The chattering problem in sliding mode systems

J. Guldner V. I. Utkin
BMW Technik GmbH Dept. Of Electrical Eng., The Ohio State Univ.
Hanauer Strasse 46 205 Dreese Lab, 2015 Neil Ave.
D-80788 München Columbus, OH 43210-1272
Germany USA
juergen.guldner@bmw.de utkin@ee.eng.ohio-state.edu

Keywords: Sliding mode control, chattering, observers ing, in the terminology used here, describes undesired sys-
tem oscillations with finite frequency caused by system
Abstract imperfections.
Almost ever since sliding mode ideas have been put for-
ward, the audible noise some sliding mode controllers ex- 2 Problem Analysis
hibit has irritated control engineers and often has led to
An ideal sliding mode system is depicted in Figure 1. For
resentments and even rejection of the technique. The phe-
this example, a simulation is depicted in Figure 2.
nomenon is best known as “chattering”. Two main causes
have been identified: First, fast dynamics in the control loop M
which were neglected in the system model, are often excited xd ( t ) s( t ) w(t ) x (t )
-M
by the fast switching of sliding mode controllers. Second,
digital implementations in micro-controllers with fixed
x& =...
-
sampling rates may lead to discretization chatter. The latter
was discussed e.g. in [Utkin 1993]. In this paper, the first Controller d (t ) Plant
section, analyses the chattering phenomenon due to the first
cause in detail. The subsequent sections discuss four solu-
tions. Mathematical details can be found in [Utkin et al. Figure 1: Block diagram of ideal sliding mode con-
1999] trol loop. A discontinuous controller forces the out-
put x ( t ) of the plant to exactly track the desired
1 Introduction trajectory xd ( t ) . No chattering occurs since the
The term “chattering” describes the phenomenon of finite- control loop is free of unmodeled dynamics.
frequency, finite-amplitude oscillations appearing in many (a) Output and desired output
sliding mode implementations. These oscillations are 1

caused by the high-frequency switching of a sliding mode 0.8

controller exciting unmodeled dynamics in the closed loop. x(t)
State space

0.6
‘Unmodeled dynamics’ may be those of sensors and actua-
0.4
tors neglected in the principal modeling process since they
x d(t)
are generally significantly faster than the main system dy- 0.2

namics. However, since ideal sliding mode systems are 0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
infinitely fast, all system dynamics should be considered in Time t
the control design. (b) Control inputs and sliding variable

Fortunately, preventing chattering usually does not require a 2

detailed model of all system components. Rather, a sliding
Control space

1 ueq(t)
mode controller may be first designed under idealized as- 0 s(t)
sumptions of no unmodeled dynamics. In a second design −1
u(t)
step, possible chattering is to be prevented by one of the −2
methods discussed Section 3 of this contribution. The solu- −3
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
tion of the chattering problem is of great importance when Time t
exploiting the benefits of a sliding mode controller in a real-
life system. To some extend, chattering, without proper Figure 2: Ideal sliding mode in first order system.
treatment in the control design, has been a major obstacle State x ( t ) converges to desired xd ( t ) in finite time,
i.e. s( t ) = 0 after t ≈ 0.45 sec . Thereafter, control
for implementation of sliding mode to a wide range of ap-
plications. It should be noted that the switching action itself
as the core of a continuous-time sliding mode system is not
u(t ) switches with infinite frequency and shows as a
referred to as chattering since in the ideal case, the switch- black area. Equivalent control ueq ( t ) , the average of
ing is intended and its frequency tends to infinity; chatter- discontinuous control, is drawn as a dashed line.
In practical applications, unmodeled dynamics in the closed 2 1 1
loop like actuators often prevent ideal sliding mode to occur V& = x(− v − 2 w + 2 u ) (3)
and cause fast, finite amplitude oscillations. Figure 3 shows µ µ µ
a block diagram of the closed control loop including the for small magnitudes of v and w . This means that the
previously neglected actuator dynamics. An example simu-
lation is shown in Figure 4. motion is unstable in an ε(µ) -order vicinity of the mani-
fold s ( x ) = x = 0.
xd (t ) s( t ) M u (t ) w(t ) x (t ) The motion equations (1) may also be written as
µ2w
&& = (..) x& =...
-
-M
x& * = − M sign( x)
(4)
Controller Actuator d (t ) Plant
µ 2 &x& + 2µx& + x = x * .
Figure 3: Control loop with actuator dynamics neglected in Sliding mode can not occur in the systems since the time
ideal control design. Sliding mode does not occur since the derivative x& is a continuous time function and can not have
actuator dynamics are excited by the fast switching of the its sign opposite to x in the vicinity of the point x = 0
discontinuous controller, leading to chattering in the loop. where the control undergoes discontinuities.

The value of x& is bounded and, as follows from the singu-

(a) Output and desired output
*

1 x(t)
lar perturbation theory (see e.g. [Kokotovic et al. 1976] or
[Kokotovic 1984]), the difference between x and x* is of
State space

0.5
µ -order. The signs of x and x* coincide beyond the
ε(µ) -vicinity of s ( x) = x = 0 , hence the magnitudes of
x (t)
d

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x* and x decrease, i.e. the state trajectories converge to
Time t
(b) Control inputs and sliding variable
this vicinity and after a finite time interval t1 the state re-
4
mains in the vicinity. According to the analysis of the equa-
2 u(t) tions (1)-(3), the motion in the vicinity of x = 0 is unsta-
Control space

s(t)
ble.
0

w(t) The fact of local instability explains why chattering may

−2
appear in the systems with discontinuous controls at the
−4 presence of unmodeled dynamics. The high frequency os-
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time t cillations in the discontinuous control system may be ana-
lyzed in time domain as well (see also Figure 4): The brief
Figure 4: Chattering in first order system with sec- periods of divergence occur after switches of the control
ond order actuator dynamics under discontinuous input variable u (t ) when the output w( t ) of the actuator is
control. After switches in control u( t ) , actuator unable to follow the abrupt change of the control command.
output w( t ) lags behind, leading to oscillatory The proposed solutions to the chattering problem thus focus
system trajectories. on either avoiding control disontinuities in general or move
the switching action to a controller loop without any un-
To qualitatively illustrate the influence of unmodeled dy-
modeled dynamics. The remainder of this paper discusses
namics on the system behavior, consider a simple case with
various types of chattering prevention schemes and exam-
x d (t ) = 0 and motion equations
ines their respective benefits.

x& = w 3 Chattering Suppression Methods

w& = v (1) The next sections discuss four solutions:
2 1 1 •
v& = − v − 2 w + 2 u. The boundary layer solution: a continuous approxima-
µ µ µ tion of the discontinuity.
• The observer-based solution: generating sliding mode
in a observer loop without unmodeled dynamics.
For the control u = − M sign(x) , the sign-varying Ly- • The regular form solution: limiting sliding mode to an
apunov function inner control loop of a cascaded control structure.
• The disturbance rejection solution: generating integral
V = xv − 0.5w2 (2) sliding mode in an auxiliary control loop.
has a negative time-derivative
3.1 Boundary layer solution
The boundary layer solution, proposed e.g. by [Slotine and (a) Output and desired output
Sastry 1983] and [Slotine 1984], seeks to avoid control
discontinuities and switching action in the control loop. The 1 x(t)

discontinous control law is replaced by a saturation function

State space
which approximates the sign(s) term in a boundary layer of
the sliding manifold s(t ) = 0. Numerous types of saturation
0.5
x (t)
d
functions sat(s) have been proposed in the literature.
“In the large”, i.e. for s( t ) > ε , sat( s) = sign( s) . How-
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time t
ever, in a small ε -vicinity of the origin, the so-called (b) Control inputs and sliding variable

boundary layer, sat( s) ≠ sign( s) is continuous. As an

2
illustrative example, consider a simple linear saturation

Control space
1 u(t)
function s(t)
0

 −1

Msign (s(t) ) for s (t) > ε

w(t)
−2

u (t) =  −3
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

 M
Time t

 ε s (t) for s (t) ≤ ε

Figure 6: Saturation function approximating control dis-

M continuity in boundary layer yields chattering free system
with linear proportional feedback gain7 within the
ε trajectories. State x ( t ) converges to desired xd ( t ) , but
boundary layer in the vicinity of the origin, s(t) ≤ ε , and does not track exactly as in ideal sliding mode.
symmetrically saturated by M for s(t) > ε outside the
3.2 Observer-based solution
boundary layer. A block diagram of the example system
under the above control is shown in Figure 5; simulation The boundary layer approach avoids generating sliding
results are displayed in Figure 6. mode by replacing the discontinuous switching action by a
continuous saturation function. In many applications, how-
ever, control discontinuities are inherent to the system, e.g.
xd ( t ) s( t ) M u(t ) w( t ) x(t ) in various voltage inputs of power converters or electric
-M
µ 2
w
&& = (..) x& =... drives. When implementing a continuous controller, a tech-
- nique like pulse-width modulation (PWM) has to adapt the
Controller Actuator d (t ) Plant
control law to the discontinuous system inputs. In the light
of recent advances of high-speed circuitry, it seems unjusti-
fied to by-pass a system’s discontinuous control inputs by
converting a continuous controller e.g. via a PWM scheme.
Figure 5: Saturation function replaces discontinuous con- Rather, such system specifications call for alternative meth-
troller. Instead of achieving ideal sliding mode, the system ods to prevent chattering while preserving control disconti-
trajectories are confined to a boundary layer of the mani- nuities.
fold s( t ) = 0.
An asymptotic observer in the control loop can eliminate
One of the benefits of the boundary layer approach is that chattering despite discontinuous control laws. The key idea
sliding mode control design methodologies can be exploited as proposed by [Bondarev et al. 1985] is to generate ideal
to derive a continuous controller. The invariance property sliding mode in an auxiliary observer loop rather than in the
of sliding mode control is partially preserved in the sense main control loop. Ideal sliding mode is possible in the
that the system trajectories are confined to a δ ( ε ) -vicinity observer loop since it is entirely generated in the control
of the sliding manifold s( t ) = 0, instead of exactly to software and thus does not contain any unmodeled dynam-
ics. The main loop follows the observer loop according to
s(t ) = 0 as in ideal sliding mode. Within the δ ( ε ) - the observer dynamics. Despite applying a discontinuous
vicinity, however, the system behavior is not determined, control signal with switching action to the plant, no chat-
i.e. further convergence to zero is not guaranteed. This type tering occurs and the system behaves is as if an equivalent
of control design is part of a class of robust controllers
continuous u eq (t ) control was applied. The definition of
which satisfy the “globally uniform ultimate boundedness”
condition proposed by [Leitmann 1981]. Note that no real equivalent control can be found e.g. in [Utkin et al. 1999].
sliding mode takes place since the switching action is re- Figure 7 shows a block diagram, simulation results can be
placed by a continuous approximation. found in Figure 8.
 Controller Actuator Plant
basic idea is to design a cascaded controller (see Figure 9)
xd ( t ) s$( t ) u( t ) w( t ) x(t )
M in two steps. In the first step, a continuous controller is
µ2w
&& = (..) x& =... derived for the plant under the assumption that the plant
-M
- d (t )
input(s) is/are the actual control input(s) to the overall sys-
Auxilliary Main
tem, defining ‘desired’ actuator output(s) wd ( t ) . In the
observer
loop
control
loop
second step, the actuator input(s) u( t ) , i.e. the real control
input(s) of the system, is/are used to ensure the actuator
x$ (t ) x&$ =... output(s) track the desired output(s) exactly via sliding
mode control with w( t ) = wd ( t ) . This approach is a spe-
cial case of cascaded control structures as applied the block
control principle, see e.g. [Drakunov et al. 1990] and the
Figure 7: Control loop with auxiliary observer loop. Ideal integrator backstepping method, see e.g. [Krstic et al.
sliding mode occurs in observer manifold s$ ( t ) = 0 since 1995]. See Figure 10 for simulation results.
the observer loop is free of unmodeled dynamics.
Auxiliary Discontinuous
(a) Output, desired output and observed output controller controller Actuator Plant
x d ( t ) xe wd M u (t ) w( t ) x (t )
1 x(t)
wd =... µ2w
&& = (..) x& =...
-M
State space

^x(t) - - d (t )

Actuator control loop

0.5
x (t)
d
Auxiliary control loop

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time t
(b) Control inputs and sliding variable

2
Figure 9: Cascaded controller with continuous auxiliary
Control space

1
^s(t)
0 control and discontinuous actuator control loop.
−1 w(t)
w(t) s(t)
−2
u(t)
−3
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 (a) Output and desired output
Time t
x(t)
1

Figure 8: Observer in auxiliary control loop enables chat-

State space

tering free system trajectories despite discontinuous control

switching after sliding manifold s$ ( t ) = 0 is reached. The
0.5

plant output x ( t ) follows the observer output x$ ( t ) without

x (t)
d

chattering despite discontinuous control u( t ) applied to 0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time t
main loop with actuator dynamics. (b) Control inputs and sliding variable
1

3.3 Regular form solution 0

Control space

s(t)
Both the boundary layer approach and the observer-based
−1
solution to the chattering problem assume that the ‘unmod- u(t) w(t)

eled’ dynamics are completely unknown. In practical appli- −2

cations, however, at least partial information about unmod-

−3
eled dynamics, in particular of actuators, is often available 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time t
together with measurements of the actuator outputs. For
example, for electric drives, models are readily available,
Figure 10: Linear feedback controller with feedforward of
but may contain uncertain parameters. Thus in the design of
desired trajectory, uc ( t ) , leads to inaccurate tracking of
a controller for the overall system, these dynamics can be
included into the control design to enhance the performance desired trajectory xd ( t ) , since the closed loop system
of the overall system. (8.5.3) is perturbed by unknown plant dynamics and exter-
nal disturbance d ( x , t ) .
Since the actuator dynamics and the plant dynamics are
block separated, i.e. the output(s) of the actuator(s) are the
input(s) of the plant, a cascaded control structure can be
designed following the regular form approach or the block
control principle (see also [Drakunov et al. 1990]). The
3.4 Disturbance rejection solution (a) Output and desired output

The regular form solution in the previous section relies on a 1

x(t)

State space
continuous controller to achieve tracking of the desired
trajectory xd ( t ) by the output x ( t ) of the plant. The linear
0.5
controller usually is augmented by an estimate of the distur- x (t)
d
bance. Often, such an estimate is not readily accessible. The
disturbance rejection approach discussed in this section 0
0 0.1 0.2 0.3 0.4
0.5 0.6 0.7 0.8 0.9 1
provides means to obtain an accurate disturbance estimate Time t
(b) Auxilliary sliding variable and disturbance/uncertainty estimation
while avoiding chattering in the main control loop. This 1.5
approach can be viewed as a special case of so-called inte- f(x,t)

Auxilliary control space

gral sliding mode. A more mathematical background of 1
−ud (t)
ave
integral sliding mode is described e.g. in [Utkin and Shi 0.5
1996, Utkin et al. 1999]
0
The main idea of disturbance rejection via sliding mode is z(t)
to compose the overall controller of a continuous part and −0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
of a discontinuous part. The continuous component is used Time t

to control the overall behavior of the system while the dis-

continuous component is used to reject disturbances and to Figure 12: Performance of linear feedback controller with
suppress parametric uncertainties. A block diagram is feed-forward of desired trajectory, uc ( t ) , is significantly
shown in Figure 11, with simulation results being displayed improved by disturbance rejection controller ud (t ) based
in Figure 12. on auxiliary variable z (t ) for estimating unknown plant
dynamics and external disturbances, summarized as
f ( x, t ) .
Continuous uc ( t )
controller Actuator Plant
xd ( t ) + u( t ) w (t ) x(t ) 4 Example: Automatic Steering Control
xe ( t ) + µ w
2
&& = (..) x& =...
- d (t ) Automation of vehicles, e.g. for Automated Highways Sys-
M
+ s(t)
+
Low pass
filter ud ( t )
tems (AHS), has been discussed for several decades and is
-M
z& =... studied in various programs worldwide in the framework of
z(t )
Auxilliary variable
Intelligent Transport Systems, see e.g. [Stevens 1996] or
[Tsugawa 1996].
Two control subtasks arise for automated driving: Steering
control to keep the vehicle in the lane (controlling the lat-
eral motion) and throttle/brake control to maintain speed
Figure 11: Disturbance rejection via sliding mode with and proper spacing between vehicles (controlling the lon-
auxiliary controller loop to avoid chattering. A continuous gitudinal motion). Both subtasks have been solved using
controller u c (t ) is augmented by a disturbance rejection sliding mode control. The focus in this section will be on
controller ud ( t ) , derived from a low pass filtered discon- automatic steering control. Longitudinal control was studied
tinuous controller for an auxiliary control variable z ( t ) . e.g. by [Hedrick et al. 1994] or [Pham et al. 1994].
The automatic steering system of an automated vehicle
consists of a reference system to determine the lateral vehi-
cle position with respect to the lane center, sensors to detect
the vehicle motion (typically yaw rate and lateral accelera-
tion), and a steering actuator to steer the front wheels. The
variety of employed reference systems ranges from look-
ahead systems like machine vision or radar to look-down
systems like electric wires or magnets embedded in the road
surface. ‘Look-ahead/look-down’ describes the point of
measurement of lateral vehicle displacement from the refer-
ence to be ahead of the vehicle or directly down from the
front bumper (see [Patwardhan et al. 1997] for a more de-
tailed treatment). The control design below, however, is
valid for any reference system.
Control design is usually based on the so-called single track
model, which concentrates on the main vehicle mass by
lumping the two wheels at each axle into a single wheel.
The road/tire interaction forces are responsible for generat- back/feedforward “yaw rate” controller to stabilize the first
ing planar lateral and yaw vehicle motions, with the front equation in (6) would be
wheel steering angle δ f being the input variable. A line-
arized second order model for constant speed v is given by ψ& d = −
1
lS
(
v( β + ψ e ) + Cy eS , ) (7)

 c f + cr c f l f − cr l r   cf  with linear feedback gain C > 0 . However, neither side slip

β&   − −1−  β   
Mv Mv 2
 Mv  δ f angle β nor yaw angle error ψ e can be measured and
  = µ c l − c l 2   + µ
c f l f + c r l r  ψ& 
&&
2
ψ
   c f l f  hence have to be estimated by an observer (for details, see
− −
f f r r

   
J Jv  J  e.g. [Guldner et al. 1994, Ackermann et al. 1995]. Intro-
(5) ducing auxiliary variable z = β + ψ e , an observer is de-
signed as
& . For a detailed
with states side slip angle β and yaw rate ψ
derivation of (5) see e.g. [Peng 1992]. Parameters are vehi-  yˆ& e  0 v   yˆ eS  l S  c 
cle mass M and yaw inertia J, distances l f and lr of front  &S  =     +  ψ& +  1  y eS , (8)
 zˆ  0 0  zˆ   0  c 2 
and rear axles from the center of gravity (CG), front and
rear tire cornering stiffness c f and cr , and road adhesion with feedback of the observation error yeS = yeS − y$eS via
factor µ . All parameters are uncertain within known gains c1>0 and c2 >0, chosen faster than the vehicle dy-
− +
bounds, e.g. 0 < µ ≤ µ ≤ µ ≤ 1 . namics in (5). With the help of the observed auxiliary vari-
dT dS able z$ = β$ + ψ$ e , a desired yaw rate is defined as
`r `f

Tail sensor
v
Front
sensor
ψ& d = −
1 ˆ
lS
(
v (β + ψˆ e ) + Cy eS . ) (9)
CG  f
yeT yeCG
_ y eS The second step of control design uses the steering angle
Lane reference δ f as the input to (5) to drive yaw rate error
ψ& e = ψ
&d −ψ
& to zero, e.g. by purely discontinuous sliding
mode control
Figure 13: Single track model of a vehicle following a lane & e.
δ f = δ 0signψ (10)
reference. Sensors at the front and tail bumpers measure
lateral displacements yeS and yeT , respectively. Also The stability analysis follows the previously discussed Ly-
shown are vehicle states, side slip angle β and yaw rate
apunov approach and is omitted here for brevity. Alterna-
tively to (10), a combination of continuous feed-
ψ& , input steering angle δ f , and various distances from back/feedforward and a discontinuity term could be em-
center of gravity CG. ployed, i.e.
When following a reference path with curvature ρ ref as
depicted in Figure 13, lateral vehicle displacement yeS ,
δ f = δ 1signψ& e +
1
ˆc f lˆf  (
 ˆ
) ( ) 
ˆ
µ
ˆ

  cˆ f l f − cˆr lˆr βˆ + 1 cˆ f lˆf2 + cˆr lˆr2 ψ&  + J ψ&&d 

 v 
measured at some sensor position d S ahead of CG, and (11)
angular error ψ e can be described by linearized dynamic where estimates of vehicle parameters are denoted with
model $ , stems from an
‘hats’ ( $⋅ ) , the estimate of side slip angle, β
y& eS = v (β + ψ e ) + d S ψ& , observer similar to (8) and the derivative of the desired yaw
(6) && d , can be derived from (9) by virtue of known ob-
rate, ψ
ψ& e = ψ& − vρ ref . server dynamics. Due to the continuous feed-
back/feedforward terms in (11), the gain of the discontinuity
Given (5) and (6), various control design options are possi-
ble. As an example, we present a cascaded control design term can be reduced as compared to (10), i.e. δ 1 < δ 0 .
under the assumption that vehicle yaw rate ψ & is measurable The above control design neglects the dynamics of the
by a gyroscope. The control design follows the regular form steering actuator, which will lead to chattering in practical
methodology (see Section 3.3) and considers subsystem (5) implementations. In addition to the chattering prevention
as the input to subsystem (6). Hence, the first design step methods discussed above, the introduction of an integrator
assumes yaw rate ψ & to be a direct input to (6) and derives a in the control loop proved to be a promising approach.
&
desired yaw rate ψ d . The second step then ensures that the Originally, the integrator was a physical model of the ac-
actual, measured vehicle yaw rate ψ & follows ψ & d exactly tuator dynamics (see e.g. [Ackermann et al. 1993]) with the
steering rate u being the system input as
via appropriate control design for steering angle δ f in (5),
the true system input. A suitable continuous feed-
δ& f = u , (12) 5 Discussion
In applications of sliding mode control, unmodeled dynam-
rather than the steering angle δ f itself. The additional inte-
ics in the control loop are often excited by the discontinuous
grator only requires to alter the outer control loop (10/11). switching action of a sliding mode controller, leading to
Define a second order sliding variable oscillations in the motion trajectory. Due to the acoustic
& e +ψ
s = C3 ψ && e , (13) noise such oscillations may cause in mechanical systems,
this phenomenon is also referred to as ‘chattering’. This
leading to a control law paper studied the chattering problem and presented four
solutions. All four discussed solutions to the chattering
δ f = δ 0 sign s (14)
problem reliably eliminate chattering in the control loop. In
instead of (10). The alternative feedback/feedforward con- order to successfully prevent chattering, all methods require
troller term in (11) has to be adjusted accordingly. If the some estimate of the time constant or the bandwidth of the
real steering actuator is not an integrator as in (12), but unmodeled dynamics. Instead of achieving exact tracking
features more complex dynamics e.g. of an electro- performance as in ideal sliding mode, small tracking error
hydraulic actuator, two design alternatives are left to the are tolerated.
control engineer. Either, a sliding mode controller is de- In general, the achievable performance of a control system
signed according to (10/11) with appropriate measures to depends on the performance of sensors and actuators, avail-
prevent chattering as discussed above, or an integrator like ability of knowledge about the system, i.e. the quality of the
(12) is introduced as part of the controller, i.e. realized in system model, and the availability of measurements of sys-
the controller software. The latter case follows the ideas of tem variables. For example, a system with a slow actuator
integral sliding mode by implementing sliding motion in an can not fully reject fast disturbances, regardless of the con-
integral manifold rather than directly in the control input trol design methodology used. A sliding mode controller
variable δ f . Hence the switching action of the sliding mode under ideal conditions is able to fully exploit the system
discontinuity is first filtered by integrator (12) and thus does capabilities. Under realistic conditions, a chattering preven-
not directly reach the input δ f , which inherently prevents tion scheme should be selected depending on the system
specifications to ensure good system performance.
chattering. A different integrator location in the control loop
was proposed by [Pham et al. 1994], before rather than after The first of the discussed method substitutes the discontinu-
the switching discontinuity. The interested reader is referred ity of a sliding mode controller by a saturation function and
to [Hingwe and Tomizuka 1995] for a comparison of differ- yields motion in a boundary layer of the sliding manifold
ent integrator locations in the controller loop. Experimental instead of true sliding along the manifold. Effectively,
results from this work are displayed in Figure 141. sliding mode methodology is utilized to design a continuous
0.1 5
high-gain controller which respects bounds on the control
resources.
Lateral error [m]

0.05
Steering angle [deg]

0 The second method shifts the switching action of the sliding

0 mode controller into an auxiliary observer loop, thus cir-
−0.05
−5 cumventing unmodeled dynamics in the main loop and
achieving ideal sliding mode in the observer loop. The plant
−0.1
0 15 30 45
−10
0 15 30 45 follows the ideal trajectory of the observer according to the
Time [s] Time [s]
observer performance. Since the control input to the plant is
1 10 still discontinuous, this method is ideal for systems which
8 already have an observer in the control structure or for
0.5
6
systems with inherently discontinuous control inputs like
Sliding variable

voltage inputs of electric drives. Implementation of a con-

Velocity [m/s]

0
4
tinuous controller in a system with discontinuous inputs
−0.5
2
generally requires pulse-width modulation (PWM), whereas
−1
0 15 30 45
0
0 15 30 45
direct implementation of sliding mode control with an ob-
Time [s] Time [s] server avoids the detour via PWM.

Figure 14: Experimental results of an automatic steering The third method is mainly designed for systems where
controller based on sliding mode design. some knowledge of the unmodeled dynamic and intermedi-
ate measurements are available, e.g. known actuator dy-
namics. Such systems consisting of separated blocks may be
controlled with a cascaded control structure which avoids
chattering by explicitly taking the unmodeled dynamics into
account for the control design. In this sense, they are no
longer “unmodeled”, but rather part of the overall system
model.
1
Courtesy Dr. Hingwe and Prof. Tomizuka.
The last method combines a continuous and a discontinuous Krstic, M., I. Kanellakopoulos, P. Kokotovic, Nonlinear
controller to achieved good performance without chattering. and Adaptive Control Design, Wiley-Interscience, New
The continuous part controls the overall motion whereas the York, NY, USA, 1995.
task of the discontinuous part is to reject the influence of
Leitmann, G., “On the Efficacy of Nonlinear Control in
parametric uncertainty and disturbances. This method is a
Uncertain Systems”, ASME Journal of Dynamic Systems,
special case of integral sliding mode and is especially useful
Measurement and Control, vol. 102, pp. 95-102, 1981.
for systems with large uncertainties and/or disturbances.
Patwardhan S., H.-S. Tan and J. Guldner,“A General
All four methods possess their advantages and disadvan-
Framework for Automatic Steering Control”, in Proc.
tages which depend on the system specifications. When
American Control Conf., Albuquerque, NM, USA, pp.
designing a sliding mode controller for a given system, the
1598-1602, 1997.
choice of which method to employ to prevent chattering
usually requires careful consideration of all details; unfortu- Peng, H., Vehicle Lateral Control for Highway Automation,
nately, no textbook solution exists to cope with all system in Ph.D. Dissertation, University of California (Dept. of Me-
a general manner. chanical Engineering), Berkeley, CA, USA, 1992
Pham, H., K. Hedrick, and M. Tomizuka, “Combined Lat-
eral and Longitudinal Control of Vehicles for IVHS”, in
References Proc. American Control Conference, Baltimore, MA, USA,
pp. 1205-1206, 1994
Ackermann, J., A. Bartlett, D. Kaesbauer, W. Sienel, and R.
Steinhauser, Robust Control: Analysis and Design of Linear Slotine, J.-J. and S. S. Sastry, “Tracking Control of Nonlin-
Control Systems with Uncertain Physical Parameters. Lon- ear Systems using Sliding Surfaces, with Application to
don, UK: Springer-Verlag, 1993. Robot Manipulators”, Int. Journal of Control, vol. 38, no.
2, pp. 465-492, 1983.
Ackermann J., Guldner J., Sienel W., Steinhauser R., und
Utkin V.I., “Linear and Nonlinear Control Design for Slotine, J.-J., “Sliding Controller Design for Nonlinear
Automatic Steering”, IEEE Trans. on Control Systems Systems”, Int. Journal of Control, vol. 40, no. 2, pp 421-
Technology, vol. 3, no. 1, pp. 132-143, 1995. 434, 1984.

Bondarev A.G., S. A. Bondarev, N. E. Kostyleva and V. I. Stevens, W., “The Automated Highway Systems Program:
Utkin, ”Sliding Modes in Systems with Asymptotic State A Progress Report”, in Preprints of the 13th IFAC World
Observers”, Automation and Remote Control, vol. 46, no. 6 Congress (plenary volume), San Francisco, CA, USA, pp.
(P. 1), pp. 49-64. 1985. 25-34, 1996.

Drakunov, S.V., D.B. Izosimov, A.G. Luk’yanov, V.A. Tsugawa, S., M. Aoki, A. Hosaka, and K. Seki, “A Survey
Utkin and V.I. Utkin, “Block Control Principle I, II“, Auto- of Present IVHS Activities in Japan”, in Preprints of the
mation and Remote Control, vol. 51, pp. 601-609 and vol. 13th IFAC World Congress (Vol. Q), San Francisco, CA,
52, pp. 737-746, 1990. USA, pp. 147-152, 1996.

Guldner, J., V.I. Utkin, and J. Ackermann, “A Sliding Mode Utkin, V. I., “Sliding mode control in discrete-time and
Approach to Automatic Car Steering”, in Proc. American difference systems”, in Variable Structure and Lyapunow
Control Conf., (Baltimore, MA, USA), pp. 1969-1973, Control (A. Zinober, editor), London, UK: Springer-Verlag,
1994. 1993

Hedrick, J.K, M. Tomizuka, and P. Varaiya, “Control Issues Utkin, V. I., and J. Shi, “Integral sliding mode in systems
in Automated Highway Systems", IEEE Control Systems, operating under uncertainly conditions” in Proc IEEE Conf.
1994. On Decision and Control, Kobe (Japan), pp. 4591-4596,
1996.
Hingwe, P., and M. Tomizuka, “Two Alternative Ap-
proaches to the Design of Lateral Controllers for Commuter Utkin, V.I., J. Guldner and J. Shi, Sliding mode control in
Buses based on Sliding Mode Control”, in Advanced Auto- electro-mechanical systems, Taylor&Francis, London, 1999
motive Technologies, presented at ASME Int. Mech. Eng.
Congress and Exposition, pp. 99-104, 1995.
Kokotovic, P.V., R.B. O’Malley and P. Sannuti, “Singular
Perturbations and Order Reduction in Control Theory”,
Automatica, vol. 12, pp. 123-132, 1976
Kokotovic, P.V., “Applications of Singular Perturbation
Techniques to Control Problems”, SIAM Review, vol. 6, no.
4, pp. 501-550, Oct. 1984.