1

Fuzzy Parking Manoeuvres of

Wheeled Mobile Robots
Amar KHOUKHI, Luc BARON and Marek BALAZINSKI

Department of Mechanical Engineering, École Polytechnique of Montreal

P. O. box. 6079, Succ. CV, Montreal, Quebec, Canada H3C 3A7
amar.khoukhi@polymtl.ca , luc.baron@polymtl.ca, marek.balazinski@polymtl.ca

Abstract - This work deals with the parking manoeuvres

problem for a wheeled mobile robot (WMR). The robot has tracking method is to design a control algorithm that makes a
the same non-holonomic kinematic constraint that has a car wheeled mobile robot (WMR) follow a reference trajectory.
vehicle. This constraint makes the robot having its direction The pose stabilization method is to stabilize the WMR to a
always tangent to the trajectory. Two sub-cases of parking desired final pose, while starting from a given initial pose [5-
problems are considered. These are forward and backward 8].
maneuvers, aiming to stabilize the robot at a pre-specified From kinematics point of view, parking is merely a path-
pose. The environment is assumed to be known, obstacle-free planning problem, in which a collision-free path between
and a local map of the area is already done by prior processing initial and final poses should be found, while meeting the non-
the information obtained from ultrasonic sensors mounted on holonomic constraint. Hence, many continues-curvature path
the robot. A linguistic fuzzy model to represent the robot and planners using geometric methods have been proposed, which
its environment is developed. From this model, the parking generate feasible paths that are followed by implementing a
manoeuvres are carried out by mimicking a human car driver path tracking technique. In these methods, the trajectory
using a fuzzy control system. Upon simulation tests this planner should be combined with a controller capable of
approach has been proved efficient giving very encouraging tracking the nominal trajectory. For instance, in [12] an
results. optimal open-loop steering approach had been considered
using sinusoidal trajectory parameterization.
Index Terms - Mobile Robot, Non-Holonomy, Parking Generally speaking, the kinematics equations of a car with
Manoeuvres, Fuzzy Planning non-holonomic constraints are nonlinear and time-varying
differential equations. It is almost impossible to find an auto-
I. INTRODUCTION driver, who uses a traditional control method to manoeuvre the
car to perform parking. Nonetheless, a human driver can
In recent years, autonomous parking problems of wheeled smoothly and even perfectly park a car into a garage or
mobile robots (WMR) have attracted a great deal of attention parking zone by some simple rules without the knowledge of
from robotics research community and automobile industry [1- the motion kinematics of the automobile. Indeed, consider the
3]. In robotics literature, a tremendous amount of research has behaviour of a human driver; according to available space, one
been published in the last decade on the topic [4, 5]. In the of the following cases may occur: If the parking spot is not
automotive industry, an intensive research is carried out for enough; the driver cannot park and decides for no parking. If
the development of intelligent automobiles that can make the the space is just sufficient, the driver decides to make a couple
driving procedure easier and more secure [6, 7]. Many of back-drive or head-in manoeuvres. If there is a large
government organizations and laboratories pursue advanced enough space; he decides for a forward parking.
researches on this topic [8-10]. Indeed, the parking problem Therefore, beside geometric based path planners, several
has a wider range of applications. It is used to automate soft-computing [7, 11, 13], especially fuzzy logic based-
service procedures like the loading and unloading of approaches were proposed to solve the parking problem by
warehouses, or the transfer of disabled persons, and can also emulating experience of human drivers. As a matter of fact,
be used to assist the mooring of ships to docks. most advanced control algorithms in autonomous mobile
Basically, the parking problem consists of placing the vehicle robots can benefit from fuzzy logic control. Among these
in parallel to its moving direction in a confined space that is works [7, 11], one can find that about half the parking
very little wider than the vehicle's dimensions. problems are solved using fuzzy control (FC) systems
Many researchers have examined the garage parking and In this framework, an intelligent parking using vision is
parallel-parking schemes. Most of these researches have been proposed in [14], where a color segmentation method is
concentrated on tracking and pose stabilization methods. The developed based on neural networks. They apply quintic
Manuscript received May 5, 2007. This work was supported by the Natural polynomials and symmetric pose to provide good results in
Science and Engineering Research Council of Canada (NSERC), under grants local path planning for automatic parking behaviours. Another
ES D3-317622, RGPIN-203618, RGPIN-105518 and STPGP-269579.
parking motion planning architecture based on a radial basis
 2
function network is reported in [15]. A third approach using a non-slipping. The velocity of the center of mass of the robot is
neuro-fuzzy control for car parking is described in [16]. assumed to be orthogonal to the rear wheels axis. The robot
In this context, this paper considers two parking sub- geometrical and physical properties are assumed to be
problems. These are the forward and backward manoeuvres. symmetrical with respect to its longitudinal axis.
The environment is assumed to be known, obstacle-free and a The robot kinematic motion is given as:
local map of the area is already done by prior processing the .
information obtained by ultrasonic sensors, which are x (t ) = v(t ) cosθ (t ) cos ϕ (t )
mounted on the robot. In section 2, the kinematic model of the
WMR is described. In section 3 the parking problem is .
described. In section 4, the design of the fuzzy manoeuvring y (t ) = v(t ) sin θ (t ) cos ϕ (t ) (3)
algorithm is investigated. Computer simulation results are .
v(t)
given in section 5, and section 6 concludes this paper. θ (t ) = sin ϕ (t )
L
II. ROBOT AND CONSTRAINTS MODELLING where v(t ) is the linear velocity.

The basic architecture of the WMR is displayed in Figure

2.3. Constraints Modelling:
1. The mobile robot is constituted of a chassis or robot body
shaped as a parallelepiped plate, a front wheel and a pair of
ƒ Kinematic non-holonomic constraints:
rear coaxial drive wheels represented by identical cylinders
These consist of a set of non-integrable equations (3). They
with the same radius.
impose that the direction of motion must be always tangent to
the robot trajectory.
2. 1. Pose model of the WMR
Two reference frames, namely world coordinate frame {W } ƒ Final pose satisfaction:
and robot coordinate frame {R} are considered. The origin of This pose is described by (pT, RT), with pT referring to the
the robot frame is fixed at its center of gravity. These two desired final Cartesian position vector (Fig. 2), and RT to the
frames are related to each other as follows: desired final rotation matrix described by eq. (2). These
constraints are represented by the following inequality
{x } = {x }−
'
R W
W
R R {x R }, (1)
constraints:
Robot motion
where x R denotes the directional column describing the robot
position and orientation with respect to the robot frame, xW
denotes the directional column describing the robot position
and orientation with respect to the world coordinate frame,
x'R indicates the coordinates of the origine of the robot frame
with respect to the world frame.
W
R R is the rotation matrix
defined as:
cos(θ ) − sin(θ ) 0
W
R R =  sin(θ ) cos(θ ) 0 (2)
Fig. 1 Geometric representation of the WMR
 0 0 1 
As shown in Fig. 1, the pose of the robot, moving on a
plane surface at every discrete time i is defined by the
triple ( xi , yi , θ i ) , where (x, y) is the 2-dimentional
(a) (b)
Cartesian position of the center of its front axis and θ is the
robot orientation, which is defined as the angle between the x- Fig. 2. Robot final pose: (a) desired, (b) computed
axis and the axis of motion of the robot. The steering angle
ϕ is defined as the angle between the robot motion axis and p - p T − TTolp = 0 (4)
the axis of the front caster wheel.
vect (R T' R ) − TTolOrient = 0 (5)
2. 2. Kinematic model of the WMR
The front wheel is used to change the robot direction in an
where (p, R ) describes the computed robot final pose, while
angle up to 30 degrees in both directions. It is assumed that the
vect(.) is the axial vector of its 3 × 3 matrix argument, and its
plane of each wheel is perpendicular to the ground and the
contact between the wheels and the ground is pure rolling and norm measures the absolute value of the angle of rotation
 3

between R and R T . TTolp and TTolOrient are, respectively, sign of x to head-in (if x < 0) or to backtrack (if x > 0) until x
=0. To achieve such an objective, a two-inputs/one-output
the tolerances for final pose position and orientation fuzzy control is designed. The inputs are the y-coordinate and
satisfaction. the orientation angle θ . The output is the steering angle rates.
III. THE PARKING MANOEUVRES PROBLEM

To park at a desired pose, the robot might have to perform A

some manoeuvres depending on its initial pose. As the robot
environment is obstacles-free, the manoeuvre concerns only
the parking problem.
Monitoring the behavior of a human driver, it can be seen
that parking involves two major phases. In the first phase, the
parking space and dimensions should be detected by
navigating and setting the robot parallel and ahead of the
parking lot. Then, one of the two following actions is decided.
If the space is largely enough, the vehicle is head-in while C B
turning until it is parallel to the goal point axis. If the space is
too small, no parking holds. If the space is just enough, the
vehicle makes a sequence of backward and forward
maneuvers as follows: in a first phase, the vehicle backtracks (a)- Forward-Backward maneuver
to a certain point and then the front wheel is turned so that the
robot moves towards the parking place until it reaches an A
angle of approach, while its direction axis overlaps with the
goal point axis. Furthermore, according to its initial location,
there are two cases. If the robot is in front of the final pose, it
will have to perform a forward maneuver to get to the target.
Conversely, if the robot is behind the final pose, the maneuver
will be performed backwardly (Fig. 3). The purpose in this
paper is to train the robot to achieve parking maneuvers. The
above discussion suggests that a way to automate parking B
manoeuvres is to use fuzzy control. In the next section, we C
describe a fuzzy control system to perform a manoeuvring
task.
(b)- Backward-Forward maneuver
IV. FUZZY CONTROL APPROACH
Fig. 3. Forward-backward and Backward-Forward maneuvers for a
In order to define the fuzzy control system, and without WMR path starting from a point A and ending at point B
loss of generality, it is assumed for the simulated case study
that the goal-parking place is the world frame origin, defined
by the positional coordinates (0, 0) and zero degree The first step in building our fuzzy control system is the
orientation. Other target poses are also achievable by fuzzification of the input and output variables. The y-
transforming them into the origin pose. The vehicle is first coordinate is defined within the interval [-20 20] (meters)
located at a starting point. Then two cases are considered to (Fig. 5). The associated linguistic variable is defined with 5
park using a forward manoeuvre. If the parking space is large linguistic terms, namely: R (Right), RC (Right Center), C
enough, so the robot starts turning by varying the steering (Center), LC (Left Center), L (Left) (Fig. 6). The orientation
angle, while moving with a constant linear velocity, until it angle θ is defined within the interval [-185 185] (degrees).
reaches the desired goal position and orientation. If the The associated linguistic variable is defined with 7 linguistic
parking space is just enough, the robot is to park using a series terms, namely: LD (Left Down), LU (Left Up), LH (Left
of backward and forward manoeuvres. It starts from a ready- Horizontal), H (Horizontal), RH (Right Horizontal), RU
to-reverse location, then moves backward while varying the (Right Up), RD (Right Down) (Fig. 7). One notes that these
steering angle, with a constant linear velocity, until it reaches linguistic terms cover the whole perceived area by the robot
the parking boundaries, then it navigates forward. This sensing (Fig. 8). Furthermore, the fuzzy sets are defined, such
procedure is repeated until the desired goal position and that any value of θ belongs to two fuzzy sets. This allows
orientation are reached. Figure 4 shows a flowchart dealing with inherent uncertainty of the measured orientation.
representation of the developed approach. Hence, a first step The output linguistic variable corresponds to the steering
to perform towards this objective is to bring the robot on the x- angle ϕ , which is defined within the interval [-40 40]
axis with a zero orientation angle, i.e. at the pose (x, 0, 0). (degrees). The associated linguistic variable is defined with 5
Once this pose is reached, the robot decides according to the linguistic terms, namely: NB (Negative Big), NM (Negative
 4
Middle), NS (Negative Small), AZ (Around Zero), P (Positive
Small), PM (Positive Middle) and PB (Positive Big) (Fig. 9). Y
At this level, one notes that to get a more refined control of
the steering angle, the membership functions of linguistic
terms Horizontal and Center for the orientation angle and or L
the y-coordinate are made tighter.
From the so-defined linguistic variables and fuzzy sets, a LC
Mamdani-like fuzzy control system is built, and two tables are C
deducted, each contains 35 inference rules. Tables 1 and 2
below show the rules defining the forward and backward RC
manoeuvres, respectively. For instance, one can read at the
first line of Table 1, if the robot is on the goal-left of the y-axis R
and if it is oriented right down (RD), then the steering angle
will be Positive Big. Figure 10 illustrates this rule. Fig.5. Fuzzy sets representing the y-coordinate

Fig. 6. Discourse space and membership functions

defining the y-coordinate

LD LU

LH

RH

RD RU
Fig. 7 Fuzzy sets representing the orientation angle

Fig. 4. Flowchart representing the developed approach

V. SIMULATION RESULTS
Fig. 8 Discourse space and membership functions
In this section, computer simulations on a WMR model are defining the orientation angle
given to demonstrate the effectiveness of the proposed parking
system. The robot is represented by a rectangle with 1.5 meter
long and 0.5 large. It is assumed that the steering rates are
updated to maneuver the robot with a constant velocity equal
to 0.1m/sec. The sampling time equals to 0.5 sec. Hence each
0.5 second the robot perceives its environment and decides
what action (the value of steering angle) to take. The motor
coupled with the corresponding wheels can be adjusted to
reach the desired velocity according to a lower level
controller. This process and those of accelerating the vehicle Fig. 9 Discourse space and membership functions
from rest and decelerating it to rest are not mentioned here. defining the steering angle
 5

consists of moving the WMR from the initial pose (xfi, yfi, θ fi)
Table 1. Inference Rules for forward maneuver
located at (10, 20, 150) to a final pose given as (80, 10, 00).
(Columns: θ , Lines: y, Outputs;: ϕ ) Figure 11 shows the simulation result for this trajectory.
L LC C RC R Figure 12 shows the case of parallel parking while performing
forward and backward maneuvers. The task here is to move
RD PB PB PM PM PS the robot from the initial pose (70, 30, -100) to the final pose
RU (20, 45, 00). Figure 13 shows that we can successfully control
PB PB PB PB PB
RH the vehicle to park at the desired position
PB PB PM NS NM
H PM PS ZE NS NM
LH PM PS NS NB NB
LU NB NB NM NB NB
LD y (m)
NS NM NM NB NB
70
Table 2. Inference Rules for backward maneuver 60
(Columns: θ , Lines: y, Outputs;: ϕ )
50
L LC C RC R
40
Initial Pose
RD NS NM NM NB NB 30 Final Pose
RU PS NB NM NB NB
RH PM PS NS NB NB 20
H PM PS ZE NS NM 10
LH PB PB PS NS NM
LU 0 x(m)
PB PB PM PB NS 0 10 20 30 40 50 60 70 80
LD
PB PB PM PB PS
Fig. 11. Simulation of trajectory starting from (80, 20, 300)

Hence, from a current pose ( xi , yi , θ i ) and providing next y (m)

y
steering angle ϕi +1 , the next robot pose ( xi +1 , yi +1 , θ i +1 ) is 70
calculated using motion equations (3). Figures 11 to 13
illustrate some typical examples of the WMR by applying the 60
developed schemes from different initial locations. A first 50
scenario consists of moving the WMR from the initial state
0
placed at (80, 20, 30 ) to the target located at the pose (0, 0, 40
0
0 ). Initial Pose
30
Final Pose
20

10
LD 0
0 10 20 30 40 50 60 70 80
x(m)
Fig. 12. Simulation of trajectory starting from (10, 20, 15 )
0
LH
θ
H

RH
RD

Fig. 10. Illustration of rule 1

As shown in Figure 11, the WMR is able to perform the

assigned task and get to the goal pose. A second example
 6
[10] Waterson, B.J., Hounsell, N.B., Chatterjee, K., 2001.
y (m) Quantifying the potential savings in travel time resulting from
parking guidance systems—a simulation case study. Journal of the
70 Operational Research Society 10, 1067–1077.
[11] Wang, L.X., Mendel, J.M., 1992. Generating fuzzy rules by
60 learning from examples. IEEE Transactions on Systems, Man and
Cybernetics 22, 1414–1427.
50 [12] Murray, R. M. and Sastry, S. S.: Nonholonomic motion
40 planning: Steering using sinusoids, IEEE Trans. Automat.
Control 4 (1993), 700–716.
30 [13] S. Achiche, L. Baron, and M. Balazinski, : “Real/Binary-
Initial Pose Like Coded Versus Binary Coded Genetic Algorithms to
20
Automatically Generate Fuzzy Knowledge Bases: A
10 comparative study, Engineering Applications of Artificial
Intelligence, 17, 4, pp. 313-325, June 2004.
0 [14] W. A. Daxwanger and G. K. Schmidt, “Skill-based visual
0 10 20 30 40 50 60 70 80
x(m) parking control using neural and fuzzy networks,” in Proc.
IEEE Int. Conf. System, Man, Cybernetics, vol. 2, 1995, pp.
Fig. 13. Simulation of a forward-backward parking maneuvers 1659–1664.
starting from the pose (70, 30, -100), ending at (20, 45, 00) [15] D. Gorinevsky, A. Kapitanovsky, A. Goldenberg: Neural
network architecture for trajectory generation and control of
automated car parking, IEEE Transactions on Control Systems
VI. CONCLUSIONS Technology, v 4, n 1, Jan, 1996, p 50-56
[16] R. J. Oentaryo, and M. Pasquier, “Self-Trained Automated
In this paper, we have developed a parking approach to a Parking System,” in 2004 8th International Conference on
wheeled mobile robot. This approach is based on fuzzy control Control, Automation, Robotics and Vision (ICARCV) 2004
systems. It considers non --- holonomic constraints of the robot v : 2, p : 1005 -1010
motion. Simulation results had shown the feasibility and
effectiveness of the proposed technique.

REFERENCES

[1] R. M. DeSantis, R. Hurteau, O. Alboui, B. Lesot: Experimental

Stabilization of Tractor and Tractor-Trailer like Vehicles, Proc. of the
2002 IEEE Int. Sym. on Intelligent Ctrl., p. 188-193, 2002.
[2] Young, W., Thompson, R.G., Taylor, M.A.P., 1991. A review of
urban parking models. Transport Reviews 11, 63–84.
[3] Arnott, R., Rowse, J., 1999. Modeling parking. Journal of Urban
Economic
[4] C. Ahrikencheikh and A. Seireg: “Optimized-Motion
Planning”, John Wiley & Sons, 1994.
[5] T.-H. S. Li, S. J. Chang: Autonomous Fuzzy Parking Control of a
Car-like Mobile Robot, IEEE Trans. Sys. Man & Cyb. Part A, v 33, n
4, p 451-465, July, 2003.s 45, 97–124.
[6] Paromtchik, I. E. and Laugire, C.: Motion generation and
control for parking an autonomous vehicle, in: Proc. of ’96
IEEE Conf. on Robotics and Automation, Vol. 4, Minneapolis,
MN, 1996, pp. 3117–3122.
[7] D. Teodorovic, P. Lucic Intelligent parking systems, European
Journal of Operational Research , Volume 175, Issue 3 , 16
December 2006, Pages 1666-1681
[8] Wong, S.C., Tong, C.O., Lam, W.C.H., Fung, R.Y.C., 2000.
Development of parking demand models in Hong Kong. Journal of
Urban Planning and development—ASCE 126, 55–74.
[9] Polak, J., Axhausen, K.W., 1990. Parking search behavior: A
review of current research and future prospects. Transport Studies
Units, Working Paper 540, Oxford University.