) Time (s) tire data. (a) Trajectory. (b) Velocity. (c) Trajectory offset. (d) Velocity offset.
(c) (d)

FIGURE 10. Results for double lane change simulation considering

experimental tire data. (a) Trajectory. (b) Velocity. (c) Trajectory offset. (d)
Velocity offset. 13, we can seen the results for both approaches when the
predictive horizon is set to 5.
Fig. 12(a) and Fig. 12(b) show the trajectory and velocity
tracking performance. The maximum lateral offset, approx-
imately 0.65 m, can be seen in 12(c). Moreover, we can
see that the reference velocity achieved a minimum offset in
12(d).
The control inputs are plotted in Fig. 13(a) and Fig. 13(b).
It is observed that a high torque increases the rear longitu-
MPC-Neural MPC-Pacejka
20 300 dinal friction coefficient. However, as the simulation time
250 elapses, the longitudinal friction tends to decrease, as seen in
10
Fig. 13(d). Moreover, towed front wheels remain with a low
Torque (Nm)
angle (deg)

200
Steering

0 150 level of longitudinal friction coefficient due to the absence

100 of driven torque. On the other hand, from Fig. 13(c), we can
-10
50 see that the lateral friction coefficient curve and the steering
-20 0 angle curve share the same trend, with the values increasing
0 2 4 6 8 0 2 4 6 8
Time (s) Time (s) during the maneuver.
(a) (b)
Front - MPC-Neural Front - MPC-Pacejka Rear - MPC-Neural Rear - MPC-Pacejka
From Fig. 14, we can see the total processing time for
2 1.5 both maneuvers considering the horizon prediction set to
1 1
5. MPC-Neural performed even faster than the traditional
MPC-Pacejka. The computational time decreases due to the
Longitudinal

coefficient
coefficient
Lateral
friction

friction

0 0.5 reduction of the number of neurons in the data-driven tire

-1 0
model, in this case, MLP {1 1}.

-2 -0.5
0 2 4 6 8 0 2 4 6 8
Time (s) Time (s)
(c) (d) TABLE 12. RMS and Maximum Error under different length of horizon and
control strategies for lane changes considering experimental tire data
FIGURE 11. Results for double lane change simulation considering
MPC Horizon YRM SE |Ymax | Average solver
experimental tire data. (a) Steering input. (b) Torque input. (c) Lateral friction
coefficient. (d) Longitudinal friction coefficient.
approach length (m) (m) time (s)
5 0.2383 0.6452 0.7712
Pacejka
10 0.2458 0.6593 1.6700
5 0.2395 0.6473 0.5845
Neural
10 0.2587 0.6785 1.2969

VOLUME 4, 2016 11

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

MPC-Neural MPC-Pacejka
performs better computationally when used in prediction for
30 MPC, which is essential in the scope of embedded solutions
250
20 for control laws.

Torque (Nm)
200
Simulated and experimental data were used to approxi-
angle (deg)

10
Steering

150
0 mate tire curves, precisely longitudinal and lateral friction
100
-10 coefficients. The former data was obtained from the Pacejka
50
-20 model, and the latter was obtained from an experimental tire
0
-30
0 10 20 30 0 10 20 30
test. Data-driven models were derived using artificial neural
Time (s)
(a)
Time (s)
(b)
networks, particularly MLP and RBF networks. In this case,
Front - MPC-Neural Front - MPC-Pacejka Rear - MPC-Neural Rear - MPC-Pacejka Section IV-A shows that the MLP {2 2} provided the best
2 1
fit of data for both friction coefficients with reduced average
1 simulation time (considering both friction coefficients). Once
Longitudinal

0.5
coefficient
coefficient

the best architecture is found to fit the simulated data, a pre-

Lateral
friction

friction

0
dictive controller based on a vehicle with the selected neural
0
-1 tire is designed. A comparison with the exact vehicle using
the Pacejka formula is shown in Section IV-B. The vehicle
-2 -0.5
0 10 20 30 0 10 20 30 models with neural and Pacejka tires are used to predict the
Time (s) Time (s)
(c) (d) future states of the vehicle over the finite time window (MPC
scheme). At the same time, the simulation occurs in a vehicle
FIGURE 13. Results for lane changes simulation considering experimental with simulated tire data. The results show that the predictions
tire data. (a) Steering input. (b) Torque input. (c) Lateral friction coefficient. (d)
Longitudinal friction coefficient. agree with trajectory, velocity, and control inputs (steering
angle and torque). Besides, computational time reduction is
observed using the predictive control with neural tire models.
Thus, the data-driven models proposed herein are better for
60 predictive models in MPC concerning computational use.
Double lane change
50 Further improvements in computational time and trajectory
solver time (s)
Accumulated

40 tracking can be obtained using more complex models than the

30 presented strategy. This approach is essential in embedded
20 real-time applications and is crucial to optimize hardware use
10 MPC-Neural
MPC-Pacejka
for MPC [49].
0
0 1 2 3 4 5 6 7 8
Regarding the results with experimental tire data, in Sec-
Simulation time (s)
(a)
tion V-A, metric results and average simulation time to obtain
friction coefficients indicate that the MLP {1 1} provided
250

Lane changes
the best fit of experimental data, as with the simulated data.
200
Then, an MPC controller based on a vehicle with neural tires
solver time (s)
Accumulated

150 is designed to control a vehicle with experimental tire data.

100 We can see in Section V-B the good agreement between the
50
reference trajectory and velocity throughout the simulation.
MPC-Neural
MPC-Pacejka It shows that the methodology for neural-tire approximation
0
0 5 10 15 20 25 30 for MPC herein presented gives good results even when real-
Simulation time (s)
(b) world noise corrupted data is used. The performance of the
path tracking controllers in Sections IV and IV in terms
FIGURE 14. Total processing time through simulation with experimental tire of path tracking is further verified by reported results in
data for (a) double lane-change and (b) lane changes.
literature regarding tracking control at high speed [11], [50]–
[52].

VI. OVERALL DISCUSSION AND IMPACTS TABLE 13. RMS and maximum error under different control strategies and
the proposed MPC approach
The simulated and experimental results presented in Section
IV and Section V give an insight into data-driven tire models Reference Control YRM SE |Ymax |
applied to predictive control during trajectory tracking tasks. strategy (m) (m)
[50] MPC - 0.40 - 0.50
This approach allows for capturing nonlinear tire characteris- [51] MPC 0.20 - 0.60 0.76 - 1.76
tics combined with a predictive control strategy during differ- [11] Game theory-based - 0.46 - 0.71
ent maneuvers. As we shall discuss next, we advocate that the [52] MPC 0.031 - 0.142 0.40 - 0.47
Proposed MPC 0.15 - 0.32 0.36 - 0.67
method presented herein gives advantages to current practice.
It enables the construction of data-driven models for tires
to arbitrary precision. Besides, the proposed methodology Data-driven tire models give an alternative way to verify
12 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

and analyze nonlinear aspects in tire modeling, which is lack- time.

ing in existing tire models, as the tire models are well-suited Here, Radial basis functions and Multilayer Perceptron ar-
for each application. Moreover, the results presented have chitectures are compared by the different number of neurons,
implications considering data-driven modeling for MPC. The average simulation time, and error-based metrics. Results
procedure grants computational time reduction, as observed indicate that the MLP network performed better during the
in the tests performed. Furthermore, it showed that the pro- experiments than the RBF architecture for tire data predic-
posed approach could provide good trajectory and veloc- tion. In addition, we have investigated the application of data-
ity tracking under real-world scenarios when we employed driven models with predictive control to provide nonlinear
measured data. This approach indicates that, under high- control laws. Thus, MPC controllers based on vehicles with
speed driving situations, data-driven modeling for efficient neural tires are designed to perform path tracking. Results
predictive control can correctly replace physically-inspired indicate that using neural network architectures as prediction
derived models for tires. It is interesting to note, though, that inference provided a computational time reduction of 10% to
the neural model is applied only to focus on the tire-road and 25% without accuracy losses. Results also indicate that
interaction, which is highly uncertain, and many different the error regarding the MPC controllers in terms of lateral
models exist for each specific road condition [53], [54]. Such position is similar to that presented in recent literature with
procedure is in line with most recent works for physically- Ymax varying from 0.36 m to 0.67 m. It is noteworthy that
inspired machine learning modeling [55]. the error is considered satisfactory considering the vehicle’s
Although the stability of the closed-loop system may be longitudinal velocity. A reduced longitudinal velocity can
affected by the approximation error of the neural tire model, reduce the trajectory error.
its behavior presents similarity when traditional approaches We shall implement different cost function architectures
for tire curve approximation are also applied to closed-loop for future work to enhance path-tracking control. In addition,
systems, as seen in [38]. Moreover, one of the shortcomings to further improve the method, we shall focus on reducing
of the proposed approach is that the sudden change in the pa- the computational time before implementing it in a prototype
rameters may decrease the MPC performance. This problem scaled vehicle as a testbed platform with dynamics simi-
could be solved with the application of an observer/online lar to those of a full-sized vehicle. The implementation of
training [5], [56]. However, the method presented herein predictive trajectory control with data-driven vehicle models
builds on other physical models, increasing knowledge about has been successfully implemented in recent works [5], [21],
the predictive modeling approach. In addition, this method [29].
helps to obtain a better overall data-driven model, as the
machine learning models can focus on specific aspects of the
system, that is, the tire-road interaction. REFERENCES
[1] L. Guo, P. Ge, M. Yue, and J. Li, “Trajectory tracking algorithm in a
hierarchical strategy for electric vehicle driven by four independent in-
VII. CONCLUSION
wheel motors,” Journal of the Chinese Institute of Engineers, pp. 1–12,
The trajectory tracking controller is an essential subsystem 2020.
of autonomous vehicles, responsible for maintaining the [2] X. He, Y. Liu, C. Lv, X. Ji, and Y. Liu, “Emergency steering control
vehicle on an established trajectory. Several methodologies of autonomous vehicle for collision avoidance and stabilisation,” Vehicle
System Dynamics, vol. 57, no. 8, pp. 1163–1187, 2019.
have been proposed throughout recent years. Predictive con- [3] N. H. Amer, H. Zamzuri, K. Hudha, and Z. A. Kadir, “Modelling and
troller based on models has been widely used since it allows control strategies in path tracking control for autonomous ground vehicles:
for dealing with constraints and uncertainties regarding the A review of state of the art and challenges,” Journal of Intelligent and
Robotic Systems, pp. 225–254, 2017.
system. However, the effectiveness of the MPC approach
[4] Q. Yao, Y. Tian, Q. Wang, and S. Wang, “Control strategies on path
relies on a reliable vehicle model used for future forecasts. tracking for autonomous vehicle: State of the art and future challenges,”
Simplified dynamic models can ensure accuracy for specific IEEE Access, vol. 8, pp. 161 211–161 222, 2020.
situations with low computational cost; however, the MPC [5] M. Rokonuzzaman, N. Mohajer, S. Nahavandi, and S. Mohamed, “Model
predictive control with learned vehicle dynamics for autonomous vehicle
controller’s performance can suffer due to nonlinear aspects. path tracking,” IEEE Access, vol. 9, pp. 128 233–128 249, 2021.
On the other hand, a nonlinear dynamic model may lead to an [6] R. He, C. Sandu, and J. E. Osorio, “Systematic tests for study of tire
unsuitable real-time application of MPC. Thus, data-driven tractive performance on soft soil: Part i – experimental data collection,”
approaches are a potential solution with a satisfactory trade- Journal of Terramechanics, vol. 85, pp. 59–76, 2019.
[7] A. D. Sabiha, M. A. Kamel, E. Said, and W. M. Hussein, “Ros-based
off between accuracy and computational cost. The present trajectory tracking control for autonomous tracked vehicle using optimized
work dealt with applying data-driven models to approximate backstepping and sliding mode control,” Robotics and Autonomous Sys-
simulated and experimental tire data for efficient predictive tems, vol. 152, p. 104058, 2022.
[8] C. Zhang, J. Hu, J. Qiu, W. Yang, H. Sun, and Q. Chen, “A novel fuzzy
control. The proposed approach has the potential to 1) ap- observer-based steering control approach for path tracking in autonomous
proximate simulated and experimental tire data with arbitrary vehicles,” IEEE Transactions on Fuzzy Systems, vol. 27, no. 2, pp. 278–
precision, reducing the effect of unmodelled aspects, 2) be 290, 2019.
used effectively with MPC to provide nonlinear control laws [9] X. Zhou, Z. Wang, and J. Wang, “Popov-h∞ robust path-tracking control
of autonomous ground vehicles with consideration of sector-bounded
by providing accurate trajectory predictions, 3) be more ef- kinematic nonlinearity,” Journal of Dynamic Systems, Measurement, and
ficient than traditional approaches in terms of computational Control, vol. 143, no. 11, 2021.

VOLUME 4, 2016 13

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

[10] Y.-M. Choi and J.-H. Park, “Game-based lateral and longitudinal coupling [33] J. Wang, Y. Liu, L. Ding, J. Li, H. Gao, Y. Liang, and T. Sun, “Neural
control for autonomous vehicle trajectory tracking,” IEEE Access, vol. 10, network identification of a racing car tire model,” Journal of Engineering,
pp. 31 723–31 731, 2022. vol. 2018, 2018.
[11] Q. An, S. Cheng, C. Li, L. Li, and H. Peng, “Game theory-based control [34] A. M. Ribeiro, A. Moutinho, A. R. Fioravanti, and E. C. de Paiva, “Esti-
strategy for trajectory following of four-wheel independently actuated au- mation of tire–road friction for road vehicles: a time delay neural network
tonomous vehicles,” IEEE Transactions on Vehicular Technology, vol. 70, approach,” Journal of the Brazilian Society of Mechanical Sciences and
no. 3, pp. 2196–2208, 2021. Engineering, vol. 42, no. 1, pp. 1–12, 2020.
[12] M. Rokonuzzaman, N. Mohajer, and S. Nahavandi, “Effective adoption [35] Y. Xiao, X. Zhang, X. Xu, X. Liu, and J. Liu, “Deep neural networks with
of vehicle models for autonomous vehicle path tracking: a switched mpc koopman operators for modeling and control of autonomous vehicles,”
approach,” Vehicle System Dynamics, pp. 1–24, 2022. IEEE Transactions on Intelligent Vehicles, pp. 1–12, 2022.
[13] Z. Zhang, L. Xie, S. Lu, X. Wu, and H. Su, “Vehicle yaw stability control [36] H. Wang, B. Liu, X. Ping, and Q. An, “Path tracking control for au-
with a two-layered learning mpc,” Vehicle System Dynamics, pp. 1–22, tonomous vehicles based on an improved mpc,” IEEE Access, pp. 161 064–
2022. 161 073, 2019.
[14] M. De Bernardis, G. Rini, F. Bottiglione, A. E. Hartavi, and A. Sorniotti, [37] Y. Zhao, W. Pi, W. Zhang, Q. Wang, S. Feng, H. Deng, and F. Lin,
“On nonlinear model predictive direct yaw moment control for trailer sway “A vehicle handling inverse dynamics method for emergency avoidance
mitigation,” Vehicle System Dynamics, pp. 1–27, 2022. path tracking based on adaptive inverse control,” IEEE Transactions on
[15] H. Taghavifar, “Neural network autoregressive with exogenous input as- Vehicular Technology, vol. 70, no. 6, pp. 5470–5482, 2021.
sisted multi-constraint nonlinear predictive control of autonomous vehi- [38] S. Savaresi and M. Tanelli, Active Braking Control Systems Design for
cles,” IEEE Transactions on Vehicular Technology, vol. 68, no. 7, pp. Vehicles, 1st ed. London: Springer, 2010.
6293–6304, 2019. [39] A. J. Maren, C. T. Harston, and R. M. Pap, Handbook of neural computing
[16] N. Mohajer, S. Nahavandi, H. Abdi, and Z. Najdovski, “Enhancing passen- applications. Academic Press, 2014.
ger comfort in autonomous vehicles through vehicle handling analysis and [40] S. S. Haykin, Neural networks and learning machines, 3rd ed. Prentice
optimization,” IEEE Intelligent Transportation Systems Magazine, vol. 13, Hall, 2009.
no. 3, pp. 156–173, 2021. [41] C. M. Bishop, Pattern recognition and machine learning, 1st ed. Springer,
[17] L. Wei, X. Wang, L. Li, Z. Fan, R. Dou, and J. Lin, “T-s fuzzy model 2006.
predictive control for vehicle yaw stability in nonlinear region,” IEEE [42] J. B. Rawlings, D. Q. Mayne, and M. Diehl, Model predictive control :
Transactions on Vehicular Technology, vol. 70, no. 8, pp. 7536–7546, theory, computation, and design, 2nd ed. Nob Hill, 2017.
2021. [43] A. Wächter and L. Biegler, “On the implementation of an interior-point
[18] Y. Wu, L. Wang, J. Zhang, and F. Li, “Path following control of au- filter line-search algorithm for large-scale nonlinear programming,” Math-
tonomous ground vehicle based on nonsingular terminal sliding mode ematical Programming, vol. 106, pp. 25–57, 2006.
and active disturbance rejection control,” IEEE Transactions on Vehicular [44] J. A. Andersson, J. Gillis, G. Horn, J. B. Rawlings, and M. Diehl, “Casadi:
Technology, vol. 68, pp. 6379–6390, 2019. a software framework for nonlinear optimization and optimal control,”
[19] M. Kim, D. Lee, J. Ahn, M. Kim, and J. Park, “Model predictive control Mathematical Programming, vol. 11, pp. 1–36, 2006.
method for autonomous vehicles using time-varying and non-uniformly [45] A. Hindmarsh, P. Brown, K. Grant, S. Lee, R. Serban, D. Shumaker, and
spaced horizon,” IEEE Access, vol. 9, pp. 86 475–86 487, 2021. C. Woodward, “SUNDIALS: Suite of nonlinear and differential/algebraic
[20] K. Berntorp, R. Quirynen, T. Uno, and S. D. Cairano, “Trajectory tracking equation solvers,” ACM Transactions on Mathematical Software, vol. 31,
for autonomous vehicles on varying road surfaces by friction-adaptive no. 3, pp. 363–396, 2005.
nonlinear model predictive control,” Vehicle System Dynamics, vol. 58, [46] G. Kim, S. Y. Lee, J.-S. Oh, and S. Lee, “Deep learning-based estimation
no. 5, pp. 705–725, 2020. of the unknown road profile and state variables for the vehicle suspension
[21] N. A. Spielberg, M. Brown, N. R. Kapania, J. C. Kegelman, and J. C. system,” IEEE Access, vol. 9, pp. 13 878–13 890, 2021.
Gerdes, “Neural network vehicle models for high-performance automated [47] MathWorks Student Competitions Team, “Analyzing tire test data,” Avail-
driving,” Science Robotics, vol. 4, no. 28, 2019. able at https://www.mathworks.com/matlabcentral/fileexchange/67987-
[22] W. Hirschberg, G. Rill, and H. Weinfurter, “Tire model tmeasy,” Vehicle analyzing-tire-test-data (2021/09/24).
System Dynamics, vol. 45, no. sup1, pp. 101–119, 2007. [48] E. Kasprzak and D. Gentz, “The formula sae tire test consortium-tire
[23] H. Pacejka, Tire and Vehicle Dynamics, 3rd ed. Oxford: Butterworth- testing and data handling,” SAE Technical Paper, no. 2006-01-3606, 2006.
Heinemann, 2012. [49] R. Verschueren, G. Frison, D. Kouzoupis, N. van Duijkeren, A. Zanelli,
[24] B. A. H. Vicente, S. S. James, and S. R. Anderson, “Linear system B. Novoselnik, T. Albin, R. Quirynen, and M. Diehl, “acados—a modular
identification versus physical modeling of lateral–longitudinal vehicle open-source framework for fast embedded optimal control,” Math. Prog.
dynamics,” IEEE Transactions on Control Systems Technology, vol. 29, Comp., 2021.
no. 3, pp. 1380–1387, 2021. [50] J. Cao, C. Song, S. Peng, S. Song, X. Zhang, and F. Xiao, “Trajectory
[25] L. Ljung, “Perspectives on system identification,” Annual Reviews in tracking control algorithm for autonomous vehicle considering cornering
Control, vol. 34, pp. 1–12, 2010. characteristics,” IEEE Access, vol. 8, pp. 59 470–59 484, 2020.
[26] J. E. A. Dias, G. A. S. Pereira, and R. M. Palhares, “Longitudinal [51] H. Wu, Z. Si, and Z. Li, “Trajectory tracking control for four-wheel
model identification and velocity control of an autonomous car,” IEEE independent drive intelligent vehicle based on model predictive control,”
Transactions on Intelligent Transportation Systems, vol. 16, no. 2, pp. 776– IEEE Access, vol. 8, pp. 73 071–73 081, 2020.
786, 2015. [52] Y. Chen, S. Chen, H. Ren, Z. Gao, and Z. Liu, “Path tracking and handling
[27] S. James and S. R. Anderson, “Linear system identification of longitudinal stability control strategy with collision avoidance for the autonomous
vehicle dynamics versus nonlinear physical modelling,” in 2018 UKACC vehicle under extreme conditions,” IEEE Transactions on Vehicular Tech-
12th International Conference on Control (CONTROL), 2018, pp. 146– nology, vol. 69, no. 12, pp. 14 602–14 617, 2020.
151. [53] A. Shaju and A. K. Pandey, “Modelling transient response using pac 2002-
[28] Y. Pan, X. Nie, Z. Li, and S. Gu, “Data-driven vehicle modeling of longi- based tyre model,” Vehicle System Dynamics, pp. 1–27, 2020.
tudinal dynamics based on a multibody model and deep neural networks,” [54] F. Marques, L. Woliński, M. Wojtyra, P. Flores, and H. M. Lankarani, “An
Measurement, vol. 180, p. 109541, 2021. investigation of a novel lugre-based friction force model,” Mechanism and
[29] N. A. Spielberg, M. Brown, and J. C. Gerdes, “Neural network model Machine Theory, vol. 166, 2021.
predictive motion control applied to automated driving with unknown [55] G. E. Karniadakis, I. G. K. L. Lu, P. Perdikaris, S. Wang, and L. Yang,
friction,” IEEE Transactions on Control Systems Technology, 2021. “Physics-informed machine learning,” Nature Reviews Physics, vol. 3, p.
[30] D. Fényes, B. Németh, and P. Gáspár, “A novel data-driven modeling and 422–440, 2021.
control design method for autonomous vehicles,” Energies, vol. 14, no. 2, [56] Z. Zhang, L. Zheng, H. Wu, Z. Zhang, Y. Li, and Y. Liang, “An estima-
p. 517, 2021. tion scheme of road friction coefficient based on novel tyre model and
[31] N. R. Kapania and J. C. Gerdes, “Learning at the racetrack: Data-driven improved sckf,” Vehicle system dynamics, vol. 60, no. 8, pp. 2775–2804,
methods to improve racing performance over multiple laps,” IEEE Trans- 2022.
actions on Vehicular Technology, vol. 69, no. 8, pp. 8232–8242, 2020.
[32] J. Olazagoitia, J. Perez, and F. Badea, “Identification of tire model param-
eters with artificial neural networks,” Applied Sciences, vol. 10, 2020.

14 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3212420

L. C. Sousa, H. V. H. Ayala: Nonlinear Tire Model Approximation Using Machine Learning for Efficient MPC

LUCAS CASTRO SOUSA received the B.S. HELON VICENTE HULTMANN AYALA re-
degree in mechanical engineering from the Ama- ceived the B.S. degree in control and automa-
zonas State University (UEA), and the M.S. de- tion engineering from the Pontifical Catholic Uni-
gree in mechanical engineering from the Military versity of Paraná (PUCPR), the M.S. degree in
Institute of Engineering (IME), respectively in advanced robotics from both Warsaw University
2016 and 2018. He is currently working toward of Technology and University of Genoa, and the
the Ph.D. degree in mechanical engineering in Ph.D. degree in industrial and system engineering
Pontifical Catholic University of Rio de Janeiro from PUCPR, respectively in 2009, 2012, and
(PUC-Rio). His research interests include vehicle 2016. He joined the Department of Mechanical
dynamics, system identification, machine learn- Engineering, PUC-Rio, in 2018, as a faculty mem-
ing, and predictive control techniques. ber. His research interests include system identification, advanced control,
and machine learning.

VOLUME 4, 2016 15

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/