Localization and Obstacle Avoidance Using

Fuzzy Logic and Neural Network

Chi Kai Hui, Feriyonika, Raras Tyasnurita, Pendry Alexandra, Min Fan Ricky Lee*

dependence one to another. When one of them has trouble, it

Abstract—This project is arranged to design a Fuzzy Logic and will cause mobile robot system also get into trouble. These
Neural Network control implemented in Mobile robot (AmigoBot three problems have to be solved so that mobile robot can
and P3DX) that will traverse through a maze with arbitrary avoid the obstacle well.
obstacles and transmit information to a Personal Computer (PC) The mobile robot system always obtains information from
that will graphically display the robot’s telemetry information
environment through two sensors (sonar and laser range
and maze layout. For input data, controller (Fuzzy Logic and
Neural Network) receives data from sensors (sonar and laser finder). To avoid the obstacles successfully, accurate
range finder). Although sonar is used in mobile robot navigation localization is required which is to obtain the robot’s pose
widely, but in recent years, the laser technology has also been (position and orientation). Sonar data for the global mapping
considered as a powerful and accurate scheme for can be using fuzzy sets to model environment and indentify
self-localization of the autonomous mobile robots (AMR). We will robot’s position and orientation. Make the good accuracy and
observe which the best controller (Fuzzy Logic or Neural performance of sonar data detection [2]. On other case,
Network) to used and give valid data. The empirical results show Ganapathy (2009) concluded that his proposed behaviors of
the effectiveness and the validity of the obstacle avoidance the controllers with Path Remembering make the mobile robot
behavior of the proposed Neural Network control strategy by
to be capable in reaching the desired goal even when the
using laser range finder as the sensor.
environment consists of some odd shaped acute obstacles. The
Keywords --- Mobile Robot, Sonar, Laser Range Finder, Fuzzy author also found that the best combination of the algorithms
Logic, Neural Network, Obstacle Avoidance. is obtained when the Obstacle Avoidance behavior is
controlled by Artificial Neural Network [3].
Harb (2009) proposed Neural controllers for local
I. INTRODUCTION navigation, Neural Networks for environmental recognition,
and a Fuzzy system for speed control. The system is fed
M OBILE robot has been used for various purposes in
many application fields include transportation,
exploration, industry, service, education, etc. The
off-line by a simulated model of a laser range-finder. These
major components of the control system perform a global
Neural navigation and a Fuzzy-Neural speed control that guide
working environment contained unknown and unstructured
a mobile robot to track its predefined path to arrive to its final
terrain and obstacles. To make a path planning in the free
goal through a set of sub-goals, or autonomously plan its path
space of the configuration space is one critical task in the
to arrive to the desired final goal, while avoiding obstacles that
intelligent mobile system. The uncertainties in the
are found along the way [4].  
environment affect the obstacle avoidance algorithms during
Tracking the position and orientation of sensor data is one
navigation process.
of important conditions of successful experiment.
Jensfelt(2001) present laser scanning with a Kalman
In this paper, we investigate three general problems (sensor,
filter-based approach utilizing a minimalistic environmental
technical, and algorithm problems) where they have
model. The minimalistic model paves the way for a
low-complexity algorithm with a high degree of robustness
Chi Kai Hui is with the Graduate Institute of Automation and Control,
and accuracy [5]. Laser range finder is also needed in motion
National Taiwan University of Science and Technology (e-mail: of mobile robot. The achievement which wanted is to estimate
[email protected]). the posture of robot. Fuzzy extended information filter is used
Feriyonika is with the Graduate Institute of Automation and Control, to improve the accuracy of robot’s posture estimation and also
National Taiwan University of Science and Technology (e-mail:
[email protected]).
make efficacy results in simulation and experiment [6].
Raras Tyasnurita is with the Department of Industrial Management, Motlagh (2009) prove that Fuzzy systems have the ability to
National Taiwan University of Science and Technology (e-mail: treat uncertain and imprecise information using linguistic
[email protected]). rules. They offer possible implementation of human
Pendry Alexandra is with the Department of Mechanical Engineering,
knowledge and experience, and have an advantage in that they
National Taiwan University of Science and Technology (e-mail:
[email protected]). do not require a precise analytical model of the environment
Min Fan Ricky Lee is with the Graduate Institute of Automation and [7].
Control, National Taiwan University of Science and Technology, Taipei Regarding with keeping track of the position and orientation
Taiwan (*corresponding author to provide phone: +886... ;e-mail:
[email protected]).
over time using sensor data, i.e., pose tracking, which is a
central component in many mobile robot systems, Yau-Zen
(2007) using a scanning laser range finder to acquire distance obstacle. We observe error data from sensors (sonar and laser
information for motion planning with wall following strategy range finder) and analyze them so that we get the valid data.
[8]. The valid data will ensure that given information is also valid.
There have been previous researches about obstacle 1. Sonar issue
avoidance and dynamic environments reorganization with Having explored behavior of the system, we found that all
mobile robot. Some of them are using Fuzzy Logic and Neural eight sonar data were not suitable with real condition. It
Network to recognizing environment and controlling speed of apparently caused by material reflecting the sonar wave back.
mobile robot [1]-[4]. We are challenged to implement the Maze we used consist of two materials (aluminum and wood)
method in those papers [1]-[4] to AmigoBot and P3DX and which have different characteristic in reflecting sonar’s wave.
also try to extending several conditions that different with the Figure 1 to 3 show the example of this problem.
case in the previous researches. This paper presents the error
analysis for both sensors and the approach to eliminate the In Figure 2, we can see that just sonar 4 (s4) and sonar 5
sensory data error to enhance the sensory information equality. (s5) which out of the region. For this condition, the designed
We propose to analyze behavior of the sensors data so that we system will give command to the mobile robot to go forward.
find valid data based on real environment. This part is the first The error reading, making robot is like in Figure 3 so that the
step in mobile robot control system. If this part is wrong, it designed system will command to the mobile robot to turn left.
will cause all system will wrong as well.
Next is mechanical problem which is caused by angle error
and limitation. To solve it, we observed several angles and
commanded to mobile robot to move so that we get the error.
Furthermore in mobile robot control, it is usually found
difficulties in modeling caused complexity of the system. The
uncertainty in the model of the mobile robot is contributed by
mobile robot system (slip, center of wheel alignment and
usage time) and environment (uneven, unstructured and
unknown terrain). Therefore the autonomous mobile robot
system appeared to have nonlinear characters. This part will
cause us to design the algorithm to do the obstacle avoidance.
To solve it, we propose a soft computing methodology
(Fuzzy Logic and Neural Network) based intelligent behavior Fig. 1 configuration of sonar 1 – sonar 8
control system for the mobile robot (AmigoBot and P3DX).
The obstacle avoidance behavior was realized through Neural
Network to model the environmental situation that mobile
robot encountering during the travel and followed by the
motion control through the Fuzzy Logic to infer the avoidance
action from a rule base associated with the recognized
situation. The Neural Network method to re-organize
environment is approached reference [1], but the different part
between our implement is using different sensor which is laser
range finder. He advantageous of this system is the mobile
robot can be controlled without implicit modeling of mobile
robot and environment. We investigate several methods and
strategies to design Fuzzy Logic and Neural Network to
control mobile robot. After that, we implement to the mobile
robot and comparing them to find which one is better than the Fig 2 Real position of mobile robot
other.

II. ANALYSIS
In the progress of experimenting, we can observer 3
different issues of difficulties. We try to analysis the problem
and to clarify 3 different parts about sensor issue, mechanical
issue and task control algorithm issue.
A. Sensor Issues
This part is the first step in this system, when this part is
wrong it will cause all decision and control will wrong as well.
In sensor problem observation, we construct several conditions
of the obstacles and command to mobile robot to reading the
data. The data will be analyzed and compared to the real
 Fig 6 Constructed obstacle to find the laser data error

Fig 3 Error data reading gives information to system that

B. Mechanical Issues
Mobile robot is in this position so that it will turn left (1) Power limitation
Our test confirms that both the AmigoBot and P3DX cannot
In this part, we observe whether the given data from sonar is move in a small angle (less than 3 degrees for P3DX and less
valid or not. This part is done by construct obstacle or put than 1 degree for AmigoBot). This limitation result in a
mobile robot in maze and read the data. Figure 4 shows the looping output control action. The mobile robot failed to
constructed obstacle and read the data. follow the control command to achieve the desired pose. No
actual pose change make the range sensor continuously
feedback the same range measurement to the controller which
same control command been output with no execution and
results in the mobile robot pose locked without moving.
(2) System Error
The left or right turn action from the task controller is input
to the low-level controller to control the robot to achieve the
desired steering to avoid obstacle in the maze. Preliminary test
done shows that there are errors exist between the desired and
actual orientation. The accumulation of the error results in
large orientation error in the robot and can lead to hit the
obstacle. ..
Fig 4 Constructed obstacle to find the sonar data error  

2. Laser issues C. Task control algorithm issues

Another sensor we used is laser range finder. Because we (1) Obstacle avoidance
divided its 681 data to be 18 regions, we found that sometime Fuzzy logic is used as the task control decision algorithm,
in every region have zero data (mean obstacle very close) but velocity and heading, to navigate the robot travel along in the
there was nothing in there. This problem is caused by material trajectory with obstacle presented. In Fuzzy Logic algorithm,
reflecting ability, Figure 5 shows the problem. we found difficulties of contradiction in rule design. If the rule
design happen contradiction will cause the robot confuse the
environment and hit maze. We decided that if we wanted to
use Fuzzy logic to control the mobile robot, we had to explore
and research appropriate rule so that it can control mobile
robot well.

(2) Learning
Neural Network algorithm is adopted as a training
mechanism and using Backpropagation method to train the
environment. We found the problem here that we need many
set data training to be trained to the mobile robot. This is little
bit hard to construct many conditions so that we get enough
data to be trained.
Fig 5 there are error data reading of laser data in region a3 and a14
III. METHODOLOGY
This part has same method as sonar observation. Figure 6 A. Technical Problems
shows the obstacle constructed and we read the data.
When mobile robot does not move like we want, it will
cause mobile robot hit the obstacle although system
understand where mobile robot should go. In this observation,
we need to find error angle between real turn and command so  
that we can command mobile robot well. In this part, we use
computer to type the command and protractor to check how B. Algorithm Problems
many angle error. The last problem is algorithm problem (Fuzzy logic and
We observe behavior of the mobile robot when given simple Neural Network). If all information and technical are solved
command (left turn and right turn). We design fix turn left and but wrong in command it will make mobile robot can’t avoid
turn right of the mobile robot with certain parameter. Figure 7 obstacle well. In this part we use sonar and laser range finder
is design of mobile robot turning, while Figure 8 is design of to get data, computer to type the command and measurer to
mobile robot angular velocity. look appropriate distance when mobile robot give response.
After we get the fix design, we command to the mobile We assumed that the sensor data error was found, so we use
robot to turn in certain degree and look for the error. the data as input in Fuzzy Logic and Neural Network control.
To study the problem, we construct maze to look the respond
and analyze the behavior.

1. Fuzzy logic
R
VL VR To design this control, we have to design output of system,
input membership function, use several Fuzzy Logic rules and
look the respond. These three parts will be important part in
ICR
Fuzzy Logic design. To get good output, usually it needs many
trial and error so that we get appropriate rule and system
configuration. Figure 9 is configuration of the laser range
finder with three defined region (close, middle and far). Figure
10 shows the output of Fuzzy control whether Figure 11 shows
the input. Figure 12 is the flow chart for the Fuzzy Logic
L L
control.
2 2
Fig 7 Instantaneous center of rotation (ICR)

Angular Angular
velocity velocity

WL
WR

t0 t1 t2 t3 Time t0 t1 t2 t3 Time
close
Fig 8 Design of angular velocity
middle
With far(obstacle)  
Fig 9 Configuration of laser range finder as input of control

§ Output
left left-left front left front-front front right front- front right
right -right front

-90 -80 -70 -60 -50 -40 -30 -20 -10 0 10 20 30 40 50 60 70 80 90

Fig 10 Output of fuzzy logic control
Where
§ Input membership function
 
 
       

 
 
clo se m id d le fa r start

reading data from

laser range finder

Quantization to be 18 region

0 10 20 30 40 50 60 70 80 90 100 500 cm   FUZZY CONTROL

Fig 11 Input of fuzzy logic control

§ Rule
no no
theta =0? theta < 0 ?

If a9 is far OR a10 is far OR a11 is far OR a12 is far THEN theta

is front yes yes

If a8 is close OR a8 is middle THEN theta is left front-front

go forward turn left turn right
If a9 is close OR a9 is middle THEN theta is left-left front
If a13 is close OR a13is middle THEN theta is right front-front
If a12 is close OR a12 is middle THEN theta is right-right front
If ((a15 is middle OR a15 is close) OR (a14 is middle OR a14 is
close ) OR ( a13 is middle OR a13 is close) OR (a12 is middle no
stop?
OR a12 is close) OR (a11 is middle OR a11 is close ) OR (a10 is
middle OR a10 is close ) OR (a9 is middle OR a9 is close ) OR yes
(a8 is middle OR a8 is close ) OR (a7 is middle OR a7 is close )
OR (a6 is middle OR a6 is close )) AND ((a16 is middle OR a16
is close ) OR (a17 is middle OR a17 is close ) OR (a18 is middle
stop
 
OR a18 is close ) OR (a19 is middle OR a19 is close )) THEN Fig 12 Fuzzy logic control flow chart
theta is right
If ((a15 is middle OR a15 is close) OR (a14 is middle OR a14 is 2. Neural Network
close ) OR ( a13 is middle OR a13 is close) OR (a12 is middle
In Neural Network control, we define nineteen areas (a1-a19) to
OR a12 is close) OR (a11 is middle OR a11 is close ) OR (a10 is
middle OR a10 is close ) OR (a9 is middle OR a9 is close ) OR
be eight areas (s1-s8) which is showed in Figure 13.
(a8 is middle OR a8 is close ) OR (a7 is middle OR a7 is close )
OR (a6 is middle OR a6 is close )) AND ((a5 is middle OR a5 is
close ) OR (a4 is middle OR a4 is close ) OR (a3 is middle OR
a3 is close ) OR (a2 is middle OR a2 is close )) THEN theta is
left.

Table  1  Rule decision

Decition Result Enviorment Condition
a9 is Far OR a10 is Far
Front OR
a11 is Far OR a12 is Far

Left-Front-Front a8 is close OR a8 is middle

Fig 13 Configuration from a1-a19 to s1-s8

left-Front-Front a9 is close OR a9 is middle

§ Region
To make simple in training data, we define one region (
Right-Front-Front a13 is close OR a13 is middle adjustable ). If obstacle inside the region, we set this
region to be one (1) otherwise to be zero (0). This region
means when system will start to respond, if we want
Right-Right-Front a12 is close OR a12 is middle mobile robot respond obstacle since long distance, we
just set the region bigger.
{a6 - a15 is close OR a6 - a15 is middle}
Right AND § Data training
{a16 - a19 is close OR a16 - a19 is We get data training from simulation placement of
middle}
obstacle. For example, a obstacle position is like as
{a6 - a15 is close OR a6 - a15 is middle}
showed in Figure 14.
Left AND
{a2 - a5 is close OR a2 - a5 is middle}
 bias 0 bias 1
bias 2 bias 3

S1

S2

S3

S4

  S5 OUTPUT

S6
Fig 14 Obstacle in right front S7

S8

INPUT HIDDEN LAYER 1 HIDDEN LAYER 2 HIDDEN LAYER 3  

Fig 16 The used Neural Network structure

start

φ
reading data from
laser range finder

 
Fig 15 Respond of mobile robot Quantization to be 8 region and
set to be zero/ one based on data distance

For this case the input of the system is

0 0 0 0 0 1 0 0
The output is φ degree (turn left φ degree) as showed in Neural Network control

Figure 15. For all data training we construct all obstacle

possibility and define respond of the output. This part needs
time and has to do carefully because Neural Network will no no
theta =0?
follow this rule give the output like we define in here.
theta < 0 ?

yes yes
§ Neural Network training
go forward turn left turn right
To train data and build Neural Network, we will use Matlab
Neural Network function.
In this project, we use Backpropagation with several method
e.g. gradient descend (traingd), gradient descend with adaptive
no
learning rate (traingda), Levenberg-Marquardt (trainlm), stop?
Fletcher-Reeves Update (traincgf). We try all method and look
for best method can recognize the pattern and give smallest yes

error.
First, we make new network. Then we set parameters, such
stop
 
as how many iteration in simulation, Mean Square Error Fig 17 Neural Network control flow chart
(MSE), how fast simulation learn the pattern of relationship
between input and target, and when the Matlab show the
simulation result (in this setting, Matlab will show the result IV. EXPERIMENTAL RESULTS
every 100 iteration). If goal is reached (convergent), Matlab
will stop the simulation. The final result of simulation is that A. Sensor Observation
we get the weight values connecting among input to hidden In this observation, we want to find valid data from
layers and output.     environment. We use sonar and laser range finder to get data.
After we try several algorithm methods and hidden layer Figure 18 shows the error data found in laser range finder.
amount combination, we decide that combination with 9
neurons in hidden layers 1, 8 neurons in hidden layer 2, and 6
neurons in hidden layer 3 are Neural Network structure we
use. This combination is chosen manually after trying possible
combination of hidden layer or neurons. Figure 16 shows the
Neural Network structure. Figure 17 is the flow chart for
Neural Network control.
 Laser data reading error in 118 From several method we have tried, we find that neural
3000
network with three hidden layer (with 9 neurons in hidden
2500 layer 1, 8 neurons in hidden layer 2, and 6 neurons in hidden
layers three) is the best structure of neural network. It is mean
2000
that this system can recognize the data pattern trained. Then
Distance(mm)

Figure 19 shows target and Neural network output.

1500

1000

500

0
Error data
-500
0 200 400 600 800 1000 120
0
Laser detector number

Fig 18 Laser range finder data condition

B. Technical Observation
In this part, we observe behavior of the mobile robot when
given simple command (left turn and right turn).
From this observation, we get minimum angle so that  
mobile robot cannot move and get mobile robot turning same
as we command. Table 2 shows the turning error and Table 3 Fig 19 Target output and Neural Network output
shows the minimum turning error.
V. DISCUSSION
Table 2 Turning error for this design
A. Sensor Observation
command result
In this observation we get conclusion that sonar can’t read
45 90 data well in maze structure. We get the error in sonar 2 and
90 135 sonar 6. This is caused by material of maze consisting of
aluminum and wood. Alternatively, we use laser range finder
135 270 to get the data. Although laser ranger finder also sometimes
270 315 has error data, we can eliminate it by taking average data but
in no zero point.

Table 3 Minimum turning error for this design

B. Technical Observation
From the technical problem, we observe error in turning
Degree result
mode and minimum degree where both wheel of mobile robot
commanded Left wheel Right wheel can move together. For turning mode, we get Constanta 2/3
which is times with degree we want so in program we times
1 No move move
this constant to the angle degree.
2 No move move Other observation result in technical problem is minimum
3 No move Move degree where both wheel can move together. We get that when
degree less than 10 degree, just one wheel move but more than
4 No move Move 10 degree all wheel can move together. For degree less than 10
5 No move Move degree we find constant (4) which is to time with degree
commanded to mobile robot.
6 No move Move
7 No move Move
C. Algorithm Observation
8 No move Move
1. Fuzzy logic control
9 No move Move This system does not work well. The ability to avoid the
10 No move Move obstacle is very unstable, so that makes the mobile to hit the
maze or obstacle sometimes. From Figure 20 we can see
11 move move trajectories of mobile robot when use Fuzzy Logic as
controller and find out the hitting place by marked.
C. Algorithm Observation
Neural Network
 2000
VI. CONCLUSIONS
1500
In this paper, there are three big parts which decide whether
1000 all system will work or not. They are sensor, algorithm, and
500 technical part. To make system can avoid the obstacle well, all
0
those part have to be in good performance. If one of them is
wrong, it will cause all system does not work well.
-500
We get conclusion that sonar can’t read data well in maze
-1000 structure. Alternatively, we use laser range finder to get the
Hit 2
-1500 data. Although laser ranger finder also sometimes has error
-2000
Hit 1 data, we can eliminate it by taking average data but in no zero
point.
-2500
-3000 -2000 -1000 0 1000 2000 3000

  Regarding with the proposed method to optimally avoid the

obstacle, system with Fuzzy Logic control does not work well
Fig 20 Trajectories and maze plot
in consistency. Sometimes it can avoid the obstacle but
sometimes it can’t avoid. On the other hand, system with
2. Neural Network control
Neural Network control can successfully avoid the obstacle.
This system can avoid obstacle well. The sensitivity respond
Finally, this system will work best with following
based on region we define. When we define too far, system
procedures :
will respond to fast so that make its movement is not stable.
1. System connect to server to connect to mobile robot
On the contrary, if it is too close, system will respond to slow
(AmigoBot).
so that make mobile robot hit the obstacle before it get
2. System take data from sensor (laser range finder). It is a
decision to avoid it. From our observation, we get the
quantization to be eighteen (18) region for Fuzzy control
appropriate distance for the region which is 700 mm, showed
and eight (8) region for Neural Network control.
in Figure 21. This distance is suitable for the maze system.
3. The reading data will be substituted in controller (Neural
Network control).
4. Output from controller will be used to command mobile
robot. If it is not equal to zero, it will be substituted to
turning function (turn left, turn right).

To get good respond of Neural Network, it needs further

experimental to get good data training. The more valid trained
data, the better Neural Network system we have. In Fuzzy
Logic control, it needs many investigations in building rule.
These rules will decide the respond of the Fuzzy system.
For the application in real world, this research can be used
in ruin building to find the victims. This system can find way
700mm   in complex obstacle. With improvement in recognizing
Fig 21 Distance for the sensor region when Neural Network start to consider
victims characteristic, this system can be used for that purpose.
the environment
To look the respond, we record the trajectories when it
avoids the obstacle. Figure 22 shows its trajectories when use REFERENCES
neural network as controller. [1] Beom, H. R. and H. S. Cho (1992). “A Sensor-based Obstacle
Avoidance Controller For A Mobile Robot Using Fuzzy Logic And
2000
Neural Network.” Intelligent Robots and Systems, 1992., Proceedings of
1500 the 1992 lEEE/RSJ International Conference.
[2] Demirli, K. and I. B. T¸rksen (2000). "Sonar based mobile robot
1000
localization by using fuzzy triangulation." Robotics and Autonomous
500 Systems 33(2-3): 109-123.
[3] Ganapathy, V., Y. Soh Chin, et al. (2009). “Fuzzy and Neural controllers
0
for acute obstacle avoidance in mobile robot navigation.” Advanced
-500 Intelligent Mechatronics, 2009. AIM 2009. IEEE/ASME International
Conference.
-1000 [4] Harb, M., R. Abielmona, et al. (2009). “Speed control of a mobile robot
-1500
using neural networks and fuzzy logic.” Neural Networks, 2009. IJCNN
2009. International Joint Conference.
-2000 [5] Jensfelt, P. and H. I. Christensen (2001). "Pose tracking using laser
scanning and minimalistic environmental models." Robotics and
-2500
-3000 -2000 -1000 0 1000 2000 3000 Automation, IEEE Transactions on 17(2): 138-147.
  [6] Lin, H. H. and C. C. Tsai (2008). "Laser pose estimation and tracking
Fig 22 Trajectories and maze plot using fuzzy extended information filtering for an autonomous mobile
 robot." Journal of Intelligent and Robotic Systems: Theory and
Applications 53(2): 119-143.
[7] Motlagh, O. R. E., T. S. Hong, et al. (2009). "Development of a new
minimum avoidance system for a behavior-based mobile robot." Fuzzy
Sets and Systems 160(13): 1929-1946.
[8] Yau-Zen, C., H. Ren-Ping, et al. (2007). “A Simple Fuzzy Motion
Planning Strategy for Autonomous Mobile Robots.” Industrial
Electronics Society, 2007. IECON 2007. 33rd Annual Conference of the
IEEE.