Multi-robot Decentralized Exploration using Weighted Random
Selection
Abhijith N. Balan
a
and Asokan Thondiyath
b
Department of Engineering Design, Indian Institute of Technology Madras, Chennai, India
Keywords:
Decentralized, Frontier Allocation, Frontier-based Exploration, Multi-robot System.
Abstract:
The exploration tasks using multi-robot systems require efficient coordination and information sharing be-
tween robots. The map creation is usually done through allocating frontiers,i.e., the boundary between ex-
plored and unknown regions of the map, to each robot. This paper introduces an efficient frontier allocation
method based on weighted random selection for a decentralized multi-robot system. The weights are calcu-
lated based on the size of the frontiers and the distance between a robot and the frontiers. In this strategy, each
robot identifies the available frontiers in a shared map and select the goal for exploration through a random
selection. Even though a robot is randomly picking a frontier without coordination with other robots, a collec-
tive intelligence is developed at the swarm level. The proposed method is computationally efficient and uses
minimal communication between the robots in a decentralized multi-robot system. As a result, the robots get
allocated to frontier according to the calculated weight. A comparison with the nearest frontier exploration
approach in computerized simulation demonstrates the efficiency of the proposed algorithm.
1 INTRODUCTION
Robots have come a long way from a mere fictional
entity to a useful tool that can aid or even replace
human effort in many situations. They can carry
out tasks that are difficult or impossible for a hu-
man to perform. Introduction of Multi-Robot Sys-
tem(MRS) extends the capabilities of robots by many
folds. These systems work by coordinating multiple
robots to perform a single task or multiple tasks at
the same time. These systems feature high levels of
scalability, resistance to failure, and parallelization of
tasks. Such systems are very efficient in tasks such
as search and rescue in disasters, nuclear site explo-
ration, surveillance, environment monitoring, and lo-
cating the sources of hazardous materials.
MRS is mainly classified into two categories, cen-
tralized and decentralized systems. A centralized sys-
tem would have multiple robots being controlled and
managed by a central entity. The central entity man-
ages the system’s decision making, communication,
and data storage. Failure to this unit would disturb
the entire system and the tasks can no longer be com-
pleted. A Centralized system has global information
about all the robots and their tasks. Access to this in-
a
https://orcid.org/0000-0003-4649-2429
b
https://orcid.org/0000-0002-5474-3999
formation makes coordinating and controlling a cen-
tralized robot swarm easier and efficient. A decentral-
ized system, however, lacks the central entity. In such
systems, the robot swarm as a whole takes care of
decision making through local communication alone.
Since the robots lack the information of the state of
the system(position of the individual robots and their
task status), they take decisions based on its limited
perception of the environment and its neighbors. This
makes developing strategies for information sharing
and exploration in the decentralized system difficult.
Even though a decentralized system will not be as ef-
ficient as a centralized one, it offers scalability and re-
dundancy, which plays a crucial role in robot swarms.
A robot swarm is an autonomous multi-robot sys-
tem that can only perform local sensing and commu-
nication, with no access to information about other
robot’s position or task status (Brambilla et al., 2013).
They are usually homogeneous, containing simple
robots and resembles a decentralized MRS.
Since the exploration and mapping process is par-
allelizable, multiple robots can carry it out faster and
more efficiently than a single advanced robot. A robot
system performs this task through coordination, infor-
mation sharing, and data fusion. Autonomous map-
ping and exploration can be done through exploring
frontiers which are the boundary between mapped ar-
Balan, A. and Thondiyath, A.
Multi-robot Decentralized Exploration using Weighted Random Selection.
DOI: 10.5220/0010608405230529
In Proceedings of the 18th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2021), pages 523-529
ISBN: 978-989-758-522-7
Copyright
c
2021 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
523
eas and unknown areas. A robot navigating to a fron-
tier can disclose new regions and expand the map.
In this paper, we focus on exploration using decen-
tralized swarm robot systems. The most challeng-
ing problem in such systems is the decision making.
Since global information is unavailable to an agent
in a swarm, the agent chooses a local goal based on
the information available. In a mapping process, it
will be the selection of most promising frontier to ex-
plore next. A good strategy for optimizing the local
goal can reduce redundant data collection and make
the process faster and efficient.
This paper puts forward a decentralized algo-
rithm for frontier selection which is computationally
efficient and uses minimal communication between
robots. This implicit frontier allocation algorithm al-
lows each robot to take a decision with the local in-
formation it possesses without the need of coordina-
tion with other robots. The algorithm works better
with multi-robot systems with large number of robots.
Comparing to other implicit exploration algorithm,
our method was 13.7% faster for a multi robot sys-
tem of 7 robots without adding any extra communica-
tion effort. The independence of frontier selection on
robot coordination enables the algorithm to scale eas-
ily as the number of robots increase and makes this
algorithm the right choice for robot swarms in explo-
ration and mapping.
2 PROBLEM STATEMENT
The following defines the problem addressed.
The task is to reduce the overall duration of ex-
ploration by a fleet of autonomous robots while keep-
ing the communication between the robots minimal.
The robot swarm is considered homogeneous and
is equipped with exteroceptive sensors, which allow
them to create a map of the environment. The robots
also contain devices for communication which are
crucial for cooperative exploration. The range of
communication is assumed to be less than the map-
ping range and may not cover the whole robot fleet,
i.e., all robots are not connected to each other always.
The environment is unknown, and all the robots may
not start at nearby locations. One robot might explore
a frontier even if another robot had already explored
it.
Due to the limited communication range, one
robot may not be able to communicate with all the
robots in the fleet. A robot should make a decision
based on the local information available at that time.
The algorithm should be efficient in computation so
that the system would not be affected as the number
of robots in the fleets is increased.
To limit the scope of the problem, the details
on robot localization, mapping, and map sharing are
not discussed in this paper. However, they are dis-
cussed in the work by (Balan, 2019) in which com-
plete multi-robot SLAM(Simultaneous Localization
And Mapping) simulations are addressed in both cen-
tralized and decentralized systems.
After identifying all the available frontiers to a
robot, the robot has to be allocated to one frontier
for further exploration. The inability of one robot to
know the position and status of all other robots in the
fleet results in selecting a sub-optimal frontier. This
problem is inherent in a decentralized multi-robot sys-
tem.
The following notations are used:
R is the set of all robots participating in explo-
ration. R = {R
1
,R
2
,R
3
......R
n
}, where n is the
total number of robots in the system.
F
i
is the set of all frontiers available to the R
i
.
F
i
= {F
1
i
,F
2
i
,F
3
i
......F
m
i
}, where m is the total
number of frontiers available to R
i
.
α
i
is the assignment vector for R
i
.
α
i
j
=
(
1 if F
i
j
is assigned to R
i
0 if otherwise
(1)
Since this is a decentralized approach, the available
frontiers and the assignment vectors are different for
each robot. Now the problem has simplified to finding
the assignment vector α for each robot using F at any
time.
3 STATE OF THE ART
In this section, we discuss the previously proposed
methods in multi-robot exploration.
3.1 Nearest Frontier Exploration
Nearest frontier exploration was a popular frontier
based exploration strategy formulated by (Yamauchi,
1997). Yamauchi recognized frontiers as the bound-
ary between the known area and the unexplored area
of the map. Navigating a robot to a frontier will en-
able the robot to identify unexplored area that conse-
quently increases the size of the explored area. In a
system with more than one robot, each robot has to
be navigated to the nearest frontier and broadcast the
map information to all other robots. The robot can
stop exploration when there are no frontiers left to ex-
plore in the map. This ensures coordination between
ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics
524
the robots. The area in the map which was already
covered would not be covered again by another robot.
This is an implicit method, i.e., after receiving the in-
formation from other robots, a robot does not need to
communicate with other robots for selecting the best
frontier to navigate. Nearest frontier exploration is
a distributed approach and requires less computation
for frontier selection. The algorithm for nearest fron-
tier exploration is given in Algorithm 1.
Algorithm 1: Nearest Frontier Exploration O(m).
Input: F
i
Output: Assignment vector α
i
of R
i
α
i
j
= 1 with j = argmin
F
i
j
F
i
distance to F
i
j
;
Nearest frontier exploration can be reduced to
O(1) using Breadth-First Search from robot position
and taking the first frontier we encounter. However,
it is possible that multiple robots may get allocated to
the same frontier if their maps are similar (when they
are close to each other. This is common in a robot
swarm). This keeps the fleet from utilizing its full po-
tential to explore the map.
3.2 MinPos
MinPos is a decentralized frontier allocation algo-
rithm for multi-robot exploration by (Bautin et al.,
2012). This algorithm strategically allocates best
frontiers to robots.each robot evaluates its relative
rank among the other robots in term of travel distance
to each frontier. Accordingly, robots are assigned to
the frontier for which it has the lowest rank. To eval-
uate this criterion, a wavefront propagation is com-
puted from each frontier giving an alternative to path
planning from robot to frontiers. This method not
only considers the distance to each frontier but also
the number of robots closer to that frontier. The algo-
rithm builds a local minimum free artificial potential
field from each group of frontier using the wavefront
propagation algorithm. This also builds the cost ma-
trix, which is used to allocate frontiers. The algorithm
for MinPos is given in Algorithm 2.
MinPos is a decentralized frontier allocation algo-
rithm. It does not require any central unit to function
like in MOARSLAM (Morrison et al., 2016). How-
ever MinPos algorithm requires the communication
range to be large enough so that it covers the whole
map,i.e., all robots should be in constant communi-
cation with all other robots at all times. MinPos re-
quires information about all the robots participating in
the cooperative mapping and their task status for the
frontier allocation to work. This would not be avail-
able in a robot swarm always. Even though MinPos
Algorithm 2: MinPos O(nm).
Input: R , F
i
,C cost matrix
Output: Assignment vector α
i
of R
i
for F
i
j
F
i
do
P
i
j =
R
k
R ,k6=1,C
k
i
<C
i
j
1
end
α
i
j
= 1 with j = argmin
F
i
j
F
i
P
i
j;
requires global information, the decisions are taken
by individual robots. It is also interesting to note that
the algorithm complexity is O(mn),i.e., the algorithm
becomes slower to allocate frontiers as the number of
robots in the swarm increases. This can cause MRS
to scale poorly.
3.3 Centralized and Decentralized
A centralized exploration strategy will have a cen-
tral server which allocates frontiers to each robot for
the maximum efficiency. In MOARSLAM (Morri-
son et al., 2016), the server stores the map created
by autonomous robots. The server guide the robots
to each location for exploration and the robots trans-
mits the data back to the server. In (Simmons et al.,
2000), the robots perform maximum likelihood esti-
mation of the map using odometry and observations.
These maps are transmitted to a server which devel-
ops the global map. A similar strategy is implemented
in (Gil et al., 2010), in which the robot sends the
virtual descriptors along with the odometry informa-
tion to the server. The server adds this information
to the global map. All these strategies represent a
single point which is in constant bi-directional com-
munication with individual robots. The performance
of the system depends on the capabilities of this cen-
tral entity. There are also limitations on the extent
of area that can be mapped (communication range of
the central unit) and the number of robots that can
communicate to the central unit at a time (communi-
cation bandwidth limitation). This inhibits the scala-
bility of the system. To make the systems decentral-
ized, one will have to remove the central server and
make the robots capable of making decisions. Nearest
frontier exploration 3.1 and Minpos 3.2, which were
discussed before, are examples for this. Another ap-
proach was proposed in (Yan et al., 2011) in which a
trade based scheme is implemented in a decentralized
multi-robot system so that each robot bids on a fron-
tier based on the limited information available. This
method can reduce computation cost through parallel
computation but induces a communication overhead
of O(mn), which does not facilitate the scalability fac-
tor.
Multi-robot Decentralized Exploration using Weighted Random Selection
525
We need a frontier allocation method which
is O(m) so that the system becomes easily scal-
able (computation cost is independent of number of
robots) and should work in the absence of global in-
formation.
4 PROPOSED METHOD
We propose an approach to the frontier allocation
problem based on weighted random selection. In this
approach, we calculate a weight for each frontier en-
countered based on its size and distance and allocate a
frontier to a robot according to it. This is a distributed
frontier allocation method that will divide the robot
swarm according to the size of the discovered fron-
tier (number of frontier points) and the distance to it.
This algorithm is computationally efficient and does
not require robot communication to select the frontier
for exploration. The algorithm performance is inde-
pendent of the number of robots in the fleet, which
makes the system to be scaled easily to any number
of robots.
In the proposed method, each robot performs the fol-
lowing steps. (More details on the robot system mod-
ules and their interaction are given in the flowchart
(Figure 1))
1. Each robot explores freely in the environment.
2. At robot rendezvous, the i
th
robot broadcasts its
local map data to other robots.
3. Also, the i
th
robot receives map data from other
robots.
4. The robot R
i
creates an updated map.
5. R
i
performs weighted random selection for fron-
tier selection.
6. The R
i
navigates to the selected frontier
Figure 1: A schematic diagram showing the robot systems
module and their interactions. Note that the communica-
tion is only required for the Map Fusion, and the frontier
selection is performed within the robot system.
Although this is an implicit method, i.e., robots
do not coordinate with each other at the frontier se-
lection stage, the fleet as a whole takes a meaningful
decision by dividing and allocating itself to different
frontiers in the map. This technique is similar to the
swarm intelligence present int the nature (e.g. Ants
finding the shortest path to food and Termites build-
ing their nests). These systems also contain various
forms of decision making with the help of limited lo-
cal information. Our technique enables a single robot
following simple local rules to give rise to a collective
intelligence in the swarm level.
4.1 Map Sharing
The robots upon rendezvous, share their map data and
update their local map. After the map update, all the
robot’s local maps share a common set of features.
Since the communication range is assumed to be less
than the mapping range, the complete map of the en-
vironment has been developed using these common
map features. Each robot fuses the received maps to
its local map to obtain the updated map result. The
image-based map merging algorithm, as illustrated in
(H
¨
orner, 2016) has been used for map sharing.
4.2 Frontier Classification
After the robot develops the updated map, the robot
searches the map for frontier points. Frontier points
in the occupancy grid map are found by searching for
the points in between the explored and unknown parts
of the map. After every frontier points are identified,
they are clustered into individual frontiers (set of fron-
tier points) using a clustering algorithm based on their
position. After finding the frontiers, the shortest dis-
tance Euclidean distance (D
i
j
) from the robot R
i
to the
frontier F
i
j
, and the frontier size (S
i
j
, the number of
frontier points in the frontier F
i
j
) are calculated for
each of the frontiers. These parameters will be used
to select the frontier for exploration.
4.3 Frontier Allocation
In this stage, the robot is left with the frontiers and
their parameters (distance and size). We create the
weight of each frontier from these to parameters as
W
i
j
=
S
i
j
D
i
j
(2)
The weight indicates the tangent of the angle cre-
ated by a frontier in robot’s vision. Refer Figure 2. An
imaginary frontier is considered at the frontier point
ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics
526
Figure 2: Diagram showing how the weight corresponds to
the angle made by the frontier in robot vision.
closest to the robot, aligned perpendicular to the sens-
ing direction, with the same frontier size. This fron-
tier makes an angle α in the view of the robot. Note
that tangent of the angle α is the weight that frontier
has. A large frontier located far away from the robot
gets lower weight. So the weights indicate the angles
each frontier makes in the vision of the robot. Refer
Figure 3 for example of such frontier selection.
Figure 3: A schematic diagram showing the concept of
weight used in the proposed approach. Each color (ma-
genta, green, blue, black) indicate an available frontier, and
the red circle indicates the robot R
i
. Here note that the an-
gle made by the frontier colored green makes the biggest
angle in the robot’s view. This frontier will have the largest
weight among all frontiers. Image credits: (Virtual Labs,
IIIT Hyderabad).
Now we select a frontier (or calculate α
i
j
) from
the set of frontiers through weighted random selection
using the weights we calculated.
α
i
j
=
(
1 if F
i
j
= Random Choice(F
i
) according to W
i
0 if otherwise
(3)
Algorithm for the proposed method is given below
Algorithm 3: Proposed Method O(m).
Input: F
i
Output: Assignment vector α
i
of R
i
for F
i
j
F
i
do
W
j
i
=
S
i
j
D
i
j
end
SelectedFrontier
Random Choice(F
i
) with weights W
i
α
i
j
= 1 where F
i
j
= SelectedFrontier
Through individual robots performing a weighted
random selection for choosing frontier, we can divide
a robot swarm to multiple frontiers according to its
proximity and size. It is interesting to note that this di-
vision is happening without explicit coordination be-
tween the robots.
5 EXPERIMENTS AND ANALYSIS
The experiments are done in a computer simulation
using Robot operating system. The main concentra-
tion is on the time taken for thoroughly exploring the
environment. The proposed approach is compared
with the other implicit frontier allocation algorithm
Nearest Frontier Exploration.
5.1 Simulation
For simulation stage simulator (Figure 5) is used with
cave map accommodating all the robots used for ex-
ploration. The robots have 180
vision for mapping
with parameterized communication range. The map
is scaled appropriately (30m X 30m) to simulate a
decentralized decision-making process with all the
robots. The robot is approximately the size of a 0.4m
cube.
The Figure 4a shows the comparison between the
map exploration time and the number of robots for
both the methods, Nearest Frontier Exploration, and
the proposed method. As both algorithms are dis-
tributed, with implicit frontier allocation, they can be
compared. The results shown are the average of 10
iterations on each robot count in both the algorithms.
Figure 4b compares the map exploration time
while varying the local communication range. The
local communication range denotes the distance to
Multi-robot Decentralized Exploration using Weighted Random Selection
527
(a) The trend in exploration completion time with respect to
the number of robots. (Local communication range = 5m)
(b) The dependence of the algorithms to the local communi-
cation range. (For five robot fleet)
Figure 4: Simulation result using the cave environment. The figure shows a comparison between Map completion time
and Number of robots(a) and Local communication range(b). The proposed algorithm is compared with Nearest Frontier
Exploration Algorithm.
Figure 5: Simulation of 3 robots (blue boxes) in Cave map
loaded in stage simulator.
which the communication between the robots is pos-
sible.
5.2 Analysis
From the simulation results, we can see that the pro-
posed method, on an average, is 13.7% faster com-
pared to the other algorithm on high robot count. Al-
though both algorithms have similar results in low
number of robots, the proposed method shows a clear
advantage as the number of robots increases. We
can also see that the algorithm works better when
the local communication range is small. This in-
dicates the advantage our algorithm has in covering
large areas compared to the robot size and its sens-
ing range. We could observe that the robots tend to
move together in Nearest Frontier Exploration when
they were spawned very close to each other. Having
similar maps and frontier distances were forcing the
robots to flock together. However, in the proposed
method, the robot system was allocated across the
frontiers even when they were outside the communi-
cation range. A larger frontier was getting assigned
more robots even if the robots are outside each other’s
communication range. This shows that our method
developed a collective intelligence in choosing the
frontiers for exploration without the need for any
communication between the robots. This proves that
the proposed algorithm can work better in a swarm
system containing a large number of robots.
6 CONCLUSION
In this paper, we addressed the problem of explor-
ing an unknown environment with the decentralized
multi-robot system. An algorithm was proposed for
this problem, which allocates the frontiers to the
robots efficiently so that the exploration time can be
reduced. The algorithm addresses the limitations of
other previous approaches and attains good results
with reducing the computation load on the robots and
also using less communication between the robots.
The proposed algorithm uses the concept of
weighted random selection for assigning frontiers
to robots. In the discussed scenario, i.e., mapping
an unknown environment, the weights are calcu-
lated according to the size and proximity of frontiers.
Other scenarios might require different parameters for
weight calculation (e.g., likelihood of finding a tar-
ICINCO 2021 - 18th International Conference on Informatics in Control, Automation and Robotics
528
get in search and rescue operations), but the underly-
ing concept of weighted random selection remains the
same.
The performance of this algorithm is compared
with nearest frontier allocation through computer
simulation. The proposed algorithm appeared to be
significantly more efficient in robot systems with
higher number of robots. Furthermore, our algorithm
has lower complexity than other multi-robot explo-
ration algorithms and is independent of the number
of robots in the fleet. These properties make our
approach a good option in robot swarms for explo-
rations.
REFERENCES
Balan, A. (2019). Strategies for multi-robot slam using
robot swarms.
Bautin, A., Simonin, O., and Charpillet, F. (2012). Minpos
: A novel frontier allocation algorithm for multi-robot
exploration. In Su, C.-Y., Rakheja, S., and Liu, H.,
editors, Intelligent Robotics and Applications, pages
496–508, Berlin, Heidelberg. Springer Berlin Heidel-
berg.
Brambilla, M., Ferrante, E., Birattari, M., and Dorigo, M.
(2013). Swarm robotics: A review from the swarm
engineering perspective. Swarm Intelligence, 7:1–41.
Gil, A., Reinoso,
´
O., Ballesta, M., and Juli
´
a, M. (2010).
Multi-robot visual slam using a rao-blackwellized par-
ticle filter. Robotics and Autonomous Systems, 58:68–
80.
H
¨
orner, J. (2016). Map-merging for multi-robot system.
Morrison, J. G., Gavez-Lopez, D., and Sibley, G. (2016).
Scalable multirobot localization and mapping with
relative maps: Introducing moarslam. IEEE Control
Systems Magazine, 36(2):75–85.
Simmons, R. G., Apfelbaum, D., Burgard, W., Fox, D.,
Moors, M., Thrun, S., and Younes, H. L. S. (2000).
Coordination for multi-robot exploration and map-
ping. In Proceedings of the Seventeenth National Con-
ference on Artificial Intelligence and Twelfth Confer-
ence on Innovative Applications of Artificial Intelli-
gence, pages 852–858. AAAI Press.
Virtual Labs, IIIT Hyderabad. Frontiers. Virtual Labs.
Yamauchi, B. (1997). A frontier-based approach for au-
tonomous exploration. In CIRA.
Yan, Z., Jouandeau, N., and Ch
´
erif, A. A. (2011). Multi-
robot decentralized exploration using a trade-based
approach. In ICINCO.
Multi-robot Decentralized Exploration using Weighted Random Selection
529