You are on page 1of 6

Proceedings of the 2nd IFAC

Workshop on Fractional Differentiation and its Applications


Porto, Portugal, July 19-21, 2006

A FRACTIONAL ADAPTATION SCHEME FOR


LATERAL CONTROL OF AN AGV

Jos Ignacio Surez ;1 Blas M. Vinagre ;1

YangQuan Chen ;2

GRASP, University of Extremadura, Badajoz, Spain


CSOIS, ECE Dept. of Utah State Univ., Logan, USA

Abstract: The lateral control of an autonomous guided vehicle (AGV) is highly


inuenced both for the longitudinal speed and the position input command
(magnitude of the reference signal) of the vehicle. For that reason, a suitable
strategy to govern the vehicle would be to use an adaptive and robust controller.
In this paper an adaptive scheme is proposed which combines a model reference
approach and a fractional order adjustment rule for a feedforward gain adjustment.
Two parameters can be tuned to obtain robustness against speed and magnitude
of the reference signal variations: adaptation gain, and derivative order of the
adjustment rule. A model is developed for the vehicle, the design procedure is
exposed, and simulation results are obtained to show the advantages of using the
proposed fractional adaptation scheme.

Keywords: autonomous vehicles, adaptive control, fractional calculus

1. INTRODUCTION Fractional calculus is a 300-years-old topic, being


the theory of fractional-order derivative devel-
The lateral control of an autonomous vehicle oped mainly in the 19-th century. Recent books
(AGV) is highly inuenced both for the longi- (Podlubny, 1999) provide a good source of refer-
tudinal speed and the position input command ences on fractional calculus. Applying fractional-
(magnitude of the reference signal) of the vehicle. order calculus to dynamic systems control is nowa-
For that reason, a suitable strategy to govern the days an increasing focus of interest (see (Vinagre
vehicle would be to use an adaptive and robust and Chen, 2002)). In (Vinagre et al., 2002) two
controller (Ackermann, 1997), (Netto et al., 2004). ideas were presented to extend the conventional
MRAS method by using fractional order para-
The model reference approach was developed
meter adjustment rules and fractional order ref-
around 1960 and MRAS (Model Reference Adap-
erence models. A step further, in this paper a
tive Systems) has become a standard part in text-
fractional adaptation scheme is proposed for the
books on adaptive control (strm and Witten-
lateral control of an AGV. This control scheme
mark, 1989). The well known MIT rule for MRAS
combines a model reference approach and a frac-
is to adjust or update the unknown parameter
tional order adjustment rule for a feedforward
using gradient information.
gain adjustment. The control scheme consists of
two loops. The inner loop uses a Model-Following
1 Partially supported by Spanish Research Grant
Controller (MFC) to obtain a stable system with
DPI2002-04064C05-03 (MCYT). gain depending on the vehicle speed, and the outer
2 Supported in part by the TCO Bridging Fund of Utah loop uses a fractional order adjustment rule for
State University (2005-2006).
2.2 Vehicle Dynamic Model

The kinematic model is not enough for simulation


purposes. We also must take into account the
dynamic behavior of the vehicle. Some authors
presents only a lateral dynamic model ((Chaib et
al., 2004), (Tsugawa et al., 1999)), but they do
not consider the dynamic of the steering system.
Fig. 1. Bicycle kinematic model The latter has signicant eects on the vehicle
performance that cannot be ignored during the
feedforward gain adjustment. By doing so, two controller design ((Brennan et al., 1998)). Some
parameters can be tuned to obtain robustness authors use the steering system actuator dynamic
against speed and magnitude of the reference ((Heredia et al., 1998), (Rodriguez-Castao et
signal variations: adaptation gain, and derivative al., 2003)) as a simplied model with good results.
order of the adjustment rule. The following equation gives a rst order actuator
This paper is organized as follows. In Sec. 2, the dynamic:
vehicle model is developed. Sec. 3 presents in de-
tail the proposed fractional adaptation scheme for _= Ka
lateral control of the vehicle. In Sec. 4 illustrative + f (3)
simulations are given to demonstrate the eec- with,
tiveness of the proposed control scheme. Sec. 5
concludes this paper with some remarks on future : the time constant of the steering actuator,
research. Ka : the gain (supposed Ka = 1) and
f : the front wheels input command signal.

2.3 The Whole Vehicle Model


2. VEHICLE MODEL
To obtain a lateral linear model of the vehicle we
2.1 Vehicle Kinematic Model can use the small angle approximations

A simple four wheel car model is the known


cos 1
as the bicycle kinematic model shown in Fig.
sin (4)
1 ((Rajamani et al., 2003)). For simplicity we
tan
suppose that the two front and two rear wheels
are lumped together. The vehicle is supposed to and by making use of (2), we can obtain
move in the X-Y plane and the so called guidance
point is located in the middle of the rear wheel. x_ = v
The equations that determine the vehicle plane y_ = v (5)
movement are the following: _ = v
L
x_ = v cos By combining (3) and (5) and by applying the
y_ = v sin (1) Laplace transformation we obtain the simplied
_ =v linear dynamic model of the lateral movement of
the vehicle:
where
(x; y): are the coordinates of the guidance Y (s) v 2 =L
point in the plane X-Y, Gv (s) = = 2 (6)
f (s) s ( s + 1)
v: is the longitudinal velocity of the vehicle
(or vehicle speed) at the guidance point, Note in (6) that the DC gain depends on the
: is the yaw angle vehicle velocity and also note the two poles at the
: is the instantaneous vehicle curvature. origin that make the vehicle unstable.
The relationship between the steering angle in
Fig. 1 and the curvature is given by:
3. CONTROLLER SCHEME

tan 3.1 Adaptation of a Feedforward Gain


= (2)
L
where L is the distance from the front to the rear The lateral position of the vehicle, and more
axle. precisely the gain, depends on the vehicle velocity
adjusted based on the error between the reference
model (ym ) and the vehicle (y) outputs:

e=y ym (10)

The classical objective of the MRAS is to min-


imize the error signal by means of an adjust-
Fig. 2. MRAS scheme for a feedfoward gain adap- ment mechanism: the MIT rule, Lyapunov func-
tation tions, etc. Instead of using one these minimization
method, we propose to design the MFC for a given
vehicle velocity, and then to use the adaptation
of the feedforward gain for improving the vehicle
performance at other speeds. The design of a MFC
can be looked at as a simultaneous zero and pole
placement approach. The controller has two terms
(Fig. 3): a feedforward term that anticipates the
Fig. 3. MFC structure that stabilizes the vehicle
vehicle response, intended to cancel its dynamics,
as shown in (6), so we use the adaptive scheme and a feedback term for compensating for any
depicted in Fig. 2 and proposed, with = 1, error in the feedforward design.
by (strm and Wittenmark, 1989). This is a
We can represent the lateral position of the vehicle
simple MRAS in which the problem is to adjust
by:
the vehicle feedforward gain to the value 0 .
From the Fig. 2, with = 1, we can observe that B(s)
the error signal is given by y= u (11)
A(s)
where u is the command input, and A(s) and B(s)
0
e=y ym = G(s)( )uc (7) are two polynomial in the domain of the complex
and hence variable s. It is assumed that:
A and B are relatively prime, i.e. they do not
@e ym have any common factors,
= uc G(s) = 0 (8) A is monic, i.e. the coe cient of the highest
@
power in A is unity,
By applying the MIT rule we obtain the following and deg(A) deg(B).
formula for adjusting the feedforward gain:
We want the vehicle to behave like:
d 0 @e 0 ym
= e = e 0 = eym (9) Bm (s)
dt @ y= uc (12)
Am (s)
where = 0 = 0 is the adaptation gain that
where Am (s) and Bm (s) are the desired denomi-
determines the adaptation rate.
nator and numerator of close-loop system, respec-
tively, and uc is the lateral position command.

3.2 Design of the Model Following Controller The desired close-loop behavior is obtained by
means of the following controller (Fig. 3):
The transfer function of the plant should be
stable, but as shown in (6), it is unstable due T (s) S(s)
u= uc y (13)
to the double integrator. This would make the R(s) R(s)
MRAS not to work properly. For solving this
di culty we propose to implement a MFC that where T =R is the feedforward term and S=R is
makes the system stable. Then we will have two the feedback term. By substituting (11) in (13)
control loops: the outer loop for the adaptation we obtain:
of the feedforward gain and the inner loop for
making the system stable by means of a MFC BT
y= uc (14)
(Fig. 3). AR + BS
In the MFC design, we must nd a controller that By comparing (12) with (14) we can obtain the pa-
fullls two conditions: rst, the performance of the rameters of the controller T , S and R. A detailed
vehicle is specied by a reference model so that the method for obtaining those parameters with the
closed loop system behaves like a reference model. lowest degree controller is described more thor-
And secondly, the parameters of the controller are oughly in (strm and Wittenmark, 1989).
In our case, we have supposed the transfer func- 3). The traditional scheme has an adaptation law
tion of the vehicle given in (6). Let us dene, given by (9) in which the rate of change of the
parameter depends solely on the adaptation gain
T (s) . As proposed by (Vinagre et al., 2002), it is
t0 (s) = possible to make the rate of change depending on
R(s) R(s)=1
(15) both the adaptation gain and the derivative order
S(s)
s0 (s) = , by using the adaptation law:
R(s) R(s)=1

Then, taking into account (13), the closed-loop d


= eym (22)
transfer function is: dt
where is a real number denoting the fractional
order derivative. The Grnwald-Letnikov deni-
Y (s) v 2 =L t0 (s) tion of the fractional dierentiation is
G(s) = = 2 (16)
Uc (s) s3 + 1 s2 + vL s0 (s)
[t a=h]
For obtaining a causal controller, see (strm and X
Wittenmark, 1989), we must fulll the condition: a Dt f (t) = lim h ( 1)j f (t jh)
h!0
j=0
j
(23)
deg(Am ) deg(Bm ) deg(A) deg(B) (17) where [ ] means the integer part.

what implies, in our case, that the pole excess,


deg(Am ) deg(Bm ), of the reference model must 4. SIMULATIONS
be at least 3. Then we choose the reference model
given by the following transfer function: Next we show the simulation results carried out
with the adaptive scheme depicted in Fig. 2 for
Ym (s) p3 = 1 and = 0:85 (not an optimum value).
Gm (s) = = 3 = In all the simulations, all the parameters remain
Uc (s) (s + p) (18) unchanged except for . The simulations consists
p3
= 3 of introducing a square input command, hence
s + (3p)s2 + (3p2 )s + p3
the vehicle must do lane changes alternatively, so
This has a triple pole at s = p. This system has that we can see the performance and the evolution
no overshoot on its step response and its dc gain of both the integer and non-integer adaptation
is unity. mechanisms. We have chosen 0 = 1, = 0:5 and
hence p = 0:667.
By comparing (16) with (18) we get t0 with order
0 and s0 with order 1. So s0 must be like this:
4.1 First Test
s0 (s) = as + b (19)
This test shows the inuence of the speed in
Then we obtain: the vehicle performance. Firstly, a set of test at
dierent speeds, from 20 to 80 Km=h (in steps of
Lp3 20), were carried out with:
b = t0 (s) = (20)
v2 = 0:01,
3 Lp2 uc , a square wave from 3 to +3 metres,
a= (21)
v2 v = 20, 40, 60 and 80 Km=h,
with the necessary condition p = 1=(3 ). MFC parameters calculated for 20 Km=h.

Note that from (20) and (21) the controller para- Fig. 4 and 5 show good performance for both the
meters depend on the vehicle speed, so we must fractional (FC) and the integer controller (IC). In
set those parameters up at a desired speed. In our the rst cycle, the IC is faster than the FC, but
simulations, unless otherwise indicated, we have in the next cycles the IC has a longer overshoot
calculated the controller parameters for 20 Km=h. than the FC. With the increasing speed, the FC
is less aected.

3.3 Fractional Adaptation Scheme


4.2 Second Test
As shown in Fig. 2 the reference model is given
by (18) with 0 = 1, and the plant (16) is now Now we present several tests at the same speed
composed of the vehicle and the MFC (see Fig. of 40 Km=h, but with dierent input signals. The
Fig. 4. Test 1 at v = 20 Km=h with = 0:01 and
MFC calculated for 20 Km=h Fig. 7. Test 2 with uc = 3, = 0:1, v = 40
Km=h and MFC calculated for 20 Km=h
sponse gets worse. So it is interesting to calculate
the larger value of the input command signal for
which the vehicle becomes unstable.
From Fig. 2 we can see that,

ym = 0 Gm (s)uc
(24)
y = G(s)uc
where, in the case of perfect model following we
Fig. 5. Test 1 at v = 80 Km=h with = 0:01 and have G(s) = Gm (s). Then substituting (18) in
MFC calculated for 20 Km=h (24) it follows, in the time domain, that:

...
a0 y m + a1 ym + a2 y_ m + a3 ym = p3 0 uc
... (25)
a0 y + a1 y + a2 y_ + a3 y = p3 uc
where
a0 = 1,
a1 = 3p,
a2 = 3p2
and a3 = p3
Fig. 6. Test 2 with uc = 1, = 0:1, v = 40 And, for the case of the IC, the adaptation law is
Km=h and MFC calculated for 20 Km=h
aim is to show the inuence of the magnitude of _= (y ym )ym (26)
the reference signal in the vehicle behavior. The
conditions of the test are: By dierentiating the second equation of (25) and
substituting in (26) we get,
= 0:1,
uc , three dierent square waves: 1 (from 1 .... ...
to +1), 2 and 3 metres, a0 y + a1 y + a2 y + a3 y_ + p3 uc ym y =
(27)
v = 40 Km=h, = p3 u c ym
2
+ p3 u_ c
MFC parameters calculated for 20 Km=h. which is a time-varying linear dierential equa-
Fig. 6 shows a similar performance to that of the tion.
rst test, with faster response in the IC in the Now, let us do an experiment following (strm
rst cycle and poorer behavior in the next cycles. and Wittenmark, 1989). Firstly, assume that the
As depicted in Fig. 7, the IC tends to show a adaptation law is disconnected and the input com-
greater oscillatory response, with the increasing in mand signal uc is a constant u0c . Then, the refer-
the input command, than the FC. Note that is ence model output will go towards an equilibrium
not chosen for an optimum performance, but only value ym0
. Now consider that the adaptation law
for showing that an improvement in the vehicle is connected when the equilibrium is reached. In
behavior is possible. this special situation, (27) becomes a dierential
equation with constant coe cients and has the
equilibrium solution
4.3 Limits of Stability
0
From the results in the second test we can observe y = ym (28)
that by increasing the input signal the vehicle re- which, by applying the Rouths test, is stable if
for optimal tuning of the couple fractional order-
adaptation gain.

REFERENCES
Ackermann, Juergen (1997). 1996 Bode prize lec-
ture: Robust control prevents car skidding.
IEEE Control System Magazine 17(3), 2331.
Fig. 8. Response of the vehicle with IC and FC strm, K. J. and B. Wittenmark (1989). Adap-
when u0c = 2:43 tative Control. Addison-Wesley.
Brennan, S., A. Alleyne and M. DePoorter
s
8p (1998). The Illinois Roadway Simulator -
0
uc < (29) A Hardware-in-the-Loop Testbeb for Vehicle
9 0
Dynamics and Control. In: American Control
0 Conference. pp. 493497.
Then, for = 0:1, p = 0:667 and = 1 we obtain Chaib, Salim, Mariana S. Netto and Said Mam-
mar (2004). H1 , adaptative, PID and fuzzy
u0c < 2:43 (30) control: A comparison of controllers for ve-
hicle lane keeping. In: 2004 IEEE Intelligent
To check this results we have simulated the system Vehicles Symposium. University of Parma.
with the following conditions: Parma, Italy. pp. 139144.
Heredia, G., A. Ollero, F. Gordillo and J. Aracil
= 0:1,
(1998). Stability Analysis of Fuzzy Path
uc = 2:43 metres,
Tracking Using a MIMO Frecuency Response
v = 40 Km=h,
Technique. In: Preprints of the IFAC Work-
parameters of the MFC calculated for 40
shop on "Intelligent Components for Vehi-
Km=h.(instead of 20 because the simulation
cles" (Anibal Ollero, Ed.). pp. 1520.
is at 40Km=h).
Netto, M. S., S. Chaib and S. Mammar (2004).
In Fig. 8 we can check that the vehicle with the Lateral adaptative control for vehicle lane
IC is in its limit of stability, oscillating around keeping. In: 2004 American Control Confer-
the limit value u0c = 2:43. But what it is more ence. Boston, Massachusetts. pp. 26932698.
important is that with the FC using the same Podlubny, I. (1999). Fractional Di erential Equa-
parameters the vehicle performance is still stable tions. Academic Press. San Diego.
and good. This demonstrates that we can choose Rajamani, R., C. Zhu and L. Alexander (2003).
an adequate value of that guarantees a proper Lateral control of a backward driven front-
performance for a given range of input amplitudes steering vehicle. Control Engineering Practice
(uc ). 11, 531540.
Rodriguez-Castao, A., A. Ollero, B. M. Vinagre
and Y.Q. Chen (2003). Fractional controller
5. CONCLUSIONS AND FUTURE WORKS for guidance of autonomous ground vehicles.
In: Preprints 5th IFAC International Sym-
A new fractional adaptation scheme is proposed in posium on Intelligent Components and In-
this paper for the lateral control of an AGV which struments for Control Applications - SICICA
combines MFC and a fractional order adjustment 2003. pp. 97100.
rule for a feedforward gain adjustment. We can see Tsugawa, S., H. Mori and S. Kato (1999). A Lat-
from the simulations results that the fractional eral Control Algorithm for Vision-Based Ve-
adaptation scheme has a good inuence in the hicles with a Moving Target in the Field of
vehicle response. The transient performance of IC View. In: 1998 IEEE International Confer-
is dependent on the model uncertainties and dis- ence on Intelligent Vehicles. Vol. 1. Stuttgart,
turances, while FC is less dependent on that. By Germany. pp. 4145.
using a fractional order adjustment rule, we can Vinagre, B. M., I. Petr, I. Podlubny and Y. Q.
vary the rate of change of the adaptation mech- Chen (2002). Using fractional order adjust-
anism without changing the adaptation gain . ment rules and fractional order reference
The simulations results demonstrate that we can models in model-reference adaptive control.
spread the variation margins of both, the vehicle Nonlinear Dynamics 29(1-4), 269279.
speed and the input command, improving the rela- Vinagre, B.M.. and Y.Q.. Chen (2002). Lecture
tive stability of the system, or even preserving the notes on fractional calculus applications in
absolute stability. The work in progress includes automatic control and robotics. In: IEEE
the analysis of stability bounds with fractional Conference on Decision and Control, Tutorial
order adjustment rules and the design of strategies Workshop 2. pp. 1310.

You might also like