1540 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO.

4, MAY 2013

Vehicle Optimal Torque Vectoring Using

State-Derivative Feedback and
Linear Matrix Inequality
Saber Fallah, Amir Khajepour, Barış Fidan, Shih-Ken Chen, and Bakhtiar Litkouhi

Abstract—A controller assistant system is developed based on to the dynamics of the vehicle. This additional yaw moment is
the closed-form solution of an offline optimization problem for obtained from the differential driving/braking forces between
a four-wheel-drive front-wheel-steerable vehicle. The objective of the left and right wheels of a vehicle [1]. With the development
the controller is to adjust the actual vehicle attitude and motion
according to the driver’s manipulating commands. The controller of electric and hybrid electric vehicles, the adoption of such
takes feedback from acceleration signals, and the imposed condi- advanced stability control technologies is considerably possible
tions and limitations on the controller are studied through the con- because of the electric drive motors in such vehicles.
cept of state-derivative feedback control systems. The controller Different control algorithms for the aforementioned active
gains are optimized using linear matrix inequality (LMI) and ge- systems are available in the literature. For example, the sliding
netic algorithm (GA) techniques. Reference signals are calculated
using a driver command interpreter module (DCIM) to accurately mode control theory was used in [2] and [3] for the tracking
interpret the driver’s intentions for vehicle motion and to allow of a target yaw rate depending on the driving situation. To
the controller to generate proper control actions. It is shown that regulate the yaw rate, the brake forces were properly distributed
the controller effectively enhances the handling performance and to the wheels with a control system proposed in [2], whereas
stability of the vehicle under different road conditions and driv- the variable torque distribution technique for an all-wheel-drive
ing scenarios. Although controller performance is studied for a
four-wheel-drive front-wheel-steerable vehicle, the algorithm can vehicle was used in [3]. Using the linear–quadratic Gaussian
also be applied to other vehicle configurations with slight changes. control, a vehicle-handling assistant control system was devel-
oped in [1]. The controller utilized an independent axle torque
Index Terms—Linear matrix inequality (LMI), optimal control,
state-derivative feedback, torque vectoring. biasing technology to achieve a regulatory torque distribution.
In [4], an active steering system using the concept of potential
I. I NTRODUCTION field functions and constructive Lyapunov analysis is designed.
In [5]–[7], an adaptive technique for vehicle stabilization is

A CTIVE stability systems play important roles in enhanc-

ing stability and handling performance of cars and ve-
hicles. The most important stability systems integrated into
used. A combined four-wheel steering and torque vectoring sys-
tem was used in [8]–[11] to improve the vehicle performance
globally. In [12], an optimized control methodology for torque
modern vehicles are active steering and active differential sys- vectoring using tire slip ratio adjustments in braking situations
tems. Active steering systems help drivers maintain the vehicle is proposed. It was concluded that the controller could not only
beyond its stability limits during critical handling situations by control a vehicle during emergency situations but also under
changing the steering angle based on the expected and actual normal driving situations.
vehicle response. This paper proposes an innovative controller based on the
Active differential actuation technologies such as torque vec- solution of an offline optimization problem. The controller
toring can effectively enhance vehicle handling performance takes feedback from acceleration signals instead of velocities.
when tire forces approach to nonlinear regions. The concept of Since the controller uses acceleration signals as the feedback,
these control systems is to provide an additional yaw moment it is categorized as a state-derivative feedback system with
which the controller is studied. Additionally, the linear matrix
Manuscript received October 26, 2011; revised March 1, 2012, July 11, inequality (LMI) technique is used to formulate the stability
2012, and October 5, 2012; accepted October 16, 2012. Date of publication constraints to study the system robustness and to optimize the
December 11, 2012; date of current version May 8, 2013. This work was
supported in part by Automotive Partnership Canada, by the Ontario Research
controller gains for different road surface conditions. Since the
Fund, and by General Motors. The review of this paper was coordinated by control actions are optimized offline, the controller is highly
Prof. J. Wang. computationally efficient and is proper for real-time implemen-
S. Fallah is with the Department of Mechanical Engineering Sciences,
University of Surrey, Surrey GU2 7XH, U.K. (e-mail: s.fallah@ surrey.ac.uk).
tation. The proposed controller is developed to ensure that the
A. Khajepour and B. Fidan are with the Department of Mechanical and actual vehicle’s attitude and motion is adjusted according to
Mechatronics, University of Waterloo, Waterloo, ON N2L 3G1, Canada the driver commands. Thus, to allow the controller to generate
(e-mail: [email protected]; [email protected]).
S.-K. Chen and B. Litkouhi are with the Global Research and Development proper control actions, it is necessary to accurately interpret
Center, General Motors Company, Warren, MI 48090-9055 USA (e-mail: the driver’s intention for the vehicle motion. Hence, a driver
[email protected]; [email protected]). command interpreter module (DCIM) is developed for this spe-
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. cific purpose. The controller performance in improving vehicle
Digital Object Identifier 10.1109/TVT.2012.2232947 handling and stability is investigated using the CarSim vehicle

0018-9545/$31.00 © 2012 IEEE

Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.
FALLAH et al.: VEHICLE OPTIMAL TORQUE VECTORING USING STATE-DERIVATIVE FEEDBACK AND LMI 1541

Fig. 2. Overall structure of the controlled vehicle system

Fig. 1. Vehicle motion trajectory.

simulation package. Generally speaking, the main contributions

of this paper include controller formulation, gain optimization,
and development of DCIM.
Fig. 3. CG and tire forces.
The rest of this paper is organized as follows. The controller
formulation procedure is described in Section II. The stability The errors between the desired and actual forces are considered
of the closed-loop system is studied through the concept of as the input of the controller. The controller generates proper
the state-derivative feedback systems in Section III. Section IV adjustment of tire forces to assist the driver in handling the
deals with the optimization formulation for controller gain vehicle.
tuning using LMI and genetic algorithm (GA) optimization In Fig. 2, Q and δ represents the driver torque and wheel
solvers. The controller gains are optimized using a bicycle vehi- steering angle inputs. F ∗ is the vector of desired CG forces.
cle model in Section V, whereas Section VI is devoted to devel- f is the tire forces vector, whereas δf represents the vector
opment of a module that functions to accurately interpret driver of control actions. μ is the road surface coefficient, and Rt
commands. In Section VII, the accuracy of the DCIM and the is the effective tire radius. E is the error between the desired
contribution of the controller in enhancing of vehicle handling and actual CG forces. M represents the vehicle mass matrix.
performance are verified through CarSim simulations. Finally, The measured output signals are the vehicle CG accelerations.
Section VIII features the paper’s summary and conclusion. The details of the controller formulation and overall control
structure are described in the subsequent sections.
II. C ONTROL F ORMULATION
A driver controls the vehicle motion using the steering wheel, A. Problem Definition
gas, and brake pedals (torque) based on his/her interpretation The desired CG forces (see Fig. 1), which have been ob-
of the road conditions. However, as Fig. 1 demonstrates, the tained from driver’s inputs (i.e., steering wheel angle and
driver may encounter difficulty controlling the vehicle or, in driving/braking torque values), are defined by
some cases, may experience loss of control due to unexpected  T
road and driving conditions (e.g., poor road conditions). Typ- F ∗ = Fx∗ , Fy∗ , G∗z (1)
ically, the primary objective of a vehicle control system is to
provide the driver with a normal driving experience in the case where, Fx∗ , Fy∗ , and G∗z are the desired CG longitudinal force,
of unexpected driving conditions and to enhance the vehicle lateral force, and yaw moment, respectively.
stability range. Likewise, the purpose of the proposed controller The actual forces and moment acting on the CG of the
in this paper is to assist the driver in handling the vehicle in such vehicle represent the vehicle motion and are mainly direct
situations. resultants of the tire forces (see Fig. 3). Therefore, each CG
The vehicle motion is mainly a function of tire contact forces force components is a function of all tire forces as
and can be controlled by properly adjusting these forces. Hence,
F = [Fx , Fy , Gz ]T (2)
the proposed control methodology in this paper uses the tire
forces as the control actions. where
Fig. 2 shows the overall structure of the proposed control
system. In this structure, a module, i.e., the DCIM, calculates Fx = Fx (fx1 , . . . , fx4 , fy1 , . . . , fy4 ) (3)
the desired forces at the vehicle’s center of gravity (CG) based Fy = Fy (fx1 , . . . , fx4 , fy1 , . . . , fy4 ) (4)
on driver inputs and according to normal driving conditions. Gz = Gz (fx1 , . . . , fx4 , fy1 , . . . , fy4 ). (5)
Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.
1542 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO. 4, MAY 2013

Here, Fx , Fy , and Gz are the actual longitudinal force, lateral where WE and Wdf are semipositive definite diagonal weight
force, and yaw moment acting on the CG of the vehicle. fxi , i = matrices on the CG force error vector and the control action,
1, . . . , 4 is the longitudinal tire force on each tire of the vehicle, respectively, and are defined by
whereas fyi , i = 1, . . . , 4 is the lateral tire force on each tire of ⎡ ⎤
the vehicle. Equations (3)–(5) reveal a possible way to achieve wx 0 0
the desired CG forces by controlling the tire forces. WE = ⎣ 0 w y 0 ⎦ (14)
Define the tire force vector as 0 0 wG
⎡ ⎤
f = [fx1 , fx2 , fx3 , fx4 , fy1 , fy2 fy3 , fy4 ]T . (6) wdf1 0 0 0
⎢ 0 w df2 0 0 ⎥
Wdf = ⎣ ⎦ (15)
The corresponding adjusted CG forces to make the error 0 0 wdf3 0
between the desired F ∗ and F (f ) zero can be represented by 0 0 0 wdf4

F (f + δf ) ≈ F (f ) + ∇F (f )δf. (7) where, wx , wy , and wG are weights on the longitudinal, lateral,

and yaw moment errors. wdfi is the weight on the control action.
The vector of control actions required to reduce the error The objective function in (13) is used to minimize the CG force
between F ∗ and F (f ) is defined by error while minimizing the control action.
Since WE and Wdf are semipositive definite, the quadratic
δf = [δfx1 , δfx2 , δfx3 , δfx4 , δfy1 , δfy2 , δfy3 , δfy4 ]T . (8)
objective function (13) is also positive definite. Thus, P is
In practice, the control actions for δfxi are provided by minimum with respect to δf when
torque vectoring, whereas the control actions δfyi are provided
∂P  
by active steering. In the current controller design, we limit = Wdf +(∇F T WE )∇F δf −(∇F T WE )E = 0. (16)
ourselves to torque vectoring; hence, there is no adjustment to ∂δf
δfyi or
The minimization of the objective function P results in
δfyi = 0, i = 1, . . . , 4. (9)
 −1
δf = Wdf + (∇F T WE )∇F (∇F T WE )E. (17)
Therefore, in (8), it is assumed that there is no means to directly
adjust tire lateral forces.
To obtain the control actions δf , let us define the error Accordingly, the controller gain K is
between the desired CG forces at t + Δt and actual CG forces  −1
at t as K = Wdf + (∇F T WE ∇F ) (∇F T WE ). (18)

E = F ∗ − F (f ) = [Ex , Ey , Ez ]T . (10) The proposed controller has a closed-form solution for an

offline optimization problem, thereby making its real-time im-
The control actions δf for time t + Δt can be then calculated
plementation more effective than model-based controllers avail-
from F ∗ − F (f + δf ), where F (f + δf ) is the CG forces at
able in literature. It is noted that the actual CG forces can be
t + Δt. Using (7), (9), and (10), one has
calculated by multiplying the acceleration signals to the mass
F ∗ −F (f +δf ) ≈ F ∗ −F (f )−∇F (f )δf = E −∇F (f )δF (11) matrix. Thus, the controller feedback is acceleration signals
instead of velocities. In practice, the acceleration signals are
where measured directly by accelerometers or inertial measurement
⎡ ∂Fx (f ) ⎤ ⎡ ⎤ units.
∂Fx ∂Fx ∂Fx ∂Fx
∂f ∂fx1 ∂fx2 ∂fx3 ∂fx4 Remark: Acceleration sensors are known to be noisy; hence,
⎢ ∂Fy (f ) ⎥ ⎢ ∂Fy ∂Fy ∂Fy ∂Fy ⎥
∇F = ⎣ ∂f ⎦=⎣ ∂fx1 ∂fx2 ∂fx3 ∂fx4 ⎦. (12) some kind of filters will be required to smooth the signals.
∂Gz (f ) ∂Gz ∂Gz ∂Gz ∂Gz However, the design of such filters is beyond the scope of this
∂f ∂fx1 ∂fx2 ∂fx3 ∂fx4
paper.
In the following, a quadratic objective function is defined to Remark: Since there are no conditions on the tire forces
optimally find δf to make the error zero. during the optimization procedure of the controller formulation,
the proposed controller is effective as long as the tire forces are
inside the friction ellipse. The controller will be more effective
B. Optimization if it is combined with other safety controllers such as antilock
To formulate the required δf , an objective function that braking systems.
includes the weighted combination error between actual and Since the controller takes feedback from acceleration signals,
desired vehicle CG forces and control action is defined. The the controller is classified as a state-derivative feedback control
mathematical representation of the objective function is system. Some theoretical aspects of state-derivative feedback
systems and related control design considerations may be found
1  in [14]–[17]. The properties and stability of the closed-loop part
P= (E −∇F δf )T WE (E −∇F δf )+δf T Wdf δf (13)
2 of the proposed controller are discussed in the following.
Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.
FALLAH et al.: VEHICLE OPTIMAL TORQUE VECTORING USING STATE-DERIVATIVE FEEDBACK AND LMI 1543

Proposition: Assuming (I + BKM ) and A are invertible,

system (21) is L2e input–output stable if and only if (I +
BKM )−1 A is Hurwitz.
Proof: Consider the following Lyapunov function for sys-
tem (21):
1 T
V = x x ⇒ V̇ = y T x. (23)
2
Fig. 4. Block diagram of the closed-loop control system.
Since (I + BKM ) and A are invertible, system (21) can be
III. C LOSED -L OOP A NALYSIS OF rewritten as (see [20] for details)
THE P ROPOSED C ONTROLLER
x(t) = A−1 (I +BKM )ẋ(t) − (A−1 BK)R(t) − (A−1 B)f (t).
The nonlinear representation of the closed-loop part of the (24)
system (see Fig. 4) is
By defining B = [(A−1 BK)(1/Rt )(A−1 B)], and u =
M ẋ = g(x, δ) + h(uc , f, δ) (19) R(t)
, (24) can be reshaped to
f (t)

where x is the state vector, and uc shows the controller signals. x(t) = A−1 (I + BKM )ẋ(t) − B̄ ū. (25)
By defining the state vector as x(t) = [vx , vy , r]T , the lin-
earized dynamics of the vehicle system around an operating Substituting (25) in (23), one obtains
point with a given wheel steering angle can be represented as  
V̇ = y T x = y T A−1 (I + BKM )y − B̄ ū
follows:
 
= y T A−1 (I + BKM ) y − y T B̄ ū. (26)
M ẋ = Av x + Bv (uc + f )
y = ẋ Since (I + BKM )−1 A is Hurwitz, A−1 (I + BKM ) is also
Hurwitz [21]. Therefore, H1 = A−1 (I + BKM ) < 0.
uc = K(R − M y) (20) By defining H1 = −H1 , H2 = −B̄u, and c ≥ 0, one has
λi (H1 ) = −λi (H1 ), where λi stands for the eigenvalue of
where Av is the linearized model of the vehicle, and Bv is
the matrix H1 . In addition, a = min(λi (H1 )) ≥ 0 and b =
the input matrix. y = [v̇x v̇y ṙ]T is the vector of the output
H2  ≥ 0. Then
(measured) signals, and R = F ast is the vector of the reference
inputs. Rt is the wheel radius. Thus, using (20), the dynamics V̇ = − y T H̃1 y + y T H̃2 ū ≤ −y T H̃1 y + y T H̃2 ū + cūT ū
of the closed-loop vehicle system becomes
≤ − ay22 + bū2 y2 + cū22
−1 −1
ẋ(t) = (I + BKM ) Ax(t) + (I + BKM ) (BK)R(t) 1 b2 a
=− (bū2 − ay2 )2 + ū22 − y22 + cū22
+ (I + BKM )−1 Bf (t) (21) 2a 2a 2

b2 + 2ac a
where A = M −1 Av , and B = M −1 Bv . ≤ ū22 − y22 ⇒ V̇
2a 2
The conditions given in the following may apply to the
2

aforementioned dynamic system for stability and performance b + 2ac a
analysis [18], [19]: ≤ ū22 − y22 . (27)
2a 2
1) The closed-loop system is well posed if (I + BKM ) is
nonsingular (det(I + BKM ) = 0). Integrating over [0, τ ] one has
2) The pair of (A, B) is controllable.  τ  τ
b2 +2ac
3) If det(A) = 0, then the system cannot be stabilize. V̇ dt ≤ ū(t)22 dt
0 2a 0
4) If (I + BKM )−1 A is Hurwitz, the linear system is 
asymptotically stable. a τ
− y(t)22 dtV (x(τ ))−V (x(0))
First, condition 1 for the controller is investigated. Since Bv 2 0

in (20) is the input matrix related to the total longitudinal forces, b2 +2ac a
it is equivalent to ∇F . Thus, according to (18), one has < (ūτ L2e )2 − (yτ L2e )2 . (28)
2a 2
 T
BKM = ∇F Wdf + (∇F T WE )∇F (∇F T WE )M. With consideration of V (x(τ )) > 0 and taking the square
(22) roots, one arrives at

Since WE , Wdf , and M are semipositive definite diagonal b2 + 2ac 
matrices, BKM is also semipositive definite. Accordingly, yτ L2e < ūτ L2e + 2V (x(0)) /a (29)
a
(I + BKM ) is strictly positive definite, and the matrix is
not singular. Conditions 3 and 4 are studied by the following According to the definition of L2e stability given in [22],
proposition. system (21) is L2e stable.
Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.
1544 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO. 4, MAY 2013

IV. ROBUSTNESS AND G AIN O PTIMIZATION By defining θ as the vector of the uncertain and time-varying
parameters, dynamic system (21) may be rewritten as
LMI is a powerful approach that solves control problems by
defining a convex constraint on the state vector x. Here, a stabil- ẋ(t) = Ã(θ)x(t) + B̃(θ)R (39)
ity constraint is formulated for the controller gain optimization
using the LMI technique and the concept of L2 gain theory for where
linear systems based on the method described in [23]. Since the
reference trajectory performance is the primary concern of this Ã(θ) = [I + B(θ)KM ]−1 A(θ) (40)
analysis, it is assumed that the driver torque input vector is zero B̃(θ) = [I + B(θ)KM ]−1 B(θ)K. (41)
(f = 0), according to the superposition principle.
To analyze this type of uncertainty, the polytopic technique
The L2 gain of the dynamic system (21) is defined as the
is used as described in the following.
following quantity:
Polytopic Technique: The linear differential equation (21)
y2 belongs to a subset of Rn , i.e., Ω, as follows:
sup . (30)
R2 =0 R2
ẋ ∈ Ωx. (42)
The supremum is taken over all nonzero trajectories of the When Ω is a polytope, viz., the convex hull of a finite set of
system (21), starting from x(0) = 0. Now, suppose there is a model parameters (vertices), the linear differential equation is
quadratic function V (ς) = ς T P ς and γ ≥ 0, such that for all t, called polytopic [24], [25]. In this case, we have
we have    
d Ã(θ), B̃(θ) ∈ Co (Ã1 , B̃1 ), . . . , (ÃN , B̃N ) (43)
V (x) + y T y − y 2 RT R ≤ 0. (31)
dt
where Co{(Ã1 , B̃1 ), . . . , (ÃN , B̃N )} is the convex hull of
Then, the L2 gain of the linear system from output y to the (Ãi , B̃i ), i = 1, . . . , N .
input R is less than γ [23]. Definition: For any subset S of a linear vector space X, the
By defining convex hull Co(S) is convex and consists precisely of all convex
combinations of the elements S.
à = [I + BKM ]−1 A (32)
Thus, it is possible to write (Ã(θ), B̃(θ)) as a convex combi-
−1
B̃ = [I + BKM ] BK (33) nation of the vertices of the polytope as follows:
 
then one has Ã(θ), B̃(θ) = α1 (Ã1 , B̃1)+α2 (Ã2 , B̃2 )+· · ·+αN (ÃN + B̃N)
d (44)
V (x) = ẋT P x + xT
dt where {(Ã1 , B̃1 ), (Ã2 , B̃2 ), . . . , (ÃN , B̃N )} are known matri-
T T T T
P ẋ = x [Ã P + P Ã]x + R B̃ P x + x P B̃R T
(34) ces, and α1 , . . . , αN are positive scalars that satisfy
T T
y y = [Ãx + B̃R] [Ãx + B̃R] 
N
αi = 1. (45)
= xT ÃT Ãx + xT ÃB̃R + RT B̃ T Ãx + RT B̃ T B̃R. i=1

(35) Thus, the LMI (38) can be generalized for a polytope of

uncertainties as follows.
Substituting (34) and (35) in (31) results in
Proposition: For the parametric linear differential equation
d (39), condition (31) holds if and only if
V (x)+y T y−γ 2 RT R = xT [ÃT P +P Ã+ ÃT Ã]x+xT
dt
⎡ P > 0, γ > 0,and T ⎤
(P + Ã)T B̃R+RT B̃ T (P + Ã)x+RT (B̃ T B̃ −γ 2 )R ≤ 0. (36)
à (θ)P +P Ã(θ)+ ÃT (θ)Ã(θ)
T
P + Ã(θ) B̃(θ)
⎣   ⎦
By defining η = [x R]T , the above inequality may be B̃ T (θ) P + Ã(θ) B̃ T (θ)B̃(θ)−γ 2 I
rewritten as follows: ≤ 0. (46)
 T T T
 
T (Ã P + P Ã + Ã Ã) (P + Ã) B̃ x
[x R] ≤ 0. Inequality (46) holds if and only if [23]
B̃ T (P + Ã) B̃ T B̃ − γ 2 I R   
(37) ÃTi P + P Ãi + ÃTi Ãi (P + Ãi )T B̃i
This is satisfied if P > 0, γ > 0,
 T  B̃iT (P + Ãi ) B̃iT B̃i − γ 2 I
(Ã P + P Ã + ÃT Ã) (P + Ã)T B̃ ≤ 0 i = 1, . . . , N. (47)
≤ 0. (38)
B̃ T (P + Ã) B̃ T B̃ − γ 2 I
The smallest upper bound on the L2 gain of the dynamic
Inequality (38) has the form of an LMI problem. It is noted system can be computed by minimizing γ over the variables P
that the matrices A and B of the dynamic system (21) have and γ satisfying conditions in (46). Accordingly, the following
some uncertain parameters that vary in different situations. optimization problem can be formulated.
Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.
FALLAH et al.: VEHICLE OPTIMAL TORQUE VECTORING USING STATE-DERIVATIVE FEEDBACK AND LMI 1545

Problem: Considering control signal δf = K(R − M ẋ), needs to satisfy stability requirements. If the controller gains
the gain matrix K is found, such that the closed-loop system satisfy the stability conditions, then the minimization problem
(21) is asymptotically stable, and the L2 gain from the output (48) is solved by the LMI solver, and the resulting γ is assigned
vector y(t) to the reference input R(t) is minimized. Thus, the as the value of the cost function of the GA. Otherwise, if the
following optimization problem requires to be solved: closed-loop system is unstable or if the feasible solution of the
LMI cannot be found, a large number is considered for the cost
min γ (48) function.
For the stability condition, the closed-loop vertex subsystems
subject to and the closed-loop nominal system must be strictly Hurwitz.
   Therefore, the minimization loop is as follows [31]:
ÃTi P + P Ãi + ÃTi Ãi (P + Ãi )T B̃i
≤0
B̃iT (P + Ãi ) B̃iT B̃i − γ 2 I Call gains (wx , wy , wG ) from MATLAB GA Toolbox
If Remax {λ(ÃN )} < 0
i = 1, . . . , N
If max{Remax (λ(ÃN ))} < 0 i = 1, 2
P > 0 and γ > 1. (49) Solve the optimization problem (48)
Else
Remark: In this paper, we have tried to minimize the up- Assign a large number to the cost function
per bound on L2 gain between output and input signals to End
have a better performance. In other words, controller gains End
(wx , wy , wG ) need to be calculated such that they guarantee
the minimum value of γ. In addition, the constraint y > 1 is Here, λ is the eigenvalue function. Ãi denotes the closed-
added since the upper bound that is less than 1 can deteriorate loop vertex subsystem, whereas ÃN = (I + BΣ K1 )−1 AΣ
the performance. is the closed-loop nominal system, r=2 in which AΣ =
Since Ã1 is a function of K [see (40)], the optimization (1/r) r=2i=1 Ãi and B Σ = (1/r) i=1 B̃i stand for the
problem given earlier is bilinear due to the multiplication of nominal system matrix and the nominal input vector,
the optimization parameters P and K. Some important results respectively. The given algorithm continues until the minimum
about the solution of such control problems in special cases of the cost function γ is found, and consequently, an acceptable
are available in the literature [26]–[28]. However, there is no controller gain is achieved.
general analytical solution to this optimization problem because
the dynamics of the control system subject to polytopic uncer-
tainties cannot be reduced to a convex optimization problem V. S IMULATION
[29]. Hence, the GA is adopted to find the optimization solution
A. Bicycle Model
according to the GA stochastic search capability.
Here, a combination of the LMI and the GA is employed to For stability analysis and optimization problem, a bicycle
find the optimized controller gain by minimizing the L2 gain of model is used to accord with the objective of this paper. The
the system. The approach proposed in [30] and [31] is adopted bicycle model represents the vehicle dynamics in the linear
to find an optimized gain for the controller with the minimized region accurately and is suitable for the study of the proposed
L2 gain of the system. controller. Assuming that the tires are in their linear region, the
It is noted that if the controller gain K is known, the functions, defined in (19) for a front-wheel-steerable vehicle
inequalities in (49) are linear and can be solved easily by an bicycle model, are given in (50)–(52), shown at the bottom of
LMI solver. Moreover, controller gain K is a function of wx , the page, where i = 1 corresponds to the front wheel, whereas
wy , and wG . Thus, using the GA, a random set of controller i = 2 shows the rear wheel. ms and Iz are the mass of the
gains (wx , wy , wG ) is selected. Each set of controller gains vehicle and its yaw moment of inertial around the z axis,

⎡ ⎤
ms 0
0
M=⎣ 0 0⎦
ms (50)
⎡ 0 0
Iz    ⎤
−μC1 δ − tan−1 y vx 1
v +L r
sin(δ) + ms rvy
⎢      ⎥
⎢ v −L r ⎥
g(x, δ) = ⎢ μC1 δ − tan−1 y vx 1 cos(δ) − μC2 tan−1 y vx 2 − ms rvx ⎥
v +L r
(51)
⎣      ⎦
v −L r
μC1 δ − tan−1 y vx 1 cos(δ)L1 + μC2 tan−1 y vx 2 L2
v +L r
⎡ ⎤
(f1 cos(δ) + f2 )
h(f, δ) = μ ⎣ f1 sin(δ) ⎦ (52)
f1 sin(δ)L1
Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.
1546 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO. 4, MAY 2013

TABLE I The GA settings are as follows. The population size is set

DYNAMIC S YSTEM P ROPERTIES
equal to 50, whereas the number of generations is set equal
to 100. The crossover fraction is set at 0.9, and the Gaussian
function is chosen for mutation. Running the optimization
algorithm achieves the following results:

γ = 29.08

wx = 50, wy = 0.6, wG = 1450.

respectively. Li is the distance from the vehicle CG, and μ VI. D RIVER C OMMAND I NTERPRETTER M ODULE
represents the road surface coefficient. Ci is the tire stiffness
coefficient. The dynamic equations (50)–(52) are highly non- Usually, this module has been defined as an upper controller
linear. Thus, they are linearized around an operating point as that generates the desired signals to keep the vehicle in stable
(vxo , vyo , ro ) = (50 km/h, 0 km/h, 0.5171 rad/s) and for δ = regions. However, this concept may result in over-restricting the
5π/180 rad. This results in (53) and (54), shown at the bottom driver in emergency cases. In these cases, the controller may
of the page. keep the vehicle in the stable region at the cost of overriding
control capability on the vehicle, leading to avoidable accidents.
Furthermore, to have the vehicle stable within the limitations
B. Gain Optimization of both the vehicle and the road, the reference signals should
allow the vehicle to be reactive to the commands of the driver.
Here, the optimization problem (48) is solved using the
To address these issues, the objective of DCIM is to generate
proposed GA/LMI algorithm for the bicycle vehicle model,
target signals based on commands provided by the driver under
assuming small values for δ. Next, the optimized gains are
normal driving conditions.
applied to a full vehicle model in CarSim. In the optimization
A driver typically manipulates the vehicle through the steer-
problem (48), the surface road coefficient μ is considered to
ing wheel and the gas/brake pedal to achieve his/her intended
be an uncertain parameter. The parameters considered for the
directional motion. The objective of the controller is to provide
analysis of the system are shown in Table I.
the driver with the similar driving feel to what he/she would
The eigenvalues of the linearized system are λ = (−0.1985,
experience on dry road conditions, while tires have normal
−3.9672 + 0.3300i, −3.9672−0.3300i). Hence, det(A) = 0,
characteristics. As such, it is assumed that, in DCIM, the road
and the system is well posed and stabilizable. Additionally, pair
surface coefficient μ is equal to 1 and that the tire characteristics
(A, B) is full rank and controllable.
are not subject to aging and wearing. Efficiency and accuracy
For optimization purposes, the lower and upper bounds on
are two conflicting objectives that have to be considered when
the optimization variables are defined as
developing DCIM. The module should be computationally
0 ≤ wx ≤ 100, 0 ≤ wy ≤ wdf , 0 ≤ wG ≤ 2000. efficient enough for real-time implementation and be accurate
enough to avoid overestimating/underestimating the driver’s
Remark: The lateral motion can be modified slightly by the intent. Considering the physical limitations, a nonlinear vehicle
proposed controller since adjustments can only be made to bicycle model is utilized to map the driver manipulation signals
longitudinal forces. Thus, to reduce the transfer of the error of to controller reference signals F ∗ while maintaining reasonable
the lateral motion to other directions, it is recommended that limits to the controller reference signals. The controller gener-
0 ≤ wy ≤ wdf . This constraint is mathematically investigated ates the proper control actions based on these data, as needed.
for a bicycle vehicle system in the Appendix. Accordingly, reference signals are calculated from functions
The lower and upper bounds on wx and wG are defined given in (51) and (52) with the assumption of μ = 1 and perfect
based on the eigenvalue analysis of the system. The bounds on tire conditions as follows:
the uncertain parameter are 0.5 ≤ μ ≤ 1, whereas wdfi is set
equal to 1. F ∗ = g(x, δ)μ=1 + h(f, δ)μ=1 . (55)

⎡ μC1 δ(vyo +L1 ro ) ⎤

− 2
vxo
μC1 δ
vxo + ms ro
μC1 L1 δ
vxo + ms vyo
⎢ μ[μC1 (vyo +L1 ro )+μC2 (vyo −L2 ro )] μ(C2 L2 −C1 L1 ) ⎥
Av = ⎢ − μ(Cv1xo ⎥
+C2 )
− ms ro − ms vxo (53)
⎣ 2
vxo vxo ⎦
μ[C1 L1 (vyo +L1 ro )−C2 L2 (vyo −L2 ro )] μ(C2 L2 −C1 L1 ) μ(C1 L21 +C2 L22 )
2
vxo vxo vxo
⎡ ⎤
1 1
Bv = μ ⎣ δ 0⎦ (54)
L1 δ 0
Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.
FALLAH et al.: VEHICLE OPTIMAL TORQUE VECTORING USING STATE-DERIVATIVE FEEDBACK AND LMI 1547

For the calculation of F ∗ , the tire forces fxi and fyi are
calculated using a nonlinear combined slip tire model proposed
in [32], i.e.,

fxi = fX (Ki , Fzi ) ∗ φx (Ki , αi ) (56)

fyi = fY (αi , Fzi ) ∗ φy (Ki , αi ) (57)
Fig. 5. DCIM.
where i = 1, 2. κi and αi represents slip ratio and slip angle,
respectively, and

fX (Ki , Fzi ) = c1xi Fzi sin [c2xi tanh(c3xi ∗ Ki )] (58)

fY (αi , Fzi ) = c1yi Fzi sin [c2yi tanh(c3yi ∗ Ki )] (59)
 
φx (αi , Ki ) = 1 − λxi sec h(γxi ∗ Ki ) tanh(βxi ∗ αi ) (60)
2

 
φx (αi , Ki ) = 1 − λyi sec h(γyi ∗ Ki ) tanh(βyi ∗ αi )2 (61)
v − rωi
Ki = (62)
max(v, rωi )

v + Li r
αi = δi − tan−1 − Crolli φi (63)
u
Fig. 6. Applied torque on vehicle wheels.
mg
Fzi = . (64)
2
In the given equations, c1xi , c2xi , c3xi , c1yi , c2yi , c3yi , λxi ,
γxi , βxi , λyi , γyi , and βyi are tire constant values. Fzi is the
vertical tire force. Croll stands for the roll steering coefficient.
In (63), the slip angle relation is modified by adding the effect
of roll angle as proposed in [33]. The roll angle can be estimated
by a single degree of freedom roll dynamic model, i.e.,
1  
φ̈ = −mHRC (v̇+ru)+mgHRC φ−Kφ φ−Cφ φ̇+Ixz ṙ
Ix
(65)
where φ, HRC , Kφ , and Cφ are the roll angle, the height of the
roll center, the roll stiffness, and the roll damping, respectively. Fig. 7. Driver’s wheel steering input.
In addition, the dynamics of wheels are modeled as
value, the upper bound will be considered to be the target value.
1 Fig. 5 shows the DCIM block, whereas Fig. 2 shows the overall
ω̇i = (Qi − fxi Rt ) (66)
Iw block diagram of the system.
where, ωi is the angular velocity of the wheel, and Qi is the
applied torque on each wheel. Iw shows the wheel inertial VII. S IMULATIONS
moment. Based on the tire/road conditions, defining an upper
bound on desired forces calculated by (55) are required to avoid Here, a full-vehicle model in CarSim is used to investigate
generating an excessive side and longitudinal slips. In a steady the accuracy of DCIM, followed by a study into the perfor-
state and in stable conditions, the upper bound are defined as mance of the controlled system shown in Fig. 2 for some
follows [2], [34]: scenarios. For simulation purposes, the driver’s torque input
and steering wheel angle shown in Figs. 6 and 7 are applied.
 2  The signals include sharp wheel steering and hard braking
∗ Qi
Fx ≤ μ and accelerating conditions, which are important for control
i=1
Ri
performance evaluation. In addition, it is assumed that the
Fy∗ ≤ μmg initial velocity is 100 km/h.
μg
G∗z ≤ − .
vx A. DCIM Verification
It is noted that road surface coefficient μ needs to be es- The function of DCIM is to provide reasonably accurate
timated from an estimation module. However, its estimation estimation of CG forces to the controller subject to perfect
procedure is beyond the scope of this paper. In cases where any road/tire contact and stable conditions. Assuming dry road
of the target signals calculated by (55) are greater than the upper conditions, the CG forces obtained from CarSim are compared
Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.
1548 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO. 4, MAY 2013

Fig. 8. Longitudinal CG forces obtained from CarSim and DCIM.

Fig. 11. Target path following for μ = 0.75.

Fig. 9. Lateral CG forces obtained from CarSim and DCIM.

Fig. 12. Target path following for μ = 0.5.

Fig. 10. Yaw moment obtained from CarSim and DCIM.

Fig. 13. Control actions (torque adjustments) on front wheels.
with those estimated by DCIM and plotted through Figs. 8–10.
As shown, DCIM can accurately estimate the forces. Since the As the figures show, the controller significantly improves the
aerodynamic and rolling resistance forces are ignored in DCIM, handling performance of the vehicle in both cases. However,
there is a small discrepancy between the longitudinal force from in cases where the road surface coefficient is 0.5, the controller
CarSim and DCIM. In addition, there is a small discrepancy in stabilizes the vehicle while also assisting the driver in following
lateral forces because of the simplicity of roll angle estimation the path. To be more precise, for μ = 0.75, control actions on
by (65). These discrepancies in lateral and longitudinal forces front and rear wheels are shown in Figs. 13 and 14. Likewise,
slightly affect yaw response. the CG force errors of controlled and uncontrolled vehicle
systems are compared through Figs. 15–17. The adjustment
torque values are calculated by multiplying the control actions
B. Controller Performance Investigation
by tire radius Rt . For μ = 0.75, control actions on front and
It is assumed that the target path is the path on which the rear wheels are shown in Figs. 18 and 19. Moreover, the total
driver would like to follow on a dry road, even if the road torque values applied on each wheels in the case of μ = 0.5 are
conditions are different from the driver’s perception. Subjected shown in Figs. 20–23. As these figures show, the controller just
to the inputs shown in Figs. 6 and 7, the vehicle performances provides adjustment to the nominal torque applied by the driver.
with and without the controller are shown in Figs. 11 and Moreover, the controller significantly reduces the CG force
12 for road surface coefficients 0.75 and 0.5, respectively. errors, resulting in better handling performance and superior
Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.
FALLAH et al.: VEHICLE OPTIMAL TORQUE VECTORING USING STATE-DERIVATIVE FEEDBACK AND LMI 1549

Fig. 14. Control actions (torque adjustments) on rear wheels. Fig. 17. Error of yaw moments on CG.

Fig. 18. Control actions (torque adjustments) on front wheels.

Fig. 15. Error of longitudinal CG forces.

Fig. 19. Control actions (torque adjustments) on rear wheels.

Fig. 16. Error of lateral CG forces.
controller is independent of the vehicle model, thus providing
driving comfort. As aforementioned, the performance of the reliable real-time implementation. Additionally, the controller
controller can be improved by coordinating with other active is in the class of the state-derivative feedback systems, which
stability controllers such as active steering systems, active takes feedback from acceleration signals instead of velocities.
suspension systems, or active braking systems. The stability of the controller was studied through the concept
Remark: The controller performance is limited to actuator of the state-derivative feedback systems, whereas robustness
constraints. For example, for low surface coefficients, the con- analysis was completed through the LMI technique. Using the
troller generates the control actions with a large amount of concept of L2 gain theory, an LMI constraint was formulated
magnitudes that the actuators are not capable of generating. for controller gain optimization. Since the optimization prob-
lem was bilinear (due to the multiplication of the optimiza-
tion variables), a GA/LMI algorithm was developed to solve
VIII. C ONCLUSION
the optimization problem. The performance of the optimized
An innovative controller was developed to improve vehicle controller on improving handling performance was studied
handling performance under different road surface conditions. using a full-car vehicle model in the CarSim software. The
The control methodology was obtained by minimizing the ob- simulation results showed that the controller improves handling
jective function, which is defined based on CG force errors. The performance.
Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.
1550 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO. 4, MAY 2013

A PPENDIX
Lateral motion can be slightly modified by the proposed
controller since only the longitudinal forces can be adjusted.
To have a stable system, it is recommended that 0 ≤ wy ≤
wdf . The above constraint is mathematically investigated for a
bicycle vehicle system in the following.
Assuming the case that wy and wG are set to zero, the control
actions are
wx
δfxf = Ex (I.1)
(2wx + wdf )
wx
δfxr = Ex . (I.2)
Fig. 20. Adjusted torque on front-left wheel. (2wx + wdf )

Thus, the closed-loop dynamic equations in x- and y-directions

become
 2wx
mv̇x = fx + Ex (I.3)
(2wx + wdf )
 wx δ
mv̇y = fy + Ex (I.4)
(2wx + wdf )
 
where fx and fx are the summation of tire forces in the
x- and y-directions, respectively.
(wx /(2wx + wdf )) always has a magnitude that is smaller
than one for all values of wx while δ 1. Thus, (wx δ/(2wx +
wdf )) 1, and the effect of the transferred error Ex to the y-
Fig. 21. Adjusted torque on front-right wheel.
direction can be ignored. In contrast, if wx and wG are set to
zero, the control actions are

wy δ
δfxf = Ey (I.5)
(wy δ 2 + wdf )
δfxr = 0. (I.6)

Again, the closed-loop dynamic equations in x- and y-

directions are as follows:
 wy δ
mv̇x = fx + Ey (I.7)
(wy δ 2 + wdf )
 wy δ 2
mv̇y = fy + Ey (I.8)
(wy δ 2 + wdf )
Fig. 22. Adjusted torque on rear-left wheel.

In this case, (wx δ/(wy δ 2 + wdf )) is not smaller than 1 for

all values of wy . In other words, the transferred error Ey to the
x-direction may be magnified by some values of wy . To reduce
the effects of lateral motion error on the x-direction and to
diminish the transfer of the lateral error to the x-direction by
factor δ, a new constraint is defined as
wy
≤ 1 ⇒ wy ≤ wy δ 2 + wdf
wy δ 2 + wdf
⇒ wy (1 − δ 2 ) ≤ wdf . (I.9)

Since δ 1, one has wy ≤ wdf . Thus, to have a stable

system and reduce the transfer of the lateral error to the
Fig. 23. Adjusted torque on rear-right wheel. x-direction by factor δ, the given condition should be satisfied.
Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.
FALLAH et al.: VEHICLE OPTIMAL TORQUE VECTORING USING STATE-DERIVATIVE FEEDBACK AND LMI 1551

ACKNOWLEDGMENT [23] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix

Inequalities in System and Control Theory. Philadelphia, PA: SIAM,
The authors would like to thank Prof. K. Morris from the 1994.
Faculty of Mathematics, University of Waterloo, for her helpful [24] A. Jadbabaie, “Robust, Non-Fragile Control Synthesis Using Model-
Based Fuzzy Systems: A Linear Matrix Inequality Approach,” M.S. The-
advice on controller investigation. sis, Univ. New Mexico, Albuquerque, NM, 1997.
[25] C. Scherer and S. Weiland, Linear Matrix Inequalities in Control. Delft,
The Netherlands: Delft Center Syst. Control, Delft Univ. Technol., Com-
R EFERENCES pilation, Nov. 16, 2004.
[26] J. Daafouz, P. Riedinger, and C. Iung, “Stability analysis and control
[1] H. Yu, W. Ming, and R. McGee, “Vehicle handling assistant control synthesis for switched systems: A switched Lyapunov function approach,”
system via independent rear axle torque biasing,” in Proc. Amer. Control IEEE Trans. Autom. Control, vol. 47, no. 11, pp. 1883–1887, Nov. 2002.
Conf., St. Louis, MO, pp. 695–700. [27] L. El Ghaoui, F. Oustry, and M. Ait Rami, “A cone complementarity
[2] J. S. Jo, S. H. You, J. Y. Joeng, K. I. Lee, and K. Li, “Vehicle stability con- linearization algorithm for static output-feedback and related problems,”
trol system for enhancing steerability, lateral stability, and roll stability,” IEEE Trans. Autom. Control, vol. 42, no. 8, pp. 1171–1176, Aug. 1997.
Int. J. Autom. Technol., vol. 9, no. 5, pp. 571–576, Oct. 2008. [28] J. Dong and G.-H. Yang, “Static output feedback control synthesis for
[3] R. J. Rieveley and B. P. Minaker, “Variable torque distribution yaw mo- linear systems with time-invariant parametric uncertainties,” IEEE Trans.
ment control for hybrid powertrains,” presented at the Soc. Autom. Eng., Autom. Control, vol. 52, no. 10, pp. 1930–1936, Oct. 2007.
Detroit, MI, 2007, Paper 2007-01-0278. [29] G. J. Pereira and H. X. de Araujo, “Robust output feedback con-
[4] E. J. Rossetter and J. C. Gerdes, “Lyapunov based performance guaran- troller design via genetic algorithms and LMIs: The mixed H2 /H∞
tees for the potential field lane-keeping assistance system,” J. Dyn. Syst., problem,” in Proc. Amer. Control Conf., Boston, MA, Jun. 30–Jul. 2,
Meas., Control, vol. 128, no. 3, pp. 510–522, Sep. 2006. 2004, pp. 3309–3314.
[5] J. Ahmadi, A. K. Sedigh, and M. Kabganian, “Adaptive vehicle lateral- [30] C. C. Sun, Y. H. Chung, and W. J. Chang, “H2 /H∞ robust static output
plane motion control using optimal tire friction forces with saturation feedback control design via mixed genetic algorithm and linear matrix
limits consideration,” IEEE Trans. Veh. Technol., vol. 58, no. 8, pp. 4098– inequality,” J. Dyn. Syst., Meas., Control, vol. 127, no. 4, pp. 715–722,
4107, Oct. 2009. Dec. 2005.
[6] H. Fujimoto, K. Fujii, and N. Takahashi, “Vehicle stability of elec- [31] M. S. Fallah, R. Bhat, and W. F. Xie, “Optimized control of semiactive
tric vehicle with slip-ratio and cornering stiffness estimation,” in Proc. suspension systems using H∞ robust control theory and current signal
IEEE/ASME Conf. Adv. Intell. Mechatron., Zurich, Switzerland, pp. 1–6. estimation,” IEEE/ASME Trans. Mechatron., vol. 17, no. 4, pp. 767–778,
[7] D. Bianchi, A. Borri, M. D. Di Benedetto, and S. Di Gennaro, “Adap- Aug. 2012.
tive integrated vehicle control using active front steering and rear torque [32] S. K. Chen, N. Moshchuk, and B. Litkouhi, “Corner Dynamics Control:
vectoring,” Int. J. Veh. Auton. Syst., vol. 8, no. 2/3, pp. 85–105, Oct. 2009. Slip-Based Optimization Approach,” General Motors, Detroit, MI, GM
[8] S. C. Baslamisli, I. E. Kose, and G. Anlaç, “Handling stability improve- Res. Rep., ECI-428, 2010.
ment through robust active front steering and active differential control,” [33] J. Ryu, N. K. Moshchuk, and S. K. Chen, “Estimation and Prediction
Int. J. Veh. Syst. Dyn., vol. 49, no. 5, pp. 657–683, May 2011. of Vehicle States for Rollover Avoidance System: VehSim Analysis,”
[9] E. Ono, Y. Hattori, Y. Moraghishi, and K. Koibuchi, “Vehicle dynamics General Motors, Detroit, MI, GM Res. Rep., ECI-258, 2005.
integrated control for four-wheel-distributed traction/bracking systems,” [34] H. E. Tseng, B. Ashrafi, D. Madau, and T. Allen Brown, “The develop-
Int. J. Veh. Syst. Dyn., vol. 44, no. 2, pp. 139–151, Feb. 2006. ment of vehicle stability control at Ford,” IEEE/ASME Trans. Mechatron.,
[10] H. Chou and B. Dandrea-Novel, “Global vehicle control using differential vol. 4, no. 3, pp. 223–234, Sep. 1999.
braking torques and active suspension forces,” Int. J. Veh. Syst. Dyn.,
vol. 43, no. 4, pp. 261–284, Apr. 2005.
[11] S. H. Yu and J. J. Moskwa, “A global approach to vehicle control: Coop-
eration of four wheel-steering and wheel torques,” J. Dyn. Syst., Meas.,
Control, vol. 116, no. 4, pp. 659–667, Dec. 1994. Saber Fallah received the B.Sc. degree from
[12] C. Hamada, K. Fukatani, K. Yamaguchi, and T. Kato, “Development of Isfahan University of Technology, Isfahan, Iran; the
vehicle dynamics integrated management,” presented at the Soc. Autom. M.Sc. degree from Shiraz University, Shiraz, Iran;
Eng. World Congr. Exhib., Detroit, MI, Apr., 2006, Paper 2006-01-0922. and the Ph.D. degree from Concordia University,
[13] M. R. Moreira, E. I. Mainardi Junior, T. T. Esteves, M. C. M. Teixeira, Montreal, QC, Canada, all in the field of mechanical
R. Cardim, E. Assuncao, and F. A. Faria, “Stabilizability and disturbance engineering.
rejection with state-derivative feedback,” J. Math. Prob. Eng., vol. 2010, He was a Research Associate and a Postdoctoral
pp. 123 751-1–123 751-12, Jan. 2010. Research Fellow with the Waterloo Center for Auto-
[14] E. Reithmeier and G. Leitmann, “Robust vibration control of dynamical motive Research, University of Waterloo, Waterloo,
systems based on derivative of the state,” Arch. Appl. Mech., vol. 72, ON, Canada, and a Research Assistant with the
no. 11/12, pp. 856–864, Jun. 2003. Concordia Centre for Advanced Vehicle Engineer-
[15] F. A. Faria, E. Assuncao, M. C. M. Teixeira, and R. Cardim, “Robust state- ing, Concordia University. He is currently a Lecturer (Assistant Professor) with
derivative feedback LMI-based designs for linear descriptor systems,” the University of Surrey, Surrey, U.K. His research interests include vehicle
J. Math. Prob. Eng., vol. 2010, pp. 927 362-1–927 362-15, Jan. 2010. dynamics and control, electric and hybrid electric vehicles, intelligent vehicles,
[16] G. Duan and X. Zhang, “Regularizability of linear descriptor systems via and vehicle system design and integration.
output plus partial state derivative feedback,” Asian J. Control, vol. 5,
no. 3, pp. 334–340, Sep. 2003.
[17] T. H. S. Abdelaziz and M. Valasek, “Pole-placement for SISO linear
systems by state-derivative feedback,” Proc. Inst. Elect. Eng.––Control
Theory Appl., vol. 151, no. 4, pp. 377–385, Jul. 2004. Amir Khajepour is currently a Professor of me-
[18] W. Michiels, T. Vyhlidal, H. Huijberts, and H. Nijmeijer, “Stabilizability chanical and mechatronics engineering with the Uni-
and stability robustness of state derivative feedback controllers,” SIAM J. versity of Waterloo, Waterloo, ON, Canada, and
Control Optim., vol. 47, no. 6, pp. 3100–3117, Nov. 2009. the Canada Research Chair in Mechatronic Vehicle
[19] T. H. S. Abdelaziz, “Robust pole assignment for linear time-invariant Systems. He has developed an extensive research
systems using state-derivative feedback,” J. Syst. Control Eng., vol. 223, program that applies his expertise in several key
no. 2, pp. 187–199, Mar. 2009. multidisciplinary areas. His research interests in-
[20] S. Kwak, G. Washington, and R. K. Yedavalli, “Acceleration feedback- clude system modeling and control of dynamic sys-
based active and passive vibration control of landing gear components,” tems. His research has resulted in several patents
J. Aerosp. Eng., vol. 15, no. 1, pp. 1–9, Jan. 2002. and technology transfers. He is the author of more
[21] Y. Tseng, “Optimal tracking control of nano-positioner in reciprocal state than 250 journal and conference publications and
space framework,” in Proc. 2th ICCAE, Singapore, Feb. 26–28, 2010, four books.
pp. 56–61. Prof. Khajepour is a Fellow of the American Society of Mechanical Engi-
[22] H. K. Khalil, Nonlinear Systems, 3rd ed. Englewood Cliffs, NJ: Prentice- neers and the Canadian Society of Mechanical Engineering. He has received
Hall, 2002. three Best Paper Awards.
Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.
1552 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO. 4, MAY 2013

Barış Fidan received the B.S. degrees in electri- Bakhtiar Litkouhi received the B.Sc. degree in
cal engineering and mathematics from Middle East mechanical engineering, the M.Sc. degree in applied
Technical University, Ankara, in 1996; the M.S. mathematics, and the Ph.D. degree in systems sci-
degree in electrical engineering from Bilkent Univer- ence, specializing in controls.
sity, Ankara, in 1998; and the Ph.D. degree in elec- He was an Assistant Professor with Oakland Uni-
trical engineering from the University of Southern versity, Rochester, MI. He was the Acting Director
California (USC), Los Angeles, in 2003. with the Electrical and Controls Integration Labo-
In 2004, he was with USC as a Postdoctoral Re- ratory, Global Research and Development Center,
search Fellow. From 2005 to 2009, he was with the General Motors Company, Warren, MI. He was a
National ICT Australia and with the Research School Program Manager for several large-scale projects on
of Information Sciences and Engineering, Australian intelligent vehicle systems, human–machine faces,
National University, Canberra, Australia. He is currently an Assistant Professor systems engineering, and integrated vehicle control, where he has made many
with the Department of Mechanical and Mechatronics Engineering, University contributions through numerous patents, publications, and presentations. He
of Waterloo, Waterloo, ON, Canada. His research interests include autonomous is currently the Manager of perception and vehicle control systems with the
multiagent dynamical systems, sensor networks, cooperative target localization, Global Research and Development Center. He is also a Program Manager
adaptive and nonlinear control, switching and hybrid systems, mechatronics, with the General Motors–Carnegie Mellon University Autonomous Driving
and various control applications, including vehicle and transportation control, Collaborative Research Laboratory and a member of the Board of Directors
high-performance and hypersonic flight control, semiconductor manufacturing with the Waterloo Center for Automotive Research.
process control, and disk-drive servo systems. Dr. Litkouhi is a board member of the Intelligent Transportation Society of
Michigan.

Shih-Ken Chen received the B.S. degree from

National Taiwan University, Taipei, Taiwan, in
1985 and the M.S. degree from the University of
Wisconsin, Madison, in 1990 and the Ph.D. de-
gree from Massachusetts Institute of Technology,
Cambridge, in 1996, both in mechanical engineering.
He then joined the Research and Development
Center of General Motors (GM) Corporation. He is
currently a Staff Researcher with the Global R&D
Center, GM, Warren, MI. His previous research inter-
ests include collision avoidance systems, electronic
stability control, active all-wheel-steer control, and rollover avoidance. His
current research interests include integrated chassis and vehicle control for
both conventional and electric drivelines, driver-in-the-loop vehicle control, and
vehicle active safety systems.

Authorized licensed use limited to: ITA - Instituto Tecnológico de Aeronáutica. Downloaded on May 02,2024 at 17:02:30 UTC from IEEE Xplore. Restrictions apply.