Multi-robotic System with Self-organization for Search of Targets in
Covered Area
Jolana Sebestyénová and Peter Kurdel
Institute of Informatics, Slovak Academy of Sciences, Dúbravská cesta 9, Bratislava, Slovakia
Keywords: Area Coverage, Multi-robotic System, Emergent Behaviour, Self-organization, Cellular Automata, Search
and Surround of Targets.
Abstract: The paper introduces the multi-robot area coverage problem, wherein a group of robots must inspect every
point of a 2-dimensional test environment and surround all contaminations (or enemies) found. Self-
organizing robotic systems are able to accomplish complex tasks in a changing environment, using local
interactions among individual agents and local environment without an external global control. Our interest
in this area is motivated by an involvement in a project with a goal to solve tasks of difficult area coverage
and surveillance by a large team of small autonomous robots. In the paper, the architecture to achieve this is
described, and simulation results are presented to compare efficiency of coverage of the area and surround
of found targets using robots groups having different sensors ranges.
1 INTRODUCTION
Self-organization is one of the most important
features observed in social, economic, ecological
and biological systems. Self-organizing robotic
systems are supposed to be able to accomplish
complex tasks in a changing environment through
local interactions among individual agents and local
environment without an external global control.
Developing such self-organizing systems, where
desired global behaviours can emerge through local
interactions among individuals and with the
environment is a challenging task (Meng, 2011).
Team of robots could help to minimize
hazardous work for humans. Efficient search and
cooperative completion of the task is possible via
sophisticated communication methods. A swarm of
small mobile robots is a set of inexpensive robots
that explore a dangerous environment with aim to
locate enemies or other targets. In non-
communicative swarming, the swarm comprises
homogeneous and anonymous robots, i.e. robots able
to recognize other robots but un-capable to identify
them individually.
Communicative swarming is distinctively more
efficient than non-communicative one as it increases
the swarm control ability. In communicative
swarming, the swarm robots interchange information
concerning their environment, which enables to
arrive to information-aware conclusions. Moreover,
the robots make use of the information received
from each other, which enables to control
cooperative behaviors as e.g. cooperative area
coverage or cooperative search / exploration. Multi-
robot systems communication can be direct or
indirect. Indirect interaction uses passive or active
mechanism of indirect coordination between agents
or actions (stigmergy).
A swarm is defined as a massive collection that
moves with no group organization, much like a
swarm of bees or a flock of birds. Similar is a
formation, the distinction is made in that it maintains
a global structure, like a flock of geese or a
marching band (http://roboti.cs.siue.edu). Robot
formations have been applied to applications such as
automated traffic cones, while swarm behavior
control has been applied to urban search-and-rescue
robotics.
The majority of existing multi-robot systems for
pattern formation rely on a predefined pattern, which
is impractical for dynamic environments where the
pattern to be formed should be able to change as the
environment changes. In addition, adaptation to
environmental changes should be realized based
only on local perception of the robots. In (Jin, Guo
and Meng, 2012), a hierarchical gene regulatory
network for adaptive multi-robot pattern generation
and formation in changing environments is
451
Sebestyénová J. and Kurdel P..
Multi-robotic System with Self-organization for Search of Targets in Covered Area.
DOI: 10.5220/0005059704510458
In Proceedings of the 11th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2014), pages 451-458
ISBN: 978-989-758-040-6
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
proposed.
The traditional artificial intelligence approach to
robot control is known as deliberative control. In the
sense-plan-act paradigm, the robot senses its
environment and, taking into account a model of that
environment, decides to start the appropriate action.
The weak point of the deliberative control is
possible failure in case of unexpected change of the
environment. Reactive control does not need a
model of the environment or traditional planning, as
it relies on a number of simple behaviors.
In the scope of bio-inspired soft robotics
behavior is orchestrated rather than controlled
(Pfeifer, Lungarella and Iida, 2010). Different bio-
inspired multi-robot coordination systems have been
developed (http://wyss.harvard.edu): distributed
robots for search and rescue, environmental
monitoring by highly agile autonomous robots, etc.
Agent-based models consist of dynamically
interacting, rule-based agents
(en.wikipedia.org/wiki/Agent-based_model).
Area coverage is one of the emerging problems
in multi-robot coordination (Fazli, 2010). In this task
a team of robots is cooperatively trying to observe or
sweep an entire area, possibly containing obstacles,
with their sensors or actuators. In barrier coverage
robot guards are deployed to prevent intrusion
(Kloder, 2008).
The foundations of automata theory in swarm
systems come predominantly from the cellular
robotics systems.
Cellular automata (CA) are abstract models of
complex natural systems having large quantities of
identical, locally interacting simple components.
Modeling based on CA leads to extremely simple
models of complex systems. It carries discrete lattice
of cells, generally in more dimensions, where each
cell in the lattice contains a number of cells. Each
cell can interact with the cells located in its
neighborhood. Though the CA's construction is
simple, its behavior can be very complex. CA
modeling is a young domain of Computer Science,
where the investigations proceed in two intersecting
forms: theoretical study of CA-models as dynamical
systems, and development of methods and tools for
computer simulation using CA-models
(http://ssd.sscc.ru/en/projects).
A cellular automaton consists of a chain (1-
dimensional) or lattice (2-or-3-dimensional) of
computational cells, each cell being in one of a
given set of states that evolve through discrete time
steps. The dynamic behavior of the automaton is
determined by a set of rules that govern the change
of state of an individual cell with respect to its
neighbors.
One of practical implications that need to be
considered is increasing risk of collisions when two
robots attempt to move to the same unoccupied grid
cell.
The paper introduces a multi-robot area coverage
problem, wherein a group of robots must inspect
every point of a 2-dimensional test environment and
surround all targets found. Fig. 1 illustrates start
positions of robots and positions of searched targets
(contaminations or enemies) in the test area.
Similar methods making use of cellular automata
do only area coverage or only move on patrol around
a given target, e.g. a building. Other methods
enabling search for target and its encircling, such as
morphogenetic swarm robotic systems (Meng, 2011)
use ingenious estimation of shapes and resulting
formation of appropriate encircling robots patterns.
Figure 1: Start position of robots (circles) and searched
targets (#) in covered area.
2 COLLECTIVE EMERGENT
BEHAVIOURS
Emergence and its accompanying phenomena are a
widespread process in nature (Jin and Meng, 2012).
Despite of its prominence, there is no agreement in
the sciences about the concept and how to define or
measure emergence.
The behaviour-based approach (Banzhaf, 2012)
has become very popular to cope with several
robotic applications, also including service robotics.
It refers to direct coupling of perception to action as
ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics
452
a specific technique which provides time-bound
responses to robots moving in dynamic, unstructured
and partially unknown environments.
The behaviour is defined to be a control law for
achieving and/or maintaining a particular goal.
Usually, robot agents have multiple goals. This
requires robot agents to be equipped with a number
of behaviours.
A behaviour-based approach assumes a robot to
be situated in, and surrounded by, its environment.
This means that a robot interacts with the world on
its own, without any human intervention, i.e. its
perspective is different from that of the observer.
The distinction between collective and
cooperative behaviour is made on the basis of
communication. Cooperation is a form of interaction
based on some form of communication.
The first, essential step enabling the emergence
of a collective behaviour is a careful design of the
behaviours that any individual robot agent will
contain. Further, one has to specify which tasks a
group of individual robots can accomplish. Last but
not least, a mechanism to initialize the cooperative
behaviour, eventually considering the level of
cooperative strategies the robots must follow to
collectively solve given tasks, is necessary. The
result of the actions provided by the individual
agents, whose activities must be coordinated to
cooperate and solve the global task, will be
emergence of a collective behaviour.
A large number of simple robots with limited
computational and communication capabilities can
be joined to form a multi-robot system (MRS).
Robots in an MRS can together fulfil difficult tasks
surpassing the capability of a single robot. As they
can be made robust, adaptable and still low cost,
there have been a large number of successful
applications, such as cooperative localization and
mapping, collaborative search and rescue, collective
construction, etc.
Rescue robots are useful for rescuing jobs in
situations that are hazardous for human rescuers
(http://emdad1.20m.com). They can enter into gaps
and move through small holes, which is impossible
for humans and even trained dogs. Robots should
explore in collapsed structure, extract the map,
search for victims and report the location of victims
in map and way that rescue team can reach them.
The main task of rescue robots is to acquire
information about damaged area and victims
(Akiyama, Shimora, Takeuchi and Noda, 2010).
Getting the reliable information is given the first
priority in rescue activity for disaster mitigation.
An additional potential application of the
proposed model is for cordoning off hazardous
materials.
In order to traverse through a complex
environment, swarm robotic systems need to self-
organize themselves to form different yet suitable
shapes dynamically, to adapt to unknown
environments (D’Angelo, 2007). Insects are
particularly good at cooperatively solving multiple
complex tasks. For example, foraging for food far
away from the nest can be solved through relatively
simple behaviours in combination with
communication through pheromones. As task
complexity increases, however, it may become
difficult to determine the proper simple rules which
yield the desired emergent cooperative behaviour, or
to know if any such rules exist at all.
3 PROBLEM FORMULATION
A large range of research has been done by imitating
ideas from nature for designing control algorithms
for multi-robot system.
Multi-robot shape construction and pattern
formation, a typical task for MRSs, has been widely
studied. Algorithms in this research field can be
roughly divided into three groups: leader/neighbour-
following algorithms, potential field algorithms, and
nature-inspired algorithms.
Leader/neighbour-following algorithms require
that individual robots follow neighbours or leader
that knows the aim or target to which the team needs
to go. These following robots should get behind a
leader's root in a specific geometric relationship with
the ones they follow. The second group of multi-
robot shape construction algorithms is based on
potential field method. The basic idea of this group
of algorithms is that each robot moves under the
governance of the gradients of potential fields,
which are the sum of virtual attractive and repulsive
forces. The third group is represented by nature-
inspired algorithms.
The problem we are addressing is to entrap
stationary (in future also mobile) targets using a few
mobile robots, i.e. coordination mechanisms for the
distributed contamination boundary coverage
problem with a swarm of miniature robots. In the
proposed model, field vector-based area coverage is
used in combination with search and surround of
some targets distributed in the area. Basic simple
behaviours of the robots are:
area coverage
collision avoidance
search for a target
Multi-roboticSystemwithSelf-organizationforSearchofTargetsinCoveredArea
453
Figure 2: Basic computation flowchart of robot moves based on behaviours.
walk around the target found
standing on guard at the found targets.
Fig. 2 shows basic simulation steps (on
informational level).
3.1 Assumptions
The following assumptions have been made
(Sebestyénová and Kurdel, 2013):
1) All robots move with equal speed.
2) There is a base station containing a sufficient
number of robots.
3) All robots have a limited sensing range.
4) The communication range between robots is
limited. Robots can communicate information
such as targets’ location with their immediate
neighbours (distance between the two robots is
within the communication range).
5) Communication between the robots and the base
station is not limited. This assumption will take
place but in future, as the model presented in
this paper runs only in computer simulation.
6) The robot should distinguish between obstacle
and boundary.
3.2 Model and Basic Rules
One way to simulate a 2D cellular automaton (k = 2)
ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics
454
is with an infinite sheet of graph paper along with a
set of rules for the cells to follow. Each square is
called a “cell” and each cell has several possible
states. There are several possible lattices and
neighbourhood structures for 2D cellular automata.
This paper considers square lattices. At start, the
robots are arranged in one of the corners of the area
(lower left corner on Fig.1). Number of robots and
number of rows in which the robots are ordered are
selectable. All robots are oriented to Nord at start,
and speed of robots equal. From two most common
2-D CA neighbourhood templates (Moore
neighbourhood and von Neumann neighbourhood –
can be extended) Moore neighbourhood (eight
surrounding cells, n = 8) is used in the model. State
of a cell is from a set: empty, robot is in it, target is
in it.
Neighbourhood size in the model as well as
sensors range (e.g. for contamination detection) is
two cells distance (r = 2). The model can be further
generalized by increasing the possible
neighbourhood size to more than two cells distance
and by enabling different sensors ranges for different
kinds of sensors. For all cells, attractions at start are
equal and changes are computed according to robots
moves and targets found.
Extra states are used to code the robot's current
direction, as well as for remembering cells where
some robot already appeared, which is then used for
slow forgetting of the robots position history
(Sebestyénová and Kurdel, 2014). All cells
remember whether and when any of the robots
visited the cell. State transition is fired by a set of
rules.
Each robot looks at the attractions of the nearby
cells and its own direction and then applies the
transition rule, specified in advance, to decide its
move in the next clock-tick. All the cells change at
the same time. Each robot moves to empty
neighbour cell with maximal attraction. It tries to
move in direction in which it is facing. If this is not
possible, the robot direction is rotated clockwise.
Some delicate configuration requiring good
decision may on certain occasions happen, e.g. the
robot must decide if it is more convenient, or even
possible (e.g. sliding along a wall), to turn around
the obstacle, instead of passing through, and which
direction to select for this turnaround. Walls have
been considered as particular kinds of obstacles, too.
A serious problem may arise if both of two opposite
directions are blocked due to some difficult
configuration. In this case the robot does not move
for a while, waiting the other robot’s move.
Basic rule for robots moves is specified as
→
(1)
where a is attraction, l is location of robot, c is cell
to which the robot will move, t is time, and r is
range. All used data are specified and/or evaluated in
subsequent simulation steps in multidimensional
cells representing the area (area width × area length
× number of used data types, in our case 40×40×8):
Attraction field - at start, attractions of cells
are equal (specified maximal attraction value).
Contamination positions (targets) are input
data of a simulation tool.
Robot identifiers at positions (start and actual
positions) and their directions; number of
robots and their starting positions are input
data. Robot speed is 1 cell per 1 simulation
step.
Cell occupied by any of robots is an obstacle;
no other robot can take the place.
Just released cell will set zero attraction.
Forgetting a visit of a robot - in subsequent
simulation steps cell forgets the visit (in each
step a small value, and after many steps cell
forgets the visit completely). Using these
values, the attraction of the cell again raises.
Obstacles in area (now only area boundaries
are considered)
Positions of found targets
Guarded targets as well as positions and IDs
of robots guarding on them.
If the robot views the target (or senses the
target according to used sensor), it needs
information whether this target is already
guarded on by any other robot:
If not, the robot remains to stay (it starts to
guard on and raises the attraction of the target
cell and its outskirts).
If yes, robot continues in walking around the
target (one target may cover more cells).
One robot can guard on more than one target
cells according to its sensors ranges.
The robots guarding on found target cells not
only need to see the target, they also need to
see each other to form a secure surround.
Repulsion - the robot starting to guard
increases the repulsion of its position’s cell.
4 SIMULATION RESULTS
Targets are static objects in the environment that
need to be encircled by robots. Robot standing on
guard refers to the robot that detects at least one
Multi-roboticSystemwithSelf-organizationforSearchofTargetsinCoveredArea
455
target in the environment. Searching robot refers to
the robot not having detected any target in the
environment, therefore doing area coverage.
Searching robot can become robot standing on guard
if it detects a target not yet guarded by any other
robot.
Fig. 3 depicts positions of robots after several
simulation steps using pseudo-colour plots (starting
positions of robots is presented in Section 1 in
Fig. 1):
a) in step 25, robot with identifier 9 stands on guard
at 3 target cells found in these few steps in the centre
of covered area;
b) in step 27, robot with ID 8 takes guard on two
target cells previously guarded by robot with ID 9;
c) robots with ID 5, 7, 8, 9 and 10 start guard on
some target cells (in step 70);
d) final surround of the target cells in the centre of
area covered by 4 robots with ID 5, 7, 10 and 11 (in
step 80).
The series of individual figures illustrates changes in
attraction field: cells visited by any of the robots
exhibit decreased attraction, whereas cells in close
vicinity of the found targets acquire increased
attraction.
Movement behaviour of robots not detecting any
target is governed by the area coverage, avoid
collision, and search for target behaviours.
One major advantage of here presented
approach, compared to existing multi-robot pattern
formation algorithms is that it provides an adaptive
mechanism being able to dynamically generate an
appropriate surround pattern adapted to
environmental changes. Most existing MRSs for
pattern formation rely on a predefined pattern, which
is not applicable to changing environments.
Entrapping multiple targets is closely related to
multi-robot target tracking, where multiple robots
are used to track the positions of single or multiple
targets. Algorithms for multi-robot target tracking
can be divided into two groups. The first group is the
region-based approach, in which the robots
cooperate to cover a certain region. This way, all the
targets in the whole region can be detected and
tracked. The advantage of this type of algorithm is
that the robots do not need to know the target
distribution information. One implicit assumption
here is that there is always a sufficient number of
robots available to cover the whole region.
The second group of multi-robot target tracking
algorithms is the target-oriented approach. In
contrast to region-based algorithm used here, the
robots will continuously update the number and
location of targets (but they do not have to form
a) in step 25
b) in step 27
c) in step 70
d) in step 80
Figure 3: Position of robots after some steps with ID of
exchanging robots guarding on central target.
ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics
456
patterns to entrap the targets). As a result, a large
number of targets can be tracked without covering
the whole area, which is not acceptable in most of
application domains.
The model was made more general with respect
to previously published one by implementing the
following behaviour: if the robot founds not guarded
target, it starts to guard it. If the target found is
already guarded by other robot, it will walk around
the target. If the target appears to be a
contamination, the kind of detected contamination is
given by a type of sensors carried by the robot. As
the target obviously covers more than one cell,
probability to find more cells with not guarded target
increases. Each robot can guard more targets in its
neighbourhood range (robots sensors range r = 2).
Figure 4: Positions of robots standing on guard around
found contaminations in step 94 (6 robots continue in area
coverage).
All of the targets in covered area are found and
guarded on in step 94. Positions of robots standing
on guard around found targets are illustrated in
Fig. 4. From the group of 18 robots in the example,
12 robots were enough to guard on all targets in the
area, 6 robots continued in area coverage.
For comparison, Fig. 5 illustrates simulation
results of the previously presented model, in which
the robots sensors range was r = 1 and increase of
attractions around found targets was smaller. All
targets were found and surrounded in step 280, and
only 5 robots could continue in area coverage (to
distinguish the cases, the left part of figure is
depicted with attraction field changes, right part
without them).
5 CONCLUSIONS
This paper introduces the multi-robot area coverage
problem, wherein a group of robots must inspect
Figure 5: Positions of robots standing on guard around
found contaminations in step 280 (5 robots continue in
area coverage).
every point of a 2-dimensional test environment and
surround all targets found. Although there are many
publications on area coverage and there are some -
not so many - publications on surrounding of found
targets, very little of them combine these two
research areas. Similar methods making use of
cellular automata do only area coverage or only
move on patrol around a given building. Other
methods enabling search for target and its encircling
use estimation of shapes and resulting formation of
appropriate encircling robots patterns. The main new
feature of the proposed model compared to existing
work is that the target search and surround pattern
need not be predefined and is adaptable to
environmental changes, e.g., the number and
location of the targets to be entrapped. It should be
pointed out that successful entrapping of the mobile
targets is conditioned on the assumption that the
movement speed of the robots is faster than that of
the targets. The paper presents new results of the
authors compared to their previous work. The main
difference in robots behaviours concerns the
possibilities arising from longer sensor ranges of
robots with respect to the previously published
work. Furthermore, robot doing area coverage and
search for targets having found the target already
guarded by other robot can take over guarding and
release the previous robot, which can change its
behaviour and continue in area coverage; each robot
is now able to guard on more than one target cell;
attraction field changes and some other parameters
were modified in order to achieve better results. A
comparison with previously published work has
been made, using the same initial positions of robots
and targets. The final goal was this time achieved in
smaller number of steps and by smaller number of
robots needed for standing on guard. The approach
presented here can be generalized in several ways.
First, synchronized moves of robots in the group can
be replaced by asynchronous ones, so that each robot
of the non-homogeneous group will be able to move
Multi-roboticSystemwithSelf-organizationforSearchofTargetsinCoveredArea
457
with its own speed. Likewise, sensor ranges can be
made non-uniform. And finally, communication
between members of a real mini-robot swarm, as
well as with base station can work asynchronously,
whenever a new relevant information will be
available. In the future, one can investigate in detail
the conditions under which the whole system is able
to keep encircling the moving targets. The presented
model could be used for military applications;
another potential application of this model is to
cordon off hazardous materials.
ACKNOWLEDGEMENTS
This work is supported by
projects APVV-0261-10
BioMRCS, ASFEU Crisis ITMS
26240220060, VEGA
2/0054/12, and VEGA
2/0194/13.
REFERENCES
Agent-based_model. Available at: http://en.wikipedia.org/
wiki/Agent-based_model.
Akiyama H., Shimora H., Takeuchi E., Noda I., 2010.
Integrated Map Generation using Multi-robot Sensory
Data in RoboCup Rescue. Available at: https://
stinckwich.users.greyc.fr/share/ROSIN10/rosin10_sub
mission_5.pdf.
Banzhaf W., 2012. Genetic Programming and Emergence.
In Frontiers of Natural Computing workshop, York,
UK, 2012, Available at: http://www.elec.york.ac.uk/
events/ncfrontiers/
Bioinspired Robotics. Available at: http://
wyss.harvard.edu/viewpage/204/bioinspired-robotics.
Cellular Automata for Control and Interactions of Large
Formation of Robots: Context, Background, and
Literature Review, Swarms vs. Formations. Available:
http://roboti.cs.siue.edu/projects/formations/literature.
php, roboti.cs.siue.edu, cs.gmu.edu.
Cellular Automata Simulation Available at: http://
ssd.sscc.ru/en/projects.
D’Angelo A. 2007. How a Cooperative Behavior can
emerge from a Robot Team. In Distributed
Autonomous Robotic Systems 6, 2007.
Fazli P., 2010. On Multi-robot Area Coverage. In A.
Farzindar and V. Keselj (Eds.): Canadian AI 2010,
LNAI 6085, Springer-Verlag Berlin Heidelberg 2010,
pp. 384–387.
Kloder S., 2008. Barrier Coverage: Deploying Robot
Guards to Prevent Intrusion. In Thesis 2008,
University of Illinois.
Meng Y., 2011. Tutorial on Morphogenetic Robotics: A
New Emerging Field of Self-Organizing Robotic
Systems. In IEEE Int Conf on Robotics and
Automation, Shanghai, China, Available at:
http://www.ece.stevens-tech.edu/~ymeng/ICRA11-
Tutorial.htm, www.ece.stevens-tech.edu.
Pfeifer R., Lungarella M., Iida F., 2010. The Challenges
Ahead for Bio-Inspired ‘Soft’ Robotics. Available at:
http://shanghailectures.org/sites/default/files/uploads/p
76-pfeifer.pdf, doi:10.1145/2366316.2366335.
Rescue Robot. Available at: http://emdad1.20m.com/,
http://www.expo21xx.net/automation21xx/15799_st2_
university/default.htm, www.expo21xx.net,
en.wikipedia.org.
Sebestyénová J., Kurdel P., 2013. Self-organizing robotic
system for area coverage and surround of
contamination found. In Proc. IEEE 9th International
Conference on Computational Cybernetics. -
Piscataway : IEEE Operations Center, 2013, p. 309-
314, ISBN 978-1-4799-0060-2.
Sebestyénová J., Kurdel P., 2014. Area Coverage and
Surround of Found Targets by Self-organizing Multi-
Robotic System. In EUROPMENT Int. Multiconf. on
Systems, Control, Signal Processing and Informatics,
Prague, Proc. of the 2014 Int. Conf. on Education and
Educational Technologies EET´14. Eds. Dondon Ph.
et al., pp. 123-129, ISBN 978-1-61804-234-7,
Available at: http://www.europment.org/library/
2014/prague/bypaper/ECS-EET/ECS-EET-18.pdf.
Yaochu Jin, Hongliang Guo, and Yan Meng, 2012. A
Hierarchical Gene Regulatory Network for Adaptive
Multirobot Pattern Formation. In IEEE Trans. on
Systems, Man, and Cybernetics, Part B, Vol.42, 2012,
pp. 805-816, DOI 10.1109/TSMCB.2011.2178021,
ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=61
44058, Available at: http://www.soft-computing.net/
SMCB2012Draft.pdf, www.soft-computing.net.
Yaochu Jin, Yan Meng, 2012. Morphogenetic Robotics:
An Emerging New Field in Developmental Robotics.
Available at: http://sclab.yonsei.ac.kr/courses/
12TCS/%EC%9D%B8%EC%A7%80_120529_%EC
%9D%B4%EB%8F%99%ED%98%84.pdf.
ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics
458
