2012 12th International Conference on Control, Automation, Robotics & Vision

Guangzhou, China, 5-7th December 2012 (ICARCV 2012) Fr14.3

A Neural-Network Based Autonomous

Navigation System Using Mobile Robots
Teng Zhao Ying Wang
The School of Computing and Software Engineering The School of Engineering
Southern Polytechnic State University Southern Polytechnic State University
Marietta, Georgia, USA Marietta, Georgia, USA
[email protected] [email protected]

Abstract—This paper presents an autonomous navigation system In this paper, we develop and validate a neural-network based
based on neural networks using mobile robots. While this kind of autonomous robot navigation system using mobile robots. The
navigation system has many applications, there are two main related work in this area is shown in Section II, and the
challenges: the learning capability of the robot, as well as a
problem formulation is shown in Section III. The neural
complex and dynamic navigation environment. The main
contribution of this paper is to develop a navigation system with network architecture for autonomous robot navigation is
learning capability to adapt to an unknown environment. In proposed in Section IV. Finally, the experimental results are
particular, the neural network model is specifically designed for provided in Section V to verify the proposed approach.
our autonomous robot navigation system, and a series of training
samples are developed to train the neural network for the robot. II. RELATED WORK
In addition, we incorporated sonar sensors with the neural Neural networks can be applied to many different fields.
network to solve the problem of autonomous robot navigation.
This approach is validated with the simulation and experimental For example, in neural network related research, there are two
results. It was shown that the robot with the well-trained neural popular areas: applications of neural networks in real world;
network can navigate out of a specifically designed maze and improvement of neural network architecture and related
successfully. algorithms. In the former, researchers compared the
performance of neural network and other statistical methods in
Keywords—Autonomous Navigation, Neural Networks, Mobile the application of medical treatment prediction. In particular,
Robots.
they pointed out the advantages and disadvantages of each
I. INTRODUCTION method, which are described in [1]. In the latter, for example,
Feng et al. proposed an approach to combine the clustering
An autonomous robot navigation system usually consists of
technique with neural networks to improve the neural network
an intelligent mobile robot and various sensors to detect the
model in the application of financial time series prediction,
outside world. Based on the data captured by the sensors, the
which is described in [2].
robot can autonomously navigate in a dynamic and unknown
Several approaches have been proposed to improve robot
environment without the assistance of humans. This technique
navigation. Typically, behavior-based robot navigation is the
can be applied to many fields, such as space exploration,
most common research direction. H. Seraji used a fuzzy logic
robotic wars, as well as home service robots.
based approach to accomplish robot navigation, which is
The challenges of autonomous robot navigation can be
shown in [3]. Using radio frequency identification (RFID), W.
summarized into two aspects. Firstly, the robot will face a
Gueaieb presented an innovative robot navigation technique
significantly complex environment with a variety of obstacles.
[4]. M. Meng proposed an approach of combining vision
In order to complete navigation missions, the robot has to have
analysis with neural networks to accomplish vision-guided
the ability to avoid obstacles or avoid entering a danger zone.
robot navigation [5].
Secondly, the robot will have to own learning capabilities to
In this paper, a modified neural network model is developed
overcome dynamic and unforeseen situations. For example,
to accomplish autonomous robot navigation in a complex
when the robot is running in a designated route, a dynamic
environment, and the neural network provides the robot with
obstacle may appear, and the robot may hit that obstacle.
learning capabilities to meet the challenges of dynamic and
However, the robot can learn lessons for previous experiences
unknown environments. The robot is equipped with range-
and will not hit obstacles again in similar situations.
detecting sonar sensors to obtain immediate data from the
environment. In addition, the neural network will generate
commands based on the training samples and real-time sonar

978-1-4673-1872-3/12/$31.00 ©2012 IEEE 1101

data to navigate the robot out of a maze. between the left edge of the robot and the nearest obstacle on
the left side, ݀ʹ means the distance between the right edge of
III. PROBLEM FORMULATION the robot and the nearest obstacle on the right side, and ݀͵
In Fig.1, a spiral maze is designed to simulate typical means the distance between the front edge of the robot and the
situations which the robot may encounter in the real world. nearest obstacle on the front side. In order to make inputs fit
The robot is required to move out of the maze as soon as for the form of training samples' inputs, three distances are
possible through its sonar date. In addition, the width of the converted further before they are input to the neural network.
road is twice as the width of the robot. As a result, the robot In particular, if the sonar sensors obtain three distances:
will have to make a left-turn when it runs straight and is very 500mm, 300mm, and 200mm, the converted inputs will be 2.5,
close to the right side. 1.5 and 1.
In order to calculate the number of nodes in the hidden
layer, we choose an experience-based method: ݉ ൌ ʹ݊ ൅ ͳ
where ݊ is the number of nodes in the input layer and ݉ is the
number of nodes in the hidden layer [6]. In the neural network
shown in Fig.2, the number of nodes in the input layer is 3, so
there are 7 nodes in the hidden layer.
The output layer of the network is made up of three nodes
including Node ‫ͳݕ‬, Node ‫ ʹݕ‬and Node ‫͵ݕ‬, which are shown
in Fig.2. Each output means a command for the robot.
Especially, ‫ ͳݕ‬means to ask the robot to go straight, ‫ʹݕ‬,
turning left and y3, turning right. In addition, the output node
with the maximum value will be selected as the final
command for the robot, which is the best option relatively
Fig.1. The spiral maze for the robot navigation.
based on our experiments. For example, if the outputs are 0.9,
In order to evaluate the performance of this autonomous 0.4, and 0.1, the robot will go straight.
robot navigation, the navigation time for the robot to get out of
the spiral maze is recorded. B. BP Algorithms
IV. NEURAL-NETWORK BASED NAVIGATION In this paper, the BP algorithm [6] is used to train the neural
In order to help the robot to move out of the maze network, which is shown in Fig.3.
autonomously, a neural network is developed to select optimal
actions for the robot. In particular, the neural network will
accept the sonar measurement data as its input, and select an
action for the robot so that it can navigate in the maze
successfully. The proposed neural network architecture is
shown in Fig. 2.

A. Network Architecture

Fig.3. The pseudo code of BP algorithms [6].

Fig.2. The neural network architecture for robot navigation. The process of the BP algorithm for training neural network
can be divided into six steps:
In the input layer of the network, there are three nodes Step 1: Initialize ܹ݄݁݅݃‫ݏݐ‬ሺܹሻ and ܹ݄݁݅݃‫ݏݐ‬ሺܸሻ with
including Node ‫ͳݔ‬, Node ‫ ʹݔ‬and Node ‫͵ݔ‬, as shown in Fig.2. random float numbers ranging from -0.01 to 0.01.
The robot obtains three original data of distances from the Step 2: Train the network with samples one by one. For
sonar sensors which are ݀ͳ, ݀ʹ and ݀͵. ݀ͳ means the distance each training sample, execute from Step 3 to Step 6 until the

1102
neural network is well trained.
Step 3: Calculate the value of ‫ݎ݁ݕܽܮ݊݁݀݀݅ܪ‬ሺܼሻ and
ܱ‫ݎ݁ݕܽܮݐݑ݌ݐݑ‬ሺܻሻ based on the ܹ݄݁݅݃‫ݏݐ‬ሺܹሻ and
ܹ݄݁݅݃‫ݏݐ‬ሺܸሻ.
Step 4: Calculate the reverse error οܸ and οܹ.
Step 5: Adjust ܹ݄݁݅݃‫ݏݐ‬ሺܹሻ and ܹ݄݁݅݃‫ݏݐ‬ሺܸሻ based on οܸ
and οܹ.
Step 6: Check the value of οܸ and οܹ. If their absolute
values are smaller than 0.0001, then it means the network is
well trained, and the training ends.
After the training process, we gain the trained
ܹ݄݁݅݃‫ݏݐ‬ሺܹሻ and ܹ݄݁݅݃‫ݏݐ‬ሺܸሻǤ Once the neural network is
(a) The weight history of ܹ݄݁݅݃‫ݏݐ‬ሺܸሻ.
trained, the network can calculate the values of
ܱ‫ݎ݁ݕܽܮݐݑ݌ݐݑ‬ሺܻሻ for the specific inputs with the trained
weights [7].
C. Training Samples
Nine training samples are designed to train the neural
network. They are listed as follows:
(1) 1 1 3 1 0 0
(2) 1 2 2 1 0 0
(3) 1 3 1 0 1 0
(4) 1 2 1 0 1 0
(5) 1 3 2 0 1 0
(6) 3 1 2 0 0 1
(7) 2 1 2 0 0 1 (b) The weight history of ݄‫ݏݐ‬ሺܹሻ .
(8) 3 1 1 0 0 1 Fig.4. The weight history when number of nodes in ‫ݎ݁ݕܽܮ݊݁݀݀݅ܪ‬ሺܼሻ is
(9) 2 1 1 0 0 1 7.
For each sample, the first three integers represent the left- In Fig.5, the results of the second experiment are shown.
side distance, the right-side distance, and the front-side Here, the number of nodes in the ‫ݎ݁ݕܽܮ݊݁݀݀݅ܪ‬ሺܼሻis 5. This
distance respectively while the last three integers represent the experiment is used to compare with the first experiment. As
values of the three output nodes, which mean the robot shown in Fig.5(b), ‫Ͳݓ݂݋ݐ݄݃݅݁ݓ݄݁ݐ‬, the top solid line, and
actions: going-straight, turning-right, or turning-left. In the last ‫ʹݓ‬, the bottom dashed line, do not approach a constant value,
three integers, "0" means false and "1" means true. which means the network is not well-trained due to less
The first two samples are designed to teach the robot when number of nodes in the hidden layer.
it should go straight. The following three samples are used to
tell the robot the situations where it should turn right. Finally,
the last four samples are for turn-left situations.
D. Simulation Results
1) Weight History
In order to prove that our neural network model is
successfully designed, we have conducted three experiments
with different number of nodes in the ‫ݎ݁ݕܽܮ݊݁݀݀݅ܪ‬ሺܼሻ.
Fig.4 shows the simulation results for the first experiment. The
number of nodes in the ‫ݎ݁ݕܽܮ݊݁݀݀݅ܪ‬ሺܼሻis 7. This is the
neural network model we use in our autonomous robot
navigation system. Based on the recorded weights, the four
(a) The weight history of ݄‫ݏݐ‬ሺܸሻ .
curves in Fig.4(a) represent the changing trend of the first four
weights in ܹ݄݁݅݃‫ݏݐ‬ሺܸሻ array, while the history of the first
four weights in ܹ݄݁݅݃‫ݏݐ‬ሺܹሻ array is shown in Fig.4(b). As
shown in Fig.4, all weights are finally approaching a constant
value which means the neural network is well trained.

1103
 (b) The weight history of ݄‫ݏݐ‬ሺܹሻ . Fig.7. The maze and the robot in the simulations.
Fig.5. The weight history when the number of nodes in ‫ݎ݁ݕܽܮ݊݁݀݀݅ܪ‬ሺܼሻ
is 5. The average time for the robot to get out of the simulation
maze is approximately 2.67 minutes. The minimum time is
In Fig.6, the number of nodes in the ‫ݎ݁ݕܽܮ݊݁݀݀݅ܪ‬ሺܼሻ is 2.43 minutes and the maximum time is 3.01 minutes. Fig.8
10. As shown in Fig.6, all weights are approaching a constant represents the probabilistic distribution of the navigation time
value which means the network converge when the number of of the robot in 30 simulations.
nodes in the ‫ݎ݁ݕܽܮ݊݁݀݀݅ܪ‬ሺܼሻ is larger than 7.

(a) The weight history of ݄‫ݏݐ‬ሺܸሻ .

FIG.8. THE PROBABILISTIC DISTRIBUTION OF THE NAVIGATION TIME OF THE
ROBOT IN 30.

V. SIMULATIONS EXPERIMENTAL RESULTS

A. AmigoBotTM
As shown in Fig.9, the AmigoBotTM robot is used in our
experiment. It is a small, light-weight, differential drive and
intelligent mobile robot which has 8 range-detecting sonar
sensors: one on the left side, one on the right side, four on the
front side and two on the back side, which make it have the
ability to detect obstacles in all directions. The robot has a
(b) The weight history of ™‡݄݅݃‫ݏݐ‬ሺܹሻ . width of 28cm and a length of 33cm. In addition, it can be
Fig.6. The weight history when number of nodes in ‫ݎ݁ݕܽܮ݊݁݀݀݅ܪ‬ሺܼሻ is 10.
connected a wireless Ethernet.
2) Simulation Results
The simulation environment is an ASUS laptop with
Intel(R) Core(TM) 2 Duo CPU P7350 @ 2.00GHz and 2.00
GB memory @ 1.60GHz. The operating system is Microsoft
Windows XP Professional sp3. In addition, our development
tool is Microsoft Visual Studio 2008.

Fig.9. The AmigoBot robot[8].

1104
B. Experiment Test bed
The experiment test bed (the maze) is shown in Fig.10. which
is built with some traffic cones. However, due to the conical
shapes of the cones, it is found that there are significant
measurement errors of the sonar sensors. As a result, we have
to invert the cones alternatively to increase accuracy, as shown
in Fig.10. The total length of the route in the maze is 1.98m,
and the width of the road is 54cm.

Fig.12.The outputs of the neural-network when receiving three inputs in

Fig.11.

Fig.10. The experiment test bed.

Fig.13 shows the history of the front forward distance, left-
side distance and right-side distance in one experiment. It is
C. Test Results obvious that the robot always avoid obstacles successfully with
a minimum distance greater than 25 cm.
The experiment was conducted for 30 times to validate our
approach's feasibility in a real environment. In this
environment, we analyzed the robot's location and distance
from the walls in three directions, as shown in Fig.11. The
measurements were taken while the robot was running. In the
current location of the robot shown in Fig.11, the front-
forward distance is 37.4cm, the left-side distance is 12.6cm
and the right-side distance is 13.4cm. These three data are the
raw data which will be processed and passed to the neural
network to generate the final command.

Fig.13.The history of three distances.

In Fig.14, it shows the history of actions which the robot

has taken in one experiment. In this graph, "0" means the
Fig.11. Measured data in the real environment. initial state. "1" means "Go Straight." "2" means "Turn Right."
"3" means "Turn Left.".
In Fig.12, a complete process procedure for the raw data in
Fig.11 is shown. The original distances are converted to three
input values: 1.00, 1.06 and 2.96. After conversion, these three
inputs will be passed through the neural network, and generate
the final command, "Go Straight."

1105
 Fig.16. The probabilistic distribution of the navigation time in the real
environment.

VI. CONCLUSIONS
Fig.14. The history of the action of the robot. In this paper, an autonomous robot navigation based on neural
networks was developed so that the robot could navigate in a
A video of our autonomous robot navigation is recorded and complex, dynamic and unknown environment. First of all, a
six screenshots of the video was captured, which are shown in neural network model was designed for the autonomous robot
Fig.15. navigation. Then the model was trained by using 9 specially
designed training samples to deal with different situations
which may appear in the real world. Finally, the trained neural
network was tested with simulations and experiments which
validate its feasibility.
REFERENCES
[1] H. B. Burke, D. B. Rosen, and P. H. Goodman, "Comparing artificial
(a) (b) neural networks to other statistical methods for medical outcome
prediction," , IEEE World Congress on Computational Intelligence, pp.
2213-2216, 1994.
[2] L. Feng, D. Peng, W. Fangfei, Q, Jun, "Use clustering to improve neural
network in financial time series prediction,” Third International
Conference on Natural Computation, pp. 89-93, 2007.
[3] H. Seraji and A. Howard, "Behavior-based robot navigation on
challenging terrain: A fuzzy logic approach," IEEE Transactions on
Robotics and Automation, Vol. 18, No. 3, pp. 308-321.
(c) (d)
[4] W. Gueaieb and M. S. Miah, "An intelligent mobile robot navigation
technique using RFID technology," IEEE Transactions on
Instrumentation and Measurement, Vol. 57, No. 9, pp. 1908-1917.
[5] M. Meng and A. C. Kak, "Fast vision-guided mobile robot navigation
using neural networks," IEEE International Conference on Systems,
Man and Cybernetics, Vol. 1, pp. 111-116, 1992.
[6] M. T. Hagan, H. B. Demuth, and M. H. Beale, Neural Network Design,
Martin Hagan, 2002.
(e) (f) [7] E. Alpaydin, Introduction to Machine Learning, 2nd edition, The MIT
Fig.15. Neural-Network autonomous navigation in a real environment. Press, 2009.
[8] AmigoBot User's Guide, ver. 1.9, ActivMedia Robotics, LLC, Amherst,
In Fig.15(a), the robot is starting out and going straight. In NH, 2003.
Fig.15(b), the robot is detecting the road situation around the
corner of the road and preparing to make a right turn. In
Fig.15(c), the robot is making a right turn and running across
the corner. In Fig.15(d), the robot is going straight because
there are no obstacles detected by the sonar sensors. In
Fig.15(e), the robot is making a right turn again to avoid
collision. Finally in Fig.15(f), the robot is getting out of the
designed maze.
As shown in Fig.16, the average time for the robot to get
out of the real maze is approximately 41.16 seconds. The
minimum time is 31.00 seconds and the maximum time is
55.00 seconds.

1106