Peni 2006

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Robust Model Predictive Control for Controlling Fast Vehicle Dynamics

T. Péni, J. Bokor

Abstract— This paper proposes a simple robust model pre- tracking problem of road vehicles. The simulation results are
dictive algorithm for discrete time uncertain systems having presented and analyzed in section 4. At the end of the paper
relatively fast dynamics, i.e. the time required to compute the the most important conclusions are drawn.
next control action is the multiple of the sampling time. The
method is based on the following concept: at time k the set of II. ROBUST MPC ALGORITHM USING L INEAR M ATRIX
all possible k + N -th states is determined, and in the next N
I NEQUALITIES
time steps an appropriate feedback gain is constructed, which
is capable of robustly stabilizing the system from time k + N . Our method is mainly based on the model predictive
At time k + N the procedure is then repeated. The method was control algorithm proposed by [1], which will be briefly
derived from the robust MPC algorithm proposed by Kothare introduced in this section.
et. al. in [1]
Following their concept we assume that the system to be
I. I NTRODUCTION controlled is linear, time invariant with structured uncertainty
The model predictive control (MPC) is an optimization placed in the feedback loop, i.e.
based control method, where, assuming discrete time case, xk+1 = Axk + Buk + Bp pk
an open loop optimal control problem is solved at each
y(k) = Cxk
sampling instant, using the actual state as initial state, and
the first element of the obtained control sequence is applied q(k) = Cq xk + Dqu uk
to the plant. For the theoretical background of MPC see pk = k q(k) (1)
e.g. [2], [3]. The model predictive approach has several
or in more compact form:
advantages: it is able to handle complex systems where off-
line computation of the control law is difficult, and in contrast xk+1 = (A + Bp k Cq )xk + (B + Bp k Dqu )uk
to other techniques it is able to handle hard constraints (2)
prescribed for the states and control inputs, even in the
presence of disturbances and plant uncertainties ([4], [1], where xk ∈ Rnx , pk ∈ Rnp , uk ∈ Rnu and k represents
[5], [6], [7], [8]). At the same time, the solution of the an arbitrary time varying, block diagonal matrix
optimization problem can be very time consuming, which  1 
k
limits the applicability of this method. This is a problem  2k 
 
also in the case of autonomous vehicle control. Due to limits k =  .. 
of the actuators in the vehicle, there are hard constraints  . 
prescribed on the state and control input, which motivates rk
the application of the model predictive approach. On the
with the following property: ik 2 ≤ 1 ∀i, k. The imme-
other hand there are generally fast maneuvers prescribed for
diate consequence of this property is the following relation
the vehicles (tracking a trajectory at high speed, avoiding
between the perturbation pk and the state xk
unexpected obstacles, keeping a closed formation, etc.),
which requires small sampling times and fast controllers. pTk pk ≤ xTk (Cq + Dqu F )T (Cq + Dqu F )xk (3)
To overcome this problem a simple robust model predic-
tive algorithm is proposed in this paper, which is able to provided that uk = F xk .
control systems with relatively fast dynamics even if the time We concentrate on the robust regulation problem, i.e. steer-
required to compute the next control action is the multiple ing the state from an arbitrary initial value x0 to the origin,
of the sampling time. Our method is a slight modification of in the presence of the uncertainties above. The MPC solution
the successful robust MPC algorithm proposed by [1]. to this problem proposed by [1] involves the following min-
The paper organized as follows. After the introduction max optimization problem:
the robust MPC method of [1] is introduced; it serves as min max Jk∞ ,
a basis for the further work. In the Section 3 this original uk+i|k ,i=0,1,...,m k+i ,i≥0
∞ 

method is modified to be applicable for systems with fast
dynamics. The obtained method is then tested on the velocity Jk∞ = xTk+i|k Q1 xk+i|k + uTk+i|k Ruk+i|k (4)
i=0
T. Péni and J. Bokor are with the Systems and Control Laboratory, where Jk∞ is a prescribed infinite horizon cost function and
Computer and Automation Research Institute of the Hungarian Academy of
Sciences, H-1518 Budapest P.O. Box 63, Hungary, fax : + 361 466 7503. xk+i|k , uk+i|k denote the predicted state and control action
E-mail of the corresponding author T. Péni: pt@scl.sztaki.hu at time instant k + i, both based on the state measurement
xk = xk|k . This optimization has to be performed at each input uk has to be norm bounded i.e. uk+i|k 2 ≤ umax is
sampling instant with the actual state measurements to obtain prescribed, the following LMI has to be introduced:
the next control input. Since this problem is computationally
2
umax I Y
hard, the following idea has been applied: if it is possible ≥0 (12)
YT Q
to find a quadratic function V (xk ) = xTk Pk xk , which gives
an upper bound on the robust performance objective Jk∞ , For further details see [1].
the min-max problem can be replaced by a minimization The following Theorem summarizes the main result [1]:
of this quadratic function over the sequences of possible Theorem 1. If at time k there exists Q, Y, γ solving the
control moves. This can be easily solved by using linear problem (9) then the control input calculated by using the
matrix inequalities [9]. For V (xk ) to be an upper bound it feedback gain Fk = Y Q−1 will
has to satisfy the following inequality [1]: • be feasible for all times t > k
• robustly stabilizes the system
V (xk+i+1|k ) − V (xk+i|k )
 • for all times t > k satisfies the prescribed state and
≤ − xTk+i|k Q1 xk+i|k + uTk+i|k Ruk+i|k (5) input constraints.

Since in this case III. M ODIFIED MPC ALGORITHM


As it has been mentioned in the introduction, one main
−V (xk ) ≤ −Jk∞ ⇒ max Jk∞ ≤ V (xk )
k+i 2 ≤1,i≥0 disadvantage of the model predictive control is the high
(6) computational time generally needed at each sampling instant
to calculate the next control input. For example, in our case
holds. Using this upper bound, (4) can be replaced by the a convex optimization problem ((9)) has to be solved with
following simpler problem: large number of decision variables over a set of linear matrix
∞ inequality constraints ((10),(11)). This can be too time-
min max Jk ≤ min V (xk )
uk+i|k ,i=0,1,...,m k+i 2 <1,i≥0 uk+i|k ,i=0,1,... consuming for several practical applications. For this, we
= xTk Pk xk (7) modify the original algorithm in this section to be applicable
to controlling systems with fast dynamics.
If uk+i|k is chosen to be uk+i|k = Fk xk+i|k it can be shown Suppose that the system (1) was obtained by sampling a
(see [1]) that the solution (Fk , Pk ) of (7) can be obtained in continuous time system with sampling time Ts , where Ts is
the following form: small, compared to the computation time Tc of (9), i.e.:
Fk = Y Q−1 , Pk = γQ−1 (8) (N − 1) · Ts < Tc ≤ N · Ts N > 1 (13)

where Q > 0, γ > 0, Y are the solutions of the following Thus the presented MPC algorithm in its original form can
linear objective minimization problem: not be used, since the control input uk = Fk xk calculated
from the kth state measurement xk becomes available only
min γ (9) at the k + N th sampling instant. To overcome this difficulty
γ,Q,Y
let us modify the original concept in the following way: at
subject to the k-th sampling instant start to compute Fk+N and not Fk

as before. In this way Fk+N will be available at time k + N
1 xTk
≥0 (10) when the state xk+N is measured. Between time instants k
xk Q
and k + N let the control actions be defined as follows:
and
 1 1 uk = Fk xk ,
Q Y TR2 QQ12
 1 uk+1 = Fk xk+1
 R2 Y γI 0
 1 ..
 Q1 Q
2
0 γI .

Cq Q + Dqu Y 0 0 uk+N −1 = Fk xk+N −1
AQ + BY 0 0
 uk+N = Fk+N xk+N (14)
QCqT + Y T Dqu
T
QAT + Y T B T
0 0  The only problem with this concept is how to design an
0 0 ≥0
 optimal state feedback Fk+N by using the control design
Λ 0
0 Q − Bp ΛBpT procedure presented in the previous section without knowing
Λ = diag(λ1 In1 , λ2 In2 , . . . , λr Inr ) > 0 (11) the state xk+N of the system. Indeed, at step k we do not
know xk+N but we can easily determine the set Xk+N of
If there are constraints prescribed for the input and/or output, all possible values of xk+N . Then, if we reformulate the
they can be easily taken into consideration by expressing LMIs (10), (11) so that the relations (5) will be valid for all
them as LMIs and attaching them to the constraints (10), possible values of xk+N an optimal and feasible feedback
(11) in the optimization problem above. For example, if the gain Fk+N can be determined.
To set up the modified procedure we first give an iterative
algorithm, which determines Xk+N from xk and Fk . For his
we use the relation (3) to calculate an upper bound on the
perturbation pi . Let G = Cq + Dqu Fk , then

xk+1 = Axk + Buk + Bp pk


∈ (A + BFk )xk + Bp Ωk = Xk+1
Ωk = {z|z T z ≤ xTk GT Gxk } (15)

xk+2 = (A + BFk )xk+1 + Bp pk+1


∈ (A + BFk )Xk+1 + Bp Ωk+1 = Xk+2
Ωk+1 = {z|z T z ≤ ck+1 ,
ck+1 = max sT GT Gs} Fig. 1. Vehicle model and its parameters
s∈Xk+1
..
. (16)
Consequently (10) has to be replaced with the following M
thus LMIs:


1 zjT
xk+N = (A + BFk )xk+N −1 + Bp pk+N −1 ≥ 0, j = 1 . . . M
zj Q
∈ (A + BFk )Xk+N −1 + Bp Ωk+N −1 = Xk+N
Xk+N = co{z1 , z2 , . . . , zM } (19)
Ωk+N −1 = {z|z T z ≤ ck+N −1 ,
where co(·) denotes the set of all convex combinations of
ck+N −1 = max sT GT Gs}
s∈Xk+N −1 its arguments. If we use the bounding box of Xk+N then
(17) M = 2nx .

At time k = 0 an arbitrary stabilizing feedback F0 is needed It is obvious, the greater N is the more conservative the
to initiate the procedure. It can be designed by any method: result the calculated feedback gain produces (provided that
pole placement, LQR, etc. a solution to the optimization (9)-(19)-(11) exists at all in
case of relatively big N ).
Remark 1. Since A + BFk is stable and the sets of
perturbations Ωk ,, Ωk+1 , . . . , Ωk+N −1 are convex, Xk+i Now we can state the following Theorem:
will be convex for all i. Theorem 2. The statements of Theorem 1 remain true for
Fk+N calculated by the modified algorithm above.
Remark 2. The operations with the sets can be greatly Proof : The only difference between the original and this
simplified if we use the bounding box (the smallest cube modified procedure is that we use a set Xk+N in (10) instead
containing the set) of every Xi in the calculation of Xi+1 . of a particular state xk+N . Since, due to (19), the LMI (10)
is satisfied for all possible values of xk+N the statements of
Remark 3. Notice that, for the calculation of each set Theorem 1 remain true.
Xi+1 a convex function sT GT Gx, (GT G ≥ 0) has to be
maximized over a convex set Xi . Consequently for the corner IV. V ELOCITY TRACKING CONTROL OF ROAD VEHICLES
points sj , j ∈ {1, 2, . . . , 2nx } of the bounding box Xibb of To test and present the applicability and dynamic proper-
Xi the following statement holds [9]: ties of the modified MPC procedure we solve in this section
the robust velocity tracking control problem of autonomous
∀j sTj GT Gsj ≤ γ ⇔ sT GT Gs ≤ γ ∀s ∈ Xibb (18) road vehicles (For the motivation to this problem see [10]
and [11]). Suppose the vehicle dynamics is described by the
i.e. the function sT GT Gs takes its maximum at one of
simplified single-track model, given in the following form:
the corner points. Thus, the maximization problem above
reduces to finding the maximal element of the finite set a11 a12 b1
β̇ + ψ̇ = β+ 2 r+ δ
{zj |zj = sTj GT Gsj } v v v
a11 a b1
12
β̇ = β+ − 1 r+ δ
To obtain a feasible Fk+N the constrained optimization v v2 v
a22
(9) has to be performed for all possible values of xk+N . ṙ = a21 β + r + b2 δ
Since only LMI (10) depends on the state it is the only point v
y = β+ψ (20)
where the original procedure has to be modified. The LMI
(10) will be satisfied for all elements of the convex set Xk+N where β + ψ, v are the angle and norm of the velocity vector
if and only if it is satisfied for all corner points of Xk+N . and β, r, ψ are the slideslip angle, yaw rate and orientation
0.15
respectively (see Fig.1). For simplicity we assume that v is 0.1
dcf

constant so the dynamics above is linear. The control input 0.05

is the steering angle (δ). The state variables are supposed to 0

be measured or estimated by appropriate sensors and filters. −0.05


0 1 2 3 4 5 6 7 8 9 10

The remaining parameters of the model are constant and can 0.2
dcr
0.1
be calculated as follows.
0

cf + cr cr lr − cf lf −0.1
a11 = − a12 =
m m −0.2
0 1 2 3 4 5 6 7 8 9 10

cr lr − cf lf cr lr2 + cf lf2 2

a21 = a22 = −
dJ

J J 1

cf cf lf 0
b1 = b2 = (21)
m J −1
0 1 2 3 4 5 6 7 8 9 10

where m is the mass of the vehicle, cr , cf are the rear


Fig. 2. Values of δcf , δcr , δJ used in the simulation
and front cornering stiffness, J is the inertia, lr , lf are the
distances of the center of mass from the rear and front axle.
We assume that m, lr , lf are known precisely but cr , cf , J 0.7
delta
can differ from the nominal value even by 20%:
0.6

m = 10344, lf = 2.53, lr = 2.07


0.5

cf = 6.848 · 105 · (1 + δcf )


cr = 8.373 · 105 · (1 + δcr ), J = 54170 · (1 + δJ )
0.4

−0.2 ≤ δcf , δcr , δJ ≤ 0.2 0.3

The numerical values of the parameters above were 0.2

borrowed from [12]. 0.1

The control problem is formulated as follows. There 0

is given a constant reference velocity angle yref , which −0.1


0 1 2 3 4 5 6 7 8 9 10
has to be tracked by the vehicle, in the presence of the
disturbances above so that the control input remains in the Fig. 3. Control input (δ) [rad] plotted as a function of time [sec]
following range: −0.5 ≤ δ ≤ 0.5

In order to apply our method the system was sampled


with Ts = 0.1s and the obtained discrete time system was The parameters of the MPC procedure were chosen as
considered: follows: Q1 = diag(20, 1, 1), R = 0.1.
   
1.0000 −0.7704 0.0294 0.3863
A= 0 0.2296 −0.0229 B = 0.2828 The initial stabilizing feedback F0 was a discrete-time
0 0.0003 0.2296 1.6747 LQR controller, which was designed with state and control
(22) weights Q = I, R = 20.

The effect of the parameter uncertainties was modelled in The system was started from x0 = [0, 0.01, 0.2]T with
the discrete time system as a full 3 dimensional uncertainty v = 10 and it had to make a 180-degree-turn, i.e. yref = π.
block, i.e. r = 1, nr = 3 in (1). The matrices Bp , Cq , Dqu
were set according to the uncertainties of the elements of A
and B as follows: The simulation results can be seen in Figures 2, 3, 4,
    5. Fig. 3 shows the applied control input δ. It can be
0.1 0 0 0 0 0 seen that in the first 5 sampling instants the control input
Bp =  0 0.1 0  , Cq = 0 0.5 0  violates the prescribed constraint. This caused by F0 , since
0 0 0.5 0 0 0.1 the initial feedback gain is a standard LQR controller, in
T
Dqu = 0.4 0 0 (23) the design of which the constraints can not be taken into
consideration. From the 5th timestep, when the optimized
controllers are used, the constraint is no more violated.
Fig. 5 shows the estimated sets X5 , X10 , . . . , X100 (or,
We assumed that Tc = 0.5s is enough to calculate the more precisely, their bounding boxes) of the possible values
next feedback gain, i.e. N = 5 in (13). of x5 , x10 , . . . , x100 . The sets are constructed at sampling
instants k = 0, 5, . . . , 95.
3.5
angle of velocity [2] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scoakert,
“Constrained model predictive control: Stability and optimality,” Au-
3
tomatica, vol. 36, no. 3, pp. 789–814, 2000.
[3] J. M. Maciejowski, Predictive Control with Constraints. Prentice
2.5 Hall, 2002.
[4] W. Langson, I. Chryssochoos, S. V. Rakovic, and D. Q. Mayne,
2
“Robust model predictive control using tubes,” Automatica, vol. 40,
pp. 125–133, 2004.
[5] F. A. Cuzzola, J. C. Geromel, and M. Morari, “An improved approach
1.5
for constrained robust model predictive control,” Automatica, vol. 38,
pp. 1183–1189, 2002.
1 [6] L. Chisci, J. A. Rossiter, and G. Zappa, “Systems with persistent dis-
turbances: predictive control with restricted constraints,” Automatica,
0.5
vol. 37, pp. 1019–1028, 2001.
[7] D. Mayne, M. Seron, and S. Rakovic, “Robust model predictive control
of constrained linear systems with bounded disturbances,” Automatica,
0
0 1 2 3 4 5 6 7 8 9 10 vol. 41, no. 2, pp. 219–224, 2005.
[8] A. Richards, “Robust constrained model predictive control,” Ph.D.
Fig. 4. The angle of velocity y = β + ψ and the prescribed reference dissertation, Department of Aeronautics and Astronautics, MIT, 2005.
yref = π in [rad] plotted as a function of time [sec] [9] C. Scherer and S. Weiland, Linear Matrix Inequalities in Control,
http://www.cs.ele.tue.nl/ SWeiland/ lmi.pdf, 2000, vers. 3.0.
[10] T. Péni and J. Bokor, “Dynamic inversion based velocity tracking con-
trol of road vehicles,” NOLCOS, 6th IFAC Symposium on Nonlinear
Control Systems, 2004.
[11] ——, “Trajectory tracking control for a class of lpv systems based on
2 X5
dynamic inversion and passivity,” WSEAS International Conference on
1.5
Applied Mathematics, 2004.
X15
X20 X25 [12] G. Rödönyi, “Vehicle models for steering control,” System and Control
1
Laboratory, Computer and Automation Research Institute, Tech. Rep.,
2003, sCL-4/2003.
x3

0.5 X10

−0.5
0.3
X100
0.2 0.5
0
0.1 −0.5
−1
−1.5
0 −2
−2.5
−0.1 −3
x2
x1

Fig. 5. Sets X5 , X10 , . . . , X100

V. C ONCLUSIONS
In this paper the model predictive procedure proposed by
Kothare et. al in [1] has been modified to be able to control
systems with fast dynamics. It was proven that the resulted
control policy preserves the advantageous properties (opti-
mality, feasibility, robust stability, constraint satisfaction) of
the original method. Through numerical simulations it was
also shown that this modified algorithm can be successfully
applied in vehicle control.

ACKNOWLEDGEMENT
This work has been supported by the Hungarian Science
Fund (OTKA) through grant T 043111 and the Hungarian
National Office for Research and Technology through the
project "Advanced Vehicles and Vehicle Control Knowledge
Center" (No: OMFB-01418/2004) which is gratefully ac-
knowledged.

R EFERENCES
[1] M. V. Kothare, V. Balakrishnan, and M. Morari, “Robust constrained
model predictive control using linear matrix inequalities,” Automatica,
vol. 32, no. 10, pp. 1361–1379, 1996.

You might also like