You are on page 1of 6

The 11th IEEE International Workshop on

Advanced Motion Control


March 21-24, 2010, Nagaoka, Japan

Optimal tracking control of two-wheeled mobile


robots based on model predictive control
Shinya Akiba Tadanao Zanma Muneaki Ishida
Graduate school of Engineering, Graduate school of Engineering, Graduate school of Engineering,
Mie University Mie University Mie University
1577 Kurimamachiya, Tsu, Mie, 1577 Kurimamachiya, Tsu, Mie, 1577 Kurimamachiya, Tsu, Mie,
514-8507 Japan 514-8507 Japan 514-8507 Japan
Email: akiba@cs.elec.mie-u.ac.jp Email: zanma@elec.mie-u.ac.jp Email: ishida@elec.mie-u.ac.jp

Abstract—This paper presents an optimal tracking control continuous or time invariant feedback control, though it is
using model predictive control for a two-wheeled mobile robot controllable in nonlinear system[10].
with a nonholonomic constraint. In the proposed method, the In recent years, several the model predictive control based
optimized control strategy can be obtained while satisfying some
constraints imposed on input and state. The dynamics is modeled approaches of two-wheeled mobile robots are reported in
as a hybrid dynamical system so that the model predictive control [11], [12], [13], [14], [15]. In [11] and [12], nonlinear model
can be applied to it. Some simulation and experimental results predictive control (NMPC) is adopted. In the NMPC, consid-
illustrate the effectiveness of the proposed method. erable computational effort is required in solving the optimal
problem, since the nonlinear programming to be solved on-
I. I NTRODUCTION line is nonconvex in general. Hence, it is hard to implement
the NMPC in real-time. For the above disadvantage, MPC
Recently, various intelligent transport systems (ITS) are strategies using a linearization of an error model are proposed
developed for automobile technologies. An adaptive cruise in [13], [14], [15]. The optimization problem is transformed
control (ACC) as one of the ITS aims to achieve tracking to a quadratic programming problem (QP), which allows to
for one or multiple vehicles in order to improve transport obtain an optimal solution in real-time due to the convexity.
efficiency or to reduce driving effort while satisfying safety However, the method is valid only when the following robot is
and comfort. For the purpose, a distance specified a priori close enough to the leading robot. Moreover, it is known that
should always be kept for safety under existence of disturbance the resulting model of a mobile robot is uncontrollable when
or variation in the circumstance while abrupt acceleration linearized at any equilibrium. Therefore, the obtained model
should be avoided for comfort driving. A method of ACC approximated in a single equilibrium point is considerably
based on a model predictive control (MPC) [1], [2], [3], [4] is restrictive.
reported in [5], [6], [7] in order to solve the problem above. For the problem, this paper proposes a novel optimal
The MPC has been regarded as a highly successful control tracking control method of the two-wheeled mobile robot
method in the process industries since the last two decades. with the nonholonomic constraint. The two-wheeled mobile
The MPC yields an optimal control strategy by using a model robot is modeled as a piecewise affine (PWA) system in
of the system to obtain an optimal control sequence with which which each subsystem linearized at a different equilibrium
an evaluation for a given objective function is minimized. In point is switched according to state of the system. Then,
the MPC, in addition, it is possible to handle constraints on we reformulate the system as a mixed logical dynamical
state and input/output for which a resulting control strategy (MLD) system which has much potential in synthesis since the
should meet. On the other hand, the MPC often suffers from representation is similar to the linear discrete time state space
computational burden, which, in the past years, limited its representation[16]. The solution of the design is obtained by
applications only to sufficient slow dynamics. However, the solving an optimization problem with help of the MPC.
development of increasingly faster processors and efficient Some simulation and experimental results are demonstrated
numerical algorithms allow to apply the MPC to a wide range to verify the effectiveness of the proposed method.
of with faster dynamics.
In [6], although the effectiveness of the method has been II. P RELIMINARIES
verified by some simulations for a limited directional course, This section reviews mixed logical dynamical systems
no case for a two-dimensional course has been considered. For which are suitable when applying MPC.
two-wheeled mobile robots in the two dimensional course, the
nonholonomic constraint[8], [9] has to be taken into account. A. Mixed logical dynamical systems
It is well known that a system with nonholonomic constraints Several modeling frameworks have been introduced for
appeared in mobile robots can be stabilized no longer by discrete-time hybrid systems. Among them, mixed logical

978-1-4244-6669-6/10/$26.00 ©2010 IEEE 454


output output Video camera Position coordinates

q(1|t) q(1|t)
q(k|t)
q(t) q(t + 1)

t t+k time t+1 time


N N
leader
predictive horizon
input input
PC
u(t) u(t + 1) follower
Bluetooth
communication

Fig. 2. Experimental environment.


t time t+1 time
N N
w
Y leader
control input horizon

Fig. 1. Outline of model predictive control.


Y
(wxlead ,wylead ) → (xlead , ylead )
dynamical (MLD) systems are described by linear dynamic follower X
equations subject to linear mixed-integer inequalities. Discrete Moving coordinate system
w
events such as mode switch included in process are represented θ
by using logic variables. One of the advantages is in that the
logical formulae can be described by linear inequalities so that
MPC can be applied to the MLD system. (wx,wy) → O
The general MLD form is given as follows, w w
O World coordinate system X

q(t + 1) = Aq(t) + B1 u(t) + B2 δ(t) + B3 z(t), (1) Fig. 3. Coordinate transformation.


E2 δ(t) + E3 z(t) ≤ E1 u(t) + E4 q(t) + E5 , (2)

where q(t) and u(t) are state and input as in the standard III. P ROBLEM STATEMENT
linear systems, respectively. Additionally, δ(t) and z(t) are
introduced as auxiliary binary and continuous variables, re- In this section, the dynamics of a two-wheeled mobile robot
spectively. A, Bi and Ei are appropriate matrices. As a result, is modeled as a piecewise affine system. For the system, a
Eqs. (1) and (2) consist of the standard linear discrete time model predictive control based control method is proposed.
state space representation with the auxiliary variables of δ
and z, and the set of linear inequalities which represents A. Experimental environment and coordinate transformation
the transformed logic formulae and other constraints on the Fig. 2 illustrates the schematic diagram of the experimental
system. Eq. (2) is also used to determine δ or z uniquely. environment. The position coordinate of robots can be saved
to a computer in real time via a video camera equipped above
B. Model predictive control the road surface. The computer calculates an optimal input by
Fig. 1 shows an outline of the model predictive control. The MPC based on the position coordinate and commands it to the
basic MPC algorithm is described as follows. First, the state robot by wireless communication via Bluetooth.
q(t) is observed at time t and k-step ahead predicted state In this paper, the moving coordinate is adopted as shown in
q(k|t) is calculated based on Eqs. (1) and (2). An optimal Fig. 3. The positions of the follower (wx, wy) and the leader
input sequence U ∗ (t) = {u(0|t), u(1|t), . . . , u(N −1|t)} over (wxlead , wylead ) observed in the world coordinate system wX-
the control input horizon [t, t + N ] is determined so that the w
Y via the video camera. The state of follower at time t is
predicted trajectory approaches the reference for the predictive transformed into zero in a new coordinate system X-Y in
horizon [t + 1, t + N ]. Then, only the first element u(0|t) of order to simplify the idea in applying the proposed method.
the optimal input sequence U ∗ (t) is applied to the system The transformed position of the leader in the new coordinate
and held until t + 1. When q(t + 1) is observed at the next system X-Y is given as follows.
time t + 1, the input is similarly determined by shifting the     
predictive horizon to one step ahead. Then, the procedure is xlead cos wθ sin wθ wxlead −wx
= . (3)
repeated. y lead − sin wθ cos wθ wy lead −wy

455
Y
moving direction Y leader

dsafe
θ ε
(xlead , ylead )
l
d
r (xc , yc )
θe (xe , ye )

O X X
follwer
Fig. 4. Two-wheels mobile robot(left) and its schematic(right).

Fig. 5. Problem setting.

B. Two-wheeled mobile robot


as follows.
Fig. 4 shows a two-wheeled mobile robot and its schematic.
The dynamics of the robot is given as the following nonlinear
model. q(t + 1) = Ai q(t) + Bi u(t) if θ ∈ Ωi , (6)
     
ξ(t) Ad Bdi Bdi
⎡r ⎤ q(t) = , Ai = , Bi = ,
2 cos θ
r
2 cos θ ω(t − 1) 0 I I
˙ = ⎣ r sin θ
ξ(t) r
sin θ ⎦ ω(t), (4) i = 1, . . . , L.
2 2
− rl r
l

    Moreover, we rewrite Eq. (6) as the corresponding MLD


where ξ(t) = x y θ and ω(t) = ωL ωR are the state system. As appeared in Eq. (6), the evolution of q(t) is
and input, respectively. Denoted by ωL and ωR are angular dependent on i. Therefore, the auxiliary binary variables are
velocity of the left and right wheels, respectively. Due to introduced in order to describe the switching of subsystems.
the nonholonomic constraint of ẋ cos θ − ẏ sin θ = 0, the The variables are associated with as follows.
two-wheeled mobile robot can be driven only in the moving
direction, i.e., the nonslipping condition in the lateral direction
holds. [δ1 (t) = 1 ∨ δL (t) = 0] → [i = 1], (7)
It is necessary to approximate Eq. (4) in order to utilize [δj−1 (t) = 0 ∧ δj (t) = 1] → [i = j],
MPC. Thus, Eq. (4) is linearized since, in general, it is j = 2, . . . , L.
difficult to solve a nonlinear optimization problem in real
time. However, note that the resulting linear system at any Thus, Eq. (6) can be expressed as the corresponding MLD by
equilibrium point is uncontrollable. Therefore, the obtained introduction some auxiliary continuous variables. Note also
model in a single equilibrium point is considerably restrictive. that Eq. (7) can be transformed into the corresponding linear
For the reason, we model the two-wheeled mobile robot as inequalities.
a piecewise affine (PWA) system in which each subsystem is
switched according to state of the system. The discrete-time
model with a sampling time Ts is given as follows.
C. Problem statement

ξ(t + 1) = Ad ξ(t) + Bdi ω(t) if θ ∈ Ωi , (5) Let us explain an outline of the problem setting using the
⎡ ⎤ ⎡r ⎤
1 0 0 r figureshown in Fig.  5. In the problem setting, the reference
2 Ts cos θi 2 Ts cos θi
Ad = ⎣0 1 0⎦ , Bdi = ⎣ 2r Ts sin θi 2r Ts sin θi ⎦ , qe = xe ye θe is set as follows; (xe , ye ) is put behind
0 0 1 − rl Ts r the leader by dsafe with an enough small tolerance ε > 0 in
l Ts
order to consider modeling error or disturbance, and θe is set as
i = 1, . . . , L,
the angle between x-axis and the leader. In a typical tracking
control, a follower is controlled to track to a leader while
where L is the number of the subsystems, θi = 2πi−Lπ L satisfying constraints on the state and input of the system.
(i = 1, . . . , L) is the i-th linearized equilibrium point, and Specifically, the distance between the leader and follower is
Ωi = {θ : 2πi−(L+1)πL ≤ θ < 2πi−(L−1)π
L } (i = 1, . . . , L) always kept more than dsafe . On the other hand, the constraints
is associated with each subsystem. Now, define u(t) := of the limit of the angular velocity of each wheel which is the
ω(t) − ω(t − 1), then the augmented model of Eq. (5) is given constraint on the input, has to be met.

456
TABLE I
PARAMETERS . 1.3

1.2
wheel radius r 0.0205 [m] 1.1
x
distance between the wheels l 0.053 [m]
1
sampling time Ts 0.11 [s]
0.9 y
horizon N 1, 2

x, y [m]
safe distance dsafe 0.15 [m] 0.8
number of the modes L 8 0.7
number of δs 15 0.6
number of zs 30
  0.5
maximam input ω 2π 2π [rad/s]
  0.4
minimam input ω −2π −2π [rad/s]
0.3
margin ε 0.006 [m] 0 2 4 6 8 10 12 14 16 18
Q in Eq. (8) diag(1, 1, 10−1 ) time [s]

R in Eq. (8) diag(0, 0), diag(5 × 10−5 , 5 × 10−5 )


Fig. 6. State (simulation), (solid:follower, dashed:reference), (R =
diag(0, 0), N = 1).

Thus, the following optimization problem is formulated.


J(q(t)) = 6
ωR

N −1 5
( q(k + 1|t) − qe (k + 1|t) 2Q +  u(k|t) 2R ) (8)
4

ωL , ωR [rad/s]
k=0

U (t) = argmin J(q(t)) (9) 3 ωL
U (t)

subject to 2

q(k + 1|t) = Aq(k|t) + B1 u(k|t) + B2 δ(k|t) + B3 z(k|t), 1


(10)
0
E2 δ(k|t) + E3 z(k|t) ≤ E1 u(k|t) + E4 q(k|t) + E5 , (11) 0 2 4 6 8
time [s]
10 12 14 16 18

ω ≤ ω(k|t) ≤ ω, (12)
  Fig. 7. Input (simulation), (R = diag(0, 0), N = 1).
x(k + 1|t) − xc (k|t) y(k + 1|t) − yc (k|t)
 
· xlead (k|t) − xc (k|t) ylead (k|t) − yc (k|t) ≤ 0. (13)
constraints on input u is satisfied. In addition, in the result, the
In the optimization problem above, Eq. (10) and Eq. (11) are
safe distance d is also satisfied. However, abrupt acceleration
the transformed predictive model as the MLD system and the
has arisen at times of rising and falling. Such an acceleration
corresponding linear mixed-integer inequalities, respectively.
may not be desirable.
In this problem, q(0|t) := q(t). Eq. (12) represents the input
constraint. Eq. (13) expressed by the inner product of the Then, the weighting matrix in Eq. (8) is set to R = diag(5×
10−5 , 5 × 10−5 ) in order to avoid the abrupt acceleration.
vectors from (xc (k|t), yc (k|t)) to the predicted position and
from (xc (k|t), yc (k|t)) to the leader position ensures that the Figs. 8 and 9 show the simulation results which are compatible
follower does not enter the meshed area in Fig. 5. with Figs. 6 and 7. From Fig. 8, it can be observed that the
follower can be tracked to the reference. From Fig. 9, it can
IV. S IMULATION AND E XPERIMENTAL RESULTS
In this section, the effectiveness of the method proposed in
the previous section is demonstrated by applying it to tracking 1.3

1.2
control of the two-wheeled mobile robot shown in Fig. 4. All
1.1
of the results are executed with a commercial optimization x
1
toolbox IBM ILOG CPLEX ver. 11.0. y
0.9
x, y [m]

A. Simulation results 0.8

0.7
The parameters used in the simulation are listed in Table I.
0.6
Fig. 6 shows the simulation result of a trajectory of the
0.5
follower for a reference given to it. The weighting matrix in
0.4
Eq. (8) is now set to R = diag(0, 0). From Fig. 6, it can be 0.3
observed that the follower can be tracked to the reference on 0 2 4 6 8
time [s]
10 12 14 16 18

the two-dimensional plane. The corresponding control input is


shown in Fig. 7. Note that the leader decelerates at the time Fig. 8. State (simulation), (solid:follower, dashed:reference), (R = diag(5×
indicated by the dotted line. From Fig. 7, it can be seen that the 10−5 , 5 × 10−5 ), N = 1).

457
1.3
6
ωR
1.2

5 1.1
x
1
4
ωL , ωR [rad/s]

y
0.9

x, y [m]
0.8
3 ωL
0.7
2 0.6

0.5
1
0.4

0 0.3
0 2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18
time [s] time [s]

Fig. 9. Input (simulation), (R = diag(5 × 10−5 , 5 × 10−5 ) N = 1). Fig. 12. State (experiment), (solid:follower, dashed:reference), (R =
diag(0, 0), N = 1).

1.3

1.2 6
ωR
1.1
x
5
1
y
0.9
4

ωL , ωR [rad/s]
x, y [m]

0.8

0.7 3 ωL
0.6
2
0.5

0.4 1
0.3
0 2 4 6 8 10 12 14 16 18
time [s] 0
0 2 4 6 8 10 12 14 16 18
time [s]

Fig. 10. State (simulation), (solid:follower, dashed:reference), (R =


diag(5 × 10−5 , 5 × 10−5 ), N = 2). Fig. 13. Input (experiment), (R = diag(0, 0), N = 1).

be seen that the constraints on input u is satisfied. In addition, acceleration appeared in Fig. 7 and the overshoot appeared in
the abrupt acceleration appeared in Fig. 7 is avoided as shown Fig. 9 are avoided as shown in Fig. 11.
in Fig. 9. However, the overshoot has still arisen at times of B. Experimental results
falling.
The parameters and condition in experiments are the same
Then, the predictive horizon is expanded to N = 2. Figs. 10
as those in the simulation. Fig. 12 shows the experimental
and 11 show the simulation results which are compatible with
result of the trajectory of the follower and the reference when
Figs. 8 and 9.
R = diag(0, 0) in Eq. (8). From Fig. 12, it can be observed
From Fig. 10, it can be observed that the follower can be
that the follower can be tracked to the reference on the two-
tracked to the reference. From Fig. 11, it can be seen that
dimensional plane. Figs. 13 show the corresponding control
the constraints on input u is satisfied. In addition, the abrupt
input. In Figs. 12 and 13, the dotted line indicates the time
at which the leader decelerates. From Figs. 13, it can be seen
that the constraints on input u is satisfied as well as in the
6
ωR
simulation.
Figs. 14 and 15 show the experimental results taking into
5
account variation of acceleration. From Fig. 14, it can be
4 observed that the follower can be tracked to the reference.
ωL , ωR [rad/s]

From Fig. 15, it can be seen that the constraints on input u


3 ωL
is satisfied. In addition, compared with the result shown in
2 Fig. 13, the input variation is relaxed as shown in Fig.15.
Figs. 16 and 17 show the experimental results when the
1
predictive horizon is expanded to N = 2. From Fig. 17,
0 it can be observed that the overshoot appeared in Fig. 15 is
0 2 4 6 8 10 12 14 16 18
time [s] avoided as shown in Fig. 17.
The results above validate the effectiveness of the proposed
Fig. 11. Input (simulation), (R = diag(5 × 10−5 , 5 × 10−5 ), N = 2). method.

458
1.3
6
ωR
1.2

1.1 5
x
1
4

ωL , ωR [rad/s]
y
0.9
x, y [m]

0.8
3 ωL
0.7

0.6 2

0.5
1
0.4

0.3 0
0 2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18
time [s] time [s]

Fig. 14. State (experiment) (solid:follower, dashed:reference), (R = Fig. 17. Input (experiment), (R = diag(5 × 10−5 , 5 × 10−5 ), N = 2).
diag(5 × 10−5 , 5 × 10−5 ), N = 1).

of friction, an optimal tracking control for varieties of paths.


6
ωR
R EFERENCES
5
[1] C. E. Garcia, D. M. Prett and M. Morari, “Model predictive control:
4 theory and practice – a survey,” Automatica, vol. 25, no. 3, pp. 335–348,
ωL , ωR [rad/s]

1989.
3 ωL
[2] M. Morari and J. H. Lee, “Model predictive control: past, present and
future,” Comp. and Chem. Eng, vol. 23, pp. 667–682, 1997.
2 [3] B. D. Schutter and T. J. J. V. D. Boom, “MPC for continuous piecewise-
affine systems,” Systems and Control Letters, vol. 52, no. 3-4, pp. 179–
1 192, 2004.
[4] J. Richalet, “Why Predictive Control?,” Journal of The Society of Instru-
0 ment and Control Engineers, vol. 43, no. 9, pp. 654–664, 2004.
0 2 4 6 8 10 12 14 16 18
time [s]
[5] D. Corona, M. Lazar and B. D. Schutter, M. Heemels, “A hybrid MPC
approach to the design of a smart adaptive cruise controller,” Proceedings
of the 2006 IEEE International Conference on Control Applications,
Fig. 15. Input (experiment), (R = diag(5 × 10−5 , 5 × 10−5 ), N = 1). pp. 231–236, 2006.
[6] D. Corona and B. D. Schutter, “Comparison of a linear and hybrid
adaptive cruise for a Smart,” Proceedings of the 46th IEEE Conference
on Decision and Control, pp. 4779–4784, 2007.
V. C ONCLUSION [7] D. Corona and D. Schutter, “Adaptive cruise control for SMART car: A
comparison benchmark for MPC-PWA control methods,” IEEE Transac-
The tracking control of a two-wheeled mobile robot was tions on Control Systems Technology, vol. 16, no. 2, pp. 365–372, 2008.
proposed in this paper. The dynamics of the two-wheeled [8] I. Kolmanovsky and N. H. McClamroch, “Developments in Nonholo-
mobile robot with the nonholonomic constraint was modeled nomic Control Problems,” IEEE Control Systems Magazine, vol. 15, no. 6,
pp. 20–36, 1995.
as a mixed logical dynamical system so that the model [9] M. Sanpei, “Feedback Control of Nonholonomic Systems,” J. Society of
predictive control was applied to the tracking control. The Instrument and Control Engineers, vol. 36, no. 6, pp. 396–403, 1997.
effectiveness of the proposed method was illustrated by some [10] R. W. Brockett, “Asymptotic Stability and Feedback Stabilization,”
Differential Geometric Control Theory, pp. 181–191, 1983.
simulation and experimental results. [11] F. Kuhne, W. F. Lages and J. M. G. da Silva, Jr, “Point stabilization of
Further researches include robust analysis in the presence mobile robots with nonlinear model predictive control,” Proceedings of
2005 IEEE International Conference on Mechatronics and Automation,
vol. 3, pp. 1163–1168, 2005.
[12] K. Kanjanawanishkul, M. Hofmeister and A. Zell, “Smooth Reference
1.3 Tracking of a Mobile Robot using Nonlinear Model Predictive Control,”
1.2 Proceedings of European Conference on Mobile Robots 2009, pp. 161–
1.1
166, 2009.
x [13] F. Kuhne, W. F. Lages and J. M. G. da Silva Jr, “Model Predictive
1
Control of a Mobile Robot Using Linearization,” Proceedings of the
y
0.9 Mechatronics and Robotics 2004, pp. 525–530, 2004.
x, y [m]

0.8 [14] D. Gu and H. Hu, “Receding horizon tracking control of wheel mobile
robots,” IEEE Transaction on Control Systems Technology, vol. 14, no. 4,
0.7
pp. 743–749, 2006.
0.6 [15] G. Klancar and I. Skrjanc, “Tracking-error model-based predictive
0.5 control for mobile robots in real time,” Robotics and Automation Systems,
0.4
vol. 54, no. 6, pp. 460–469, 2007.
[16] A. Bemporad and M. Morari, “Control of systems integrating logic,
0.3
0 2 4 6 8 10 12 14 16 18 dynamics, and constraints,” Hybrid Systems: Computation and Control,
time [s] vol. 385, no. 3, pp. 407–427, 1999.

Fig. 16. State (experiment), (solid:follower, dashed:reference), (R =


diag(5 × 10−5 , 5 × 10−5 ), N = 2).

459

You might also like