Neural Network-Based Path Planning

for Mecanum-Wheeled OmniDirectional Robot

Ing Irving Adi Nugraha Ismail (10041882)

Mochammad Rizky Diprasetya (10053435)

Contents
1.

Introduction ........................................................................................................................................... 3

2.

Mobile Robot ........................................................................................................................................ 3

3.

Task ....................................................................................................................................................... 4

4.

3.1

Kinematics .................................................................................................................................... 5

3.2

Environment .................................................................................................................................. 8

3.3

Path Planning Concept .................................................................................................................. 9

References .......................................................................................................................................... 12

1. Introduction
Nowadays, the development of technology is growing rapidly. Researchers are racing to invent a
new breakthrough in order to make human life easier. One of the focus in the recent years is mobile
robots. Mobile robots are widely used in modern industries because of the benefits it offers.
One of the key function of a mobile robot in an industry is to move workpieces. In the old days,
humans used forklifts in order to move workpieces between workstations. Now, many industries are
looking into mobile robots to replace forklifts. This is because mobile robots simple structure, easy to
control, efficient and moves smoothly. It can significantly reduce the amount of time needed to move
workpieces around the industry and also reduce the number of accidents during production process.
Despite having all this advantages, a conventional mobile robot has some limitations. The
movement of a conventional mobile robot is limited when travelling in a small, narrow space. In order
to overcome this obstacle, a new type of mobile robot is developed. This so called new type of mobile
robot is called omni-directional robot. It is more mobile than a conventional mobile robot because it
can move in any direction without turning its body.
This research was made to improve a Mecanum-wheeled omni-directional robot by introducing
path planning in it. The path planning was based on neural network.

2. Mobile Robot
The Mecanum-wheeled omni-directional robot in this research was designed by our colleague,
Kshitij. The robot has four Mecanum wheels. Figure 1 shows the 2D drawing of the Mecanum-wheeled
robot.

Fig. 1. Technical Drawing of Mecanum-wheeled Omni-directional Robot

Eight ultrasonic sensors will be attached to the robot, two at each side. The ultrasonic sensors
were used for detecting obstacles near the robot.
The basic idea of controlling the Mecanum-wheeled omni-directional robot movement direction
is by manipulating rotational direction of each wheel. Each wheel is driven by a single motor which
are controlled separately, so we can manipulate each wheel behavior one by one. Figure 2 shows the
resulting movement direction of the robot from different sets of wheel behavior.

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

= wheel
= wheel rotational direction
= mobile robot movement direction

Fig. 2. Movement Direction of The Robot Based on Different Wheel Rotation Direction
(a) Forward, (b) Backward, (c) Right, (d) Left, (e) left forward, (f) left backward, (g) right
forward, and (h) right backward

3. Task
Our task is to develop a path planning using neural network. In order to achieve this goal, we must
have a solid understanding about the kinematics of the Mecanum-wheeled omni-directional robot. We
also must choose the suitable environment for the operating area of the robot. Finally, the neural
network can be developed based on the previous understanding.

3.1 Kinematics
The Mecanum wheels were rotated by electric motors. The rotation of the whole wheel caused each
angled roller to rotate. Because of the angle of the roller, the rotational force generated by the motor is
translated into two different direction: the direction of the rotational force and normal to the rotational
force. We can control the direction which the robot is moving to by manipulating the rotational direction
of each wheel (clockwise or counter-clockwise) and the velocity of each wheel. This made the robot be
able to move in any direction smoothly without hardware changes.
As mentioned above, the forces generated by each wheel can be broken down into two different
force vectors as shown in Figure 3. The resultant net force direction from each of the four wheels is the
direction the robot is moving to.
V1r
Y

V2r

V2W
X

V1W

Wheel 1

L
X

Wheel 2

vx

vy
V3W
X

COG

V3r

z
V4r
X

V4W

Y
Wheel 3

Wheel 4
l

Fig. 3. Kinematics of Mecanum Wheel

ViW (i = 1,2,3,4) R is the velocity vector from the wheel rotation at each wheel.
Then,
ViW = RW* iW
where,
RW = radius of the wheel
iW = angular velocity of the wheel

ViR (i = 1,2,3,4) R is the tangential velocity vector of the roller touching the surface.

Then we can get the net velocity vector on each wheel in coordinate X and Y,
ViX = ViW + ViR/cos 45 , ViY = ViR/cos 45

(1)

From the whole robot point of view, the robot itself has its own velocity in direction X vx and Y vy
and angular velocity z thus resulting:
ViX = vx l. z , ViY = vy + L. z

(2)

By doing mathematical manipulation on equation (1) and (2), we get:

V1W = vx vy (L+1).z

(3)

V2W = vx + vy + (L+1).z

(4)

V3W = vx + vy (L+1).z

(5)

V4W = vx vy + (L+1).z

(6)

We got the inverse kinematics by combining equation (3) (6):

V W = J * V0
1 1 ( + )
1 1
( + )
= [
]
1 1 ( + )
1 1 ( + )
where:
V0 = [vx, vy, z]T R3*1 is the velocity vector in Cartesian coordinates
VW = [V1W, V2W, V3W, V4W] T R4*1 is the wheel velocity vector corresponding to the
angular velocity.

The forward kinematics is derived as

Computing this matrix results to:

in global reference coordinates (X, Y, Z).

The direction of the resultant motion, :

in a stationary coordinate axis (x, y, z).

Based on the kinematics equations mentioned above, we can get the position, direction, and the
velocity of the mobile robot. We can calculate the velocities needed on each of the four wheels to do
desired mobile robot motion by using the inverse kinematics equation. We use the forward kinematics
equation only if we know the velocities of each of the four wheels to predict the motion of the mobile
robot.

3.2 Environment
The area of operation of the robot is shown in figure 4. The area is made up by 3 different labs.
There are obstacles in the area marked by green color.

Lab 1

Lab 2

Lab 3
= Obstacle
= Wall

Dimension in meters (m)

Fig. 4. Current Map of the Environment

The current map of the environment is then converted to binary occupancy grid as shown in
figure 5. This is developed to be used in MATLAB Simulation.
Binary Occupancy Grid

Fig. 5. Binary Occupancy Grid of the Environment

3.3 Path Planning Concept

The neural network will be built up based on the sensors of the robot and the environment. In this
case, we will have 2 neural network. The first neural network is for determining space without obstacle.
The second neural network is for determining if the free space within the goal position.
As mentioned before, the mobile robot has 8 ultrasonic sensors. Each of the sensors will be the
input for our first neural network. Let s1 to s8 are the ultrasonic sensors. S1 to s8 can be only 1 or 0.
Each sensor has its own threshold to determine value of 1 or 0. Figure 6 shows the location of each
sensor.

Fig. 6. Location of Ultrasonic Sensors

The output of the first neural network is a free space which the mobile robot can goes through that
space. Each space will be determined by the combination of the sensors. Figure 7 shows the space
fragment.

Fig. 7. Space Fragments

For example, if s1 and s2 detect an obstacle, then space a1 is obstructed. If only s1 detects an
obstacle, then space a12 is obstructed. The combination of the sensors and the output can be seen in
table 1.
Table 1
Possible Combination of Sensors Value Input and the Obstacle Location Output
Sensor Value (Input)

Obstacle Location (Output)

s1

s2

s3

s4

s5

s6

s7

s8

Front

Right

Rear

Left

FFL

FFR

BBL

BBR

FLL

FRR

BLL

BRR

The first neural network can be built up based on above concept. Figure 8 below shows the neural
network model.

Fig. 8. The First Neural Network Model

The second neural network will determine if the free space has the same direction with the goal
location. Let g1 to g12 are the goal segment for each space a1 to a12. For example, if the a1 and a2 are
not obstructed, but only a1 has the goal fragment then the g1 is 1, the other goal fragment is 0. The robot
should move in g1 direction. The output of the second neural network will be the movement of the robot.
The robot has 8 movements; Forward, backward, rotate, move at angle 90, move at angle 22.5 and
move at angle 67.5. These outputs will be sent to the motor controller.

Fig. 9. The Second Neural Network Model

4. References
1. Kiran Wakchaure, KINEMATICS MODELLING OF MECANUM WHEELED MOBILE
PLATFORM. [Online] Available:
http://www.academia.edu/4557426/KINEMATICS_MODELLING_OF_MECANUM_WHEELE
D_MOBILE_PLATFORM. Accessed on: Nov. 14 2016.
2. X. X. Wu, X. L. Zhang, H. D. Zhou, W. S. Chou, "Modeling and Simulation of Omni-Directional
Mobile Robot with Mecanum Wheel", Applied Mechanics and Materials, Vol. 624, pp. 417-423,
2014
3. L. Miclea, Ed., 2014 IEEE International Conference on Automation, Quality and Testing,
Robotics: 22-24 May 2014, Cluj-Napoca, Romania. Piscataway, NJ: IEEE, 2014.