Robotics and Autonomous Systems 72 (2015) 4858

Contents lists available at ScienceDirect

Robotics and Autonomous Systems

journal homepage: www.elsevier.com/locate/robot

Navigation of multiple mobile robots in a highly clutter terrains using

adaptive neuro-fuzzy inference system
Jayanta Kumar Pothal , Dayal R. Parhi
Mechanical Engineering Department, NIT Rourkela, Odisha, India

article info abstract

Article history: In recent years, the interest in research on robots has increased extensively; mainly due to avoid human to
Received 1 July 2014 involve in hazardous task, automation of Industries, Defence, Medical and other household applications.
Received in revised form Different kinds of robots and different techniques are used for different applications. In the current
7 April 2015
research proposes the Adaptive Neuro Fuzzy Inference System (ANFIS) Controller for navigation of single
Accepted 13 April 2015
Available online 23 April 2015
as well as multiple mobile robots in highly cluttered environment. In this research it has tried to design
a control system which will be able decide its own path in all environmental conditions to reach the
Keywords:
target efficiently. Some other requirement for the mobile robot is to perform behaviours like obstacle
Adaptive neuro-fuzzy inference system avoidance, target seeking, speed controlling, knowing the map of the unknown environments, sensing
Fuzzy logic different objects and sensor-based navigation in robots environment.
Neural network 2015 Elsevier B.V. All rights reserved.
Navigation
Mobile robot
ANFIS

1. Introduction
ANFIS is the combination of both Fuzzy Logic and Neural Net-
work. An important property of a fuzzy system is that knowledge
is represented in the form of the easily comprehensible logical
rules. On the other hand, neural networks have the ability to use
knowledge and experience. By combining both fuzzy and neural
paradigms, neuro-fuzzy controllers can be developed. Fuzzy logic
approach to incorporate the logical inputs of human and take de-
cision much faster. Neural controllers learn the discrete input rep-
resentations, which may cause such system unstable. Hui et al. [1]
have developed a hybrid technique for navigation system of mobile
robot. They used the neuro-fuzzy controller to find the path plan-
ning strategy with obstacle avoidance behaviour in a dynamic en-
vironment. Cavalcanti et al. [2] have applied different techniques as
neural networks, genetic algorithms, and Nano robotics concepts
to solve the problem of automation of mobile robots for the ap-
plication in medicine. Mbede et al. [3] have used intelligent mo-
tion control strategy that makes possible the integration of fuzzy
obstacle avoidance, multisensory based motions, and robust recur-
rent neural network control. The simulation results, validate under
real conditions, clearly demonstrate that their proposed strategy
is an effective approach. Hegazy et al. [4] have presented a novel
Corresponding author. Tel.: +91 9173417227.
E-mail address: jkpothal@gmail.com (J.K. Pothal).
http://dx.doi.org/10.1016/j.robot.2015.04.007
0921-8890/ 2015 Elsevier B.V. All rights reserved.
neuro-fuzzy technique that can control the path planning strategy
to reach the target in an unknown environment. Rusu et al. [5] have
developed neuro-fuzzy control system for navigation of a robot.
They have carry out experiments with the mobile robot in indoor
condition. The autonomous mobile robot uses infrared and con-
tact sensors for detecting targets and avoiding collisions. Er et al.
[6,7] have developed hybrid control system of fuzzy logic and
neural network techniques. They implemented their technique in
Khepera II mobile robot. Lee [8] has used a neural network for be-
haviour decision controller. The input of the neural network was
decided by the distance and existence of other robots. The output
result determines the directions in which the robots to move. The
connection weight values of this neural network were determined
as genes, and the fitness individuals were encoded using a genetic
algorithm. The fitness values imply how much group behaviours
fit adequately to the goal and can express group behaviours.
They have validated the system through simulation. Pulasinghe
et al. [9] have developed a fuzzy-neural network methodology for
controlling machines using spoken language commands with the
intention of (i) interpreting natural language words with fuzzy
implications and (ii) screening out-of-vocabulary words, which
gives the ability to converse freely without restrictions. Mash-
hadany [10] has been developed the navigation system for a spher-
ical wristed robot using adaptive neuro-fuzzy inference system.
The author has found the solution for kinematic problem, local-
ization and navigation. The simulation has been carried out us-
ing Simulink/MATLAB. Baturone et al. [11] have been developed
the low-cost control system for navigation of mobile robot using
 J.K. Pothal, D.R. Parhi / Robotics and Autonomous Systems 72 (2015) 4858 49

neuro-fuzzy technique. They have been applied heuristic and Tak- fi is the linear consequent functions defined in terms of the
agiSugeno fuzzy method in this. The controller gives good result inputs (x1 , x2 , x3 and x4 ).
for obstacle avoidance and target seeking. Vatankhah et al. [12] qi , ri , si , ti and ui are the consequent parameters of the ANFIS
have been developed the control system by following undirected fuzzy model.
network and local communications. The neuro-fuzzy technique A, B, C and D are the fuzzy membership sets defined for the
has applied for achieving behaviour control and target seeking. input variables x1 , x2 , x3 and x4 .q1, q2 , q3 and q4, are the number of
Wang et al. [13] have been developed a system for motion con- membership functions for the fuzzy systems of the inputs x1 , x2 , x3
trol and location identification for mobile robotic system in an and x4 respectively. fi is the linear consequent functions defined
unknown and dynamic environment. They have been designed in terms of the inputs (x1 , x2 , x3 and x4 ). pi , ri , si , ti and ui are the
the controller using RFWNNs (recurrent fuzzy wavelet neural net- consequent parameters of the ANFIS fuzzy model. In case of ANFIS,
works). The controller gives good stability as well as location in- the nodes of same layer perform similar functions. The output
formation for the mobile robots. AbuBaker [14] has been proposed signals from the nodes of the previous layer are the input signals
a new control system to achieve collision free navigation system for the current layer [15]. The output obtained with the help of the
using hybrid neuro-fuzzy methodology. He has used large num- node function will be the input signals for the subsequent layer
ber of ifthen rule and selected the optimum number of activation (Fig. 1).
rules to reduce the time in practical application. The simulation and The input layer receives signal from arrays of sensors x1 , x2 x3 ,
experiments have been carried out to test the performance of the
x4 , which defines the static as well as moving obstacles, and target
controller.
poisons from the target tracker robot. The target position measured
according to the target coordinates. The coordinates are given
2. Analysis of adaptive neuro-fuzzy inference system (ANFIS) to the robots during navigation. The robot measures its global
position according to its wheel movements during navigation.
ANFIS is a fuzzy inference system implemented in the frame-
work of adaptive networks. The main aim of ANFIS is to borrow the Layer 1: Every node in this layer is an adaptive node (square
learning capability from the Neural Network and put this in Fuzzy node) with a particular fuzzy membership function (node function)
Inference System, i.e. superimpose of Fuzzy Inference System over specifying the degrees to which the inputs satisfy the quantifier.
a Neural layer architecture. Then the FIS will able to learn the train- For four inputs the outputs from nodes are given as follows:
ing procedure as in Neural Network and after the learning mecha- L1,g = Ag (x) for g = 1, . . . , q1 (For input x1 )
nism the Fuzzy Inference System will behave exactly as the Neural
Network. In NN there will be input layer, output layer and multiple where x is the input to node i, and A is the linguistic label (small,
hidden layers inside. The number inputs to the system are equal to large, etc.) associated with this node function. In other words, L1, g
the number of neurons that the network has. Each neuron behaves is the membership function of A and it specifies the degree to which
as a processing element. In case of FIS, the input has given some the given x satisfies the quantifier Ag.
membership function. The way of the input membership functions
combines or interacting among each other, is called as the rule. The L1,g = Bg (x) for g = q1 + 1, . . . , q1 + q2 (For input x2 )
rules have two parts; (a) antecedent and (b) consequent. Both of L1,g = Cg (x) for g = q1 + q2 + 1, . . . , q1 + q2 + q3
these parts are handled separately and then fusion and integration (For input x3 )
of all the rules give the output. The output is defuzzyfied to get the L1,g = Dg (x) for g = q1 + q2 + q3 + 1, . . . ,
result. q1 + q2 + q3 + q4 (For input x4 ).
ANFIS is five layer architecture. It is a specially grown multilayer
perception of limited connectivity, not the detailed architecture. Here the membership function for A, B, C and D considered are the
ANFIS is an integrated system of both artificial neural network bell shaped function and are defined as follows:
(ANN) and fuzzy inference system (FIS). The ANFIS analysed here
is a first order TakagiSugeno Fuzzy Model (Fig. 1). In the current 1
Ag (x) = 2 bg ; g = 1 to q1 ; (2)
analysis there are four inputs and one output. They are as follows:

x cg
1+ ag
Inputs
(1) Front obstacle distance (x1 ) 1
(2) (2) Right obstacle distance (x2 )
Bg (x) = 2 bg ; g = q1 + 1 to q1 + q2 (3)
x cg
(3) Left obstacle distance (x3 ) 1+ ag
(4) Target angle (x4 )
1
Output Cg (x) = 2 bg ;
x cg
(1) Steering angle 1+ ag

The rule for the ANFIS can be written as below; g = q1 + q2 + 1 to q1 + q2 + q3 (4)

Rule: 1
IF x1 is Aj , x2 is Bk , x3 is Cm and x4 is Dn THEN fi = pi x1 +ri x2 +si x3 + Dg (x) = 2 bg ;
ti x4 x cg
1+ ag
where;
g = q1 + q2 + q3 + 1 to q1 + q2 + q3 + q4 (5)
fi = pi x1 + ri x2 + si x3 + ti x4 + ui ; for steering angle (1)
j = 1 to q1 ; k = 1 to q2 ; m = 1 to q3 ; n = 1 to q4 and where ag bg and cg are the parameters for the fuzzy membership
function. The membership function used is ball-shaped function,
i = 1 to q1 .q2 q3 .q4
which changes its pattern as per the change of the parameters.
q1 , q2 , q3 and q4, are the number of membership functions for the This change will give the various contour of bell shaped function
fuzzy systems of the inputs x1 , x2 , x3 and x4 respectively. as needed in accord with the data set for the problem considered.
50 J.K. Pothal, D.R. Parhi / Robotics and Autonomous Systems 72 (2015) 4858
Fig. 1. The Adaptive Neuro-Fuzzy Inference System (ANFIS).
Layer 2: Every node in this layer is a fixed node (circular) labelled
as . The output denoted by L2,i is the product of all incoming
signal.
L2,i = Wi = Ag (x) , Bg (x) , Cg (x) , Dg (x) ;
for i = 1, . . . , q1 , q2 , q3 , q4 and g = 1, . . . ,
q1 + q2 + q3 + q4 .
The output of each node of the second layer represents the firing
strength (degree of fulfilment) of the associated rule. The T -norm
operator algebraic product {Tap (a, b) = ab}, has been used to
obtain the firing strength (Wi ).
Layer 3: Every node in this layer is a fixed node (circular) labelled
as N. The output of the ith node is calculated by taking the ratio of
firing strength of ith rule (Wi ) to the sum of all rules firing strength.
Wi
L3,i = W i = r =q .q .q .q
1 2 3 4
. (7)
Wr
r 1
This output gives a normalized firing strength.
Layer 4: All the nodes in this layer are adaptive node (square node)
with a node function.
L4,i = W i fi = W i (pi x1 + ri x2 + si x3 + ti x4 + ui )
where W i is a normalized firing strength form (output) from
layer 3 and {pi , ri , si , ti , ui } is the parameter set for steering angle.
Parameters in this layer are referred to as consequent parameters.
Layer 5: The single node in this layer is a fixed node (circular)
labelled as , which computes the overall output as the
summation of all incoming signals.
i=q1 .q2 .q3 .q4
i=q1 .q2 .q3 .q4 Wi f i
i0
(6) L 5 ,i = W i fi =
i=q1 .q2 .q3 .q4
. (9)
i 0
Wi
i0
In the current developed ANFIS structure there are six
dimensional space partitions and has q1 q2 q3 q4 regions.
Each region is governed by a fuzzy ifthen rule. The first layer is
the input layer. The first layer (consists of premise or antecedent
parameters) of the ANFIS and is dedicated to fuzzy sub space. The
second and third layers are fixed nodes (circular) labelled as
and N. The fourth layer has consequent parameters and is used to
optimize the network. The first order TakagiSugeno model is used
for difuzzyfication in fourth and fifth layer. During the forward pass
of the hybrid learning algorithm node outputs go forward till layer
four and the consequent parameters are identified by least square
method.
3. Simulation for mobile robot navigation
3.1. Problem description
(8)
Here the main aim is to modelling and implementation of
ANFIS controller for the navigation of mobile robots in a dynamic
environment without collision with obstacle and reach the target
with shortest path and within minimum time. The environment is
having obstacles of different shapes and may be static or dynamic.
During motion of robots, for one robot, other robots behave as
 J.K. Pothal, D.R. Parhi / Robotics and Autonomous Systems 72 (2015) 4858
Fig. 2. Showing the position of obstacles, starting point and target.
dynamic obstacles. The controller should easily detect the position
of obstacles and take decision quickly. Here the map is 2-D square
and the map consists of 10 10 grids.
3.2. Proposed solution
Here the ANFIS controller has designed and implemented in
Khepera robots for navigation. The simulation results are obtained
with the help of ROBNAV software being developed in Robotics lab.
The obstacle avoidance and target seeking behaviours are achieved
for four mobile robots using this software. This software is effective
Fig. 3. (i) Robots avoiding obstacles searching the target (stage-1); (ii) Robots avoiding obstacles and heading towards target (stage-2); (iii) Avoiding obstacles the robots
reaching the target (stage-3); (iv) All robots reached the target (stage-4).
51
up to 1000 mobile robots. In the proposed control strategy, reactive
behaviours are formulated and trained by ANFIS.
The obstacle avoidance behaviour is activated when the robot
reaches a minimum distance given in the program which is
continuously measured by sensors. When robots detect the
obstacle they simply diverted from their path making certain angle
towards the target. When the obstacle is very close to the robot
it just returns in opposite direction for collision avoidance. It can
avoid static as well as dynamic obstacles. According to the position
of target the robot always fixes the target angle so that it can reach
the target.
3.3. Simulation results
Here the experiments are done in various environments,
consisting of different size and number of obstacles different
number of mobile robots and different number of iterations. The
positions of obstacles are selected randomly. Fig. 2 shows the
position of obstacles along with the positions of robots and target.
3.3.1. Effect of different types of obstacles during navigation
The simulation has been carried out with four robots with
different shape and types of obstacles with one target. The robots
show the obstacle avoidance behaviour when they reach in the
range of certain distance. (see Fig. 3).
3.3.2. Wall following behaviour of Mobile robots
3.4. Comparison of results with other researchers
In the similar environments the developed ANFIS technique is
found to be more efficient in reference to the navigational path. The
52 J.K. Pothal, D.R. Parhi / Robotics and Autonomous Systems 72 (2015) 4858

Fig. 6. Initial stage for navigation of 90 mobile robots, showing different types of
obstacles and two targets.

developed ANFIS controller shows smooth motion with travelling

less distance. The sensors output gives the feedback of obstacle
position perfectly and early to the controller. Accordingly the
controller plans the path to the target easily with avoiding multiple
obstacles (see Fig. 4). Fig. 5(c) and (d) shows that the robot easily
navigates in highly cluttered environment.

3.5. Navigation of multiple mobile robots with many targets

Fig. 4. (a) Simulation result of Nefti et al. [16] (b) Simulation result of proposed The experiments have been carried out taking 90 robots into
controller with showing Wall following Behaviour. consideration with two target points (see Fig. 6). The simulation

Fig. 5. (a) and (b) Simulation result Garbi et al. [17]. (c) and (d) Results of proposed ANFIS approach during simulation in highly cluttered environment.
 J.K. Pothal, D.R. Parhi / Robotics and Autonomous Systems 72 (2015) 4858 53

Fig. 7. (a) Intermediate stage of navigation of multiple mobile robots; (b) All robots reached the target.

Fig. 8. Shortest path travelled by individual robots to reach the target. Fig. 9. Shortest path travelled by single robot with ANFIS.

shows that the robots reach the target efficiently and choose the
nearest target (see Fig. 7).

3.6. Navigation of four mobile robots with single target

The experiments have been carried out taking four mobile

robots and single target. The experiment continued for several
times and after that the shortest distance travelled by individual
robots has been shown in Fig. 8. Initially robots travel randomly
to reach the target. Then during next experiments they travel by
different path. The robots always find the shortest path to reach the
target. After several experiments the robots travel in a particular
path as shown in Fig. 9.
Fig. 10 shows the shortest path travelled by individual robot
as per the information of robot controller (see Figs. 1115). This
is the case when individual robots are taken into consideration.
Here the blocks are stationary obstacles and the robots are dynamic
obstacle for other robots. During the navigation of mobile robots,
ANFIS controller, controls all the mobile robots. The mobile robots Fig. 10. Shortest path travelled by four robots with ANFIS.
are always in a co-ordination with each other. The information is
continuously analysed in the controller. After several experiments,
the mobile robots find the shortest path collectively. This path is particular environment with same scenario and obstacles. This is
automatically feed to the controller of the individual mobile robots. the proposed shortest path for four mobile robots in a particular
Subsequently the robots try to follow this shortest path in the environment. The simulation result is useful for comparison of
54 J.K. Pothal, D.R. Parhi / Robotics and Autonomous Systems 72 (2015) 4858

Fig. 11. Path length for different number of robots of 50 trials.

Fig. 12. Comparison of path length for simulation and experimental results by single robot using ANFIS.

Fig. 13. Comparison of the time taken for simulation and experimental results by single robot using ANFIS.

Fig. 14. Comparison of path length for simulation and experimental results by 4 robots using ANFIS.
 J.K. Pothal, D.R. Parhi / Robotics and Autonomous Systems 72 (2015) 4858 55

Fig. 15. Comparison of the time taken for simulation and experimental results by four robots using ANFIS.

Fig. 16. Path followed by mobile robots to reach the Target.

results of other techniques described in this research work in same

environment (see Tables 15).

4. Experimental results

The experimental results are found using four Khepera

mobile robots (see Fig. 16). These robots allow confrontation
to the real world of algorithms developed in simulation for
trajectory execution, obstacle avoidance, pre-processing of sensory
information, hypothesis on behaviours processing. These robots
have on-board control system, which can easily control the DC
motors coupled with the wheels. It is user friendly with different
software. So it is easy to program and reprogram as per the
requirement. The robot has eight sensors around it, which can
easily sense the obstacle around it and give feedback to the
controller. So it is useful for detecting the obstacle position within
a fraction of seconds, and very useful for obstacle avoidance.
Assumptions for robot motion are as follows:
(a) There are no flexible parts in mobile robot. (b) There is no Fig. 17. Shortest path followed by Khepera robots during experiment.
transitional slip between the wheel and the surface. (c) There is
enough rotational friction between the wheels and the surface. The inputs for these experiments are front obstacle distance,
(d) The wheels should rotate without any disturbance. left obstacle distance, right obstacle distance, target angle and
56 J.K. Pothal, D.R. Parhi / Robotics and Autonomous Systems 72 (2015) 4858

Table 1
Average path length (in pixelsa ) of 50 trials.
Case no. The number of robots
1 4 10 100 500

Case 1 25.6453 23.7634 22.7445 18.5467 17.4356

Case 2 25.8734 23.9546 22.3366 19.3420 17.7645
Case 3 25.7356 23.9845 23.4356 18.7509 17.5467
Case 4 25.8563 24.3284 23.9459 19.7345 18.5467
Case 5 25.7560 24.0943 22.7534 18.7345 18.0334
Case 6 24.9645 23.6454 22.0554 19.2964 18.0245
Case 7 25.7435 24.5470 23.8453 19.6479 18.4398
Case 8 24.9530 23.9455 22.6334 19.4358 17.5464
Case 9 25.5436 24.4366 23.4536 19.6556 18.3436
Case 10 25.6470 24.4649 22.9675 18.6624 17.5464
Case 11 25.0436 23.5480 22.3352 19.2064 17.6790
Case 12 25.8547 24.0442 23.4350 19.4743 18.6345
Case 13 25.6457 24.4336 23.6454 19.6445 18.5467
Case 14 24.8556 23.8902 22.7698 20.0985 18.4356
Case 15 25.7543 23.9436 22.9604 19.5349 18.0546
Case 16 25.6453 24.5447 22.5334 19.4359 18.4309
Case 17 25.7456 23.9432 22.4965 19.5432 18.4356
Case 18 25.6345 23.9433 22.6489 19.4309 17.9567
Case 19 25.9646 24.5478 22.8735 20.4362 18.5468
Case 20 25.7586 24.4357 22.9865 19.5434 18.4357
Average path length (in pixel) 25.5810 24.1220 22.9428 19.4078 18.1192
a
One pixel is equivalent to 10 cm during experiment.

Table 2
Comparison of path length for simulation and experimental results by single robot using ANFIS.
Case number Simulation result (in pixel) Experimental result (in pixel) Percentage error

Case 1 25.7645 28.6300 10.09

Case 2 25.5346 28.4300 13.24
Case 3 25.1376 27.8500 12.87
Case 4 25.7835 28.7500 10.32
Case 5 25.9467 28.7600 09.78
Case 6 25.4276 28.4200 10.53
Case 7 24.9645 27.6400 09.68
Case 8 25.2867 28.5600 11.46
Case 9 24.8546 27.8200 11.61
Case 10 25.6457 28.5300 10.11
Average path length (in pixel) 25.4346 28.3390 10.25
a
One pixel is equivalent to 10 cm during experiment.

Table 3
Comparison of the time taken for simulation and experimental results by single robot using ANFIS.
Case no. Simulation result (in Experimental result (in Percentage error
second) second)

Case 1 16.38 18.30 10.49

Case 2 16.16 18.00 10.22
Case 3 16.29 18.20 10.49
Case 4 16.40 18.30 10.38
Case 5 15.92 17.80 10.56
Case 6 16.18 18.10 10.60
Case 7 16.26 18.20 10.66
Case 8 16.08 18.00 10.67
Case 9 16.32 18.20 10.33
Case 10 16.06 17.90 10.28
Average time taken (in 16.21 18.10 10.47
second)

the output is steering angle. Using ANFIS controller during the
experiment different navigational paths are traced (see Fig. 17) and
are compared with the simulation results.
5. Comparison of simulation and experimental results
Comparison of simulation and experimental results has been
shown in Fig. 18. The experiments have been carried out for four
mobile robots. The best result of 50 trails during experiment and
simulation has been shown in the figure. To compare the results,
several experiments have been carried out considering the same
position of obstacles. The behaviour of individual mobile robots as
well as in group has also been observed. During experiment and
simulation the path length decreases with increase in trials. But
after some time the optimal path is detected by the mobile robots
and after that they follow the same path for further experiment for
same controller.
5.1. Conclusion
On the basis of the above results (simulation and experimental),
it has been found ANFIS controller can be useful for mobile robots
 J.K. Pothal, D.R. Parhi / Robotics and Autonomous Systems 72 (2015) 4858 57

Table 4
Comparison of path length for simulation and experimental results by 4 robots using ANFIS.
Case number Simulation result (in pixel) Experimental result (in pixel) Percentage error

Case 1 23.6487 26.5400 12.22

Case 2 23.8635 26.4300 11.71
Case 3 23.8456 26.8600 11.22
Case 4 23.6598 26.1800 12.95
Case 5 23.8456 26.9500 11.52
Case 6 24.2647 27.1400 11.89
Case 7 23.8645 26.8400 11.09
Case 8 24.2867 26.8400 12.76
Case 9 23.8634 26.5300 13.32
Case 10 24.5835 27.6200 9.02
Average path length (in pixel) 23.9726 26.79 10.53
a
One pixel is equivalent to 10 cm during experiment.

Table 5
Comparison of the time taken for simulation and experimental results by four robots using ANFIS.
Case no. Simulation result (in Experimental result (in Percentage error
second) second)

Case 1 15.84 17.70 10.51

Case 2 15.76 17.60 10.45
Case 3 15.92 17.80 10.56
Case 4 15.80 17.60 10.23
Case 5 15.61 17.50 10.80
Case 6 15.58 17.40 10.46
Case 7 15.64 17.50 10.63
Case 8 15.88 17.70 10.28
Case 9 15.52 17.40 10.80
Case 10 16.00 17.90 10.61
Average time taken (in 15.76 17.61 10.53379
second)

References

[1] N.B. Hui, V. Mahendar, D.K. Pratihar, Time-optimal, collision-free navigation of

Fig. 18. Comparison of simulation and experimental results for shortest path
travelled by four mobile robots in 50 trials.
for avoiding static as well as dynamic obstacles. The obtained
results from above architecture was analysed in a number of
simulated experiments using ROBONAV simulation model and
it was found that the results compromise with satisfaction the
obstacle avoidance and moving towards the target. As desired
ANFIS is showing the character to take less time to reach the target
and continuously defines the direction of robot motion. During
training, ANFIS controller gradually learns the rules and gives the
output closer to desire one as the training progresses. The sensors
used for navigation detect the obstacles properly. The simulation
and experimental results show good agreement.
a car-like mobile robot using neuro-fuzzy approaches, Fuzzy Sets and Systems
157 (16) (2006) 21712204.
[2] A. Cavalcanti, R.A. Freitas Jr., Nanorobotics control design: A collective
behavior approach for medicine, IEEE Trans. Nanobiosci. 4 (2) (2005) 133140.
[3] J.B. Mbede, P. Ele, Chantal-Marguerite Mveh-Abia, Y. Toure, V. Graefe, S.
Ma, Intelligent mobile manipulator navigation using adaptive neuro-fuzzy
systems, Inform. Sci. 171 (2005) 447474.
[4] O.F. Hegazy, A.A. Fahmy, O.M.E. Refaie, An intelligent robot navigation system
based on neuro-fuzzy control, in: Lecture Notes in Computer Science, vol.
3157, 2004, pp. 10171018.
[5] P. Rusu, E.M. Petriu, T.E. Whalen, A. Cornell, H.J.W. Spoelder, Behavior-based
neuro-fuzzy controller for mobile robot navigation, IEEE Trans. Instrum. Meas.
52 (4) (2003) 13351340.
[6] M.J. Er, T.P. Tan, S.Y. Loh, Control of a mobile robot using generalized dynamic
fuzzy neural networks, Microprocess. Microsyst. 28 (9) (2004) 491498.
[7] M.J. Er, C. Deng, Obstacle avoidance of a mobile robot using hybrid learning
approach, IEEE Trans. Ind. Electron. 52 (3) (2005) 898905.
[8] M. Lee, Evolution of behaviors in autonomous robot using artificial neural
network and genetic algorithm, Inform. Sci. 155 (2003) 4360.
[9] K. Pulasinghe, K. Watanabe, K. Izumi, K. Kiguchi, Modular fuzzy-neuro
controller driven by spoken language commands, IEEE Trans. Syst. Man
Cybern. B 34 (1) (2004) 293302.
[10] Y.I.A. Mashhadany, ANFIS-inverse-controlled PUMA 560 workspace robot with
spherical wrist, Procedia Eng. 41 (2012) 700709.
[11] I. Baturone, A. Gersnoviez, A. Barriga, Neuro-fuzzy techniques to optimize an
FPGA embedded controller for robot navigation, Appl. Soft Comput. 21 (2014)
95106.
[12] R. Vatankhah, S. Etemadi, A. Alasty, G. Vossoughi, Adaptive critic-based neuro-
fuzzy controller in multi-agents: Distributed behavioral control and path
tracking, Neurocomputing 88 (2012) 2435.
[13] Y.N. Wang, T.L. Mai, J.X. Mao, Adaptive motion/force control strategy for non-
holonomic mobile manipulator robot using recurrent fuzzy wavelet neural
networks, Eng. Appl. Artif. Intell. 34 (2014) 137153.
[14] A. AbuBaker, A novel mobile robot navigation system using neuro-fuzzy rule-
based optimization technique, Res. J. Appl. Sci. Eng. Technol. 4 (15) (2012)
25772583.
[15] Jyh-Shing Roger Jang, ANFIS: Adaptive-network-based fuzzy inference system,
IEEE Trans. Syst. Man Cybern. 23 (3) (1993) 665685.
[16] S. Nefti, M. Oussalah, K. Djouani, J. Pontnau, Intelligent adaptive mobile robot
navigation, J. Intell. Robot. Syst. 30 (2001) 311329.
[17] Giuliani Paulineli Garbi, Victor Orlando Gamarra Rosado, Francisco Jos
Grandinetti, Multivalued adaptive neuro-fuzzy controller for robot vehicle, in:
International Conference on Intelligence, System and Knowledge Engineering,
ISKE 07, China, 2007.
58 J.K. Pothal, D.R. Parhi / Robotics and Autonomous Systems 72 (2015) 4858

Jayanta Kumar Pothal was born in India. He has received
the Ph.D. degree in Mechanical Engineering from National
Institute of Technology, Rourkela. He is presently working
as Senior Scientist in Council of Scientific and Industrial
Research (CSIR), Ministry of Science and Technology,
Government of India.
Dayal R. Parhi, was born in India. He has received the Ph.D.
degree in Robotics and Artificial Intelligence from Cardiff
Univ. and presently working as Professor, Department of
Mechanical Engineering, National Institute of Technology
Rourkela, India.