Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

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

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

IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 1

Design and Comparison of Robust Nonlinear


Controllers for the Lateral Dynamics of
Intelligent Vehicles
Gilles Tagne, Reine Talj, and Ali Charara

Abstract—This paper focuses on the lateral control of intelligent or partially to nearly 90% of road accidents. As result, various
vehicles; the aim is to minimize the lateral displacement of the research laboratories and firms are progressively stimulated by
autonomous vehicle with respect to a given reference trajectory. the development of autonomous driving applications. Some
The control input is the steering angle, and the output is the
lateral error displacement. We present here an analysis of com- examples can be seen in [1]–[3]. This research field is in
monality of three lateral nonlinear adaptive controllers. The first expansion and one of the current major challenges is to warrant
controller is a higher order sliding-mode controller (SMC). The a high speed autonomous driving.
second controller is based on the immersion and invariance (I&I) An autonomous navigation can be completed in three manda-
principle. The design of this controller led us to prove a very strong tory steps: the perception and localization, the path planning
stability criterion of the closed-loop system for all controller gains
chosen to be positive. Thereafter, some interesting characteristics and the control. The vehicle control can be divided into two
of passivity of the systems were proved following this development. tasks: longitudinal control and lateral control. The objective of
Hence, the third controller is a passivity-based controller (PBC), this paper is the lateral control of intelligent vehicles, which is a
an adaptive PI controller based on the feedback of a passive very active research field that has been studied since the 1950s.
output. To validate our control laws, tests have been performed on Lateral control consists on automatically maneuvering the
SCANeR Studio, a driving simulation engine, according to several
real driving scenarios. A comparison of these different controllers vehicle using the steering wheel to track the reference tra-
is made to highlight the advantages and drawbacks of each control jectory. Considering the high non-linearity of the vehicle dy-
approach in lateral tracking of a reference trajectory. namics on one hand, and the uncertainties and disturbances
Index Terms—Lateral control, sliding mode control (SMC), in automotive applications on the other hand, robustness can
immersion and invariance (I&I) control, passivity-based con- be considered as a key issue in control design. The controller
trol (PBC), reference tracking, autonomous vehicles, intelligent should be able to reject disturbances and handle parameter’s
vehicles. uncertainties and variations.
I. I NTRODUCTION Lately, significant research has been carried out to provide
lateral guidance of autonomous vehicles. In the literature, sev-
A. Motivations and Problem Statement eral control strategies have been developed. In [4] and [5], a
simple PID controller has been proposed. We also find a nested
S EVERAL contests such as the DARPA (Defense Advanced
Research Projects Agency) Challenges in the USA (2004,
2005, and 2007); the Korean Autonomous Vehicle Competi-
controller in [6]. Furthermore, other classical approaches have
been used such as: state feedback [7]; Linear Quadratic (LQ)
tions (AVC) in Korea (2010, 2012, and 2013) and many others approach [8]; H∞ control [9]; Lyapunov stability based control
have been organized worldwide to favor the development of [10]; adaptive control [11]; fuzzy logic control [12], [13]; fuzzy
autonomous intelligent vehicles. The establishment of such ve- Takagi-Sugeno LQ [14]; backstepping based approach [15] and
hicles would give rise to various advantages that will diminish many others. Model Predictive Control (MPC) seems to be well
road accidents. The objective is to provide an autonomous suited to the trajectory tracking [16], [17]. Nonetheless, the
system more reliable and faster to react than human drivers. It is computation time of nonlinear MPC is the main drawback of
important to notice that the driver’s mistakes contribute entirely this approach. In [18]–[21], Sliding Mode Control (SMC) has
been applied. This control strategy is known for its robustness
against uncertainties and its capacity to reject disturbances.
However, its main drawback is the chattering.
Some comparisons between existing controllers can be found
Manuscript received March 11, 2015; revised July 28, 2015; accepted
September 24, 2015. This work was supported by the French Government and in the literature. In [22], a comparison is made between pro-
carried out in the framework of Labex MS2T through the program Investments portional, adaptive, H∞ and fuzzy controllers. Snider also
for The Future managed by the National Agency for Research under Grant presented and compared several path tracking methods which
ANR-11-IDEX-0004-02. The Associate Editor for this paper was J. E. Naranjo.
The authors are with Heudiasyc Laboratory, Sorbonne universités, Université where principally developed during the DARPA Challenges
de technologie de Compiègne, CNRS, Heudiasyc UMR 7253, CS 60 319, 60 [23]. More recently, in [24], the authors have compared two
203 Compiègne, France (e-mail: gilles.tagne@hds.utc.fr). emergency trajectory tracking controllers. In [25], continuous-
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. time and discrete-time switched H∞ controllers are compared.
Digital Object Identifier 10.1109/TITS.2015.2486815 As a conclusion, we can notice that it is difficult to make
1524-9050 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

an objective classification, but it is clear that different results


pointed out the adaptive controller’s class as a very promising
approach for such uncertain and non-linear application.
Over the years, extensive control strategies based on Sliding
Mode Control (SMC) method have been proposed to design
better and more reliable controllers for the lateral dynamics
of intelligent vehicles. The first-order SMC is used to control
the vehicle with small displacement error by compensating
the uncertainties and encountered disturbances [18]–[20]. It
also provides experimental results comparable to or better than
linear auto-tuning controllers. This control law also possesses
Fig. 1. Bicycle model.
the advantage of being simple, compared to other more complex
approaches of robust control. The major drawbacks of the SMC
method are the following: 1) it needs a prior knowledge on the lateral dynamics in the case of the road coefficient of friction
maximum bounds of the disturbances and uncertainties, 2) it is equal to one (μ = 1) [27]. In the present paper, we extend this
not robust outside the sliding surface, and 3) the chattering. The study by considering μ as a varying parameter.
chattering phenomenon often leads to undesirable results, for The passivity of several input-output maps of the system
example; high vibrations of moving mechanical parts and even have been proved. We have shown that the sliding surface s
passengers discomfort [19]. For those reasons, it is necessary (SM C controller) that has been chosen equal to the—off-the-
to control and maintain the amplitude of oscillations at a low manifold variable—z (I&I controller) is a passive output. To
level. Thus, [21] presented a higher-order SMC method to maintain the same structure with respect to the two previously
alleviate the chattering. The super-twisting algorithm is used developed ones, we propose as third controller, an adaptive PI
to minimize the lateral displacement with respect to a given controller with non-linear gains based on the feedback of this
reference trajectory for an autonomous vehicle. In the present passive output. Finally, a comparison of the three mentioned
paper, we will remind the main lines of the development of this controllers is made to highlight the advantages and drawbacks
controller for a better comparison with the proposed controllers. of each approach in lateral tracking of a reference trajectory.
To design the controllers, we consider that the vehicle is fitted
B. Contributions with sensors and/or observers to measure sideslip angle, yaw
rate, lateral error and its derivative. It is well known that, to
Given the implicit resemblance between SMC and the Im- control the trajectory of an autonomous vehicle tracking using
mersion and Invariance (I&I) principle for designing non- only the lateral error is difficult [28]. Therefore, the use of
linear and adaptive controllers, we develop a controller based other dynamic variables in our control laws such as, the yaw
on the I&I approach in order to overcome the drawbacks of rate, the sideslip angle or the equivalent term helps to quickly
SMC. A preliminary work related to the I&I controller design bring the operating point to the desired equilibrium thus facil-
was presented in [26]. To enhance the performances of this I&I itating the control. To validate the control strategies, we used
controller, in the present paper we use a different law of con- SCANeR Studio [29], a driving simulation engine.
vergence of the—off-the-manifold variable—z to bring up an
integral term in the control law so as to improve the robustness
with respect to parametric uncertainties and disturbances. An C. Paper Structure
analysis of commonality in mathematical frameworks is made; This paper is organized as follows. Section II presents the
both controllers had similar structures. Additionally, the study dynamic models of the vehicle and the control problem def-
of the I&I controller shows a very strong stability criterion inition. Section III presents the SMC design; and the design
of the closed-loop system, for all controller gains chosen to of the I&I controller is presented in Section IV. Then, the
be positive. This strong result of stability led us to study the design of the PBC-PI controller is presented in Section V.
intrinsic properties of passivity of the vehicle lateral dynamics. Section VI presents results. We conclude in Section VII, with
Passivity is a concept that expresses a very interesting stabil- some remarks and prospects.
ity property of some physical systems. Indeed, passive systems
are dynamical systems in which the energy exchanged is the
point of interest. Therefore, a passive system is not able to II. V EHICLE DYNAMIC M ODELS AND C ONTROL
store more energy than it has been supplied; this reflects a P ROBLEM D EFINITION
strong intrinsic stability criterion. Passivity has been studied by
A. Vehicle Dynamic Models
analyzing the frequency behavior of several input-output maps,
in order to determine the passive outputs. Then, passive outputs To design the controllers, a simple and widely used dynamic
can be used to provide robust stability and good performances bicycle model [7] is considered (see Fig. 1). This model is used
by the development of Passivity-Based Controllers (PBC). This to represent the lateral vehicle behavior (lateral acceleration
is particularly relevant in this application given the parametric ay , yaw rate ψ̇, sideslip angle β) and assumes that the vehicle
variations and uncertainties (speed, curvature, road friction is symmetrical, and tire sideslip angles on the same axle are
coefficient, etc...). Very recently, we studied the passivity of the equal. The roll and pitch dynamics are neglected and angles
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TAGNE et al.: ROBUST NONLINEAR CONTROLLERS FOR LATERAL DYNAMICS OF INTELLIGENT VEHICLES 3

TABLE I where,
V EHICLE PARAMETERS AND N OMENCLATURE (B ICYCLE M ODEL )
⎡ μ(Cf + Cr ) μ(L C −L C )

− −1 − f f r r
0 0

mVx  2 mV2x2  ⎥
⎢ μ(Lf Cf −Lr Cr ) μ L Cf +Lr Cr ⎥
A=⎢− − f
0 0⎥
⎢ Iz Iz Vx
μ(L Cf −Lr Cr ) ⎥
⎣ − μ(Cf + Cr ) − f mV 0 0⎦
m x
0 0 1 0
⎡ μCf ⎤ ⎡ ⎤
mVx 0
⎢ μLf Cf ⎥ ⎢ 0 ⎥
⎢ Iz ⎥
B1 = ⎢ ⎥ B2 = ⎢ ⎥
⎣−Vx2 ⎦ . (5)
⎣ μCf ⎦
m 0
0

are assumed to be small (steering δ, sideslip β, yaw ψ). With The control input is the steering angle δ̃ and the output is
a linear tire force model we obtain a linear parameter varying the lateral error displacement e. The lateral control aims to
(LPV) model, where the longitudinal velocity Vx is considered eliminate the lateral error. Then, for a given curvature ρ and
as a varying parameter. Dynamic equations in terms of slip longitudinal velocity Vx , the desired behavior corresponds to
angle and yaw rate of the bicycle model are given by: ė = e = 0. Hence, it is easy to prove that the desired equilib-
⎧   rium point is [7]:
⎨β̇ = − μ(Cf +Cr ) β − 1 + μ(Lf Cf −L r Cr ) μC
ψ̇+ mVfx δ
 2 x 2 
mVx mV 2

⎩ μ(L C −L C ) μ L Cf +Lr Cr μL C (β, ψ̇, ė, e) = (β  , ψ̇  , 0, 0)


ψ̈ = − f fIz r r β − f
Iz Vx ψ̇ + Ifz f δ
(1) with
where β and ψ represent respectively the sideslip angle and the  
Lf mVx2
yaw angle of the vehicle. Table I presents vehicle parameters β = Lr − ρ
and nomenclature. Note that the cornering stiffness values of μCr (Lf + Lr )
the dynamic bicycle model (Table I) are twice the value of a ψ̇  = Vx ρ. (6)
data sheet since the two tires are being treated as one. For more
details on the model and parameters see [23] or [18]. To validate At the equilibrium point, the control input is:
in simulation the proposed controllers, we used SCANeR studio
simulator [29]. This simulator represents the vehicle with its mVx2 (Lr Cr − Lf Cf )
δ  = (Lf + Lr )ρ + ρ. (7)
full dynamical model. μCf Cr (Lf + Lr )

We define the new error variables:


B. Control Problem Definition ⎧
⎨β̃ = β − β
⎪ 
The lateral control’s aim is to minimize the lateral dis-
ψ̃˙ = ψ̇ − ψ̇  (8)
placement of the autonomous vehicle with respect to a given ⎪

reference path. The lateral error dynamics at the centre of δ̃ = δ − δ  .
gravity of the vehicle, with respect to a reference path, is
given by: Note that the desired reference state (Equations (6), (7)) is the
steady-state solution to the bicycle model. Hence, using this as
ë = ay − ayref (2) a reference input (Equation (8)) can ignore the natural transient
dynamics of the model.
where ay and ayref represent respectively the lateral accelera- The error dynamics of the system (4) whose origin is the
tion of the vehicle and the desired lateral acceleration along the ˙
equilibrium point (β̃, ψ̃, ė, e) = (0, 0, 0, 0) become:
reference trajectory. Assuming that the desired lateral acceler-
ation of the vehicle can be written as ayref = Vx2 ρ; where ρ is x̃˙ = Ax̃ + B1 δ̃ (9)
the curvature of the road which we assume constant. Given that
ay = Vx (β̇ + ψ̇) [7], then: where A and B1 have been defined above in (5).
ë = Vx (β̇ + ψ̇) − Vx 2 ρ. (3)
III. SMC C ONTROLLER D ESIGN
Adding the equation of the lateral error dynamics (3) in
the system (1), the new system state variables become x = The Sliding Mode Control (SMC) has been developed since
(β, ψ̇, ė, e) , corresponding to the sideslip angle β, the yaw rate 1950s and is recognized as one of the most promising tech-
ψ̇, the lateral error e and its derivative ė. The system dynamics niques for robust control. The principle of SMC is to constrain
can be written as follows: the system trajectories to reach and remain on a sliding surface
in finite time. However, its main drawback is the chattering.
ẋ = Ax + B1 δ + B2 ρ (4) In this subsection, we recall the main lines of the design of
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

the SMC controller. We used the super-twisting algorithm to case of a more wide class of adaptive non-linear controllers,
minimize the lateral displacement. For more details, see [21]. known by the principle of “Immersion and Invariance.”
The objective is to cancel the lateral displacement error. To
this end, choosing the sliding variable s as follows,
IV. I&I C ONTROLLER D ESIGN
s = ė + λe (10) The Immersion & Invariance (I&I) principle is a relatively
new method for designing non-linear and adaptive controllers.
with λ a positive constant, we obtain ṡ = ë + λė. Replacing ë
The idea consists on achieving the control objective by immers-
by its expression (3), we obtain:
ing the system into a target dynamics that ensures the objective
μ(Cf +Cr ) μ(Lf Cf −Lr Cr ) μCf of the command. Indeed, as mentioned before, the SMC is a
ṡ = − β− ψ̇−Vx2 ρ+ δ+λė. particular case of the I&I method where the sliding surface is
m mVx m
(11) the target dynamics and the law of convergence towards this
The variable s has a relative degree r = 1 with respect to the latter may be discontinuous (and in finite time).
input δ. We can write ṡ of the form: To design the I&I controller, some assumptions should
be verified: the system and the target dynamics should both
ṡ(t, s) = φ(t, s) + ϕ(t, s)δ (12) have their equilibrium at the origin. The immersion of the
system is achieved by rendering the manifold attractive and
with,
invariant. Therefore, the command consists on bringing all the

μ(C +C ) μ(Lf Cf −Lr Cr ) system state trajectories towards the manifold. Astolfi et al. [30]
φ(t, s) = − fm r β − mVx ψ̇ − Vx2 ρ + λė
μCf present the main stabilization principle of the Immersion and
ϕ(t, s) = m .
Invariance method.
(13) Considering the vehicle lateral dynamics model (9), the main
It is assumed that there exist positive constants s0 , bmin , bmax , objective of the steering controller is to eliminate the lateral
C0 such that ∀ x ∈ Rn and |s(t, x)| < s0 , the system satisfies error displacement with respect to a given trajectory, so that
the following conditions: e = ė = 0 at the equilibrium. The target system is consequently

⎪ chosen to be the image of the system (9) when e = ė = 0. More
⎨|δ| ≤ δmax
precisely, the target state vector is (ξ1 , ξ2 ), where ξ1 and ξ2
0 < bmin ≤ |ϕ(t, s)| ≤ bmax (14)

⎩ represent β̃ and ψ̃˙ respectively, when e = ė = 0. Note that for
|φ(t, s)| < C0 .
e = ė = 0, we also have ë = 0. With ë = 0, the control input of
So, the SMC control based on super-twisting algorithm is the system (9) is
given by: Cf + Cr Lf Cf − Lr Cr ˙
 1
δ̃0 =
Cf
β̃ +
Cf Vx
ψ̃. (19)
δ1 = −α1 |s| 2 sign(s)
δST = δ1 + δ2 (15)
δ̇2 = −α2 sign(s) Hence, the target dynamics can be expressed as follows,
    
with α1 and α2 positive constants. The finite time convergence ξ˙1 0 −1 ξ1
= μCr (Lf +Lr ) μLr Cr (Lf +Lr ) . (20)
to the sliding surface is guaranteed for all: ξ˙2 I − I V
ξ2
⎧ z z x

⎨α2 > C0

bmin This target dynamics is quite simply the dynamics of the
(16)
⎩α1 ≥ 4C 0 (bmax α2 +C0 )
. sideslip angle and yaw rate. It is stable and has been studied
b2 (bmin α2 −C0 )
min
in the literature [31]. With this new formulation, the target
To avoid important peaks in transient phases, we add an equiva- dynamic can be expressed directly as a function of the set point
lent command δeq obtained by solving the equation ṡ = 0. This ρ (reference curvature) as input. These dynamics are given by
term has the role of a feedforward that brings the system near the following equation:
to the sliding surface, and is given by:       
β̇ 0 −1 β Vx
= μCr (Lf +Lr ) + Lf mVx2 ρ
Cf + Cr Lf Cf − Lr Cr mVx2 mλ ψ̈
μL C (L +L )
− r Irz Vxf r ψ̇
δeq = β+ ψ̇ + ρ− ė. Iz Iz
Cf Cf Vx μCf μCf (21)
(17)
where, for a given curvature ρ, β, and ψ̇ converge respectively
Hence, the steering angle representing the control input of the
to β  and ψ̇  .
system is defined as follows:
Proposition 1: The target model (20) has a globally asymp-
δSMC = δST + δeq . (18) totically stable equilibrium at the origin (0, 0).
Proof 1: The dynamics of the state vector ξ = [ξ1 , ξ2 ] can
The sliding mode is a kind of non-linear PI control that forces be written in the form ξ˙ = Cξ, with
the state trajectories of the system towards a sliding surface,  
rendered invariant by the use of discontinuous “sign” functions. 0 −1
C = μCr (Lf +Lr ) μL C (L +L ) . (22)
Hence, the principle of the sliding mode control is a particular Iz − r Irz Vxf r
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TAGNE et al.: ROBUST NONLINEAR CONTROLLERS FOR LATERAL DYNAMICS OF INTELLIGENT VEHICLES 5

Then, after some simple calculations, we obtain: The subsystem S2 combines and represents the interactions
between the dynamics of the two off-the-manifold variables z
μLr Cr (Lf + Lr ) μCr (Lf + Lr )
det(sI − C) = s2 + s+ . and e. This lateral error dynamics in closed-loop (S2 ) does not
Iz Vx Iz ˙ it is given by :
(23) depend on the variables β̃ and ψ̃,
For any uncertainties and variations encountered, the model’s t
parameters (Cf , Cr , Lf , Lr , Iz , μ) are always positive. The ë = −(K1 + λ)ė − (K1 λ + K2 )e − K2 λ e dt. (28)
matrix C satisfies the Routh-Hurwitz stability criterion (Vx > 0), 0
which leads to the desired result.
Proposition 2: Consider the system (9) whose equilibrium Proposition 3: The subsystem S2 has a globally asymptoti-
point is at the origin. Moreover, the subsystem (20), which cally stable equilibrium at the origin.
is the image of the system (9) for e = ė = 0, has a globally Proof 3: The dynamics of the subsystem S2 have the
asymptotically stable equilibrium at the origin. The system (9) characteristic equation:
is I&I-stabilizable with target dynamics (20).
Proof 2: We now define the off-the-manifold variable P (s) = s3 + (K1 + λ)s2 + (K1 λ + K2 )s + K2 λ. (29)

z = ė + λe, s.t. λ > 0. (24) This equation satisfies the Routh-Hurwitz stability criterion
for all the gains K1 > 0, K2 > 0 and λ > 0, and hence the
Note that we choose z = s, the sliding variable defined above attractivity of manifold is ensured.
for the design of the SMC controller. Finally it is clear that when S2 converges to zero, the input ũ
We have to select a control input δ̃ such that the trajectories of the subsystem S1 (27) converges to zero. The system then
of the closed-loop system remain bounded and z converges to converges to the target subsystem, whose dynamics (20) are
zero. It should be noticed that when z → 0, e converges to zero, stable and converge to (0, 0). We have already proved above
yielding to the desired result. that this target dynamics have a globally asymptotically stable
As a way of addressing this problem, in [26], we have equilibrium at the origin (0, 0). Consequently, the trajectories of
imposed an exponential convergence of z (we used ż = −Kz, the closed-loop system are bounded, they immerse into the tar-
with K > 0). In the present paper, we added an integral term get dynamics inside the invariant manifold, and then converge
that changes the asymptotic convergence behavior of the control to the desired equilibrium, yielding to the desired result. For
law, so as to improve the robustness with respect to parametric more details on the proof of global asymptotic stability of the
uncertainties and disturbances. To this end, we define a stable closed-loop system, see [32], where it is shown that the cascade
dynamics for z such that: interconnection of Global Asymptotic Stable (GAS) systems
with a bounded trajectory gives a globally asymptotically stable
t
system.
ż + K1 z + K2 z dt = 0, with K1 > 0, K2 > 0. (25) The proposed I&I controller guarantees the stability of the
0 system, for all its gains λ > 0, K1 > 0 and K2 > 0. This is a
The off-the-manifold variable z converges to 0, then e and very strong result that reveals some interesting characteristics
ė converge to 0. The system is thus immersed into the target of the system. The parameters K1 > 0, K2 > 0 and λ have to
dynamics, where it converges to the origin. be chosen based on other practical considerations related to the
Replacing z by its expression (24) in the equation (25), system performance and transient behavior. Finally, the control
and after some calculations, the corresponding control input is input applied to the system (4) is:
given by:
m(K1 + λ) m(K1 λ + K2 ) mK2 λ
δI&I = δ̃ + δ  = − ė− e−
t μCf μCf μCf
m(K1 + λ) m(K1 λ + K2 ) mK2 λ
δ̃ = − ė − e− e dt t
μCf μCf μCf Cf + Cr Lf Cf − Lr Cr mVx2
0 × e dt + β+ ψ̇ + ρ. (30)
Cf Cf Vx μCf
Cf + Cr Lf Cf − Lr Cr ˙ 0
+ β̃ + ψ̃. (26)
Cf Cf Vx
The I&I formulation assumes the exact knowledge of the
t model parameters. However, the controller is designed to deal
Defining ũ = (K1 + λ)ė + (K1 λ + K2 )e + K2 λ 0 e dt, the
with some model errors. The structure of the proposed con-
˙
dynamics of β̃ and ψ̃ in closed-loop become: troller guarantees the convergence to the neighborhood of the
      desired state, even though the model presents some weaknesses
˙ β̃ − V1x
β̃ to represent the system in some regions. The control law can be
¨ = C ψ̃˙ + − Lf m ũ. (27)
ψ̃ Iz interpreted as a dynamic state feedback controller plus a PID
controller (with gains depending on system parameters), for
The closed-loop system can be interpreted as the intercon- robustness. The dynamic state feedback provides compensation
˙
nection of two subsystems: S1 with the state variables β̃ and ψ̃, of some system dynamics, in order to rapidly converge to a
and S2 with the state variables ė, e and the integral of e( e). neighborhood of the desired state. On the other hand, the PID
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

Fig. 3. Cascade interconnection of two passive systems.

B. Passivity Analysis of the Lateral Error Dynamics


Fig. 2. Feedback interconnection of passive systems scheme. Theory of passivity is a framework for analyzing physical
systems and designing controllers using a description of the
controller helps to ensure local convergence and to reject the
input-output relationship based on energy considerations. The
effect of uncertainties and / or disturbances. This control law
Kalman-Yakubovich-Popov (KYP) lemma is considered to be
is very close to the SMC control law (the equivalent term).
one of the pillars for control and systems theory. To demonstrate
The difference being at the level of the robust term which is
the passivity of an output, this lemma is considered. It estab-
an adaptive PID for the I&I and a nonlinear PI for the SMC.
lishes an equivalence between the conditions in the frequency
domain (positive realness of a system), in the time domain (an
V. PASSIVITY-BASED C ONTROLLER D ESIGN
input-output relationship of the system), and in the state-space
A. Passivity-Based Control Method representation where conditions on the matrices describing the
system should be verified. The Lemma is given as follows:
As mentioned before, the strong stability result obtained with
Lemma 1: (KYP lemma) [34]
the I&I controller for all positive gains, led us to study the
In the frequency domain: Consider the transfer function
intrinsic characteristics of the system. Indeed, this study reveals
H(s) and the operator returning the real part (). H(s) is
the presence of several input-output passive maps in the system,
Positive Real (PR) if and only if:
which is a very interesting and promising result that allows us

to design Passivity-Based Controllers suitable to improve the H(s) is stable
robustness of such uncertain non-linear system. (31)
[H(jω)] ≥ 0, ∀ ω ∈ R.
The Passivity-Based Control (PBC) was introduced in 1989
to define a control methodology that aims to make the closed- In the time domain: A system with input u and output y,
loop system (system + controller) passive [33]. These passivity where u(t), y(t) ∈ Rm , is passive if there is a positive constant
properties guarantee the stability and robustness of the closed- υ such that
loop system. Two approaches are used in the literature to design
PBC controllers: T
y T (t)u(t) dt ≥ υ (32)
• Controller design, choosing an appropriate energy func- 0
tion which will guarantee the passivity of the closed-loop
system (so that the closed-loop system is made passive by for all function u and all T ≥ 0.
the command) [33]. In the state-space representation: Let us consider a sys-
• Controller design, based on the passivity properties of tem described by the following state-space representation: ẋ =
the system as well as the interconnection of passive Ax + Bu and y = Cx + Du, where x ∈ Rn , u, y ∈ Rm with
systems [34]. n ≥ m. The transfer function H(s) = C(sI − A)−1 B + D,
with A ∈ Rn×n , B ∈ Rn×m , C ∈ Rm×n , D ∈ Rm×m is PR
The interconnection of passive systems has been studied with H(s) ∈ Rm×m , if and only if there exists matrices P > 0,
in the literature, particularly the re-looping and cascading. P ∈ Rn×n , L ∈ Rn×m and W ∈ Rm×m such that:
The following corollary (corollary 4.1 in [34]) reminds a ⎧

⎨P A + A P = −LL
T T
very useful property of stability of two systems on feedback
P B − C = −LW
T (33)
interconnections. ⎪
⎩ T T
Corollary 1: Considering the diagram of Fig. 2, this feedback D + D = W W.
system with finite gains is stable if either of the following
statements is true: Remark: In view of the above conditions, it is clear that
unstable systems and non-minimum phase systems are not
• S1 is Passive (P) and S2 is Strictly Passive (SP). positive real.
• S1 is Strictly Passive (SP) and S2 is Passive (P). We will analyze the passivity properties of the lateral error
dynamics (ë, ė and z) with respect to the steering input δ̃.
Corollary 2: Considering the diagram of Fig. 3, the system 1) Passivity of the map δ̃ → ë:
resulting from this cascade interconnection is passive if one of Proposition 4: The map δ̃ → ë is Strongly Strictly Passive
the following statements is true (satisfied): for a road coefficient of friction μ = 1.
Proof 4: The proof of strong strict passivity is established
• S1 is Passive (P) and S2 is Strictly Passive (SP). showing that the transfer function H00 (s) of the map δ̃ → ë is
• S1 is Strictly Passive (SP) and S2 is Passive (P). Strongly Strictly Positive Real (SSPR).
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TAGNE et al.: ROBUST NONLINEAR CONTROLLERS FOR LATERAL DYNAMICS OF INTELLIGENT VEHICLES 7

Assume that the road coefficient of friction μ = 1. Consider- this second term is positive, implying that f is always positive
ing the error dynamic equations (9), after some calculations, (which is the case of most commercial vehicles). For oversteer
one can find that the transfer function H00 (s) between δ̃ as vehicles, f can theoretically be negative in case the vehicle has
input and ë as output is given by: atypical parameters (extremely large speed or mass). In this
work, we assume that f is positive, which is usually the case
ë(s) a0 s2 + b 0 s + c0
H00 (s) = = (34) for the conventional cars.
δ̃(s) s2 + d0 s + f0 According to the criterion of Routh-Hurwitz, zeros and poles
of H0 (s) are strictly stable, so H0 (s) is minimum phase and
where,
Hurwitz.
⎧ C
⎪ a = f Setting s = jω, the frequency response function becomes

⎪ 0 Lmr Cf Cr (Lf +Lr )



⎪b0 = −aω 2 + bjω + c RN 0 + jIN 0


mIz Vx
H0 (jω) = = (39)
⎪ C C (L +L )
⎪c0 = f rmIfz r −ω 2 + djω + f

⎨  RD + jID
(Cf +Cr ) L2 Cf +L2r Cr
d0 = mV + f
(35) where RN 0 , IN 0 , RD , and ID are the real and imaginary parts

⎪ x Iz Vx

⎪ Cf Cr (Lf +Lr )2
of the numerator and the denominator respectively, given by:
⎪ 1
⎪ f = mIz Vx2


⎪f2 = (Lr Cr −Lf Cf )


⎪ Iz RN 0 = c − aω 2 , IN 0 = b ω
⎩ C Cr (Lf +Lr )2 (L C −L C )
f0 = f mI zV
2 + r rIz f f = f1 + f2 .
x
RD = f − ω , 2
ID = dω.
The transfer function H00 (s) has a relative degree equal to 0.
According to the KYP Lemma applied in the frequency domain, Hence,
we have shown that the transfer function has a positive real part RN 0 RD + IN 0 ID
for a road coefficient of friction μ = 1 (see [27]). [H0 (jω)] = 2 + I2 . (40)
RD D
We extend this study by considering μ as a varying
parameter. Knowing that RD2 2
+ ID > 0, [H0 (jω)] is positive if and only
Proposition 5: The map δ̃ → ë is Strongly Strictly Passive if RN 0 RD + IN 0 ID > 0.
for all μ > μ0 .
Proof 5: The proof of strong strict passivity is established RN 0 RD + IN 0 ID = aω 4 + (bd − af − c)ω 2 + cf. (41)
showing that the transfer function H0 (s) of the map δ̃ → ë is
Strongly Strictly Positive Real (SSPR). The transfer function Given the order of magnitude of the parameters mentioned
H0 (s) between δ̃ as input and ë as output for all μ
= 1 is above and whatever is Vx > 0, one can find that:
given by: ⎧

⎨a > 0
ë(s) as2 + bs + c
H0 (s) = = 2 (36) cf > 0 (42)
δ̃(s) s + ds + f ⎪

bd − af − c is not always positive.
where,
⎧ Given that (bd − af − c) is not always positive, let x = ω 2 ,


μC
a = mf = μa0 hence



⎪ μ2 Lr Cf Cr (Lf +Lr )

⎪b= = μ2 b 0

⎨ 2
mIz Vx RN 0 RD + IN 0 ID = f (x) = ax2 + (bd−af −c)x + cf. (43)
μ Cf Cr (Lf +Lr )
c= = μ2 c 0 (37)


mIz   The derivative of this function is zero for

⎪ μ L2 Cf +L2r Cr

⎪d =
μ(Cf +Cr )
+ f
= μd0

⎪ mVx Iz Vx
bd − af − c

⎩f = μ2
C C (L +L ) 2
μ(L C −L C ) x0 = .
f r f
mIz V 2
r
+ r rIz f f = μ2 f1 +μf2 . −2a
x

The transfer function H0 (s) has a relative degree equal to 0. So, the unique extremum of the function is
Consequently, according to the KYP Lemma applied in the fre-
(bd − af − c)2
quency domain, the proof of SSPR is tantamount to verifying: f (x0 ) = + cf. (44)
−4a
[H0 (jω)] ≥ ζ > 0, ∀ ω ∈ (−∞, +∞). (38)
Given that
Whatever the uncertainties and variations encountered, the
model parameters (Cf , Cr , Lf , Lr , m, Iz , and μ) are always lim [H0 (jω)] = a > 0 (45)
ω→∞
positive. Thus, the coefficients a, b, c, d of H0 (s) are always c
positive. Considering the expression of f , it can be positive lim [H0 (jω)] = >0 (46)
ω→0 f
or negative. In fact, the second term of f (f2 ) can be positive
or negative depending on the type of vehicle steering system it is sufficient to have f (x0 ) > 0 in order to prove that the
(understeer or oversteer). In the case of understeer vehicles, transfer function H0 is strongly strictly positive real.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

TABLE II
C OEFFICIENT OF M AXIMUM A DHESION μ FOR
D IFFERENT PAVEMENT C ONDITIONS

Fig. 5. Transfer function Hz (s) of the map δ̃ → z.

Strict passivity of H0 guarantees that,


T
ë(t)δ̃(t)dt > 0 for all T. (51)
0

Also, it is easy to show that:

Fig. 4. Cascade connection scheme. T


1
ė(t)ë(t)dt ≥ − ė(0)2 = υ. (52)
2
0
To have f (x0 ) > 0 in the equation (44), this is equivalent
to have : Finally, we have:
T
4c0 a0 f1 μ3 + 4c0 a0 f2 μ2 + (b0 d0 − a0 f1 )μ − af2 − c0 > 0. ė(t)δ̃(t)dt ≥ υ. (53)
(47) 0

This inequality shows clearly that the passivity of the lateral The map δ̃ → ė is passive, what is the desired result. For more
error depends on the road coefficient of friction μ. details see [34].
In [27], we have shown that for μ = 1, f (x0 ) > 0. The 3) Passivity of the map δ̃ → z:
Equation (47) clearly shows that when μ → 0, f (x0 ) < 0. In Proposition 7: Consider an output defined by:
addition, f (x0 (μ)) is an increasing continuous function of μ.
Thus there exist a μ0 such that f (x0 (μ0 )) = 0. z = s = ė + λe. (54)
Thus, we have shown that ∀ μ > μ0 defined above:
The map δ̃ → z is Passive.
Proof 7: The proof of the passivity of z is established
[H0 (jω)] ≥ ζ > 0, ∀ ω ∈ (−∞, +∞). (48) by showing that it is a cascade interconnection of two sys-
tems respectively passive S1 and strictly passive S2 (see the
To have an idea on the order of magnitude of μ0 , we Corollary 2).
considered the parameters of the vehicle of the Table I for Consider H1 (s) the transfer function between δ̃ and ė.
a longitudinal velocity of Vx = 40 m/s, then we obtain μ0 Hence, the transfer function Hz (s) between δ̃ and z can be
0.012. Table II provides a list of values of μ for different given by the Fig. 5.
pavement conditions [35]. We have shown that S1 = H1 is passive (see Proposition 6).
It is important to note that the value of μ0 is relatively
s+λ
low, corresponding to a very icy pavement. In other words, for S2 (s) =
s
driving on dry pavement, wet or snowy, the transfer function
H0 (s) is strongly strictly passive. S2 defines a Strictly Passive (SP) map; it is easy to show
2) Passivity of the map δ̃ → ė: that [34]:
Proposition 6: The map δ̃ → ė is Passive (P) for all μ > μ0 .
Proof 6: The transfer function H1 (s) of the output ė with [S2 (jω)] ≥  > 0, ∀ ω ∈ (−∞, +∞).
respect to the input δ̃ is given by:
So the map δ̃ → z is passive, what yields to the desired result.
This result partially explains the stability result obtained
1
H1 (s) = H0 (s). (49) previously in the literature [26]. Indeed, we have previously
s developed different control approaches based on the variable
z, and the stability obtained was robust. In fact, z was chosen
The transfer function H1 (s) is a cascade connection of an in-
as the sliding variable for the SMC in [21], and as the off-
tegrator with the strongly strictly positive real transfer function
the-manifold variable for the I&I controller in [26]. Now, this
H0 (s) for all μ > μ0 . (see Fig. 4). passivity result explains and allows to extend our results by us-
Considering the scheme of the Fig. 4, we have,
ing this intrinsic property in Passivity-Based robust controllers
design for example.
T T T To summarize, considering the error system (9), the deriva-
ė(t)δ̃(t)dt = ë(t)δ̃(t)dt + ė(t)ë(t)dt. (50) tive of the lateral error ė and the variable z are passive outputs
0 0 0 (P) for a steering input δ̃ for all μ > μ0 .
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TAGNE et al.: ROBUST NONLINEAR CONTROLLERS FOR LATERAL DYNAMICS OF INTELLIGENT VEHICLES 9

Fig. 6. PBC control strategy.

C. PBC-PI Controller Design


Fig. 7. The non-linear gain function with emax = 0.25 m and k0 = 12.
Probably the main property used in PBC for passive systems
is the fact that passivity is invariant under negative feedback
interconnection. Thus, in this methodology, the controller can esat is defined as:
be designed as a passive system. However, it will also be useful 
to know that interconnections not only preserve the passivity e for |e| ≤ emax
esat = (57)
properties of the subsystems but, in certain cases, passivity can emax sign(e) for |e| > emax
be strengthened.
The aim of the lateral control of autonomous intelligent where k0 and emax are positive constants defined by the user:
vehicles is to minimize the lateral displacement e of the vehicle emax is the variation range of the maximum admissible error,
with respect to a given reference trajectory. Considering the and, k0 is the rate of reduction of the non-linear gains near to
Fig. 6, we can design the lateral controller with the purpose the equilibrium.
to control the lateral error e by controlling the variable z. The When the error between the desired and actual values of the
controller is designed to cancel the passive output z with respect controlled variable e is large, the non-linear gain is amplified
to the reference trajectory curvature ρ. The control input is the substantially so as to generate a large correction to rapidly drive
steering angle δ̃. We have to select a control input δ̃ such that the system towards its desired state. When the error diminishes,
z = ė + λe converges to zero. Notice that, when z → 0, the the non-linear gain is automatically reduced to prevent exces-
lateral error e converges exponentially to zero with the rate of sive oscillations and large overshoots in the response near to the
convergence λ, yielding to the desired result. equilibrium. Thanks to this automatic gains adjustment, this PI
Proposition 8: Consider the diagram in Fig. 6, where the controller enjoys the advantage of providing high gains far from
control is achieved by a PI controller, the closed-loop system the objective to obtain a fast response, and a reduced gain near
is stable and passive. the equilibrium so as to prevent an oscillatory behavior.
Proof 8: The proof is relatively simple, the Corollary 1 Considering the allowable lateral error to be less than
allows to show it. 25 cm, in such a way to ensure trajectory tracking when
The map δ̃ −→ z is passive (P). Therefore, any Strictly avoiding risk of collision, the Fig. 7 presents the non-linear gain
Passive (SP) controller (a simple Proportional (P) or a PI for ke (e) function of the lateral error e. The gain varies between
example) guarantees stability and passivity of the closed-loop 1 and 10 for k0 = 12.
system. The proposed PBC-PI controller guarantees the stability and
It is well-known that PI controllers, if suitably tuned, provide the passivity of the system, for all its gains kI ke (e) > 0, and
a good solution to many practical applications without requiring kP ke (e) > 0. The parameters kI and kP have to be chosen
a detailed description of the system dynamics (when their use based on practical considerations related to the system perfor-
does not destabilize the system). In the presence of strong mance. Finally, the control input applied to the system is:
non-linearities, their performances can be deteriorated and it is t
therefore necessary to re-tune the controller or to use adaptive
δP BC−P I = δ̃ + δ = −kI ke (e)

zdt − kP ke (e)z
gains. Hence, we propose a PI with non-linear gains of the form:
⎡ ⎤ 0
t mVx2 (Lr Cr − Lf Cf )
+ (Lf + Lr )ρ + ρ. (58)
δ̃ = − ⎣kI zdt + kP z ⎦ ke (e) (55) μCf Cr (Lf + Lr )
0

where kI and kP are positive gains. ke (e) is a nonlinear gain VI. R ESULTS
function of the lateral error e. Before presenting the simulation results, we remind that
The non-linear gain ke (e) is a function depending on the the control input with the SMC controller is given by the
lateral error e which is bounded. This non-linear gain is Equation (18). The finite time convergence to the sliding surface
given by: is guaranteed for the conditions (16). These conditions depend
exp (k0 esat ) + exp (−k0 esat ) on the maximum considered magnitude (C0 , bmin , bmax ) of the
ke (e) = cosh(k0 esat ) = (56) system’s variables. If these conditions are no more verified, the
2
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

Fig. 8. Track circuit. OKTAL-SCANeR Simulator.

convergence of the closed-loop system is no more guaranteed


and the performances of the SMC controller can also be signifi-
cantly deteriorated. For the SMC control law, we used the gains
λ = 8, α1 = 0.008, α2 = 0.008.
The control input with the I&I controller is given by the
Equation (30). This control law can be interpreted as a dy-
namic state feedback plus a PID controller (with adaptive gains
depending on the system parameters and variables) to ensure
robustness and reject the effect of parametric uncertainties and
disturbances. To validate this controller, we used the gains
λ = 8 as for the SMC, and K1 = 2, K2 = 0.5.
The control input with the PBC-PI controller is given by the
Equation (58). The gains used are kP = kI = 0.015.
We observe that the three control laws have almost similar
structure and can be interpreted as linear / non-linear PIDs
with a state feedback. It is important to note that in addition
to the conditions of stability, the gains of the three controllers
have also been adjusted in order to have comparable behavior
for a normal driving test with known design parameters. The
aim is to have a reliable assessment and comparison of the
performances of the controllers thereafter.
To apply our control laws, the yaw rate and the lateral error
are measured. The sideslip angle and the derivative of the lateral
error are estimated. Regarding the estimation of the sideslip
angle, we used a Kaman estimator based on the dynamic bicycle
model. For more details see [36]. Note that the same measures
are used for the SMC and the I&I controllers. The use of other
dynamic variables such as the yaw rate, the sideslip angle or an
equivalent term quickly bring the operating point to the desired
equilibrium while aiding the tracking of the reference.
To validate our control laws, tests were performed on
Fig. 9. Test of the controllers in normal driving conditions with known nominal
SCANeR Studio Driving Simulation Engine [29]. SCANeR parameters: (a) Road information and longitudinal speed. (b) Paths and errors.
Studio is a simulation environment developed by OKTAL so-
ciety. This is a complete software that meets the challenges of
driving simulation. Far from being a closed tool, it is a scal-
A. Test of Controllers During Normal Driving With Known
able simulation platform, extensible and open to the needs of
Nominal Parameters
researchers and engineers. This tool includes modules to model
and simulate the dynamics of the vehicle, traffic, pedestrian Fig. 9 presents the controllers in normal driving conditions
behaviour, etc. It offers –Human and Hardware in the Loop with known nominal parameters. Fig. 9(a) shows the profile of
(HHIL)– solutions allowing to get closer to a real experimental the vehicle longitudinal speed, curvature and slope of the road.
validation. The road curvature varies between −0.02 and +0.02 m–1 and
This simulator uses the vehicle full model; a more representa- the longitudinal speed varies between 5 m/s and 30 m/s. The
tive model to represent the vehicle coupled non-linear dynamics path followed by the controlled vehicle is shown in Fig. 9(b) for
and behaviour. Fig. 8 shows the test circuit selected. This circuit the three controllers. Moreover, the lateral error and the yaw
is used to test the controllers at high-speed. rate error are also presented in this Figure.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TAGNE et al.: ROBUST NONLINEAR CONTROLLERS FOR LATERAL DYNAMICS OF INTELLIGENT VEHICLES 11

Fig. 10. Steering angle.

Fig. 12. Robustness against uncertainties of the cornering stiffness.

B. Robustness With Respect to Vehicle


Parameter Uncertainties
The test described above shows the performances of the
controllers during normal driving. For trajectory tracking, one
of the major challenges is the robustness of the controller
against uncertainties in the vehicle’s parameters or the driving
Fig. 11. Non-linear gain value of the PBC-PI controller. environment. We then evaluate the performance of the con-
trollers against parameter uncertainties.
We observe that the SMC, the I&I and the PBC-PI con- Several parameters of the vehicle may be uncertain, as the tire
trollers can ensure the convergence of the lateral error. These stiffness coefficient and the mass of the vehicle. To assess the
non-linear controllers are robust against speed variations. The robustness of the controller with respect to parametric uncer-
lateral error remains small when the speed Vx is varying. The tainties, we have changed the nominal value of the parameter
controllers have comparable maximum errors. The displace- against which we want to verify the robustness of the control
ment from the reference trajectory does not exceed 10 cm in law. Then we have tested the controller with this incorrect
this test conditions (Fig. 9(b)). parameter value.
Fig. 10 presents the steering angle of the controllers. The It is difficult to accurately estimate the tire cornering stiff-
I&I and PBC-PI controllers provide a continuous steering ness. In addition, this parameter varies considerably depending
angle while avoiding chattering. The I&I controller provides on the type of road, the vertical load, camber, etc. Similarly,
the smoother steering angle. the mass of the vehicle may be varied or poorly estimated. It
Fig. 11 shows the variation of the non-linear PBC gain depends on the number of persons in the vehicle, the amount
function of time. Adaptive PBC-PI controller ensures a robust of fuel, etc. It is therefore important to assess the robustness of
path tracking. During the large turn; around t = 25 s, the the controller with respect to uncertainties on these parameters.
non-linear gain term improves performance by approximately Figs. 12 and 13 present the lateral error due to uncertainty from
doubling the gains of the controller. Because of this automatic +/−20% to +/−40% on the tire cornering stiffness and the
gain adjustment, this PI controller enjoys the advantage of high mass of the vehicle. For uncertainties of the order of +/−40%
gain far from the equilibrium, so as to obtain a fast convergence, regarding the value of cornering stiffness, the I&I and PBC-
and a reduced gain near to the desired state to prevent an PI controllers are able to follow the path with acceptable errors
oscillatory behaviour. in these test conditions (see Fig. 12. We assume that for safe
This test therefore shows the good performances of the three driving the maximum error should not exceed +/−25 cm).
control strategies during normal driving, at high, varying speed The SMC controller has small errors, similar to nominal
and with varying curvature, when nominal parameters are known. conditions where curvature is constant. We observe a peak in
The I&I and the adaptive PBCi-PI controllers provide equiva- the lateral error around t = 25 s. This is due to a large variation
lent performances while having a smoother steering angle. of the curvature with the vehicle at a high speed (see Fig. 9(a)).
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

TABLE III
C OMPARISON OF THE C ONTROL S TRATEGIES

uncertainties. The maximum error of the controller I&I de-


Fig. 13. Robustness against uncertainties of the vehicle mass. pends on the level of uncertainty but remains lower than the
maximum error of the SMC. The PBC-PI controller is robust
This allows us to test the robustness of the controller against to parametric uncertainties. The error remains the same in the
abrupt changes of curvature at variable high speed. The SMC presence of uncertainties. This is explained by the fact that the
controller is very sensitive to this abrupt change of curvature control law of this controller does not depend directly on the ve-
of the road. This is explained by the fact that the behaviour hicle parameters. Indeed, this control input depends only on
of the closed-loop system with SMC can be described by two passive outputs that are not dependent on uncertain parameters.
phases [37]: This controller is almost insensitive to parametric uncertainty.
Finally, it is well known that sliding mode control is unaf-
• The convergence phase: this phase corresponds to the time fected by disturbances if the matching condition is satisfied. In
interval during which the system state trajectories are not case the vehicle’s lateral dynamics do not satisfy this condition,
on the sliding surface. In this phase, the system is sensitive the behaviour of the system is not only governed by sliding
to changes in parameters. surface, but also determined by mismatched disturbances. It is
• The sliding phase: this phase corresponds to the time important to underline the fact that this drawback of SMC also
interval during which the state trajectories are confined shows the benefit provided by the I&I and PBC-PI controllers.
throughout sliding surface. The sliding system has the Even when the conditions of the SMC are respected as it is
property of insensitivity with respect to disturbances. the case in this validation (small and almost constant errors
of the SMC controller during the steady state), the I&I and
In fact, the system with the SMC controller is insensitive to
the PBC-PI approaches provide several improvements. They
disturbances only in sliding phase, but remains sensitive for the
provide smaller transient errors and a smoother steering an-
transient phase, i.e. before the sliding surface is reached. This
gle. With this new controllers (I&I and PBC-PI), we obtain
is the reason why the behaviour of the system with the SMC
performances that are similar to or better than SMC without
controller is not predictable for the transient phase and depends
chattering. In addition, we obtain a kind of robust stability of the
on the level of disturbances and uncertainties. However, during
controlled system that is required for high speed safe driving.
steady state this controller provides a lower error. With the I&I
Table III summarizes the advantages and disadvantages of each
controller, the error depends on the value of parametric uncer-
control strategy.
tainty and remains acceptable even for large variations. We note
that large variations in the parameters diminish performances,
VII. C ONCLUSION
but the stability is maintained.
We obtain similar results concerning the robustness with In this paper, an analysis of commonality of three nonlinear
respect to the uncertainties on the mass of the vehicle (Fig. 13). controllers has been done for the application of lateral vehicle
The maximum error with the SMC controller increases sig- control. The first controller is based on the sliding mode (SMC
nificantly in the presence of parametric uncertainties. This is controller). To improve the performances of this last, we have
explained by the fact that during the convergence phase, the used the principle of immersion and invariance which is more
sliding mode controller is sensitive to disturbances and param- general; it allows to have a greater degree of freedom in the
etric uncertainties. However, in steady state (sliding phase) the choice of the target dynamics and in the choice of law of conver-
error is almost nil and the controller is robust to parametric gence towards this latter. The development of this controller has
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TAGNE et al.: ROBUST NONLINEAR CONTROLLERS FOR LATERAL DYNAMICS OF INTELLIGENT VEHICLES 13

allowed us to obtain a robust stability for all the positive gains of [6] R. Marino, S. Scalzi, and M. Netto, “Nested PID steering control for lane
the controller. This strong result of stability led us to study the keeping in autonomous vehicles,” Control Eng. Practice, vol. 19, no. 12,
pp. 1459–1467, Dec. 2011.
passivity properties of the system, showing that the variable z = [7] R. Rajamani, Vehicle Dynamics and Control. Berlin, Germany: Springer-
ė + λe is a passive output for a steering input δ̃. Then we pro- Verlag, 2006.
posed an adaptive PI with non-linear gains based on this struc- [8] D. Kim, J. Kang, and K. Yi, “Control strategy for high-speed au-
tonomous driving in structured road,” in Proc. Int. IEEE ITSC, 2011,
tural property of the model. Developed controllers had similar pp. 186–191.
mathematical structure and have been compared to highlight [9] S. Hima, B. Lusseti, B. Vanholme, S. Glaser, and S. Mammar, “Trajectory
the advantages and disadvantages of each control approach. tracking for highly automated passenger vehicles,” in Proc. IFAC World
Congr., Milano, Italy, 2011, pp. 12 958–12 963.
The simulation environment was used to assess the robustness [10] A. Benine-Neto, S. Scalzi, S. Mammar, and M. Netto, “Dynamic con-
of the controllers in several scenarios representing different troller for lane keeping and obstacle avoidance assistance system,”
driving situations: constant/variable speed, curvature, strong in Proc. Int. IEEE Conf. IV Symp., 2010, pp. 1363–1368.
[11] M. Netto, S. Chaib, and S. Mammar, “Lateral adaptive control for vehicle
non-linearity. The robustness of the controllers has also been lane keeping,” in Proc. IEEE ACC, 2004, vol. 3, pp. 2693–2698.
assessed against the parametric uncertainties of the vehicle. [12] J. Naranjo, C. Gonzalez, R. Garcia, and T. de Pedro, “Lane-change
In the different tests performed, it appears that the three non- fuzzy control in autonomous vehicles for the overtaking maneuver,” IEEE
Trans. Intell. Transp. Syst., vol. 9, no. 3, pp. 438–450, Sep. 2008.
linear controllers (SMC, I&I and PBC-PI) ensure a robust [13] E. Onieva et al., “Automatic lateral control for unmanned vehicles via
tracking of the reference trajectory. The three controllers are genetic algorithms,” Appl. Soft Comput., vol. 11, no. 1, pp. 1303–1309,
robust to variations in speed and curvature. The SMC provides Jan. 2011.
[14] B. Soualmi, C. Sentouh, J. C. Popieul, and S. Debernard, “Fuzzy Takagi-
low steady state error, but it is quite sensitive in transient period Sugeno LQ controller for lateral control assistance of a vehicle,” in Proc.
causing significant peaks during rapid changes of curvature in Int. IEEE Conf. IV Symp., Alcalá de Henares, Spain, 2012, pp. 377–382.
presence of disturbances or parametric uncertainties. The con- [15] L. Nehaoua and L. Nouvelière, “Backstepping based approach for the
combined longitudinal-lateral vehicle control,” in Proc. Int. IEEE Conf. IV
troller I&I allows a better consideration of the model during Symp., Alcalá de Henares, Spain, 2012, pp. 395–400.
the controller design and provides good performances. This [16] J. Levinson et al., “Towards fully autonomous driving: Systems and al-
controller provides an improvement of performances compared gorithms,” in Proc. Int. IEEE Conf. IV Symp., Baden-Baden, Germany,
2011, pp. 163–168.
to the sliding mode for lateral control of intelligent vehicles. [17] P. Falcone, F. Borrelli, H. E. Tseng, J. Asgari, and D. Hrovat, “Linear
The PBC-PI controller provides a robust control with respect time varying model predictive control and its application to active steering
to parametric uncertainties. With its adaptive gains, it achieves systems: Stability analysis and experimental validation,” Int. J. Robust
Nonlinear Control, vol. 18, no. 8, pp. 862–875, May 2008.
very good performances. Results show that a greater consid- [18] J. Ackermann, J. Guldner, W. Sienel, R. Steinhauser, and V. Utkin, “Lin-
eration of structural features of the model during the design ear and nonlinear controller design for robust automatic steering,” IEEE
of a controller can considerably improve the robustness of the Trans. Control Syst. Technol., vol. 3, no. 1, pp. 132–143, Mar. 1995.
[19] P. Hingwe and M. Tomizuka, “Experimental evaluation of a chatter free
controller for autonomous driving applications. sliding mode control for lateral control in AHS,” in Proc. IEEE ACC,
It is important to stress that the proved results of passivity 1997, pp. 3365–3369.
are largely beyond the scope of the PBC-PI controller proposed [20] J. He, D. A. Crolla, M. C. Levesley, and W. J. Manning, “Coordination of
active steering, driveline, and braking for integrated vehicle dynamics
in the paper. As these interesting properties are proved, they control,” J. Automobile Eng., vol. 220, no. 10, pp. 1401–1420, Jan. 2006.
can be used in general to design more efficient model-based [21] G. Tagne, R. Talj, and A. Charara, “Higher-order sliding mode control for
controllers. We are studying the properties of passivity of lateral dynamics of autonomous vehicles, with Experimental validation,”
in Proc. Int. IEEE Conf. IV Symp., Gold Coast, Qld., Australia, 2013,
the lateral dynamics using a non-linear model of tire contact pp. 678–683.
forces in order to extend these results to vehicle stabilization [22] S. Chaib, M. Netto, and S. Mammar, “H inf, adaptive, PID and fuzzy
applications (non-linear region). control: A comparison of controllers for vehicle lane keeping,” in Proc.
Int. IEEE IV Symp., 2004, pp. 139–144.
We would also like to extend this comparison to other robust [23] J. Snider, “Automatic steering methods for autonomous automobile path
control strategies such as the H∞ command and the model tracking,” Carnegie Mellon Univ., Pittsburgh, PA, USA, Tech. Rep., 2009.
predictive control. [24] D. Heß, M. Althoff, and T. Sattel, “Comparison of trajectory tracking
controllers for emergency situations,” in Proc. Int. IEEE Conf. IV Symp.,
Gold Coast, Qld., Australia, 2013, pp. 163–170.
ACKNOWLEDGMENT [25] L. Menhour, D. Koenig, and B. Andr, “Continuous-time and discrete-
time switched H∞ state feedback controllers: Application for a robust
The authors would like to thank R. Lozano for their exchange steering vehicle control,” in Proc. IEEE ECC, Zurich, Switzerland, 2013,
pp. 986–991.
about the theory of passivity, which was a great help to them. [26] G. Tagne, R. Talj, and A. Charara, “Immersion and invariance control for
reference trajectory tracking of autonomous vehicles,” in Proc. IEEE Int.
R EFERENCES ITSC, The Hague, The Netherlands, 2013, pp. 2322–2328.
[27] G. Tagne, R. Talj, and A. Charara, “Passivity analysis and design
[1] B. Siciliano, O. Khatib, and F. Groen, The DARPA Urban Challenge: of passivity-based controllers for trajectory tracking at high speed of
Autonomous Vehicles in City Traffic, vol. 56. Berlin, Germany: Springer- autonomous vehicles,” in Proc. Int. IEEE Conf. IV Symp., Dearborn,
Verlag, 2009. Michigan, 2014, pp. 1151–1156.
[2] J. Wei et al., “Towards a viable autonomous driving research platform,” [28] L. Li and F.-Y. Wang, Advanced Motion Control and Sensing for Intelli-
in Proc. Int. IEEE Conf. IV Symp., 2013, pp. 763–770. gent Vehicles. Berlin, Germany: Springer-Verlag, 2007.
[3] M. Bertozzi et al., “VIAC: An out of ordinary experiment,” in Proc. Int. [29] Scaner_url, 2014. [Online]. Available: http://www.sera-cd.com, http://
IEEE Conf. IV Symp., Baden-Baden, Germany, 2011, pp. 175–180. www.scanersimulation.com
[4] A. Broggi, M. Bertozzi, and A. Fascioli, “The ARGO autonomous vehi- [30] A. Astolfi, D. Karagiannis, and R. Ortega, Nonlinear and Adaptive Con-
cle’s vision and control systems,” Int. J. Intell. Control Syst., vol. 3, no. 4, trol With Applications. London, Germany: Springer-Verlag, 2008.
pp. 409–441, 1999. [31] J. Guldner, H.-S. Tan, and S. Patwardhan, “On fundamental issues of
[5] P. Zhao, J. Chen, T. Mei, and H. Liang, “Dynamic motion planning vehicle steering control for highway automation,” Res. Rep., California
for autonomous vehicle in unknown environments,” in Proc. Int. IEEE Partners Adv. Transit Highways (PATH), Inst. Transp. Studies, Univ.
Conf. IV Symp., 2011, pp. 284–289. California, Berkeley (UCB), Berkeley, CA, USA, Tech. Rep. Mar. 1997.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

[32] P. Kokotovic and R. Marino, “On vanishing stability regions in nonlinear Reine Talj received the Electrical Engineer degree
systems with high-gain feedback,” IEEE Trans. Autom. Control, vol. 31, from Lebanese University, Lebanon, in 2005; the
no. 10, pp. 967–970, Oct. 1986. Master Research degree in automatic signal process-
[33] R. Ortega and M. Spong, “Adaptive motion control of rigid robots: A ing from the University of Paris-Sud XI and Supelec
tutorial,” Automatica, vol. 25, no. 6, pp. 877–888, 1989. Gif-sur-Yvette, France, in 2006; and the Ph.D. de-
[34] R. Lozano, B. Brogliato, O. Egeland, and B. Maschke, Dissipative Systems gree in modeling and nonlinear control of the air
Analysis and Control: Theory and Applications. London, Germany: system of fuel cells from Université de Paris-Sud XI,
Springer-Verlag, 2000. France, in 2009. Since 2010, she has been a Re-
[35] U. Kiencke and L. Nielsen, Automotive Control Systems. Berlin, searcher with HEUDIASYC Laboratory, University
Germany: Springer-Verlag, 2005. of Technology of Compiègne, Compiègne, France.
[36] M. Doumiati, A. Charara, A. Victorino, and D. Lechner, Vehicle Dy- Her research interests include control and trajectory
namics Estimation using Kalman Filtering: Experimental Validation. planning of intelligent vehicles.
New York, NY, USA: Wiley, 2012.
[37] V. Utkin, Sliding Modes in Control and Optimization. Berlin, Germany:
Springer-Verlag, 1992. Ali Charara received the Electrical Engineer degree
from Lebanese University, Lebanon, in 1987; the
Master Research degree in automatic signal process-
ing from Grenoble Institute of Technology, Grenoble,
Gilles Tagne received the Engineering degree and France, in 1988; and the Ph.D. degree in automatic
the M.Sc. degree in control engineering and indus- control from Université de Savoie, Annecy, France,
trial computing from Ecole Nationale Supérieure
in 1992. Since 1992, he has been an Assistant Pro-
d’Ingénieurs Sud Alsace (ENSISA), France, in 2011 fessor and subsequently a Full Professor with the
and the Ph.D. degree in automation and Robotics Department of Information Processing Engineering,
from Université de technologie de Compiègne (UTC), Université de technologie de Compiègne (UTC),
France, in 2014. His research interests include ro- Compiègne, France, where he has also been the Di-
botics, particularly the analysis, task planning, and rector of the Heudiasyc Laboratory (joint Laboratory CNRS/UTC) since 2008.
control of complex systems. His current research interests include intelligent vehicles, driving assistance
systems, state observers, and diagnosis of electromechanical systems.

You might also like