See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/320620251

Victims Search, Identification, and Evacuation with Heterogeneous Robot

Networks for Search and Rescue

Article · October 2017

CITATIONS READS

0 200

4 authors:

David R. Yanguas Rojas Gustavo Andres Cardona

National University of Colombia National University of Colombia
8 PUBLICATIONS   4 CITATIONS    8 PUBLICATIONS   9 CITATIONS   

SEE PROFILE SEE PROFILE

Juan Ramirez-Rugeles Eduardo Mojica-Nava

National University of Colombia National University of Colombia
3 PUBLICATIONS   2 CITATIONS    99 PUBLICATIONS   368 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Load Frequency Microgrid Control: A Cyber Physical Approach View project

Mechanism design for demand management using social incentives View project

All content following this page was uploaded by David R. Yanguas Rojas on 26 October 2017.

The user has requested enhancement of the downloaded file.

 Victims Search, Identification, and Evacuation with Heterogeneous
Robot Networks for Search and Rescue
David Yanguas-Rojas1 , Gustavo A. Cardona 2 , Juan Ramirez-Rugeles
Abstract—This paper presents a control strategy for multiple
heterogeneous robots oriented to the assistance in situations of
search and rescue from two complementary perspectives, the
real time control and the discrete task allocation. The real time
control laws are in charge of the correct execution of the tasks
assigned by the reallocation algorithm respecting constraints
in the connectivity range, the collision avoiding and the full
completion of each task. The considered tasks are exploration
of the mission environment, search and identification of victims,
delivery of medical supplies to victims unable to move and
evacuation of victims capable to move. During the development of
each task, the robots take routes optimized considering distance
metrics based on the breath-first search algorithm and an
exploration algorithm for non-convex environments. In order
to assign the tasks to the robots thorough the mission, we
present an optimized task reallocation algorithm focused on the
minimization of the time required to attend all the victims in
the mission environment. This algorithm allows to assign each
of the robots to an appropriate task considering that the robots
may differ in the capability of completing each task as well as
in their moving capabilities. This complete framework allows a
team of autonomous robots to bring a valuable assistance in a
search and rescue situation and its behavior is presented in a
non-trivial simulation scenario.
Index Terms—Multi-Robots Systems, Rescue Robots, Robot
Control
I. INTRODUCTION
The worst natural disasters and terrorist attacks of mankind
history have motivated multiple researchers around the world
to study the rescue robotics area, for example in [3] it is
presented how robotics can impact on some events such as
the attacks of the World Trade Center. Unfortunately, most of
those disaster situations cannot be detected on time, causing
populations to be unable to evict the area, to be unable to
be ready for this kind of situations and to have problems
avoiding a large number of human and animal casualties.
When a disaster occurs, it is really important to have a solid
contingency procedure as explained in the work presented
in [2] which shows how the first hours and initial days are
the most important in the rescue of survivors. Taking this
into account the use of networked heterogeneous robots to
collaborate the rescuers in a disaster area can improve the
chance of survive for the people trapped in it. Also the area
1 D. Yanguas-Rojas is from Faculty of Engineering, University National of
Colombia Bogota, Colombia
2 G. Cardona is from Faculty of Engineering, University National of
Colombia Bogota, Colombia
3 J. Ramirez-Rugeles is from Faculty of Engineering, University National
of Colombia Bogota, Colombia
4 E. Mojica-Nava is from Faculty of Engineering, University National of
Colombia Bogota, Colombia
2017 IEEE 3rd Colombian Conference on Automatic Control (CCAC)
978-1-5386-0398-7/17/$31.00
2017
c IEEE
3 and Eduardo Mojica-Nava4
can be explored faster with the assistance of the robots giving
precious time to the paramedics to attend the victims. The
use of the robots can reduce the number of victims in the
situation since more lives will not be risked when searching
due to the area can still be dangerous, so we can lain over them
to perform this tasks while the human resources are focused
in the paramedical attention and the evacuation of people in
difficult medical conditions.
This opens the door to the study of which tasks can be
performed by the robots in search and rescue of victims. Tasks
such as exploration, search, delivery of supplies, localization,
rescue of survivors, among others, are the typical tasks studied
for them. However there are multiple tasks different that can
be made for them such as evacuation of people or firefighting.
There are many theories and research work applied to this
topic of robotics that try to solve different challenges of
the previously mentioned tasks, such as the research made
in [4] that shows the need to use heterogeneous robotics
teams in search and rescue operations, in [5], it is presented
schemes of heterogeneous robots team formation, while the
work presented in [6] exposes an exploration algorithm using
heterogeneous robots. Other common approach it is the use
of swarm theory applied into exploration and recognition of
large areas a short time, compared with the time employed by
a single robot for performing the same task as presented in
[7]. It is important to mention that the work presented in [8]
and in [9] are a benchmark to our contribution on exploration
task with their use of Voronoi tessellation on homogeneous
networked robots and the use of artificial potential functions,
respectively.
In allocation tasks there are some other researchers that are
working on this area, for example in [10] there are algorithms
presented for task allocation challenges in robotics search and
rescue operations, in the same way in [11] it is analyzed
the robots task allocation with cost uncertainties for different
situations, and finally in [13] it is shown a control strategy
employable for the coordination of multi-robots to different
kinds of tasks.
The other sections that complete this paper are organized
as follows: Section 2 presents the exploration of the disaster
environment, Section 3 the optimized task allocation and
Section 4 simulation results.
II. E XPLORATION OF THE D ISASTER E NVIRONMENT
The efficiency on exploration task is one of the most
important things in SAR (Search and Rescue) because the
velocity and the way to do that, can define the survival of
humans that be in distress or imminent danger. That is the
reason why one of the major aims on search and rescue
robotics topics is to find an optimized method to explore and
map the disaster area due that frequently there may not be
information available about the state of the environment after
a disaster occurs in an urban area or rural area. The correct
development of this task allows to reduce the time and the
difficulty to recognize and localize all the victims as soon as
possible. The method contemplated in this work to exploration
task is given by:
A. Heterogeneous DisCoverage
In the present paper, we extend a previous work presented in
[1] where there are modifications to the DisCoverage algorithm
presented in [8] in order to adapt it to the heterogeneous case.
This frontier based algorithm tracks the weighted centroid of
certain area (assigned to each robot using a Voronoi based
tessellation) allowing the robots network to explore since the
weight function grows near to the map’s frontier. There, the
tessellation employed considers the different moving speeds
for the robots resulting in cells different from the traditional
Voronoi Cells previously employed with homogeneous robots
with the same speed. This happens because the criteria of
tessellation are not anymore the points that are closest to the
robot, but the points that the robot can reach the fastest.
Another consideration taken in the previous work is that,
the disaster area is a non-convex space where the euclidean
distance direct calculation may be inaccurate. This makes
necessary the implementation of another kind of distance
metric such as 4-neighbors flooding distance based on the
Breadth first search presented in [14] or the use of a optimized
distance metric considering that the robot is capable to move
on any direction how it is shown in Fig. 1 where the direct
calculus of the euclidean distance is not correct since the
obstacles do not allow to follow a direct path.
Fig. 1: Green dot: Actual position Red dot: Target position.
Red line: Euclidean distance. Blue line: Strictly optimized
distance in a non-convex environment.
An important issue boarded in the previous work that affects
the right performance of a mobile robots network is the
connectivity maintaining and the collision avoidance of the
robots. As presented in [9], artificial potential functions are
applied to keep the distance between robots within the limits
of the communication range, but with the possibility of doing
reconfigurations in order to extend the network over a bigger
area. During the mission, the network changes constantly
since the robots are moving at different speeds as they are
approaching to the exploration frontiers. This makes some
links useless and requires the creation of new links, so it
was necessary a network reconfiguration algorithm capable
of maintaining connectivity and flexibility. It is worth to
mention that considering that sometimes the robots can reach
undesired equilibrium points where they reach the centroid
of the Voronoi cell and the exploration task is not complete
yet, we use a perturbation signal which allows the robots to
escape from these equilibrium points respecting the effect of
the artificial potential keeping the network connectivity and
the collision avoidance.
B. Modifications to the Exploration Algorithm
One of the biggest changes in the algorithm is how the
breadth first search is implemented, in the paper presented
before [1], there is used a 4 neighbors flooding, which means
that the algorithm was comparing an initial cell’s value with
the cells above, under, right side and left side of the initial cell,
considering only movements in strictly horizontal and vertical
directions, what can be consider as the representation of the
Manhattan distance in a grid or when the grid tends to the
continuous case to the norm L1 . Now the algorithm is using a
8 neighbors flooding (8NF), allowing to not just going through
the directions mentioned before but also through the ones
located immediately on the diagonals (NE, NW, SW and SE).
The improvement of this change is that the distance is way
more accurate and tends to be closer to the Euclidean distance.
In the algorithm implemented for flood distance, we changed
the term that updates the value of the cell by comparing it with
the current
√ value plus 1 if it is in the 4 neighbors directions
or plus 2 if is in the diagonal directions (the corresponding
euclidean distance for step the movement). In the next image
we can realized the effect of this change (See Fig.2) where the
euclidean distance, the Manhattan distance and the presented
8NF are compared. It is worth to mention that an error can be
produced when the robot is calculating in different angles than
the ones represented by n45◦ , because the main calculus are
performed in those angles. The maximum error presented in
the calculus is when the robot sense in the angles represented
by n45◦ + 22.5◦
Fig. 2: Green line: Manhattan distance. Red line: euclidean
distance. Blue line: equivalent to 8 breadth search.
In Fig.3 it is shown the comparison between the strictly
optimal trajectory and the trajectories made by the theoretical
4NF and 8NF, (3a and 3b). Then we have the strictly optimal
trajectory against the real trajectories of the 4NF and 8NF (3c
and 3d). The real trajectories are those where the robots also
apply the net virtual potential functions to avoid collisions.
We can appreciate how the robot manage to trace a trajectory
that tends to the strictly optimal, when the robot applies the
8NF and can still avoid the walls by the potential functions
respecting the security ranges. (3d)
(a) Optimal convex distance vs (b) Optimal convex distance vs
4NF 8NF
(c) Optimal convex distance vs (d) Optimal convex distance vs
Real 4NF Real 8NF
Fig. 3: Comparison between convex optimal distance with
4NF, 8NF, real 4NF and real 8NF
Another change that is worth to mention is that when new
tasks are assigned to the robots network or just to one of
the robots, the control algorithm will perform a change in the
exploration component, replacing the exploration component
by the corresponding component of the new task. However
the net virtual potential will remain the same. This way the
control signal adopts the form ui = uti +upoti , where ui is the
control signal for the i robot, uti corresponds to a proportional
controller that tracks the mission point of interest (Voronoi
centroid, victim location or evacuation point) for the robot i
and upoti is the net artificial potential for the robot i.
The final important change made to the original algorithm
is the way the robots r that do not have frontier dS within its
own Voronoi cell Vr behave. In the previous work the Voronoi
centroid of the cell without frontier became its geometric cell
and the robot will track it. However, that leads the robots to
stay in areas of low interest since there were already explored.
Considering that, we changed this behavior for tracking the
neighbor N CF who has the frontier the closest (in terms of
nodes of the graph NN F (r)). This way if the robot has no
frontier within its cell it will go closer to the limits of the
exploration frontier and will help with the exploration process
speeding up the whole process as is presented in the Algorithm
1.
Algorithm 1 Exploration Behavior
.
1: function E XPLORATION (Robots, scene). Where Robots
- robots array, scene - matrix
2: for r ∈ Robots do
3: Heterogeneous Voronoi Partition(Robots,Scene)
4: if dS ∈ Vr 6= {φ} then
5: Track Voronoi Centroid
6: NV F (r) = 0
7: else
8: N CF ← Find min NV F (rn) for rn ∈
Robot0 sN eighbors
9: Track N CF
10: NV F (r) = NV F (N CF ) + 1
11: end if
12: end for
13: end function
III. O PTIMIZED TASK A LLOCATION
Throughout the exploration process there are multiple emer-
gent task that must be done by the rescue team. Some of
them could only be made by humans like the first aid and
paramedical tasks, however there are many kinds of tasks
that can be realized by the robots like the presented in
[12]. In this paper, we are going to consider mainly the
victims identification, the victims evacuation and the delivery
of medical supplies. Those tasks must be allocated to the
robots prioritizing the minimization of the mission time that
is minimize the time that the robots take for accomplishing
the attention to all the victims and considering that the robots
may have different capacities for doing each task.
A. Task Definition
Since the objective of the algorithms presented is to min-
imize the time for accomplish the mission, all the task that
will be developed by the robots must have an estimation of its
temporal cost associated in order to determine an optimized
allocation. Those estimations are made as follows:
1) Exploration: The first task that each robot must do is
the exploration of the scene as mentioned before, however the
cost value in time of this task is not trivial and the estimation
that we took is presented in (1) in such a way that in the early
stages of the mission the exploration takes a high priority and
in the final stages of the mission it may be even discarded if
there are not more expected victims to be found.
Am
Cie = Ke (1)
Vi Nmv
where Cie is the cost of the exploration task for the robot
i, Ke is an arbitrary constant that allows to give priority to
the attention of the already found victims or to the finding
of the expected missing victims and allows the conversion of
the result of the expression to seconds, Am is the area of
the mission environment that has not been yet explored, Vi is
the speed of the robot i and Nmv is the expected number of

missing victims to be found. The variable Nmv must be set by
a upper bound of the number of people expected to be found
in the disaster area and can be estimated considering the usual
population density of the location or logistics information
available before the emergency situation.
2) Identification: Concerning to the victims we have con-
sidered three types of them. The victims that are capable of
moving by themselves and can be evacuated, the victims that
are alive but cannot move or require human assistance and
the casualties or false victims that cannot be helped. Once a
victim is found the next task to be done is its classification and
since not necessarily all the robots can, there is the possibility
that the robot that has detected the victim is not capable of
classify or requires to be nearer into the sensing range in order
to do the classification. The cost for this task is presented
in (2) taking advantage of the distance matrix calculus with
the 8 neighbors-flooding which allows the estimation of the
minimum distance from each robot to the victim even if the
path to it is not linear.
ID di,j − ri,ID
Ci,j = Ti,ID + (2)
Vi
where Ci,jID
is the cost of the robot i for identifying the
victim j, Ti,ID is the time required for the robot i for the
identification of a victim (since they may have different types
of sensors this time may vary), di,v is the distance between
the robot i and the victim j evaluated through the distance
matrix generated using the 8 neighbors-flooding, ri,ID is the
minimum distance required for the robot i to identify a victim
and Vi is the speed of the robot i.
3) Supplies Deployment: if the victim is alive but cannot
be evacuated the robots may set a transmitter for marking
precisely its location for the human rescue teams or depending
of the situation may deliver first aid kits or gas masks among
others. This task requires that the robot reaches the location of
the victim and executes the delivery, this is presented in (3).
Del di,j
Ci,j = Ti,Del + (3)
Vi
Del
where Ci,j is the cost of the robot i for delivering a supply
to the victim j and Ti,Del is the time required for the robot i
to hand over the supply.
4) Evacuation: if the victim is alive and can be evacuated
the robot may guide it to a safe area so the robot must reach the
location of the victim going to its own speed and then when
the victim is in range they must go to the safe area going at
the smallest speed between the speed of the victim and the
speed of the robot. This is summarized in (4) considering that
the speed of the victim depends on its state and the class of
the situation, for example the victim could move faster if is
in perfect health, has good visibility and the air is clean but
could move slower if is wounded, has poor visibility and/or
there is smoke in the air.
Ev di,j dj,SA
Ci,j = Tr,Ev + + (4)
Vi min(Vi , Vj )
V
where Ci,j is the cost of the robot i for evacuate the victim j,
Tr,Ev is the time required for the robot i to communicate to
a victim to follow it, dj,SA is the distance between the victim
j and the safe area and Vj is the speed of the victim j.
B. Events and Task Allocation
During the start of the mission, all the robots are set
to explore the mission environment since the position of
the victims is unknown so the allocation is trivial. When a
victim is found or a task is concluded, the tasks available
to be done change so the previous task allocation must be
reevaluated and since generally there will be more tasks than
robots it is necessary to evaluate which task allocation is
optimal in the current setup. In order to solve this allocation
problem we utilize the extension of the Hungarian method
for rectangular matrix presented in [16] with the current cost
matrix considering that all the robots can explore (if there is
space pending to be explored) and that if there are victims that
are not satisfactorily attended the robots that can attend them
will have the possibility. Note that if a robot is not capable
of realizing determined tasks the correspondent cost will be
infinity and the Hungarian method will not assign that task to
any robot.
1) Event Triggered Reallocation: Since the mission condi-
tions change in time we call the reallocation algorithm each
time that an important event occurs. Those events are the
discovery of a potential victim, the identification of the type
of a victim, the successful deliver of a supply, the successful
evacuation of a victim and the end of the exploration task
(there is no more space to be explored). Since those are the
uniques events considered that change the tasks costs table
and the Hungarian method guarantees the optimality of the
assignation, we can be certain that in every moment of the
mission but the moments during the reallocation takes place
the task allocation will be optimal.
2) The Reallocation Propagation: In order to respond to the
changes in the mission environment the fastest as possible, the
robot that detects the change is in charge of evaluating the task
reallocation and send the message to all the other members of
the team. Note that the message will be always delivered to all
the robots of the team since the communication is guaranteed
by the artificial functions that preserve the existing links in the
network and the network reconfiguration algorithm presented
in the previous work.
It is important to mention that the events that trigger the
reallocation, in general are expected to not be simultaneous,
however if there is the case when two robots present different
task allocations based on two simultaneous events the com-
munication protocols will give an pseudo arbitrary order to
the messages and the network will accomplish the consensus
based on the latest. Other possibility that could be considered
is to run a simple exception algorithm which allows the restore
of the consensus of the mission information and generate the
correct optimal allocation. However in practical scenarios the
final behavior is expected to be very similar and the first
approach was considered in this work.
 IV. S IMULATION R ESULTS
In Fig. 4 we present the map corresponding to a complete
mission scenario where the robots have to explore the scene
in the search of the potential victims, determine the state of
each one, accomplish the tasks corresponding to each victim
and return to a Home position. The simulation setup employed
is the same that in the earlier version of the last simulation
presented in [1].
Fig. 4: Scene and Victims Location for the Simulations
A. Exploration Improvement
The main change in the exploration algorithm was made
over the behavior of a robot which has finish the exploration
task due to presenting no frontier (δS) to explore and does
not have any other task assigned, normally it will move
towards the center of its Voronoi cell. This is quite useless, the
network can be using the resources of this robot to enhance
the performance, here is where the new algorithm enters, now
the command is plain and simple this, if the robot does not
present frontier to explore, search for the nearest robot in the
graph network structure (Rni ), that has frontier, then move to
the exact location of that robot. As soon as the wanderer robot
reach the followed robot, it will have frontier at its sight so
it is going to start exploring once again, this will reduce the
exploration time that would be obtain if we had kept the old
algorithm.
In the Fig.5, which is the same presented in Fig.4 without
victims, we can appreciate the entire process, first we have the
top left corner, where the entire robot network is deploy, we
call this home position, then they start the exploring task, each
robot has its own velocity so, we can see in the top right corner
how the network expands to cover the left side of the area, at
the moment just the robot a the front has frontier, the rest has
no frontier to go to, so they will follow the one with frontier,
in the bottom left corner, we have two robots with frontier and
then two robots, surrounded by a green dot line, we will call
them wanderers, those robots are the ones asking for neighbors
with frontier, in the final photograph, bottom right corner, we
can see how almost the entire area has been discover, from the
two wanderer robots one went to the upper side of the area
to reach the fastest robot and the other one stayed in range to
keep connectivity and to help with the frontier in the bottom
side of the area, it is quite clear how the robots reach those
with frontier as soon as they do not have any more to explore,
Fig. 5: Exploration sequence of the robot network with the
exploration improvements
enhancing performance and reducing time in the exploration
task, while maintaining connectivity.
B. Full Mission Simulation
In this simulation the initial position of the robots is the
same line configuration presented in the previous subsection,
however in this case there were 10 victims randomly allocated
in the right half of the environment and with random status
(4 victims that can move, 3 victims that cannot move and 2
casualties). It is worth to mention that the victims capable
to move have simple dynamics programmed that consider the
collision evasion, follow a robot if the robot calls for it and go
to the top left corner (safe zone) once the robots takes then near
enough. They were programmed with a speed of 5 km/h that
is a typical speed for a person walking. In Fig. 6a we present a
situation when during the exploration task there were 4 robots
found and identified however the task allocation algorithm has
determined most of the robots to continue exploring and two
of them were set to deliver supplies to the non-capable to move
victims.
In Fig. 6b we present the moment when the robots found and
identified the last of the victims and did not explore anymore.
In that moment all the non-capable to move victims have
already received their medical supplies so 4 of the robots were
allocated to the evacuation task and the other were set to return
home since there were not more task pending to be done. Note
that since all the victims expected were found the map was
not fully explored (remembering that the number of victims
expected is set to be higher or equal to the real number of
victims in the scene). Later in Fig. 6c we present the process
of evacuation, in this moment the victims are being evacuated
by the robots to the top left corner of the environment (the
safe zone). Note that all the robots that are evacuating victims
are close enough to their victims, this behavior is expected

(a) First Victims Identified (b) End of Exploration Process
(c) Evacuation of capable victims (d) End of mission
Fig. 6: Victims detection, classification and attention - The
Blue Dots correspond to victims capable to move, the green
ones correspond to the victims that are unable to move and
the yellow ones to the casualties found
and will be hold even if the victim’s speed and robot’s speed
differ (actually in most cases this is true). Finally in Fig. 6d
we show the final state of the mission, here the robots are set
to an arbitrary formation, the capable victims were allocated,
the medical supplies were delivered and all the victims and
casualties locations were recorded in the robots memories.
V. CONCLUSIONS AND FUTURE WORK
In general, as it was expected, the desired behaviors of the
previous work is kept. The network connectivity was held; the
collisions between robots, victims and obstacles were avoided
(note that the victim-victim collision was not considered since
in the real world the people will not collide themselves) and
the exploration process was successfully made. Regarding to
the exploration, we aimed to reduce the execution time of the
network to explore an area and we succeeded. When the robots
do not have frontier at sight they follow their neighbor that is
the closest to the frontier speeding up the whole exploration
process. In particular the employ of the 8NF metric made the
distance calculus closer to the real distance in the non-convex
domain and allowed the determination of optimized routes for
the execution of most tasks. Note that the artificial potential
functions in charge of the connectivity maintenance and the
collision evasion change the real routes followed.
Concerning to the task definition and event based realloca-
tion proposed, they showed to be effective in the simulation
environment, since they allowed the mission completion fully
autonomous without any human assistance. It is important to
mention that the tasks cost definition can be improved to be
more accurate with the real routes and that would be an in-
teresting next step to be followed. Another important upgrade
View publication stats
interesting that can be made is a prediction based approach that
do not consider just the actual state of the environment and
victims, but estimations of the future positions of the victims
and possible locations for the remaining victims. Finally an
important consideration that can be taken in the future work
is to consider tasks interdependencies such as in [17] in a way
that allows the robots to do tasks more complex or to evaluate
sequential tasks like exploration, identification and evacuation
in a non-independent way.
R EFERENCES
[1] D. Yanguas-Rojas and E. Mojica-Nava, “Exploration with Heterogeneous
Robots for Search and Rescue”,’ IFAC 2017 World Congress of the
International Federation of Automatic Control’, 9-14 July 2017.
[2] R. C. Huder, Disaster Operations and Decision Making (1). Hoboken,
US: Wiley, 2013.
[3] J. Casper y R. R. Murphy, “Human-robot interactions during the robot-
assisted urban search and rescue response at the World Trade Center”,
IEEE Trans. Syst. Man Cybern. Part B Cybern., vol. 33, nm. 3, pp. 367-
385, jun. 2003.
[4] R. Ventura y P. U. Lima, “Search and Rescue Robots: The Civil
Protection Teams of the Future”, en 2012 Third International Conference
on Emerging Security Technologies (EST), 2012, pp. 12-19.
[5] Gunn, T., and Anderson, J. (2015). Dynamic heterogeneous team forma-
tion for robotic urban search and rescue. Journal of Computer and System
Sciences, 81(3), 553-567.
[6] Chand, P., and Carnegie, D. A. (2013). Mapping and exploration in a
hierarchical heterogeneous multi-robot system using limited capability
robots. Robotics and autonomous Systems, 61(6), 565-579.
[7] Leon J., Cardona G.A., Botello A., Caldern J.M. (2017) Robot Swarms
Theory Applicable to Seek and Rescue Operation. In: Madureira A.,
Abraham A., Gamboa D., Novais P. (eds) Intelligent Systems Design
and Applications. ISDA 2016. Advances in Intelligent Systems and
Computing, vol 557. Springer, Cham.
[8] Haumann, D., Willert, V., and Listmann, K. D. (2013). DisCoverage:
from coverage to distributed multi-robot exploration. IFAC Proceedings
Volumes, 46(27), 328-335.
[9] Zavlanos, M. M., Egerstedt, M. B., and Pappas, G. J. (2011). Graph-
theoretic connectivity control of mobile robot networks. Proceedings of
the IEEE, 99(9), 1525-1540.
[10] Turner, J., Meng, Q., and Schaefer, G. (2015, May). Increasing allocated
tasks with a time minimization algorithm for a search and rescue scenario.
In Robotics and Automation (ICRA), 2015 IEEE International Conference
on (pp. 3401-3407). IEEE.
[11] Nam, C., and Shell, D. A. (2015, May). When to do your own thing:
Analysis of cost uncertainties in multi-robot task allocation at run-time. In
Robotics and Automation (ICRA), 2015 IEEE International Conference
on (pp. 1249-1254). IEEE.
[12] Pellenz, Johannes and Jacoff, Adam and Kimura, Tetsuya and Mi-
hankhah, Ehsan and Sheh, Raymond and Suthakorn, Jackrit (2013).
’RoboCup rescue robot league’ in Lecture Notes in Artificial Intelligence
(Subseries of Lecture Notes in Computer Science) (pp. 673685). Elsevier
Verlag.
[13] Lee, S. G., Diaz-Mercado, Y., and Egerstedt, M. (2015). Multirobot
control using time-varying density functions. IEEE Transactions on
Robotics, 31(2), 489-493.
[14] Skiena, S. S. (2008). Sorting and Searching. The Algorithm Design
Manual, 103-144.
[15] Zhang, B., Peng, J. (2013). Uncertain programming model for uncertain
optimal assignment problem. In Applied Mathematical Modelling (pp.
64586468). Elsevier.
[16] F. Burgeois and J.-C. Lasalle (1971). An extension of the Munkres
algorithm for the assignment problem to rectangular matrices. Commu-
nications of the ACM, 142302-806, 1971.
[17] Brutschy, A., Pini, G., Pinciroli, C. et al. Self-organized task allocation
to sequentially interdependent tasks in swarm robotics In Autonomous
Agents and Multi-Agent Systems (2014) 28: 101. doi:10.1007/s10458-
012-9212-y