566 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 15, NO.

3, MAY 2007

Predictive Active Steering Control for Autonomous

Vehicle Systems
Paolo Falcone, Francesco Borrelli, Jahan Asgari, Hongtei Eric Tseng, and Davor Hrovat, Fellow, IEEE

AbstractIn this paper, a model predictive control (MPC) stability control systems [34] (which are also known under
approach for controlling an active front steering system in an different acronyms such as electronic stability program (ESP),
autonomous vehicle is presented. At each time step, a trajectory is vehicle stability control (VSC), interactive vehicle dynamics
assumed to be known over a finite horizon, and an MPC controller
computes the front steering angle in order to follow the trajectory (IVD), and dynamic stability control (DSC)). Essentially, these
on slippery roads at the highest possible entry speed. We present systems use brakes on one side and engine torque to stabilize
two approaches with different computational complexities. In the vehicle in extreme limit handling situations through con-
the first approach, we formulate the MPC problem by using a trolling the yaw motion.
nonlinear vehicle model. The second approach is based on suc- In addition to braking and traction systems, active front
cessive online linearization of the vehicle model. Discussions on
computational complexity and performance of the two schemes steering (AFS) systems make use of the front steering com-
are presented. The effectiveness of the proposed MPC formulation mand in order to improve lateral vehicle stability [1], [2].
is demonstrated by simulation and experimental tests up to 21 m/s Moreover, the steering command can be used to reject ex-
on icy roads. ternal destabilizing forces arising from -split, asymmetric
Index TermsActive steering, autonomous vehicles, model pre- braking, or wind [21]. Four-wheel steer (4WS) systems follow
dictive control, nonlinear optimization, vehicle dynamics control, similar goals. For instance, in [3], Ackermann et al. present a
vehicle stability. decoupling strategy between the path following and external
disturbances rejection in a four-wheel steering setup. The
automatic car steering is split into the path following and the
I. INTRODUCTION
yaw stabilization tasks, the first is achieved through the front
ECENT trends in automotive industry point in the di- steering angle, the latter through the rear steering angle.
R rection of increased content of electronics, computers,
and controls with emphasis on the improved functionality
Research on the AFS systems has also been approached
from an autonomous vehicle perspective. In [16], an automatic
and overall system robustness. While this affects all of the steering control for highway automation is presented, where
vehicle areas, there is a particular interest in active safety, the vehicle is equipped with magnetic sensors placed on the
which effectively complements the passive safety counterpart. front and rear bumpers in order to detect a lane reference im-
Passive safety is primarily focused on the structural integrity plemented with electric wire [13] and magnetic markers [36].
of vehicle. Active safety on the other hand is primarily used to A more recent example of AFS applications in autonomous
avoid accidents and at the same time facilitate better vehicle vehicles is the Grand Challenge race driving [5], [23], [30].
controllability and stability especially in emergency situations, In this paper, it is anticipated that the future systems will be
such as what may occur when suddenly encountering slippery able to increase the effectiveness of active safety interventions
parts of the road [10]. beyond what is currently available. This will be facilitated not
Early works on active safety systems date back to the 1980s only by additional actuator types such as 4WS, active steering,
and were primarily focused on improving longitudinal dy- active suspensions, or active differentials, but also by additional
namics part of motion, in particular, on more effective braking sensor information, such as onboard cameras, as well as in-
(ABS) and traction control (TC) systems. ABS systems in- frared and other sensor alternatives. All these will be further
crease the braking efficiency by avoiding the lock of the braking complemented by global positioning system (GPS) information
wheels. TC systems prevent the wheel from slipping and at the including prestored mapping. In this context, it is possible to
same time improves vehicle stability and steerability by maxi- imagine that future vehicles would be able to identify obstacles
mizing the tractive and lateral forces between the vehicles tire on the road such as an animal, a rock, or fallen tree/branch, and
and the road. This was followed by work on different vehicle assist the driver by following the best possible path, in terms of
avoiding the obstacle and at the same time keeping the vehicle on
Manuscript received November 10, 2006. Manuscript received in final form
the road at a safe distance from incoming traffic. An additional
January 12, 2007. Recommended by Associate Editor K. Fishbach. source of information can also come from surrounding vehicles
P. Falcone and F. Borrelli are with the Universit del Sannio, Dipartimento di and environments which may convey the information from the
Ingegneria, Universit degli Studi del Sannio, 82100 Benevento, Italy (e-mail:
[email protected]; [email protected]).
vehicle ahead about road condition, which can give a significant
J. Asgari, H. E. Tseng, and D. Hrovat are with Research and Innovation amount of preview to the controller. This is particular is useful
Center, Ford Research Laboratories, Dearborn, MI 48124 USA (e-mail: jas- if one travels on snow or ice covered surfaces. In this case, it is
[email protected]; [email protected]; [email protected]). very easy to reach the limit of vehicle handling capabilities.
Color versions of Figs. 1, 2, and 412 are available online at http://ieeexplore.
ieee.org. Anticipating sensor and infrastructure trends toward in-
Digital Object Identifier 10.1109/TCST.2007.894653 creased integration of information and control actuation agents,
1063-6536/$25.00 2007 IEEE
FALCONE et al.: PREDICTIVE ACTIVE STEERING CONTROL FOR AUTONOMOUS VEHICLE SYSTEMS 567

it is then appropriate to ask what is the optimum way in con-

trolling the vehicle maneuver for a given obstacle avoidance
situation.
We assume that a trajectory planning system is available and
we consider a double lane change scenario on a slippery road,
with a vehicle equipped with a fully autonomous guidance
system. In this paper, we focus on the control of the yaw and
lateral vehicle dynamics via active front steering. The control
input is the front steering angle and the goal is to follow the
desired trajectory or target as close as possible while fulfilling
various constraints reflecting vehicle physical limits and design
requirements. The future desired trajectory is known only over Fig. 1. Simplified vehicle bicycle model.
finite horizon at each time step. This is done in the spirit of
model predictive control (MPC) [14], [26] along the lines of
our ongoing internal research efforts dating from early 2000 framework for autonomous vehicle guidance. The contribution
(see [7] and references therein). and the research topic of this paper are described in details
In this paper, two different formulations of the AFS MPC and put in perspective with existing work and future research.
problem will be presented and compared. The first one follows Section IV formulates the control problem when the nonlinear
the work presented in [7] and uses a nonlinear vehicle model and the linear prediction models are used. The double lane
to predict the future evolution of the system [26]. The resulting change scenario is described in Section V, while in Section VI,
MPC controller requires a nonlinear optimization problem to the experimental and simulation results are presented. This is
be solved at each time step. We will show that the computa- then followed by concluding remarks in Section VII which
tional burden is currently an obstacle for experimental valida- highlight future research directions.
tion at high vehicle speed. The second formulation tries to over-
come this problem and presents a suboptimal MPC controller II. MODELING
based on successive online linearization of the nonlinear vehicle This section describes the vehicle and tire model used for
model. This is linearized around the current operating point at simulations and control design. This section has been extracted
each time step and a linear MPC controller is designed for the re- from [7] and it is included in this paper for the sake of complete-
sulting linear time-varying (LTV) system. The idea of using time ness and readability. We denote by and the longitudinal (or
varying models goes back to the early 1970s in the process con- tractive) and lateral (or cornering) tire forces, respectively,
trol field although it has been properly formalized only recently. and are the forces in car body frame, is the normal tire
Studies on linear parameter varying (LPV) MPC schemes can be load, is the car inertia, and are the absolute car position
found in [9], [18], [20], [22], and [35]. Among them, the work inertial coordinates, and are the car geometry (distance of
in [18] and [20] is the closest to our approach and it presents front and rear wheels from center of gravity), is the gravita-
an MPC scheme for scheduled LTV models which has been tional constant, is the car mass, is the wheel radius, is the
successfully validated on a Boeing aircraft. In general, the per- slip ratio, and are the longitudinal and lateral wheel veloc-
formance of such a scheme is highly dependant on the nonlin- ities, and are the local lateral and longitudinal coordinates
earities of the model. In fact, as the state and input trajectories in car body frame, is the vehicle speed, is the slip angle,
deviate from the current operating point, the model mismatch is the wheel steering angle, is the road friction coefficient,
increases. This can generate large prediction errors with a con- and is the heading angle. The lower subscripts and
sequent instability of the closed-loop system. We will show that, particularize a variable at the front wheel and the rear wheel,
in our application, a state constraint can be introduced in order respectively, e.g., is the front wheel longitudinal force.
to significantly enhance the performance of the system. Exper-
imental results show that the vehicle can be stabilized up to A. Vehicle Model
21 m/s on icy roads. Finally, an LTV MPC with a one-step con- A bicycle model [25] is used to model the dynamics of the
trol horizon is presented. This can be tuned in order to provide car under the assumption of a constant tire normal load, i.e., ,
acceptable performance and it does not require any complex op- . Fig. 1 depicts a diagram of the vehicle model,
timization software. which has the following longitudinal, lateral, and turning or yaw
We implemented the MPC controllers in real time on a pas- degrees of freedom:
senger car, and performed tests on snow covered and icy roads.
The last part of this paper describes the experimental setup and (1a)
presents the experimental and simulation results of the proposed (1b)
MPC controllers. It should be noted that our early work in [7]
(1c)
focuses on the vehicle dynamical model and on simulation re-
sults of the nonlinear MPC scheme only. The vehicles equations of motion in an absolute inertial
This paper is structured as follows. Section II describes frame are
the used vehicle dynamical model with a brief discussion on
tire models. Section III introduces a simplified hierarchical (2)
568 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 15, NO. 3, MAY 2007

The wheels equations of motion describe the lateral (or cor-

nering) and longitudinal wheel velocities

(3a)
(3b)
(3c)
(3d)

where and are front and rear wheel steering angle, respec-
tively, and

(4a)
(4b)

The following equations hold for rear and front axes by using
the corresponding subscript for all the variables. Longitudinal
and lateral tire forces lead to the following forces acting on the
center of gravity:

(5a)
Fig. 2. Longitudinal and lateral tire forces with different  coefficient values.
Tire forces and for each tire are given by

(6)
B. Tire Model
where , , , and are defined next. The tire slip angle
represents the angle between the wheel velocity vector and With the exception of aerodynamic forces and gravity, all of
the direction of the wheel itself, and can be compactly expressed the forces which affect vehicle handling are produced by the
as tires. Tire forces provide the primary external influence and, be-
cause of their highly nonlinear behavior, cause the largest vari-
(7) ation in vehicle handling properties throughout the longitudinal
and lateral maneuvering range. Therefore, it is important to use
The slip ratio is defined as a realistic nonlinear tire model, especially when investigating
large control inputs that result in response near the limits of the
if for braking maneuvering capability of the vehicle. In such situations, the
(8)
if for driving lateral and longitudinal motions of the vehicle are strongly cou-
where and are the radius and the angular speed of the wheel, pled through the tire forces, and large values of slip ratio and
respectively. The parameter represents the road friction coef- slip angle can occur simultaneously.
ficient and is assumed equal for front and rear wheels. is the Most of the existing tire models are predominantly semi-em-
total vertical load of the vehicle and is distributed between the pirical in nature. That is, the tire model structure is determined
front and rear wheels based on the geometry of the car model through analytical considerations, and key parameters depend
(described by the parameters and ) on tire data measurements. Those models range from extremely
simple (where lateral forces are computed as a function of slip
(9) angle, based on one measured slope at and one measured
value of the maximum lateral force) to relatively complex al-
The nonlinear vehicle dynamics described in (1)(9), can be gorithms, which use tire data measured at many different loads
rewritten in the following compact from: and slip angles.
In this paper, we use a Pacejka tire model [4] to describe the
(10) tire longitudinal and cornering forces in (6). This is a complex,
semi-empirical nonlinear model that takes into consideration
where the dependence on slip ratio and friction coefficient the interaction between the longitudinal force and the cornering
value at each time instant has been explicitly highlighted. The force in combined braking and steering. The longitudinal and
state and input vectors are and , cornering forces are assumed to depend on the normal force, slip
respectively. In this paper, is assumed to be zero at any time angle, surface friction coefficient, and longitudinal slip. Fig. 2
instant. depicts longitudinal and lateral forces versus longitudinal slip
Model (10) captures the most relevant nonlinearities associ- and slip angle, for fixed values of the friction coefficients. We
ated to lateral stabilization of the vehicle. Section II-B briefly remark that the front tire of the bicycle model represents the
describes the models of tire forces and . two front tires of the actual car.
FALCONE et al.: PREDICTIVE ACTIVE STEERING CONTROL FOR AUTONOMOUS VEHICLE SYSTEMS 569

We remark that the scheme in Fig. 3 is an oversimplified

scheme and that additional hierarchical levels could be present
both in the trajectory/mode replanning module and in the
low-level control system module. The union of the first three
modules is often referred to as guidance and navigation control
(GNC) system.
Typically, the trajectory replanner and the low-level control
system modules do not share the same information on envi-
Fig. 3. Simplified architecture for fully autonomous vehicle guidance system. ronment and vehicle. For instance, the replanning algorithms
can use information coming from cameras or radars which may
not be used at the lower level. Also, typically, the frequency at
III. HIERARCHICAL FRAMEWORK FOR which the trajectory replanning module is executed is lower than
AUTONOMOUS GUIDANCE the one of the lower level control system. The design of both
In this section, we borrow the simplified schematic architec- modules makes use of vehicle and environment models with
ture in Fig. 3 from the aerospace field [8], [24], [31], in order to different levels of detail. The fidelity of the dynamical model
explain our approach and contribution. The architecture in Fig. 3 used for the design of the two modules is dictated, among many
describes the main elements of an autonomous vehicle guidance factors, by a performance/computational resource compromise
system and it is composed of four modules: the trajectory/mode and, in the literature, there is no accepted standard on this. One
generator, the trajectory/mode replanning, the low-level con- of the possible control paradigms for the two modules con-
trol system, and the vehicle and the environmental model. The sists in using a high-fidelity vehicle model for designing the
trajectory/mode planning module precomputes offline the ve- lower level controller while the trajectory planner relies on a
hicle trajectory together with the timing and conditions for op- rougher/less detailed dynamical model of the vehicle. Clearly,
eration mode change. In the aerospace field, examples of op- the higher the fidelity of the models used at the higher level is,
eration mode selection include aeroshell parachute deployment the easier the job for the lower level control algorithm becomes.
or heatshield release, in the automotive field this could include Studies on GNC algorithms vary in 1) the focus (trajectory
switching between two or more types of energy sources (i.e., replanner and/or the low-level control system); 2) the type of
gas, electricity, hydrogen) or (in a very futuristic scenario) mor- vehicle dynamical model used; 3) the type of control design
phing between different vehicle shapes. used; and 4) inputs and sensors choice.
The trajectory and the mode of operation computed offline In [23], the trajectory replanner module is based on a receding
can be recomputed online during the drive by the trajectory/ horizon control design. The planning problem is formulated as a
mode replanning module based on current measurements, at constrained optimization problem minimizing a weighted sum
fixed points or on the occurrence of certain events (such as of arrival time, steering, and acceleration control efforts. The
tracking errors exceeding certain bounds, hardware failure, ex- vehicle model is a simple rear-centered kinematic model with
cessive wind, the presence of a pop-up obstacle). acceleration, speed, steering, steering rate, and rollover con-
The low-level control system commands the vehicle actua- straints. The lower level control module uses two separated pro-
tors such as front and rear steering angles, four brakes, engine portionalintegraldifferentials (PIDs) to control longitudinal
torque, active differential, and active suspensions based on and lateral dynamics. The longitudinal controller acts on throttle
sensor measurements, states, and parameters estimations and and brakes while the lateral controls on the steering angle.
reference commands coming from the trajectory/mode replan- The GNC architecture in [30] is similar to [23]. The trajectory
ning module. Such reference commands can include lateral and planning task is posed as a constrained optimization problem.
longitudinal positions, pitch, yaw, and roll rates. The low-level The cost function penalizes obstacles collision, distance from
control system objective is to keep the vehicle as close as pos- the precomputed offline trajectory and the lateral offset from
sible to the currently planned trajectory despite measurement the current trajectory. At the lower level, a PI controller acts
noise, unmodeled dynamics, parameteric uncertainties, and on brakes and throttle to control the longitudinal dynamics. A
sudden changes on vehicle and road conditions which are not simple nonlinear controller, instead, is used to control the lateral
(or not yet) taken into account by the trajectory replanner. dynamics through the steering angle. Details on the vehicle dy-
In particular, when a vehicle is operating near its stability namical model used in [30] are not disclosed. In [29], a scheme
limit, these additional noises, disturbances, and uncertainties similar to the one in [23] is used to design a GNC systems for a
must be considered, possibly through detecting the vehicles flight control application.
internal state, and compensated for. For example, if rear tires In [33], an explicit MPC scheme has been applied at the lower
saturate, a skillful driver would switch his/her steering input level control to allocate four wheel slips in order to get a desired
from the usual steering command for trajectory following to a yaw moment. The steering angle is not controlled.
counter-steering one for stabilizing the vehicle. It is conceivable In this paper, we assume that the path can be generated with
that an automated steering would not produce the necessary two different methods. In the first, the trajectory is established
stabilizing counter-steer if the commanded steering is only a by simply driving a test vehicle slowly along the desired path,
function of the desired trajectory and vehicles current position e.g., a double lane change manoeuvre. The actual path is
and heading (without considering additional vehicle dynamic recorded by differential GPS and then used as a desired path
states). for subsequent tests at higher speed. This method has been used
570 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 15, NO. 3, MAY 2007

in [34] to generate the reference path for a steering robot on nonlinear vehicle model (10) and the Pacejka tire model are
high . In this case, no trajectory replanning is needed and the used to predict the future evolution of the system. The mini-
contribution of the work presented in this paper is to facilitate mization of a quadratic performance index, subject to the non-
systematic and repeatable tests of safety critical emergency linear vehicle dynamics, is a nonlinear optimization problem.
manoeuvres during limit conditions, such as obstacle avoidance Such optimization problem is solved online, at each time step.
manoeuvres on slippery surfaces, i.e., snow and ice. In the This can be computationally demanding, depending on the ve-
second method, we assume that a trajectory replanning module hicles states and constraints. The second formulation, presented
is available and the trajectory is recomputed at a less frequent in Section IV-B, tries to overcome this problem. A LTV approx-
rate than the frequency of the lower level controller. For both imation of vehicle model (10) and the Pacejka tire model are
cases, we focus on the lower level control design by means of used to predict the future evolution of the system. This leads to
nonlinear and LTV MPC for the specific scenario of an active a suboptimal LTV MPC controller. In this case, a time varying
steering system. convex quadratic optimization problem is formulated and solved
As suggested in [28], there is a significant challenge involved at each time step, leading to the reduction of the computational
in obtaining the steering required to accomplish the limit ma- burden with an acceptable loss of performance. We will show
neuver considered in this paper while maintaining vehicle sta- that the MPC performance is enhanced by including a constraint
bility. By focusing on the lower level MPC controller, we also on the tire slip angle which stabilizes the vehicle at high speed.
believe that the resultant steering may mimic a skillful driver
who takes the full vehicle dynamic states into account. Com- A. Nonlinear (NL) MPC
pared to the lower level control algorithms presented in the
In order to obtain a finite-dimensional optimal control
aforementioned literature, our approach 1) is model based and
problem, we discretize the system dynamics (10) with a fixed
uses the vehicle model (10) and the highly nonlinear Pacejka tire
sampling time
model described in Section II-B; 2) includes constraints on in-
puts and states in the control design; 3) is systematic and multi-
variable and can accommodate new actuators and higher fidelity (11a)
models. Moreover, we have experimentally validated the con- (11b)
troller presented in this paper with a dSPACE AutoBox system
which is a standard rapid prototyping system used in automo- where the formulation is used, with and
tive industries [11]. .
We define the following output map for yaw angle and lateral
IV. ACTIVE STEERING CONTROLLER DESIGN position states:

In this section, we introduce the control design procedure

for the proposed path following problem via an active steering (12)
system.
Desired references for the heading angle , the yaw rate , and consider the following cost function:
and the lateral distance define a desired path over a finite
horizon. The nonlinear vehicle dynamics (10) and the Pacejka
tire model are used to predict the vehicles behavior, and the front
steering angle is chosen as control input. The rear steering
angle is assumed to be zero , the tire slip ratios and (13)
are measured, and the road friction is estimated at each where and denote the corresponding reference
time instant. The approach used in [6] can be used for the online signal. At each time step , the following finite horizon optimal
estimation of . control problem is solved:
A MPC scheme is used to solve the path following problem.
The main concept of MPC is to use a model of the plant to pre- (14a)
dict the future evolution of the system [6], [14], [17], [26], [27].
At each sampling time, starting at the current state of the ve- subj. to (14b)
hicle, an open-loop optimal control problem is solved over a fi- (14c)
nite horizon. The open-loop optimal control problem minimizes (14d)
the deviations of the predicted outputs from their references over
a sequence of future steering angles, subject to operating con-
straints. The resulting optimal command signal is applied to the (14e)
process only during the following sampling interval. At the next (14f)
time step, a new optimal control problem based on new mea- (14g)
surements of the state is solved over a shifted horizon.
In the following two different formulations of the AFS MPC
problem will be presented. Section IV-A describes the first MPC (14h)
formulation as presented in the preliminary work [7]. There, the (14i)
FALCONE et al.: PREDICTIVE ACTIVE STEERING CONTROL FOR AUTONOMOUS VEHICLE SYSTEMS 571

where is the optimization where and model (16b) and

vector at time , denotes the output vector predicted (16c) is obtained by linearizing model (11) at each time step
at time obtained by starting from the state , around the point , , for the estimated , .
and applying to system (11) and (12) the input sequence The variables and denote the outputs
. and denote the output prediction of the linearized system and the corresponding reference signal,
horizon and the control horizon, respectively. We use respectively. The variable denotes the tire slip angle varia-
and the control signal is assumed constant for all . tion and it is an additional output of the linearized model which
We assume slip and friction coefficient values constant and is only constrained and not tracked. Inequalities (16j) are soft
equal to the estimated values at time over the prediction constraints on the tire slip angle and is a slack variable. The
horizon [constraint (14d)]. term in (16b) penalizes the violation of the constraint on the
In (13), the first summand reflects the penalty on trajectory slip angle and is a weight coefficient.
tracking error while the second summand is a measure of the The optimization problem (16) can be recast as a quadratic
steering effort. and are weighting matrices of appropriate program (QP) (details can be found in [7]). We denote by
dimensions. the sequence of optimal input
We denote by the se- deviations computed at time by solving (16) for the current
quence of optimal input increments computed at time by observed states . Then, the first sample of is used
solving (14) for the current observed states . Then, the first to compute the optimal control action and the resulting state
sample of is used to compute the optimal control action feedback control law is
and the resulting state feedback control law is
(17)
(15)
At the next time step the optimization problem (16) is
At the next time step , the optimization problem (14) is solved over a shifted horizon based on the new measurements
solved over a shifted horizon based on the new measurements of the state and based on an updated linear model
of the state . (16b)(16d) computed by linearizing the nonlinear vehicle
model (11) around the new state, slip ratio, road friction coeffi-
cient, and previous input.
B. LTV MPC
We remark that model (11) is linearized around an operating
Let be the current time and and be the current point that, in general, is not an equilibrium point. Therefore,
state and the previous input of system (11) and (12), respec- the linear time-invariant (LTI) model (16b)(16d) at time is
tively. We consider the following optimization problem: used to predict the state and the output deviations from the tra-
jectories , , for , respec-
tively, computed by solving (11) with as initial condition
and for . Accordingly, the op-
timization variables represent the input
variation with respect to the previous input .
(16a) Alternatively, the vehicle model (11) can be linearized around
a nominal input and state trajectory . In this
subj. to (16b) case, (16e) would become and the
(16c) optimization variables would represent
the input variations around the nominal input. This approach
(16d) requires a nominal input and state trajectory, i.e., and
. Such trajectory could be computed from the higher
level replanning algorithm described in Section III or from the
lower level MPC controller. We remark that in this case an LTV
(16e)
model over the prediction horizon ( , , , ),
(16f) could be used at each time step instead of
(16g) the LTI model (16b)(16d).
(16h) An MPC scheme similar to the one presented in this paper
can be found in [18], [19], and [20]. In these works, a similar
(16i)
MPC formulation is used. An LTI prediction model is used to
predict the behavior of the system over the prediction horizon.
(16j) The LTI model is updated according to the values of flight con-
dition dependent scheduling parameters. In [19] and [20], the
LTI model is obtained by interpolation over a precomputed data-
(16k)
base of linearized models, while in [18] the LTI model is ob-
(16l) tained by linearizing the nonlinear kinematics around the cur-
(16m) rent measurements.
572 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 15, NO. 3, MAY 2007

When evaluating the online computational burden of the pro- additional term significantly improves the performance of the
posed scheme, in addition to the time required to solve the op- LTV MPC controller (16) and (17).
timization problem (16), one needs to consider the resources
spent in computing the linear models ( , , , ) in (16b) V. DOUBLE LANE CHANGE ON SNOW USING ACTIVE STEERING
and (c) and translating (16) into a standard quadratic program-
The MPC steering controllers described in Sections IV-A and
ming (QP) problem. Nevertheless, for the proposed application,
IV-B have been implemented to perform a sequence of double
complexity of the MPC (16) and (17) greatly reduces compared
lane changes at different entry speeds. This test represents an
to the NL MPC presented in Section IV-A. This will be shown obstacle avoidance emergency maneuver in which the vehicle
for a specific scenario in Sections VI-A and VI-B.
is entering a double lane change maneuver on snow or ice with
The stability of the presented control scheme is difficult to
a given initial forward speed. The control input is the front
prove even under no model mismatch and it is a topic of cur-
steering angle and the goal is to follow the trajectory as close
rent research. Also, robustness of nonlinear MPC schemes is an as possible by minimizing the vehicle deviation from the target
active area of research by itself. An analytical and meaningful
path. The experiment is repeated with increasing entry speeds
study of the robustness of the proposed scheme would be even
until the vehicle loses control. The same controller can be used
more difficult. The uncertainty of the tire characteristics and the
to control the vehicle during different maneuvers in different
road condition are often difficult to describe with a mathemat- scenarios [21].
ical formalism which is realistic and not too conservative.
The simulation and experimental results will be presented
It should be noted that in the MPC scheme (16) and (17), the
in Section VI. Next, we describe the reference generation and
introduction of the state constraints (16j) is needed in order to
present the experimental setup in Section V-B.
obtain an acceptable performance and it is a contribution of this
paper. As shown next in Section VI-A, such constraint arises A. Trajectory Generation
from a careful study of the closed-loop behavior of the non-
linear MPC presented in Section IV-A. In fact, extensive sim- The desired path is described in terms of lateral position
ulations have shown that the nonlinear MPC never exceeds cer- and yaw angle as function of the longitudinal position
tain tire slip angles under stable operations. By removing the
(18a)
constraints (16j) the performance of the LTV MPC controller
(16) and (17) is not acceptable and the system becomes un-
stable at high vehicle speeds. In fact, a simple linear model is
not able to predict the change of slope in the tires characteristic
(see Fig. 2). To overcome this issue, we included constraints
(18b)
(16j) in the optimization problem, in order to forbid the system
from entering a strongly nonlinear and possibly unstable re-
gion of the tire characteristic. In particular, by looking at the where ,
tire characteristics in Fig. 2, it is clear that a linear approxima- , 2, , ,
tion of the tire model around the origin is no longer valid if the , and .
slip angle exceeds certain bounds. Led by this observation and The reference trajectories (18a) and (18b), can be used di-
by a study on the closed-loop behavior of the nonlinear MPC rectly only in the nonlinear MPC formulation, being a nonlinear
presented in Section IV-A, we included the constraints (16j) in function of the longitudinal distance . In the LTV MPC for-
the optimization problem. In particular, for a given , the tire mulation, we generate the reference trajectories from (18a) and
slip angle is constrained in the mostly linear region of the lat- (18b) by assuming that the vehicle will travel a portion of the
eral tire force characteristic. By no means does the constraints desired path at a constant speed in the next steps.
(16j) enforce the dynamical system to operate in a linear region: Because of the assumption on constant travel velocity, the
system nonlinearities (11) and longitudinal tire nonlinearities method for generating the previously described trajectory can
are still relevant when constraints (16j) are included in the MPC affect the performance of the closed-loop system. In particular,
formulation. in extreme handling situations, when the tracking errors are
Note that the constraints (16j) are implicit linear constraints large due to spinning or side skidding, the computed reference
on state and input and they can be handled systematically only could lead to aggressive maneuvers. As explained in Section III,
in an MPC scheme. A soft constraint formulation is preferred to more accurate methods could be used in order to generate a
a hard constraint in order to avoid infeasibility. In fact, during smoother reference for the LTV MPC scheme by taking into
experiments the tire slip angle is estimated from IMU and GPS account the state of the vehicle.
measurements. Acceleration measurements are noisy and the
GPS signal can be lost. Moreover, as shown in (3) and (7), the B. Experimental Setup Description
tire slip angle depends on the steering angle. The latter, as ex- The MPC controllers presented in Sections IV-A and IV-B
plained in Section V-B, in our experimental setup is affected by have been tested through simulations and experiments on slip-
the drivers imposed steering angle. pery surfaces. The experiments have been performed at a test
An additional tracking error on yaw rate is included in the per- center equipped with icy and snowy handling tracks. The MPC
formance index of the LTV MPC problem (16) and (17) (com- controllers have been tested on a passenger car, with a mass of
pare (16d) to (12)). Extensive simulations have shown that this 2050 Kg and an inertia of 3344 kg/m . The controllers were run
FALCONE et al.: PREDICTIVE ACTIVE STEERING CONTROL FOR AUTONOMOUS VEHICLE SYSTEMS 573

in a dSPACE Autobox system, equipped with a DS1005 pro- TABLE I

cessor board and a DS2210 I/O board, with a sample time of MAXIMUM COMPUTATION TIME OF CONTROLLERS A AND B PERFORMING
A DOUBLE LANE CHANGE MANEUVER AT DIFFERENT VEHICLE SPEEDS
50 ms.
We used an Oxford Technical Solution (OTS) RT3002
sensing system to measure the position and the orientation
of the vehicle in the inertial frame and the vehicle velocities
in the vehicle body frame. The OTS RT3002, is housed in a
small package that contains a differential GPS receiver, inertial
measurement unit (IMU), and a DSP. It is equipped with a Controller B: LTV MPC (16) and (17) with the following
single antenna to receive GPS information. The IMU includes parameters:
three accelerometers and three angular rate sensors. The DSP 0.05 s, ; , ,
receives both the measurements from the IMU and the GPS, , , ,
utilizes a Kalman filter for sensor fusion, and calculates the , , ;
position, the orientation, and other states of the vehicle such as
longitudinal and lateral velocities. weighting matrices , ,
The car was equipped with an AFS system which utilizes an
electric drive motor to change the relation between the hand .
steering wheel and road wheel angles. This is done indepen- Controller C: Same as Controller B with .
dently from the hand wheel position, thus the front road wheel Next, the results obtained with the three controllers will be
angle is obtained by summing the driver hand wheel position described and a comparison between the simulation and the ex-
and the actuator angular movement. Both the hand wheel posi- perimental results will be given for each of them. The actual
tion and the angular relation between hand and road wheels are road friction coefficient was set manually and constant for
measured. The sensor, the dSPACE Autobox, and the actuators each experiment depending on the road conditions. This choice
communicate through a CAN bus. was driven by the study of the controller closed-loop perfor-
The autonomous steering test is initiated by the driver with a mance independently from the estimation and its associated
button. When the button is pushed, the inertial frame in Fig. 1 is error and dynamics. For each controller more simulation, exper-
initialized as follows: the origin is the current vehicle position, iments, and comments can be found in [12].
the axes and are directed as the current longitudinal and
lateral vehicle axes, respectively. Such inertial frame becomes A. Controller A
also the desired path coordinate system. Once the initialization The controller (14) and (15) with the parameters defined in
procedure is concluded, the vehicle executes the double lane Section VI has been implemented as a C-coded S-function in
change maneuver. which the commercial NPSOL software package [15] is used for
During the experiment, the hand wheel may deviate from its solving the nonlinear programming problem (14). The choice of
center position. This is caused by the difficulty the driver can NPSOL has been motivated by its performance and the avail-
have in holding the steering still, which was needed to facil- ability of the source C code.
itate autonomous behavior with that particular test vehicle. In Limited by the computational complexity of the nonlinear
our setup, this is treated as a small bounded input disturbance. programming solver and the hardware used, we could perform
Furthermore, noise may affect the yaw angle measurement due experiments at low vehicle speeds only. In fact, as the entry
to the single antenna sensor setup. Compared to a dual antenna speed increases, larger prediction and control horizons are re-
setup, a single antenna system has to learn the vehicle orienta- quired in order to stabilize the vehicle along the path. Larger pre-
tion and/or coordinate during vehicle motion. When the vehicle diction horizons involve more evaluations of the objective func-
stands still the yaw angle is computed by integrating the yaw tion, while larger control horizons imply a larger optimization
rate measurement from the IMU. This might cause the presence problem (14). In Table I, we report the maximum computation
of a small offset in the orientation measurement, while traveling time required by the Controllers A and B to compute a solution
at low speed or being still. The effects of both input disturbance to the problems (14) and (16), respectively, when the maneuver
and measurement noise will be clear later in the presented ex- described in Section V is performed at different vehicle speeds.
perimental results. The selected control and prediction horizons in Table I are the
shortest allowing the stabilization of the vehicle at each speed.
VI. PRESENTATION AND DISCUSSION OF RESULTS The results have been obtained in simulation with a 2.0-GHz
In Section VI-A, three types of MPC controllers will be Centrino-based laptop running Matlab 6.5.
presented. These controllers have been derived by the MPC During experiments, the maximum iterations number in
problem formulations presented in Sections IV-A and IV-B and NPSOL bas been limited in order to guarantee real-time com-
will be referred to as Controller A, B, and C. putation. The bound was selected after preliminary tests on the
Controller A: Nonlinear MPC (14) and (15) with the fol- real-time hardware.
lowing parameters: In Fig. 4, the simulation results for a maneuver at 7 m/s are
0.05 s, ; ; , presented. In Fig. 5, the corresponding experimental results are
, , , ; presented. In the upper plot of Fig. 5(b), the dashed line rep-
resents the steering action from the driver (i.e., the input dis-
, .
turbance) that, in this test, is negligible. The actual road wheel
574 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 15, NO. 3, MAY 2007

_
Fig. 4. Simulation results at 7-m/s entry speed. Controller A described in
( )
Section IV-A. (a) Lateral position Y , yaw angle ( ) , and yaw rate .
( ) ( ) _
Fig. 5. Experimental results at 7-m/s entry speed. Controller A described in
( )
(b) Front steering angle  , change in front steering angle (1 )
 , and
Section VI-A. (a) Lateral position Y , yaw angle
( )
steering angle  , change in front steering angle
, and yaw rate . (b) Front
(1 )
 , and NPSOL output
NPSOL output flag.
flag.

angle (RWA) is the summation of the RWA from the MPC con- and are the root mean squared (rms) yaw angle
troller and the steering action from the driver. In lower plots of and lateral position tracking errors, respectively. The values of
Figs. 4(b) and 5(b), the NPSOL output flag is reported. In our and are the maximum tracking errors on the same
tests, the flag assumed the values 0, 1, 4, and 6. The value 0 is variables.
returned when an optimal feasible solution is found. The value By comparing the simulated and the experimental steering
1 is returned when the solver does not converge to a feasible command, we notice the presence of an unmodeled rate satu-
solution. The value 4 indicates that the limit on the iteration ration in the steering response. In fact, the actual road steering
number has been reached and a feasible but nonoptimal solu- angle variation is smaller than the selected in (14g).
tion has been found. The value 6 indicates that the solution does This can be observed in Fig. 5(b), between 12 and 14 s,
not satisfy the optimality conditions [15]. In simulation and ex- where the desired change in the steering angle was limited to
perimental tests, the solver often reaches the selected iteration , while the actual road wheel steering angle
limit and returns a suboptimal solution. Yet, because of the low increased at a slower rate. This led to a larger tracking error in
vehicle speed, the performance associated to the suboptimal so- the experiment during the second lane change compared to the
lution is excellent. simulations.
By comparing the lateral position and yaw angle in the sim- The experimental results 10 m/s in [12] show that the con-
ulation and the experiment, we can conclude that the matching troller is not able to stabilize the vehicle and, around 13 s, the
between simulation and experimental results is very good. The vehicle starts to skid. The controller fails because the nonlinear
tracking errors are very small and reported in Table II, where solver does not converge to a feasible solution.
FALCONE et al.: PREDICTIVE ACTIVE STEERING CONTROL FOR AUTONOMOUS VEHICLE SYSTEMS 575

TABLE II
SIMULATION RESULTS. CONTROLLER A PRESENTED IN SECTION VI-A.
rms AND MAXIMUM TRACKING ERRORS AT 7 m/s

Fig. 6. Simulation results of Controller A at 17 m/s.

An analysis of simulation results shows that the solver does

not converge because of the limit on the number of iterations.
The work presented in [7] has shown that, the NL-MPC con-
troller is able to perform the maneuver at 10-m/s or higher speed
if the solver maximum iteration number is not limited.
Finally, in order to better motivate the introduction of the con-
straints (16j) in the MPC formulation presented in Section IV-B,
in Fig. 6, we report the simulation results of a nonlinear MPC at
_
Fig. 7. Simulation results at 10-m/s entry speed. Controller B described in
17 m/s on snow. We observe that, due to the knowledge of the ( ) ( )
Section IV-B. (a) Lateral position Y , yaw angle , and yaw rate . (b) Front
tire characteristics, the controller implicitly limits the front tire ( )
steering angle  , change in front steering angle (1 )
 , and front tire slip
slip angle in the interval . As shown in Fig. 2, angle  .
this is within the linear region of the tire characteristic for a
snow covered road . The results in Fig. 6 have been
obtained by Controller A using the same tuning parameters of solution. At each time step, the linear model (16b) and (16c) is ob-
Controller B presented in Section VI. In particular, a prediction tained by analytic differentiation of the nonlinear vehicle model
horizon of 25 steps and a control horizon of ten steps have been (11) and a numeric linearization of the Pacejka tire model.
selected. The steering angle has been constrained within the in- We remark that the computation burden of this LTV MPC
terval , while the steering rate in . We controller is reduced significantly compared to the controller
also point out that no constraint on the front tire slip angle has presented in Section VI-A, as demonstrated by the computation
been used. times reported in Table I.
Extensive simulations have shown that this phenomenon can Fig. 7 depicts simulation results at the entry speed of 10 m/s.
always be observed in extreme conditions under operations and Table III summarizes the tracking errors obtained in simulation
led us to the use of the constraint (16j) in the LTV MPC formu- for entry speeds of 10 to 21.5 m/s. In Fig. 7(a), one can ob-
lation (16). serve: 1) an undershoot in the lateral position and 2) steady-state
errors in both lateral position and yaw angle. Moreover, the
B. Controller B steady-state values of the position and the yaw angle are not co-
The controller (16) and (17) with the parameters defined in herent, being the lateral position is constant and the yaw angle
Section VI has been implemented as a C-coded S-Function, using is nonzero. These phenomena are generated by an orientation
the QP solver routine available in [32]. Such routine implements error, detected in the experimental data and introduced in the
the DantzigWolfes algorithm, has a good performance and its presented simulations in order to fairly compare the simula-
source C code is publicly available. We do not report the solver tion and the experimental results. For each simulation, the intro-
output flag since the solver always converged to an optimal duced orientation error is reported in the last column of Table III.
576 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 15, NO. 3, MAY 2007

TABLE III
SIMULATION RESULTS. CONTROLLER B PRESENTED IN
SECTION VI-B. rms AND MAXIMUM TRACKING ERRORS AS
FUNCTION OF VEHICLE LONGITUDINAL SPEED

_
Fig. 9. Experimental results at 19-m/s entry speed. Controller B described in
( )
Section VI-B. (a) Lateral position Y , yaw angle( ) , and yaw rate . (b) Front
( )
steering angle  , change in front steering angle (1 )  , and front tire slip
angle  .

experiments and simulations. We also observe a chattering phe-

nomenon in the signals and . This is a function of the
MPC tuning and of the constraints and . No
additional effort has been spent in reducing such phenomenon
since it is mostly filtered by the vehicle response and not felt by
the driver.
_
Fig. 8. Experimental results at 10-m/s entry speed. Controller B described in
( )
Section VI-B. (a) Lateral position Y , yaw angle( ) , and yaw rate . (b) Front For all experimental results presented in the following, the
( )
steering angle  , change in front steering angle (1 )  , and front tire slip slip angle has been computed using the experimental data and
angle  . (3), (4), and (7).
In Fig. 9, the experimental results at 19 m/s are presented. By
comparing the lateral positions in Figs. 8(a) and 9(a) at 10 and 19
The orientation error is due to a measurement offset in the yaw m/s, respectively, we observe a larger error in the lateral position
signal coming from the OTS sensor. The effects of this error will at 19 m/s, between 6 and 8 s at the beginning of the second lane
be clarified and described in detail in Section VI-C. In the lower change. A postprocessing of experimental data and the analysis
plot of Fig. 7(b) the front tire slip angle is reported. of simulation data have shown that the constraint (16j) on the tire
In Fig. 8, the experimental results at 10 m/s are presented. By slip angle corresponds to an implicit constraint on the maximum
comparing these results with the simulations at the same speed steering action which decreases with the higher longitudinal ve-
shown in Fig. 7(a) and (b), we observe a good matching between hicle speeds. The commanded wheel steering angle when the
FALCONE et al.: PREDICTIVE ACTIVE STEERING CONTROL FOR AUTONOMOUS VEHICLE SYSTEMS 577

TABLE IV
CONTROLLER B PRESENTED IN SECTION VI-B. EXPERIMENTAL
RMS AND MAXIMUM TRACKING ERRORS AS FUNCTION
OF VEHICLE LONGITUDINAL SPEED

slip angle is at its lower bound decreases with the ve-

hicle longitudinal speed.
We remark that in almost all experimental tests the tire slip
angle violates its constraint in a small amount. This is in agree-
ment with the use of soft constraint and makes the system robust
to driver steering action, as it can be seen in Fig. 9 between 8
and 9 s.
The presented experimental results are summarized in
Table IV. The comparison between simulation and experi-
mental results, in Tables III and IV, respectively, demonstrates
Fig. 10. Effects of the orientation error on lateral position and yaw angle.
that, in spite of a significant model mismatch at high speeds, (a) Undershoot in the lateral position. (b) Steady-state error on lateral position
vehicle stability is achieved in all experimental tests. This is and yaw angle.
enforced by the constraints (16j) which mitigates the effect of
model uncertainties. The mismatch between the closed- and
the open-loop behavior resides in the uncertainty in the tire This explains the initial undershoot in the lateral position. At the
characteristics. Such mismatch led to a conservative choice end of the maneuver (position A in Fig. 10(a)), the vehicle is
for the slip angle constraints in the experimental tests. As actually going straight, as confirmed by the constant lateral po-
a direct consequence, the steering angle has been implicitly sition and zero steering angle, while the measured nonzero yaw
overconstrained in experiments at high vehicle speed with a angle is exactly the orientation error, as illustrated in Fig. 10(a).
consequent performance degradation. This can be observed in Consider Figs. 8, 9, and 11, in spite of the steady-state errors
the experimental tests at the second lane change, where the on both lateral position and yaw angle, the MPC controller does
error in the lateral position becomes large compared to the not attempt to reduce the errors on tracking variables. The expla-
simulation results. nation is simple: because of the offset on the yaw measurement,
By comparing the simulation and experimental results at any attempt to reduce the yaw angle (lateral position) tracking
10 m/s in Tables III and IV, respectively, it can be observed error would imply an increase of the lateral position (yaw angle)
that the experimental rms lateral error is lower than the corre- tracking error according to model (2). Since the proposed MPC
sponding simulative result. In fact, a different equilibrium point does not contain any integral action, the steady-state equilibria
is reached in the two cases. Section VI-C clarifies the cause which can be observed in Figs. 8, 9, and 11 are the closed-loop
of mismatch between simulative and experimental equilibrium optimal equilibria for the designed MPC. The simulation results
point. in Fig. 7 reproduces exactly the steady-state offset observed in
As remarked in previous sections, Controller B performance experiments and confirm our explanation. We remark that an
degrades when the constraint (16j) on the tire slip angle is re- MPC controller with integral action on the tracking errors would
moved. In this case, extensive simulations have shown that the have led to an unstable vehicle behavior. In fact, due to the in-
LTV MPC controller is able to stabilize the vehicle only up to consistency between the orientation and position measurements,
10 m/s. a zero steady-state orientation error would have implied a di-
verging lateral position.
C. Steady-State Errors
In this section, we will explain the initial undershoot in the lat- D. Controller C
eral position and the steady-state errors in both lateral position The controller (16) and (17) with the parameters defined in
and yaw angle observed in the presented experimental results Section VI has been implemented as a C-coded S-function, by
of Controller B. As previously mentioned, both phenomena are using a very simple QP solver for a 2-D problem. Since
caused by an offset on the orientation measurement. Fig. 10 re- , the two optimization variables are the commanded steering
produces the scenario: the axes and are the actual lon- variation and the slack variable. We implemented a tailored QP
gitudinal and lateral vehicle axes, respectively, while and solver.
are the vehicle axes as learned by the OTS sensor. The plot of the simulation results are not included in this man-
Because of a positive orientation error, at the beginning of the uscript because of lack of space. We only report, in Table V, the
test the sensor measures a zero yaw angle while the actual ve- tracking errors at 10-, 15-, 19-, and 21-m/s entry speeds. The
hicle longitudinal axis points rightwards, as shown in Fig. 10(a). experimental results of controller C at 10, 17, and 21 m/s are
578 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 15, NO. 3, MAY 2007

TABLE VI
SUMMARY OF EXPERIMENTAL RESULTS, CONTROLLER C PRESENTED
IN SECTION VI-D. rms AND MAXIMUM TRACKING ERRORS
AS FUNCTION OF VEHICLE LONGITUDINAL SPEED

_
Fig. 11. Experimental results at 21.5-m/s entry speed. Controller B described in
( )
Section VI-B. (a) Lateral position Y , yaw angle ( ) , and yaw rate . (b) Front
( )
steering angle  , change in front steering angle (1 )  , and front tire slip
angle  .

TABLE V
SUMMARY OF SIMULATION RESULTS, CONTROLLER C PRESENTED IN
SECTION VI-D. SIMULATIVE RMS AND MAXIMUM TRACKING
ERRORS AS FUNCTION OF VEHICLE LONGITUDINAL SPEED.
THE ORIENTATION ERRORS HAVE BEEN REPRODUCED

_
Fig. 12. Experimental results at 17 m/s entry speed. Controller C described in
( )
Section VI-D. (a) Lateral position Y , yaw angle ( ) , and yaw rate . (b) Front
( )
steering angle  , change in front steering angle (1 )  , and front tire slip
angle  .

summarized in Table VI, the results at 17 m/s are presented in In order to fairly compare Controllers B and C, the orientation
Fig. 12. offset observed during experimental tests of Controller B have
The plots results do not show any orientation error, being the been reproduced in simulations of Controller C. The tracking
experiments performed in a day different from experiments of errors for such simulation are reported in Table V. Compare the
controller B. It is interesting to observe that, even if the control tracking errors of Controllers B and C in Tables III and V, re-
horizon has been tightened significantly, the controller is still spectively. It can be noticed that Controller C performs slightly
able to stabilize the vehicle, even at high speed, and the tracking worse than Controller B. However, as confirmed by the exper-
errors do not increase significantly compared to Controller B. imental tests, it is able to stabilize the vehicle at high speed.
FALCONE et al.: PREDICTIVE ACTIVE STEERING CONTROL FOR AUTONOMOUS VEHICLE SYSTEMS 579

We emphasize that such an MPC controller can be implemented [5] R. Behringer, B. Gregory, V. Sundareswaran, R. Addison, R. Elsley, W.
on a low-cost hardware and its code can be easily verified. The Guthmiller, and J. Demarche, Development of an autonomous vehicle
for the DARPA grand challenge, presented at the IFAC Symp. Intell.
maximum number of operation at each time step is also easily Autonomous Veh., Lisbon, Portugal, 2004.
computed whereas this can be very difficult for Controllers B [6] F. Borrelli, A. Bemporad, M. Fodor, and D. Hrovat, An MPC/hy-
brid system approach to traction control, IEEE Trans. Control Syst.
and C. Technol., vol. 14, no. 3, pp. 541552, May 2006.
[7] F. Borrelli, P. Falcone, T. Keviczky, J. Asgari, and D. Hrovat, MPC-
VII. CONCLUSION based approach to active steering for autonomous vehicle systems, Int.
J. Veh. Autonomous Syst., vol. 3, no. 2/3/4, pp. 265291, 2005.
We have presented a novel MPC-based approach for active [8] PC. Calhoun and EM. Queen, Entry vehicle control system design
steering control design. Experimental results showed that for the mars smart lander, presented at the AIAA Atmospheric Flight
double lane change maneuvers are relatively easily obtained as Mech. Conf., Monterey, CA, 2002.
[9] L. Chisci, P. Falugi, and G. Zappa, Gain-scheduling MPC of nonlinear
a result of the MPC feedback policy, leading to the capability systems, Int. J. Robust Nonlinear Control, vol. 13, pp. 295308, 2003.
of stabilizing a vehicle with a speed up to 21 m/s on slippery [10] T. Costlow, Active safety, SAE Int., Warrendale, PA, 2005.
[11] dSPACE GmbH, dSPACE autobox, Paderborn, Germany, 2006.
surfaces such as snow covered roads. [12] P. Falcone, F. Borrelli, J. Asgari, H. E. Tseng, and D. Hrovat, Pre-
There are three main contributions of this manuscript. These dictive active steering control for autonomous vehicle systems,
are associated with the three types of MPC controllers which Dipartimento di Ingegneria, Universit del Sannio, Benevento, Italy,
(2007). [Online]. Available: http://www.grace.ing.unisannio.it/publi-
have been designed and experimentally tested. First, a non- cation/416
linear MPC has been designed and implemented on a rapid [13] R. E. Fenton, G. C. Melocik, and K. W. Olson, On the steering of au-
prototyping system. Because of computational burden, experi- tomated vehicles: Theory and experiments, IEEE Trans. Autom. Con-
trol, vol. AC-21, no. 3, pp. 306315, Jun. 1976.
mental tests could be performed only at low vehicle speed on [14] C. E. Garcia, D. M. Prett, and M. Morari, Model predictive control:
the available hardware. Yet, to the best of our knowledge, for Theory and practice-A survey, Automatica, vol. 25, pp. 335348, 1989.
[15] P. Gill, W. Murray, M. Saunders, and M. Wright, NPSOLNonlinear
the first time a nonlinear MPC scheme has been implemented Programming Software. Mountain View, CA: Stanford Business
on a dSPACE rapid prototyping system to control the vehicle Software, Inc., 1998.
dynamics of an autonomous passenger car with a frequency of [16] J. Guldner, W. Sienel, H. S. Tan, J. Ackermann, S. Patwardhan, and T.
Bnte, Robust automatic steering control for look-down reference sys-
20 Hz. Second, an LTV MPC has been designed and imple- tems with front and rear sensors, IEEE Trans. Control Syst. Technol.,
mented. The use of a state and input constraint on the tire slip vol. 7, no. 1, pp. 211, Jan. 1999.
angle has been proposed in order to stabilize the vehicle at high [17] D. Hrovat, MPC-based idle speed control for IC engine, presented at
the FISITA, Prague, Czech Republic, 1996.
speeds. Its effectiveness has been shown through simulations [18] T. Keviczky and G. J. Balas, Flight test of a receding horizon con-
and experiments. Both approaches suffer from the difficulty in troller for autonomous uav guidance, in Proc. Amer. Contr. Conf.,
2005, pp. 35183523.
verifying the optimization code and computing the worst case [19] , Receding horizon control of an f-16 aircraft: A comparative
computational time. The third and last contribution overcomes study, Control Eng. Practice, vol. 14, no. 9, pp. 10231033, Sep. 2006.
these issues and it is represented by an LTV MPC controller [20] , Software-enabled receding horizon control for autonomous
UAV guidance, AIAA J. Guid., Control, Dyn., vol. 29, no. 3, pp.
of low order which shows acceptable performance and can be 680694, May/Jun. 2006.
easily implemented in a low cost hardware. [21] T. Keviczky, P. Falcone, F. Borrelli, J. Asgari, and D. Hrovat, Predic-
The results presented here open the route to interesting and tive control approach to autonomous vehicle steering, in Proc. Amer.
Contr. Conf., 2006, pp. 46704675.
challenging research activities, which are the current topic of [22] M. V. Kothare, B. Mettler, M. Morari, P. Bendotti, and C. M. Fali-
ongoing work. From a computational point of view, we are im- nower, Level control in the steam generator of a nuclear power plant,
proving and testing several nonlinear optimization algorithms IEEE Trans. Control Syst. Technol., vol. 8, no. 1, pp. 5569, Jan. 2000.
[23] T. B. Foote, L. B. Cremean, J. H. Gillula, G. H. Hines, D. Kogan, K. L.
which can enlarge the range of conditions for which a nonlinear Kriechbaum, J. C. Lamb, J. Leibs, L. Lindzey, C. E. Rasmussen, A. D.
MPC becomes real-time implementable. From a control design Stewart, J. W. Burdick, and R. M. Murray, Alice: An information-rich
autonomous vehicle for high-speed desert navigation, J. Field Robot.,
point of view, we are studying new control paradigms in order 2006, submitted for publication.
to achieve similar performance to the nonlinear MPC with a [24] W. M. Lu and D. S. Bayard, Guidance and control for mars at-
significant reduction of the computational burden. We are also mospheric entry: Adaptivity and robustness, Jet Propulsion Lab.,
Pasadena, CA, 1997.
studying the stability and the robustness of the proposed LTV [25] D. L. Margolis and J. Asgari, Multipurpose models of vehicle dy-
MPC scheme. From an application point of view, we are in- namics for controller design, SAE Int., Warrendale, PA, 1991.
creasing the number of control inputs and working on multi- [26] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert,
Constrained model predictive control: Stability and optimality,
variable vehicles dynamic control schemes controlling steering, Automatica, vol. 36, no. 6, pp. 789814, Jun. 2000.
brakes, and throttle. [27] M. Morari and J. H. Lee, Model predictive control: Past, present and
future, Comput. Chem. Eng., vol. 23, no. 45, pp. 667682, 1999.
[28] E. Ono, S. Hosoe, H. D. Tuan, and S. Doi, Bifurcation in vehicle dy-
REFERENCES namics and robust front wheel steering control, IEEE Trans. Control
[1] J. Ackermann, D. Odenthal, and T. Bnte, Advantages of active Syst. Technol., vol. 6, no. 3, pp. 412420, May 1998.
steering for vehicle dynamics control, in Proc. 32nd Int. Symp. [29] J. Hauser, R. M. Murray, A. Jadbabaie, M. B. Miliam, N. Petit, W. B.
Autom. Technol. Autom., 1999, pp. 263270. Dunbar, and R. Franz, Online control customization via optimization-
[2] J. Ackermann and W. Sienel, Robust yaw damping of cars with front based control, in Software-Enabled Control: Information Technology
and rear wheel steering, IEEE Trans. Control Syst. Technol., vol. 1, for Dynamical Systems. Piscataway, NJ: IEEE Press, 2003.
no. 1, pp. 1520, Mar. 1993. [30] M. Montemerlo, S. Thrun, H. Dahlkamp, D. Stavens, A. Aron, J. Diebel,
[3] J. Ackermann, W. Walter, and T. Bnte, Automatic car steering using P. Fong, J. Gale, M. Halpenny, G. Hoffmann, K. Lau, C. Oakley, M.
robust unilateral decoupling, presented at the Int. Conf. Adv. Veh. Palatucci, V. Pratt, P. Stang, S. Strohband, C. Dupont, L.-E. Jendrossek,
Control Safety, Genoa, Italy, 2004. C. Koelen, C. Markey, C. Rummel, J. van Niekerk, E. Jensen, P. Alessan-
[4] E. Bakker, L. Nyborg, and H. B. Pacejka, Tyre modeling for use drini, G. Bradski, B. Davies, S. Ettinger, A. Kaehler, A. Nefian, and P.
in vehicle dynamics studies, SAE Int., Warrendale, PA, 870421, Mahoney, Stanley the robot that won the DARPA grand challenge, J.
1987. Field Robot., accepted for publication.
580 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 15, NO. 3, MAY 2007

[31] R. R. S. Smith, K. Mease, D. S. Bayard, and D. L. Farless, Aeromaneu- Jahan Asgari has received the B.S.M.E. degree from
vering in the martian atmosphere: Simulation-based analyses, AIAA J. California State University, Sacramento, in 1983, and
Spacecraft Rockets, vol. 37, no. 1, pp. 139142, 2000. the M.S. and Ph.D. degrees from University of Cali-
[32] Model predictive control toolbox Inc. The MathWorks, Natick, MA, fornia, Davis, in 1985 and 1989, respectively.
2005. Since then, he has been working at Research and
[33] P. Tndel and T. A. Johansen, Control allocation for yaw stabilization Advance Engineering Department, Ford Motor Com-
in automotive vehicles using multiparametric nonlinear programming, pany, Dearborn, MI, where his activities include sev-
in Proc. Amer. Contr. Conf., 2005, pp. 453458. eral projects in the area of driveline and chassis mod-
[34] H. E. Tseng, J. Asgari, D. Hrovat, P. Van Der Jagt, A. Cherry, and S. eling, control, and optimization.
Neads, Evasive maneuvers with a steering robot, Veh. Syst. Dyn., vol.
43, no. 3, pp. 197214, Mar. 2005.
[35] Z. Wan and M. V. Kothare, Efficient scheduled stabilizing model pre-
dictive control for constrained nonlinear systems, Int. J. Robust Non-
linear Control, vol. 13, pp. 331346, Mar./Apr. 2003.
[36] W. Zhang and R. E. Parsons, An intelligent roadway reference system Hongtei Eric Tseng received the B.S. degree from
for vehicle lateral guidance/control, in Proc. Amer. Contr. Conf., National Taiwan University, Taipei, Taiwan, R.O.C.,
1990, pp. 281286. in 1986, and the M.S. and Ph.D. degrees from the
University of California, Berkeley, in 1991 and 1994,
respectively, all in mechanical engineering.
He is currently a Technical Leader at the Research
Paolo Falcone received the Laurea degree in com- and Innovation Center, Ford Motor Company, Dear-
puter science engineering from the University of born, MI, where he has been since 1994. His previous
Naples Federico II, Naples, Italy, in 2003. He is work includes low pressure tire warning system using
currently pursuing the Ph.D. degree in automatic wheel speed sensors, traction control, electronic sta-
control engineering at the Universita del Sannio, bility control, and roll stability control. His current
Benevento, Italy. research interests include both powertrain and vehicle dynamics control.
His research interests include model predictive
control and vehicle dynamics control.

Davor Hrovat (F07) received the Dipl. Ing. degree

from the University of Zagreb, Croatia, in 1972,
and the M.S. and Ph.D. degrees in mechanical engi-
neering from the University of California, Davis, in
Francesco Borrelli received the Laurea degree in 1976 and 1979, respectfully.
computer science engineering from the University of Since 1981, he has been with the Ford Research
Naples Federico II, Naples, Italy, in 1998, and the Laboratory (FRL), Dearborn, MI, where he is a
Ph.D. degree in automatic control from ETH Zurich, Henry Ford Technical Fellow coordinating and
Zurich, Switzerland, in 2002. leading research efforts on various aspects of ve-
He is currently an Assistant Professor at the Uni- hicle/power train control systems, where he holds
versit del Sannio, Benevento, Italy. He was a Re- numerous patents. He has served as an Associate
search Assistant at the Automatic Control Labora- Editor and as a member of the Board of Editors for a number of ASME, IEEE,
tory, ETH Zurich, and a Contract Assistant Professor IFAC, and other journals.
at the Aerospace and Mechanics Department, Univer- Dr. Hrovat was a recipient of the 1996 ASME/Dynamic Systems and Control
sity of Minnesota, MN. He is the author of the book Innovative Practice Award and the 1999 AACC Control Engineering Practice
Constrained Optimal Control of Linear and Hybrid Systems (SpringerVerlag, Award. He was recently elected to the National Academy of Engineering.
2003). His research interests include constrained optimal control, model pre-
dictive control, robust control, parametric programming, singularly perturbed
systems, and automotive applications of automatic control.
Mr. Borrelli was a recipient of the Innovation Prize 2004 from the Elec-
troSwiss Foundation.