Proceedingsof tbe 38* ThP07 17:20

Conference on Decision & Control

Phoenix, Arizona USA December1999

Dynamic Tire Friction Models for Vehicle Traction Control

CARLOSCANUDAS DE WIT^
Laboratoire d’Automatique de Grenoble, UMR CNRS 5528
ENSIEG-INPG, B.P. 46, 38 40.2 ST. Martin d’lieres, FRANCE
PANAGIOTIS TSIOTRAS2
School of Aerospace Engineering
Georgia Institute of Technology, Atlanta, GA 30332-0150, USA

Abstract The static model shown in Fig. 1is derived empirically

based solely on steady-state experimental data [lo, 11.
In this paper we derive a dynamic friction force model Under steady-state conditions, experimental data seem
for road/tire interaction for ground vehicles. The model t o support the force vs. slip curves of Fig. 1. Neverthe-
is based on a similar dynamic friction model for contact less, the development of friction force at the tirelroad
developed previously for contact-point friction prob- interface is very much a dynamic phenomenon. In other
lems, called the LuGre model [4]. We show that the words, the friction force does not reach its steady-state
dynamic LuGre friction model is able t o accurately cap- instantaneously, but rather exhibits significant tran-
ture velocity and roadlsurface dependence of the tire sient behavior which may differ significantly from its
friction force. steady-state value. Experiments performed in commer-
cial vehicles, have shown that the tirelroad forces do
not vary along the curves shown Fig. 1, but “jump”
from one value t o an other when these forces are dis-
1 Introduction played in the p - s plane [15].

The problem of traction control for ground vehicles is of In this paper, we develop new, speed-dependent, dy-
enormous importance t o automotive industry. Traction namic friction models that can be used t o describe the
control systems reduce or eliminate excessive slipping tirelroad interaction. These models have the advan-
or sliding during vehicle acceleration and thus enhance tage that are developed starting from first principles
the controllability and maneuverability of the vehicle. and are based on simple contact (punctual) dynamic
Proper traction control design will have a paramount friction models [4].Thus, the parameters entering the
effect on safety and handling qualities for future pas- models have a physical significance which allows the
senger vehicles. Traction control aims to achieve max- designer to tune the model parameters based on exper-
imum torque transfer from the wheel axle to forward imental data. The models are also speed-dependent,
acceleration. The friction force in the tirelroad inter- which agrees with experimental observations. A sim-
face is the main mechanism for converting wheel angu- ple parameter in the model can also be used to capture
lar acceleration (due t o the motor torque) to forward the road surface characteristics. Finally, our model is
acceleration (longitudinal force). Therefore, the study shown to be well-defined everywhere and hence, is ap-
of friction force characteristics at the roadltire interface propriate for control law design.
has received a great deal of attention in the automotive
literature.

A common assumption in most of tire friction models 2 Tirelroad friction models

is that the normalized tire friction p
F Friction force In this study we consider a system of the form:
p=-=
F,, Normal force
is a nonlinear function of the normalized relative ve-
locity between the road and the tire (slip coefficient s)
with a distinct maximum; see Fig. 1. It is also under-
stood that p depends also on the velocity of the vehicle where m is 114 of the vehicle mass and J , T are the
and road surface conditions, among other factors (see inertia and radius of the wheel, respectively. v is the
[3] and [lo]). The curves shown in Fig. 1illustrate how linear velocity, w is the angular velocity, U is the ac-
these factors influence the shape of p. celerating (or braking) torque, and F is the tirelroad
friction force. For the sake of simplicity, only longitu-
‘Directeur de Recherche. Corresponding author. Email: dinal motion will be considered. The dynamics of the
canudasQlag.ensieg.inpg.fr braking and driving actuators are also neglected.
‘Associate Professor. Email: p. tsiotrasaae.gatech.edu

0-7803-5250-5/99/$10.000 1999 IEEE 3746

 Wheel with Wheel with
lumped friction P distributed friction F

Figure 2: One-wheel system with: lumped friction (left),

distributed friction (right)

particular conditions of constant linear and angular ve-

locity. The Pacejka model has the form

F ( s ) = c1 sin(c2 arctan(c3s - cq(c3s - a r c t a n ( s s ) ) > ) ,

where the 4 s are the parameters characterizing this
otui 4 model. These parameters can be identified by match-
ing experimental data, as shown in Bakker et al. [l].
The parameters ci depend on the tire characteristics
(such as compound, tread type, tread depth, inflation
pressure, temperature), on the road conditions (such as
type of surface, texture, drainage, capacity, tempera-
Figure 1: Typical variations of the tire/road friction pro- ture, lubricant, i.e., water or snow), and on the vehicle
files for: different road surface conditions (top), operational conditions (velocity, load); see Pasterkamp
different vehicle velocities (bottom). Curves ob- and Pacejka [12].
tained by Harned in et al [lo].
As a n alternative t o the static F ( s ) maps, dynamic
models based on the dynamic friction models of Dah1
2.1 Slip/Force maps ["I1, can be adapted t o suitably describe the road-tire
The most common tire friction models used in the lit- contact friction. Dynamic models can be formulated as
erature are those of slip/force maps. They are defined a lumped or distributed models, as shown in Fig. 2.
as oneto-one (memory-less) maps between the friction This distinction will be discussed next.
F , and the longitudinal slip rate s, defined as:
1- if v > T W , v # 0 braking 2.2 Lumped models
A lumped friction model assumes punctual tire-road
s={ (3) friction contact. An example of such a model can be
1- & if v < T W , W # 0 driving
derived from the LuGre model2 (see Canudas et al,
[4]), i.e.
The slip rate results from the reduction of the effec-
tive circumference of the tire (consequence of the tread
deformation due t o the elasticity of the tire rubber), (4)
which implies that the ground velocity will not be equal
t o II = rw. The slip rate is defined in the interval [0,1].
When s = 0 there is no sliding (pure rolling), whereas
s = 1 indicates full sliding.
The slip/force models aim at describing the shapes
shown in Fig. 1 via static maps F ( s ) : s c+ F . They 'Dahl's models lead to a friction displacement relation that
bears much resemblance with stress-strain relations proposed in
may also depend on the vehicle velocity U , i.e. F ( s ,U), classical solid mechanics.
and vary when the road characteristics change. 'This model differs from the one in [4] in the way that the func-
1
tion g ( v ) is defined. Here we propose to use the term e-Iwp/wa12
One of the most well-known models of this type is Pace-
instead the term e - ( u ~ / u ~ ) 2as in the LuGre model in order t o
jka's model (see, Pacejka and Sharp [13] ), also known better match the pseudo-stationary characteristic of this model
as the "magic formula". This model has been shown (map s ct F ( s ) ) with the shape of the Pacejka's model, as it will
to suitably match experimental data, obtained under be shown later.

3747
where 00 is the normalized rubber longitudinal lumped Nevertheless, it is also possible t o include different nor-
stiffness, 01 the normalized rubber longitudinal lumped mal force distribution if necessary, i.e. 6Fn = f(c).
damping, 02 the normalized viscous relative damping,
p c the normalized Coulomb friction, ps the normal- Note that Eq. (6) describes a partial differential equa-
ized Static friction,(pc <
ps E [0,l]),vs the Stribeck tion (PDE), i.e.
relative velocity, F, the normal force, v, = (rw - U ) the
relative velocity, and a the internal friction state.

Dynamic friction models specifically for tires have been \-,

reported in the work of Clover and Bernard [6], where that should be solved in both: time and space.
they develop a differential equation for the slip coeffi-
cient, starting from a simple relationship of the rela- 2.4 Relation between distributed model and the
tive reflections of the tire elements in the tire contact magic formula
patch. They still use the semi-empirical static force/slip The linear motion of the differential bF in the patch
models, however, to compute the corresponding fric- frame 0, is C = rw, for positive w, and [ = -rw, for
tion force. In that respect, such models can be best negative w (the frame origin changes location when the
described as quasi-dynamic models.
wheel velocity reverses). Hence ( = rlwl. We can thus
rewrite (6) in the [ coordinates as:
2.3 Distributed models
Distributed models assume the existence of an area of
contact (or patch) between the tire and the road, as
shown in Fig. 2. This patch represents the projection
of the part of the tire that is in contact with the road. where s = v,./wr = 1 - v/wr. Assuming that U , and
The contact patch is associated t o the frame 0,, with w are constant (hence also U,, and s), the above equa-
C as the axis coordinate. The patch length is L. tion describes a linear space-invariant system having
the sign of the relative velocity as its input.
Distributed dynamical models, have been studied pre-
viously, for example, in the works of Bliman et al. [2]. The solution of the above equation over the space
In these kinds of models, the contact patch area is dis- ] ,equivalent over [CO,[I], with
interval [ [ ( t o ) , [ ( t ~ )or
cretized t o a series of elements, and the microscopic bz([O) = (0 = 0 is
deformation effects are studied in detail. In particu-
lar, Bliman at al. characterize the elastic and Coulomb
friction forces at each point of the contact patch, but
then they give the aggregate effect of these distributed
forces by integrating over the whole patch area. They Introducing this solution together with Eq. (9) in Eq.
propose a second order rate-independent model (similar (7),and integrating, we obtain
to Dahl’s model), and show that, under constant v and
w, there exist a choice of parameters that closely match
a curve similar t o the one characterizing the magic for-
mula.
and using (8) we obtain
Similar results can be obtained by using a model based
in the first-order LuGre friction model, i.e.

Finally, we have that F ( s ) ,is given as

with g(vr) defined as before and + F,O~TWS

with 7 = 1- O1lVrl/g(S) and

where, dF is the differential friction force, SFn = F,/L g(s) = pc + (ps - p c ) e-Ipw5/va14

the differential normal force, V p = (rw - v) the relative for some constant w , and s E [0, 11.
velocity, and 6z the differential internal friction state.
This model assumes that: Uncertainty in the knowledge of the function g ( v r ) , can
be modeled by introducing the parameter 0, as
0 the normal force is uniformly distributed, and
8(vr) = eg(vr) ,
0 the contact velocity of each differential state ele- where g(v,) is the nominal known function. Computa-
. ment is equal to U,. tion of the function F ( s ,e), from Eq. (12) as a function
3748
 Parameter Value Umts dzi(t) = 6zi. Similarly, we have that

0.0018 [s/m]

PS 0.9
' vs 12.5 [m/s]
where each of the
ward differences, as:
can be approximated using for-

Table 1: Data used for the plot shown in Fig. 3

-={
d dzi
4
6zi+1-62;
7
0
i = 0 , 1 , ... N - 2
i=N-1
of 8, gives the curves shown in Fig. 3. These curves Hence, for each i-th equation we have
match reasonably well the experimental data shown in
Fig. 1-(a), for different coefficient of road adhesion us-
ing the parameters shown in Table 1. Hence, the pa-
rameter 8, suitably describes the changes in the road
characteristics. Similarly,with AF,,i = F, f L , Vi, and A< = LfN, F
can be approximated as:
N -1 N-1
F= AFi = (a0 6zi + ~ 1 6 i iAFn,iAC+~2vrFn
)
i=O i=O

which simplifies to:

~ N-1

Introducing Z, as the mean value of all the 6zi, i.e.

~ N-1
z- = - 6Zi
N
i=O

we have, from Eq. (14),that

Figure 3: Static view of the distributed LuGre model, un-

der different values for l/O. Braking case, with P. = --
1 N-l (6~i+l- 6zi) TW +ZU,d - -; o~ Iv I (15)
v = 20m/s = 72Km/h. These curves show the L i=O dVP)

normalized friction p = F ( s )fFn, as a function

of the slip rate s. Noticing that E&' (Szi+l - dzi) = 6x0, and taking
620 = 0, we have that

Note that the steady-state representation of Eq. (12)

can be used t o identify the model parameters by feeding
this model t o experimental data. These parameters can F = (002 + O I + CzV,)
~ Fn (17)
also be used in the simpler lumped model, which can be
shown t o suitably approximate the solution of the PDE These equations describe the approximate behaviour of
described by Eqs. (6) and (7). This approximation is the PDE, in terms of the mean variable P . When com-
discussed next. pared to Eqs. (4)-(5), they indicate that the lumped
model can be used as a suitable approximation of the
distributed one. Therefore, parameters identified from
2.5 From distributed to lumped models the stationary behaviour shown in Fig. 3, can be used
Under the assumptions given in subsection 2.3 we can in the lumped model (4)-(5).
approximate the PDE in Eqs. (6)-(7) by a set of n
ordinary differential equations via a spatial discretiza-
tion. To this end, let's divide the contact patch into
N equally spaced discrete points, t o each we associate 3 Traction Control
the "discrete" average displacement 6zi, i.e. dzi =
6z(iL/N,t ) , V i = 0,1,. , .N - 1. The space/time scalar We consider the one-wheel model with the tirelroad
dz(C, t ) is thus approximated by the N-dimensional friction described in Eqs. (1)-(2). Using the pseudc-
time vector, 6z = [dzo, 6z1,. . . ,6zN-I]T where, for the static (or steady-state) force friction point of view, the
sake of simplicity of the notation, we have written friction force is given as an algebraic (static) function of
3749
the slip coefficient. Typical friction force vs. slip coeffi- x 10‘

cient curves are shown in Fig. 1. This figure suggests a II I

simple way to achieve maximum traction between the
road and the wheel tire. Namely, to operate at the max-
imum point of the friction/slip curve. This “extremum
seeking” control strategy requires the a priori knowl-
edge of the optimal target slip. With the exception
of [8], where the authors present a control algorithm
which does not require the a priori knowledge of the
5-:-:;yi
-2

8 3000
l?
‘ 0 05 01 0 15
Time
02 0 25
I
03

optimal slip, current literature does not seem to have 2 1000

adequately dealt with this problem. Nonetheless, slip 0
and friction estimation algorithms have been proposed 0.05 0.1 0.15 0.2 025 0.3

and verified experimentally in [12]. 11 , ,

TimO
I

A simple traction control law using this idea, and based

on sliding mode techniques, is given in [9]. For the
simplified one-wheel friction model of Eqs. (1)-(2) this 0‘ I
0 01 01 0 15 02 0 25 03
control law is given by T W

r T 1

U = [m(l-sd)
+T
I F-ksgn(S) (18) Figure 4: Static friction model,

where Sd is the desired slip coefficient, S is given by

s = (9 - S d ) TW and was performed using the dynamic friction model given
in Eqs. (4)-(5) with the values shown in Table 1. The
results of the simulation are shown in Fig. 5. In both
cases, the initial applied torque for both static and dy-
namic cases was U = 10000Nm. The history profiles of
Indeed, simple calculation shows that

s = (1 - Sd) TdJ - v (20)

x lo‘
I I I

Substituting Eq. (18) into Eq. (20) and using Eqs. (1)-
(2), one obtains
4’ I
S = -qsgn(S) (21) 150001
0.05 0.1 0.15
Tim
0.2 025 0.3
I

This implies that S + 0 after S(O)/q seconds. The

major drawback of the control law in Eq. (18) is that
is highly oscillatory due t o the zero order sliding mode
S = 0. One can reduce the chattering by smoothing 0.05 0.1 0.15 0.2 025
I
0.3
the discontinuity of sgn(.) via low-pass filtering. In the Tim

smoothed implementation of the previous ccntrol law,

the term ksgn(S) is replaced by the term ksat(S/@),
where = XQ, and where sat(.) is the saturation func-
tion. For more details on the previous control law, the
interested reader is referred t o [9]. 0 0 05 01 0 15 02 025 03
TIM

Figure 5: Dynamic friction model.

4 Numerical example

In this section we use the traction control law of the pre- the applied torque and the slip coefficient are very sim-
vious section on two different friction models. In par- ilar. The most serious discrepancy between Figs. 4 and
ticular, we are interested in differences between static 5 is the actual friction force developed between the tire
and dynamic friction models. We consider the one- and the ground for the two cases. These figures show
wheel model with the values shown in Table 1, and: clearly that the maximum friction force predicted us-
m = 500 Kg, J = 0.2344 Kgm2, T = 0.25m, F,, = mg. ing the dynamic friction model is more than three times
The first simulation was performed using the steady- the maximum of the friction force predicted using the
state LuGre model from Eq. (12). The relevant param- static friction model during the initial transient. The
eters of the LuGre friction model are shown in Table 1. steady-state value of the friction force for both cases is
The maximum traction is achieved for s d = 0.15. The almost the same. Since the main mechanism for trans-
results of the simulations, using the smoothed version ferring the axle torque to forward movement is friction
of the traction control law presented in the previous force, these results suggest that new traction control
section, are shown in Fig. 4. The second simulation algorithms using dynamic friction models may have an
3750
advantage over traditional control laws based on track- Acknowledgements
ing the optimal slip coefficient. Finally, Fig. 6 shows
the distances traveled by the wheel for each case, along The LuGre version of the dynamic friction model was de-
with the path of a point at the circumference of the veloped during the visit of the first author at the School
wheel. A complete circle indicates complete slipping of Aerospace Engineering at the Georgia Institute of Tech-
(the wheel spins without moving forward) whereas a nology during December 1998, as part of the CNRS/NSF
cycloid indicates that the relative velocity of the con- collaboration project (NSF award no. INT-9726621/INT-
tact point is zero. Because of the higher friction force 9996096). The first author would like to thank M. Sorine
developed in the dynamic friction model, the wheel has and P.A. Bliman for interesting discussions on distributed
traveled a longer distance than for the static friction friction models.
case.

References
[l] Bakker, E., L. Nyborg, and H. Pacejka, “Tyre Mod-
elling for Use in Vehicle Dynamic Studies,” Society of Au-
tomotive Engineers Paper # 870421, 1987.
[2] Bliman, P.A., T. Bonald, and M. Sorine, “Hysteresis
Operators and Tire Friction Models: Application to vehicle
dynamic Simulator,” Prof. of ICIAM. 95, Hamburg, Ger-
many, 3-7 July, 1995.
[3] Burckhardt, M., Fahnuerktechnik: Radschlupfregel-
systeme. Vogel-Verlag, Germany, 1993.
(a)Wheel trajectory with static friction model. [4] Canudas de Wit, C., H. Olson, K. J. Astrom, and
P. Lischinsky, “A New Model for Control of Systems with
Friction,” IEEE TAG, Vol. 40, No. 3, pp. 419-425, 1995.
[5] Canudas de Wit, C., R. Horowitz, R. and P. Tsio-
tras, “Model-Based Observers for Tire/Road Contact Fric-
tion Prediction,” In New Directions in Nonlinear Observer
Design, Nijmeijer, H. and T.1 Fossen (Eds), Springer Ver-
lag, Lectures Notes in Control and Information Science,
May 1999.
[6] Clover, C. L. and J. E. Bernard, “Longitudinal Tire
4 6 4 4 6 2 0 02 04 06 08 1 12 14
Dynamics,” Vehicle System Dynamics, Vol. 29, pp. 231-259,
1998.
(b) Wheel trajectory with dynamic friction
model. [7] Dahl, P. R., “Solid Friction Damping of Mechanical
Vibrations,” A I A A Journal, Vol. 14, No. 12, pp. 1675-1682,
1976.
[8] Drakunov, S., U. Ozgiiner, P. Dix, and B. Ashrafi,
Figure 6: Comparison of wheel trajectories using static “ABS Control Using Optimum Search via Sliding Modes,”
and dynamic tire friction models (note different IEEE Pansactions on Control Systems Technology, Vol. 3,
stopping points). NO. 1, pp. 79-85, 1995.
(91 Fan, Z., Y . Koren, and D. Wehe, A Simple Traction
Control for Tracked Vehicles. In Proceedings of the Ameri-
can Control Conference, pp. 1176-1177, Seatlle, WA, 1995.
[lo] Harned, J., L. Johnston, and G. Scharpf, Measure-
ment of Tire Brake Force Characteristics as Related to
5 Conclusion Wheel Slip (Antilock) Control System Design. S A E Z h n s -
actions, Vol. 78 (Paper # 690214), pp. 909-925. 1969.
Ill] Liu, Y . and J. Sun, “Target Slip Tracking Using Gain-
In this paper, we have derived a new dynamic, speed- Scheduling for Antilock Braking Systems,” In The American
and surface-dependent tire friction model for use in ve- Control Conference, pp. 1178-82, Seattle, WA, 1995.
hicle traction control design. This model captures very [12] Pasterkamp, W. R. and H. B. Pacejka, “The Tire as
accurately most of the main characteristics that have a Sensor to Estimate Friction,” Vehicle Systems Dynamics,
been discovered via experimental data. It was also Vol. 29, pp. 409-422, 1997.
shown that distributed models can collapse t o a lumped [13] Pacejka, H. B. and R. S. Sharp, “Shear Force Devel-
model, which is rich enough t o capture the main dy- opments by Psneumatic tires in Steady-State Conditions: A
namic characteristics. Since the main mechanism for Review of Modeling Aspects,” Vehicle Systems Dynamics,
transferring the axle torque t o forward movement is Vol. 20, pp. 121-176, 1991.
friction force, these results suggest that new traction [14] Wong J. Y . , Theory of Ground Vehicles. John Wiley
control algorithms using dynamic friction models may & Sons, Inc., New York, 1993.
have an advantage over traditional control laws based [15] van Zanten, A., W. D. Ruf, and A. Lutz, “Measure-
on simple, static friction models. Although for the sake ment and Simulation of Transient Tire Forces,” in Interna-
tional Congress and Ezposition, (Detroit, MI). SAE Tech-
of brevity the discussion has been restricted to traction nical Paper # 890640, 1989.
control, the results of the paper have an immediate ap-
plication t o the design of ABS systems.
3751