ArticlePDF Available

A survey on multi-robot coverage path planning for model reconstruction and mapping

Authors:

Abstract and Figures

There has been an increasing interest in researching, developing and deploying multi-robot systems. This has been driven mainly by: the maturity of the practical deployment of a single-robot system and its ability to solve some of the most challenging tasks. Coverage path planning (CPP) is one of the active research topics that could benefit greatly from multi-robot systems. In this paper, we surveyed the research topics related to multi-robot CPP for the purpose of mapping and model reconstructions. We classified the topics into: viewpoints generation approaches; coverage planning strategies; coordination and decision-making processes; communication mechanism and mapping approaches. This paper provides a detailed analysis and comparison of the recent research work in this area, and concludes with a critical analysis of the field, and future research perspectives.
Content may be subject to copyright.
published in SN Applied Sciences DOI: 10.1007/s42452-019-0872-y
https://link.springer.com/article/10.1007%2Fs42452-019-0872-y
A Survey on Multi-Robot Coverage Path Planning for Model
Reconstruction and Mapping
Randa Almadhoun1
·Tarek Taha2
·Lakmal Seneviratne1
·Yahya
Zweiri1,3
Abstract There has been an increasing interest in re-
searching, developing and deploying multi-robot sys-
tems. This has been driven mainly by: the maturity of
the practical deployment of a single-robot system and
its ability to solve some of the most challenging tasks.
Coverage Path Planning (CPP) is one of the active
research topics that could benefit greatly from multi-
robot systems. In this paper, we surveyed the research
topics related to multi-robot CPP for the purpose of
mapping and model reconstructions. We classified the
topics into: viewpoints generation approaches; coverage
planning strategies; coordination and decision-making
processes; communication mechanism and mapping ap-
proaches. This paper provides a detailed analysis and
comparison of the recent research work in this area, and
concludes with a critical analysis of the field, and future
research perspectives.
Keywords Coverage Path Planning ·Viewpoint
Sampling ·Multi Robot ·Model Reconstruction.
1 Introduction
Coverage path planning (CPP) is the process of com-
puting a feasible path encapsulating a set of viewpoints
through which a robot must pass in order to completely
scan or survey the structure or environment of interest.
Various technological developments and advancements
in sensor technology, navigational, communication and
computational systems have facilitated the increase in
the level of autonomy in multi-robot systems. There-
fore, some autonomous systems shifted over the past
Randa Almadhoun , randa.almadhoun@ku.ac.ae ·
1Abu
Dhabi, UAE, Khalifa University of Science and Technology ·
2Algorythma’s Autonomous Aerial Lab, Abu Dhabi, UAE ·
3Faculty of Science, Engineering and Computing, Kingston
University London, London SW15 3DW, UK
decade towards cooperative systems in order to achieve
(CPP) objectives more efficiently and robustly [46,3].
Different research approaches have been followed in
the past to perform CPP depending on the environ-
ment, the shape of the structure, and the level of the
required details. The two main challenging components
of CPP are viewpoints generation and coverage path
generation. Viewpoints generation defines the positions
and orientations of the sensor from where the data will
be collected, thus affecting the overall coverage. The
performance of the coverage planning approach is usu-
ally measured by the coverage completeness and its
optimality. The main contributing factors that affect
the overall multi-robot system, and model/map quality
include the information gathering method whether it
is continuous or discrete, the coverage path generation
method whether it is online or offline, and the mapping
or reconstruction methods.
Generally, the development of cooperative multi-
robot systems is concerned with a group of agents that
work collectively to achieve a common objective. Typ-
ical common objectives include surveillance, monitor-
ing, surveying, and modeling. Wide range of applica-
tions utilize cooperative multi-robot systems includ-
ing: search and rescue missions [3], forest fire monitor-
ing [50], industrial inspection [15,2], and natural disas-
ter monitoring and relief [46]. The objectives in these
applications can be achieved far more efficiently and re-
liably using a team of cooperative agents rather than a
single agent. In these kinds of applications, CPP plays
a vital role in coordinating the tasks of each of these
agents in order to achieve the main objective.
Multi-robot CPP is the process of computing a
set of feasible paths encapsulating a set of viewpoints
through which the team of robots must visit, each with
its assigned path, in order to completely scan, explore
or survey the structure or environment of interest. In
2 Randa Almadhoun1et al.
specific applications, CPP is a process used for au-
tonomous mapping and reconstruction where these gen-
erated maps or models can hold different level of in-
formation, such as temperature, occupancy, or signals
strength, and can be used for various applications.
CPP strategies were deployed on various au-
tonomous robotic systems in the literature varying from
Unmanned Aerial Vehicles (UAVs), Unmanned Ground
Vehicles (UGVs) to marine robots. Recently, the inter-
est in UAVs have grown steadily, especially due to de-
creases in their weight, size and cost, and the increase in
their actuators performance. Moreover, the availability
of cheap light-weight processing power and miniatur-
ized accurate sensors increased their level of autonomy.
Furthermore, UAVs covers a broad set of applications
that cannot be fulfilled by other types of robots due to
their agility, and ability to move in unstructured envi-
ronments.
Technically, various robotic capabilities are required
in order to perform CPP including: localization, navi-
gation and path planning, and sensing and perception.
The level of complexity of these requirements varies
based on the number of robots and the environment di-
mensionality. As such, the robots need to be equipped
with various intelligent sensing capabilities that provide
information with enhanced quality in order to recon-
struct the structure of interest or map the environment
accurately.
The main factors that could effect the performance
of a multi-robot CPP approach include: information
sharing (whether it is centralized, decentralized, or dis-
tributed), viewpoints generation, path generation, task
allocation, reacting to dynamic changes (collision avoid-
ance), and model reconstruction or mapping approach.
Majority of existing approaches in literature attempt
to: reduce the computational cost (time need to com-
pute and execute the CPP mission) [62,83], avoid col-
lision internally between team of robots and externally
with the structure or environment [10,41,4], and gather
information with sufficient resolution for mapping and
reconstruction [69,79].
Most of the existing surveys in this topic address is-
sues such as perception, exploration, guidance and con-
trol. A few surveys address the CPP problem focus-
ing on single robot CPP problems and briefly mention-
ing the multi-robot as an extension [78,28], or focusing
on multi-robot area coverage problems [76]. Although
some of the techniques performed in single robot CPP
can be extended and applied to multi-robot systems,
several additional aspects must be considered including
viewpoints generation, communication/task allocation,
robustness (failure handling), and mapping. Also, the
dimensionality of the CPP differs between area cover-
age and large structures coverage problems.
This paper presents a survey on multi-robot cover-
age path planning. Our review focuses on approaches
related to multi-robot systems applied on environ-
ments of different dimensions. The survey provides de-
tails about the main aspects of performing CPP us-
ing multi-agent systems including: viewpoints gener-
ation, path planning, communication/task allocation,
and mapping. A detailed discussion about the main as-
pects is also provided with a flow chart showing the in-
formation flow throughout a multi-robot CPP mission.
A set of research perspectives in this topic are discussed
for further future developments. The main performance
metrics are also explored in the survey and shown in a
table summarizing the most recent work in this topic
using multi-robot systems.
In this survey paper, the main components of the
multi-robot CPP process are detailed in separate sec-
tions. Section 2focuses on single and multi robot CPP.
Viewpoints generation strategies will be reviewed in
section 3.1, and multi-robot CPP approaches in sec-
tion 3.2. An overview discussion of the main aspects
and future research perspectives of CPP are presented
in section 4. Finally, conclusions are presented in sec-
tion 5.
2 Overview of CPP
Coverage path planning (CPP) is the process of explor-
ing or exhaustively searching a workspace, whether it
is a structure of interest or an environment, and de-
termining in the process the set of locations to visit
while avoiding all possible obstacles [28,77]. Some ap-
plications require achieving complete coverage using
the various CPP techniques such as structure paint-
ing, object reconstruction, lawn mowing, surveillance,
geo-spatial mapping, agricultural surveying, and floor
cleaning. Generally in CPP, either the model is recon-
structed in real time utilizing the robot’s sensing ca-
pabilities, or a reference model is provided in advance
for the structure or the environment of interest [72,73].
Extensive reviews of the various CPP methods in liter-
ature are presented in [18,28] describing their function-
alities and applications.
The main components of CPP include: viewpoints
generation, path planning, and coverage completeness
quantification. These components are directly depen-
dent on the exploration method selected for the struc-
ture or environment of interest coverage whether it is
performed offline or online. Performing offline CPP re-
quires using a reference model of the structure or en-
vironment of interest while performing CPP online re-
A Survey on Multi-Robot Coverage Path Planning for Model Reconstruction and Mapping 3
quires utilizing the sensors information to perform cov-
erage.
2.1 Model Based or Non-Model Based CPP
Several approaches exist in the literature for performing
CPP using single robot. These approaches are classified
into model and non-model based approaches. A heuris-
tic based approach is proposed in [61] which takes into
consideration length, timing, coverage, and UAV en-
durance constraints for a distributed structure inspec-
tion application. In this work, an Art Gallery Problem
(AGP) approach is used to generate the set of view-
points based on the selected structures’ mesh models.
A Travelling Salesman Problem (TSP) is used to gener-
ate optimized paths for each structure, which are then
assigned randomly based on time constraints and TSP
path for the UAV. The work presented in [64] proposed
an inspection path planning approach that formulate
the CPP as an extended TSP, where coverage and ob-
stacle avoidance are taken into account. The surface
of the mesh model of the structure of interest is used
to generate viewpoints perpendicular and at a distance
from the structure surface. The TSP is solved using
an enhanced Particle Swarm Optimization (PSO) ap-
proach developed in GPU framework.
Additional research work performing model-based
CPP and considering the UAV energy are presented
in [25] and [81]. The work presented in [25] proposed
multi-objective Evolutionary Algorithm (EA) to gener-
ate coverage path for complex structure inspection tar-
geting coverage and energy objectives. The proposed
work performs uniform sampling based on an existing
model, utilizing a predefined bounding box, and dis-
cretization resolution. The EA plans the path using
the Non determined Sorting Genetic algorithm (NSGA-
II), then it measures energy and coverage scores offline
and penalize paths with collisions. Minghan and Vol-
can in [81] proposed a CPP approach that considers
the energy of the UAV in order to perform area cover-
age. The area is represented as polygonal grid with sin-
gle charging station where the polygonal grid represents
the viewpoints. The coverage planning approach follows
a depth-first approach, where three types of robot mo-
tion are defined including Advance (move to the next
cell providing shortest path), Retreat (return to charge)
and Follow (follow current contour). Most of the pre-
sented single robot model-based CPP approaches tar-
gets area coverage and simple structures. Some of the
approaches focuses on the energy aspect without focus-
ing on the quality of the scans and the computational
cost which are critical in CPP applications [81,16].
The second approach of the single robot CPP ap-
proaches follows a non-model based approach. The work
presented in [75] proposed a CPP approach extended
from [74] by utilizing the surface information to plan
the coverage path online using Truncated Signed Dis-
tance Fields (TSDF). The search space is divided into
cuboid regions that are used to build a volumetric map
of containing surface regions and frontier regions. The
volumetric map is used in computing the Information
Gain (IG) considering the cuboid volume and path
length. Visitation order is computed for each cuboid
applying Hamiltonian path problem while the path is
generated using Generalized TSP. Another non-model
based approach is presented by Emanuele and Cyrill
in [59] where they proposed an exploration algorithm,
which selects the Next Best View (NBV) that maxi-
mizes the predicted information gain, taking into con-
sideration the cost of the distance, and battery life.
The proposed work dynamically builds a hull that sur-
rounds a predefined bounding box which is updated
based on the new information. The viewpoints are uni-
formly sampled into a fixed number where they point
to the vertical axis that passes through the centroid of
the bounding box. The planning approach follows prob-
abilistic approach with a utility function that reduces
the 3D reconstructed model uncertainty, turns in the
flight path, and produce safe path based on time lim-
its. The energy aspect in non-model based single robot
approaches is considered a critical part, especially be-
cause the CPP is performed online. Using this type of
approaches with a single robot makes it hard to achieve
a high coverage percentage and increase the computa-
tional complexity. This difficulty arise from the com-
plexity of the environments which include occluded re-
gions that are hard to be found by single robot and
consumes a lot of time of exploration. A review of other
single robot CPP approaches with critical analysis is
presented in [8]. The advantages and disadvantages of
model/non-model based approaches are highlighted in
the viewpoint generation subsections and the discussion
section.
2.2 Single or Multi-Robot
For large areas and structures, utilizing multi-robot
CPP strategies could have great advantages in achiev-
ing full coverage rapidly. Utilizing a single robot to
cover a large structure or a wide area suffers from var-
ious drawbacks including time, length, robot energy,
and quality and quantity of information [28,18]. Multi-
robot CPP approaches follow the same approaches of
single robot coverage but additional factors and require-
ments have to be considered. These factors are: coop-
4 Randa Almadhoun1et al.
eration type, information sharing, robustness of agent
failure handling, autonomy level, robot endurance, and
task allocation. This paper surveys different approaches
that utilizes multi-robot systems from the perspective
of the CPP components including the viewpoints gen-
eration and path planning approach. Viewpoint genera-
tion will be surveyed in details in section 3.1. Path plan-
ning approaches can be divided into grid based search
approaches, geometric approaches, reward based, NBV
approaches, and random incremental planners. Each of
these approaches will be surveyed in section 3.2.
3 Multi-Robot CPP
Various aspects need to be considered in order to per-
form CPP using multi-robot systems for model recon-
struction and mapping. The two main CPP related as-
pects in order to generate a feasible coverage path in-
clude viewpoints generation and path planning. The
remaining aspects are communication/task allocation
which is critical for multi-robot systems, and mapping
which is important for modeling the area or structure of
interest using the gathered data. This section provides
a detailed description of each of these aspects and the
employed approaches in each one of them.
3.1 Viewpoints Generation
In majority of the work presented in literature, the
coverage exploration methods are classified into model-
based, and non-model based exploration methods. The
model-based exploration methods depend on a refer-
ence model of the environment or the structure is
provided priori, while the non-model based methods
perform planning and exploration without having a
prior knowledge of the structure or environment [72,
73]. Based on this classification, the viewpoints are gen-
erated to form the search space of the planner.
Some of the viewpoints generation methods are uni-
form due to the dependency on the structure or re-
gion model existence, and the predetermined region or
structure model. Other types of viewpoints generators
are randomized due to the lack of knowledge about the
model of the structure or the regions.
Viewpoints generation is considered critical in
multi-robot CPP process, since it aims to output a
set of optimized paths representing a set of admissi-
ble viewpoints that covers the structure or environment
of interest. Various techniques are used in literature to
perform viewpoints generation based on the used ex-
ploration method and the coverage application.
3.1.1 Model-based
Having a CPP algorithm with uniformity characteris-
tics means that the robotic sensors are deployed to a
predefined coverage pattern [45]. Most of the algorithms
in literature are of a uniform nature especially when
the model information is already available. Having the
model simplifies the process of generating viewpoints
especially in critical areas giving higher weights to some
high priority areas of the model. It also facilitates the
process of connectivity network generation which forms
the search space for the path planners. The main limi-
tation of these approaches is the size and quality of the
existing model or region.
A good example of model-based viewpoints gener-
ation method is a grid-based sensors deployment and
the grids can be of different shapes like: triangular
lattice, square, hexagon, diamond, etc. The work pre-
sented in [57] performed grid based decomposition ac-
cording to each UAV footprint considering sensor size,
focal length, and UAV altitude. The squares of the grid
are divided into residential, empty and energy depots in
directed graph. The authors in [62] proposed a hexagon
pattern as shown in Fig. 1for their cell decomposition.
The generated 2D hexagon cells are used for 2D plan-
ning and the generated paths are transformed into 3D
adding zelevation value based on an elevation map.
The work in [27] performed subsequent Trapezoidal de-
composition (convex polygons) of the 2D environment
until it computed a set of samples (guards) where each
guard can cover one convex polygon. The samples are
used as part of graph computed using a reduced con-
strained delaunay triangulation. A Boustrophedon Cell
Decomposition (BCD) approach is used in [41] to gener-
ate either a weighted graph called Reep graph or divide
regions into passes that are used in unweighted graph.
Furthermore, the work in [56] and [30] performed grid
based decomposition for a 2D area of interest. The de-
composed area is used then in planning the path and
the assignment of the regions to the team of robots.
Similarly in [50], the area is divided into evidence grid
cells based on an elevation map where each cell holds
the value of posterior probability of fire using fire front
contours. The fire front contours are obtained using a
fire segmentation algorithm from several images. The
obtained fire fronts by each vehicle in the team of robots
are used by the central station to update the probabil-
ities of each cell of the grid.
In addition to grid based approaches, the work
in [10] performed 2D decomposition into sweeping rows,
where each row represent an edge in the graph rep-
resentation. A rotated polygon is used to define the
sweeping direction with low turns and rows. The dis-
A Survey on Multi-Robot Coverage Path Planning for Model Reconstruction and Mapping 5
Fig. 1: Hexagonal decomposition steps including (a)
Hull of area, (b) generating a set of hexagons that are
inside area in green or intersecting, and (c) calculat-
ing centroid of every hexagon. (d) shows an example of
non flight-able zones inside possible grids. Courtesy of
IEEE [62]
tance between rows is computed based on the sensor
footprint overlap. A Vehicle Routing Problem (VRP)
is then solved to divide sweeping rows among the set of
UAVs. The authors in [49] divided a 3D complex struc-
ture using horizontal planes translated vertically along
z-axiz to check number of loops and intersection points
as shown in Fig. 2. An offset is added to the struc-
ture to generate the viewpoints at a distance from the
structure. A graph theory approach is used to check the
number of loops then a clustering algorithm is used to
categorize the number of points to each loop that are
then used as viewpoints. if the structure had one slice,
then the viewpoints are categorized based on the refer-
ence yaw difference between UAVs. In the case of mul-
tiple slices, each UAV is assigned to a slice where each
one of them will have the same number of viewpoints.
The work in [45] divided the area of interest into big
voronoi cells based on the used number of agents. Fur-
thermore, the work in [47] used different representation
for the region of interest such as a 1.5Dterrain where
the altitude value for every xalong a single dimension
is returned by a function interpreting the 1.5Dterrain.
This process utilizes fixed altitude paths and the chain
visibility property of 1.5Dterrains to generate visibility
segments that are then used to extract viewpoints for
each agent.
In [7], a generalized Voronoi diagram approach was
utilized in order to perform safe uniform distribution
of many robots in complex areas for the purpose of
coverage. This work performs initially a grid decompo-
sition of the complex area which converges to the com-
puted Voronoi partitioning. The complex area is parti-
tioned using a modified Dijkstra algorithm and grow-
ing functions to provide safe distances between multi-
ple robots. The work presented in [15] proposed a grid
based decomposition approach to divide the area and
generate the viewpoints for multi climbing robots. A
set of voronoi cells were used in the decomposition to
divide a user selected area in order to be explored by
Fig. 2: Concept of plane slicing and intersection points.
Courtesy of [49]
the team of robots. In this work, either agents spread
locally and then interact for coverage globally to sweep
over the area (method 1), or agents assign areas of op-
eration cooperatively then each agent sweep over their
areas (method 2).
Majority of the mentioned papers are following 2D
grid based approaches consisting of cells. However, grid
based representations with its various shapes have lim-
itations in the coverage of boundaries areas and han-
dling partially occluded cells. On the other hand, sweep
based approaches ensure every point has been seen by
at least one robot performing progressive movements
through the environment. These type of approaches
is considered energy and time consuming approaches.
Furthermore, structures and environments of complex
shapes are hard to be decomposed and will generate
a set of slices of different shapes. Distributing these
slices among different robots is difficult in terms of
effort where some robots might have large areas and
others have small ones. Therefore, efforts distribution
is another issue that need to be maintained in multi-
robot systems considering trajectory distance, area to
be covered, time, and energy.
3.1.2 Non-Model based
The locations of the viewpoints in non-model based
CPP algorithms are not predetermined. Additionally,
it is suitable in the cases when there is no information
about the model of the structure or the region. Non-
model based algorithms are preferred in large scale cov-
erage problems where the positions and number of re-
quired viewpoints cannot be predetermined. The main
challenge in non-model based approaches is maximizing
coverage while minimizing the energy consumed by the
team of robots.
6 Randa Almadhoun1et al.
A set of approaches of randomized characteristics
are presented in literature and classified in various
ways. The work presented in [22] proposed a frontier
based exploration algorithm for a team of robots. Ini-
tially the robots are randomly located in an unknown
environment where an initial map is generated by the
first robot and utilized by the other robots to localize
themselves using particle filter algorithm. The approach
is based on identifying frontiers in the initial created oc-
cupancy grid map and performing threshold rank based
frontier allocation.
Additional work with random characteristics is pre-
sented in [24], where two-step procedure that allows
aligning a team of flying vehicles is proposed for the
purpose of surveillance in GPS denied areas. The al-
gorithm starts by generating an elevation map using
modified visual Simultaneous Localization and Map-
ping (SLAM) and meshing approach. Then, the map
used to generate a set of initial candidate positions
(viewpoints) for the team of robots. This set is ran-
domly generated at the beginning, and then it is opti-
mized using cognitive and adaptive methodology. This
methodology aims at maximizing visibility and cover-
age while minimizing the path distance. Furthermore,
Sai and Srikanth in [79] proposed an uncertainty aware
path planning approach for multiple UAVs that collab-
orate between each other for vision based localization.
This approach Involves solving an NBV by taking the
initial poses of the UAVs, then minimizing a weighted
heuristic function to generate better poses at each itera-
tion. This optimization function is solved using Covari-
ance Matrix Adaptation Evolution Strategy (CMA-ES)
algorithm which contains certain degree of randomness.
The heuristic involves different aspects including visi-
bility, span, overlap, baseline, vergence angle, collisions
and occlusions.
These articles presented non-model based ap-
proaches of randomized behavior which consumes time
and energy. Generating the candidate viewpoints at
each step for each robot consumes time and requires co-
operation to avoid repeated coverage. Non-model based
approaches are considered practical in cluttered envi-
ronments where the environment has unknown dynamic
objects like in search and rescue, or disastrous environ-
ments, but model-based approaches are more practi-
cal in applications like inspection, modeling, and fault
traceability. Non-model based approaches require no
prior knowledge and also allow performing online re-
planning in case of a failure of one of the agents. More-
over, performing decentralized non-model based explo-
ration is costly compared to centralized exploration as
interactions across and within robots increases. How-
ever, using decentralized non-model based approach
provides long term benefits of reaching high perfor-
mance.
3.2 Path Planning Approaches
Various multi-robot CPP approaches are surveyed in
literature illustrating the challenging problems and the
proposed solutions. All these approaches have simi-
lar goal which is to provide a collision-free path that
achieves full coverage of the structure or region. These
approaches are classified based on the path planning
methodology into: grid based search, geometric, reward
based, NBV, random incremental planners approaches.
3.2.1 Grid Based Search Approaches
The most important grid based search methods in coop-
erative CPP consist of Cell Decomposition (CD) meth-
ods, and tree based search methods. The cell decom-
position methods are further classified into exact and
approximate cell decomposition.
The work of Rekleitis et al. [67] extended the BCD
approach from single robots to multi-robot systems
for covering unknown arbitrary environments. Based
on the communication restrictions, two types of algo-
rithms were presented including distributed coverage
and team-based coverage. Both types utilize a sensor
based approach and assume an unknown environment.
This work minimizes repeated coverage employing an
auction mechanism which facilitates cooperative behav-
ior among the team of robots. Another work utilizing
BCD approach is presented in [17] which divides the
cells equally among the robots. Each of these cells is
covered by an atomic cycle algorithm. The main objec-
tive of this work is to minimize the path execution time.
Adiyabaater et. al [35] proposed a method of path plan-
ning for coverage that plans the minimal turning path
based on the shape and size of the grid cell. This ap-
proach provides efficient covering order over the cells
based on distance among centroids of cells. It also pro-
vides more optimal coverage path and reduces the rate
of energy consumption and working time.
Moreover, a recent study in [10] proposed a two
stage multi-robot coverage approach similar to the pre-
sented work in [82]. The first stage constructs a com-
plete graph by decomposing the area of interest into
line sweeping rows as shown in Fig. 3where the bor-
ders of the coverage rows represent the vertices of the
graph. The second stage utilizes the generated graph to
divide the sweeps among the team of robots by solving
a VRP. This approach ensures performing coverage in
short time for UAVs in contrast to other approaches as
A Survey on Multi-Robot Coverage Path Planning for Model Reconstruction and Mapping 7
claimed by the authors. Minimizing the number of re-
quired robots for completing the coverage is one of the
main objectives of this work. However, this work could
only be employed in clear areas with no obstacles. An
exact cellular decomposition method is proposed in [34]
to achieve complete coverage utilizing flow networks us-
ing single and multiple mobile robots. In this work, the
free space of the environment is divided into a number
of cells utilizing sweeping lines. These cells are approx-
imated to be either trapezoidal or rectangular which
form the flow network nodes. Then, the minimum cost
path from the generated flow network is computed us-
ing a search algorithm. Each cell in the flow network
is covered using twelve developed templates of back-
and-forth motion. The total coverage time is calculated
considering the duration of covering the cells and the
durations of the moves between the cells taking into
consideration of the order in the flow network.
Fig. 3: The back-and-forth sweep motion performed
along lines perpendicular to the sweep direction for a
rectangular area coverage. The number of turns outside
the area of interest is influenced by the sweep direction
which affects the coverage duration. Courtesy of [10]
An additional classical approach is developed by Fa-
zli et al. [26] which include several algorithms for the
problem of multi-robot repeated area coverage. The
overall proposed approach is divided into stages. The
first stage include generating a set of points called static
guards such that the entire environment of interest is
observed joining all the points. The number of guards
is assumed to be similar to the number of robots. The
second stage involves creating a graph connecting the
guards and the workspace nodes. The third and forth
stages include reducing the size of the graph and cov-
ering the graph using either cyclic coverage or cluster
based coverage. The aim of the cyclic coverage is to gen-
erate the shortest path by allocating a portion of the
generated path which passes through all the guards to
each robot in the team. However, the cluster based ap-
proach divides the graph into clusters according to the
number of robots. The authors presents several algo-
rithms to tackle the problems of each mentioned stage.
A classical grid based approach is proposed also in [11]
where the area is divided into hexagonal cells which are
then clustered. These clusters are allocated to groups of
aerial vehicles based on the battery level and position in
order to acquire data. The cells can be covered by either
centroid, square or lawnmower patterns with specific
configurations to address the constraints of magneto-
metric (geophysical) surveys. The three patterns are il-
lustrated in Fig. 4. In this work, collisions are avoided
by assigning different altitudes for the UAVs. This ap-
proach aims to minimize the survey time and to be
adaptable for different vehicles and resolutions.
Fig. 4: Different patterns of coverage including Lawn-
mower shown in (a), Square pattern shown in (b),
and Hexagon centroid shown in (c). Courtesy of
Springer [11]
Furthermore, the work in [21] proposed a greedy ap-
proach to calculate the path for exploration. The used
graphs for path search are grid graphs. The main aim
of this work is to minimize the total UAV traveled dis-
tance ensuring that every node in the path is visited
exactly once. The same amount of nodes are assigned
to every UAV to explore assuming there is no jumps.
The work presented in [40] proposed a coverage ap-
proach for multi-robot system to deal with the com-
plete, optimal, and communication-less coverage prob-
lem. The problem is defined as Min Max k-Chinese
postman problem where the main goal is to minimize
the maximum coverage cost over a group of robots. This
paper proposed two approximation heuristics for solv-
ing the multi-robot coverage problem. The first pre-
sented solution is an extension of exact CD approach
which is considered as an efficient single robot area cov-
erage algorithm. In this solution, a Eulerian graph (ev-
ery vertex in the graph has even degree) is generated,
then a Eulerian path is computed (a path where each
graph edge is visited once). The second solution is us-
ing a greedy approach (Breadth First Search) by which
the area of interest is divided into equal parts, then a
single robot coverage approach is applied to each part.
Different classical CPP approaches were used in fire
monitoring applications as described in [51,23,63,12].
In [23], a CPP approach is proposed for forest fire
surveillance. The main parts of the proposed approach
of fire monitoring involves fire detection, confirmation
and precise localization with several cooperating UAVs.
8 Randa Almadhoun1et al.
The fire detection and confirmation is achieved by ap-
plying fire segmentation to extract fire contours which
are used then to compute the fire front position. The
overall area to be surveyed by the team of UAVs is di-
vided into convex searching regions. The UAVs perform
back and forward rectilinear sweeps in order to cover
their convex regions. The area decomposition is per-
formed taking into consideration minimizing the num-
ber of sweep turns. These sweep turns consume sig-
nificant amount of time for the UAV stop, rotate and
then accelerate for the next sweep. A similar approach
to [23] was followed in [51] and focused on the percep-
tion system and the contour detection. The work in [63]
proposed a distributed coverage control approach in or-
der to monitor a wildfire and track its development in
open spaces using a team of UAVs. The area is divided
into discrete grid that include information about the fire
spread. In this work, the UAVs team follow the border
region of the wildfire as it keeps expanding while simul-
taneously maintaining the coverage of the whole wild-
fire. This work utilizes potential field control for colli-
sion avoidance, and directing the robot towards of fire
fronts. Another fire monitoring algorithm is proposed
in [12] which proposed Variable Neighborhood Search
(VNS) approach that generate the UAVs trajectories in
order to observe the fire front. This approach depends
on modeling the area of interest as grid of cells holding
information about the fire propagation. The fire front
depends on the rate of spread from one cell to another
and the main propagation direction. The fire spread
rate is calculated using Rothermel’s method. The plan-
ning algorithm of this work distribute the task implic-
itly such that the same locations are not observed by
the team of UAVs concurrently.
A typical approach used in cooperative CPP is the
use of Spanning Trees. A spanning tree approach is pre-
sented in [27] where a Multi-Prim’s algorithm is used to
partition a graph into a forest consisting of partial trees
for the purpose of inspection. The graph is generated
by performing subsequent Trapezoidal decomposition
(convex polygons) of the 2D environment until each
single guard (sample) can cover a corresponding sin-
gle convex polygon. The presented approach performs
graph reduction to reduce the time required by the
robot to traverse the graph. Then, a Constrained Span-
ning Tour (CST) method is used to build cycle on each
partial tree and assign each cycle to a robot. The algo-
rithm minimizes the redundant movements and guar-
antees the coverage completeness. Using this approach,
robustness is also guaranteed by handling failures us-
ing supportive trees. Fig. 5illustrates the described
spanning tree approach.
Fig. 5: (a) A Sample Tree (b) Double-Minimum Span-
ning Tree (DMST) (c) Revised-DMST (d) CST. Cour-
tesy of IEEE [27]
Additional work presented in [84] deals with the
problem of non-uniform traversable terrain coverage
using multi-robot systems. Their algorithm can han-
dle a terrain with locations that does not have a con-
stant traversing time. In order to handle terrains with
non uniform traversability, this algorithm extends the
Multi-robot Forest Coverage (MFC) algorithm. The
work was compared against Multi-robot Spanning Tree
Coverage (MSTC) and it generated paths with shorter
coverage time. The work in [85] proposed a Multi-
objective Genetic Algorithm (GA) with forest individ-
ual containing non-intersecting trees (Mofint) to tackle
the time-limited complete coverage problem. It finds
the least number of robots and allocate tasks to these
robots optimally to finish the mission within a time
limit. The environment is decomposed into regions
(nodes), of specific weights, and edges representing the
borders of two regions, where the weights are used to
indicate the coverage completion time. An arbitrary
spanning tree of the graph is divided by the proposed
work to find the lower and upper bounds of number of
robots. Then a time limited version of the problem is
viewed as a number fixed problem (min-max Balanced
Connected Partition (BCP) problem) which forms with
the estimated bounds a bi-objective optimization prob-
lem (Multi Objective Optimization Problem MOP).
The Mofint algorithm includes two objective functions
where they focus on the number of the spanning trees
(objective 1) and heaviest tree (objective 2). This work
outperformed the MFC in the least number of robots
used and the average completion time.
In a recent work presented in [39], the authors pro-
posed an algorithm that partitions the area of interest
fairly among a team of robots considering their initial
positions. For each robot planning, the algorithm con-
struct a Minimum Spanning Tree (MST) for all the un-
blocked nodes and then Apply the ST to the original
terrain and circumnavigate the robot. Their solution is
considered efficient in terms of computational complex-
ity and it guarantees covering the entire area of interest
without backtracking paths. Although this algorithm
tackles the problem of fair area partitioning and CPP,
the approach accounts only for fixed cell sizes for area
division while not all sub-areas have a uniform geomet-
ric distribution. Another work utilizing spanning tree
A Survey on Multi-Robot Coverage Path Planning for Model Reconstruction and Mapping 9
is proposed in [43] which focuses on time synchronized
coverage control of cooperative mobile robots. Given
a set of Locations To Visit (LTV), the proposed work
generates a Euler graph path tree to be used for cov-
erage work. In order to have a synchronized execution
time of cooperative robot coverage, time-synchronizing
velocity profiles are designed for all LTV pair.
3.2.2 Geometric Based Approaches
Geometric based approaches utilizes visibility graphs in
generating the path. The visibility graph includes a set
of points and obstacles where the nodes represent the
locations, and the edges are line segments that do not
pass though obstacles. Geometric approaches are used
in many areas such as finding shortest euclidean path,
and polygonal area coverage. The most used Geomet-
ric based method in multi-robot CPP is the Voronoi
Diagrams.
In [83], a Dynamic Path-Planning approach for het-
erogeneous Multi-Robot Sensor-Based Coverage (DPP-
MRSBC) is proposed considering energy capacities.
The environment is modeled as Generalized Voronoi Di-
agram (GVD) where the edges of the diagram need to
be covered. The proposed algorithm starts with undi-
rected graph then generates subgraphs that are di-
rected. These graphs are used to generate optimal paths
for each robot which are then turned into arcs with costs
referring to its lengths. Then, directed multi graph is
used to find the shortest path. The Ulusoy’s partition-
ing Algorithm (UA) is modified to add energy demands
(energy, coverage) and to re-plan path utilizing the re-
maining energy capacity. The work in [4] proposed a
control algorithm that forms the sensors to a Voronoi
tessellation while score function of the coverage is in-
creasing. The proposed work utilizes the Voronoi tessel-
lation defining a coverage score function as a measure
of the quality of the surveillance attained by the sensor
network.
Moreover, the work presented in [31] uses multi-
objective optimization to allocate the partitioned ar-
eas to the robots whilst optimizing robot team’s objec-
tives. The area surface is partitioned following Voronoi
partitioning approach. The main optimization objec-
tives include minimizing the overall completion time
and achieving complete coverage in addition to opti-
mizing the manipulator joint’s torque. Fotios et. al [13]
proposed a coverage approach based on partitioning a
coastal region for a team of heterogeneous UAVs. The
proposed approach combines the strategies of graph
search and computational geometry algorithms which
partition the area of interest regardless of the number
of UAVs or their relative capabilities and consider the
sensing radius and Field of View (FOV). Initially, the
area is partitioned using a growing regions approach
which perform isotropic portioning based on the UAVs’
initial locations and relative capabilities. Then, the area
and holes are defined by a forced edge constraints per-
forming a Constrained Delaunay Triangulation. The de-
scribed coverage approach is Antagonizing Wavefront
Propagation (AWP) which is a step transition algo-
rithm that involves an isotropic cost attribution. This
algorithm starts from the initial position of each UAV,
propagating towards the other UAVs or the borders of
the area.
In [69], a coverage algorithm is proposed to effi-
ciently generate a flight plan to completely cover a given
area of interest (i.e. area with complex contours or must
not be overflown). The first step of the proposed ap-
proach include the computation of the area of interest
percentage that should be covered by each robot con-
sidering the sensor footprint of each UAV. The number
of cells of the grid required for each sub-area is de-
termined using the computed percentage values. The
sub-areas are further extended using flood-fill like algo-
rithm. The second part involves performing path plan-
ning which prioritizes the selection of long straight seg-
ments (stride) to reduce the number of turns. In order
to select strides, a set of heuristics are proposed in-
cluding the selection of the neighbor cells with a higher
value, the ones located in the contour of the area, and
the ones that lead to a longer stride in the next step.
Another type of geometric approach is the work pre-
sented in [37] which proposed a partitioning scheme
that depends on creating a power diagram to cover non
convex areas. The power diagram takes into account
the different sensing capabilities available in each agent
in addition to the visibility domain of the agents. A
distributed gradient control scheme is also proposed to
lead a set of heterogeneous robots to cover the specified
area.
3.2.3 Reward-based Approaches
The reward based approaches are employed in several
recent work in multi-robot CPP due to their advantages
such as nonlinear mapping, learning ability, and parallel
processing. The most important used methods include
Neural Network (NN), Nature Inspired methods and
hybrid algorithms.
Several bio-inspired multi-robot coverage ap-
proaches have been developed based on the behaviors
that exist in nature. Ranjbar-Sahraei et al. [66] pre-
sented an example where the robots avoid entering the
boundaries of each other utilizing the behavior of the
ants. The robots in this work move in a 2D environment
10 Randa Almadhoun1et al.
in a circular motion where they deposit pheromones on
their areas’ borders mimicking the ants behavior in or-
der to reduce the intersections of borders. That is, if a
pheromone is detected by a robot, it changes its motion
direction and hence avoids another robot’s border. As a
result, this approach facilitates spreading out the robots
in the environments gradually. This approach is further
extended in the paper into two extensions. The first ex-
tension include increasing the radius of the robot’s cir-
cular motion if pheromone detection likelihood is small
and vice versa. The second extension facilitates the be-
havior change when an intruder is detected by decreas-
ing the robots territory areas. The authors in [38] pro-
posed an approach that minimizes the completion time
and number of turns of a team of robots while trying
to achieve complete coverage. The proposed work fol-
lows a grid-like decomposition approach that is based
on disks. The experimental setup consists of areas of
convex rectilinear polygons. A pattern based genetic
algorithm is used to achieve complete coverage by uti-
lizing eight neighbor-disk prioritization patterns used
to represent rectilinear moves. The results of using two
patterns is illustrated in Fig. 6. An optimized genetic
algorithm combined with total coverage method is pro-
posed in [20] for generating 3D map of large areas using
multiple UAVs. The area of interest in this work is di-
vided among the team using flood fill algorithm and
game theory, then the proposed path generation algo-
rithm is applied on each sub area. In [86], a genetic
algorithm is used to plan a path that detect the en-
tire area. Initially, this is done for one drone, then the
search space is divided among the UAVs according to
their position and endurance and multi-objective Inte-
ger programming model.
The work in [58] proposed an optimal coverage path
planning approach and coverage action controller that
performs an active selection of goals. Two cost functions
are proposed in order to allow robots to avoid obstacles,
and to find the optimal paths to the specified goals.
The method used to select the best goal considers the
coverage paths and introduce the notion of safe goal.
3.2.4 NBV Approaches
NBV approaches are usually used when no information
about the model exists priori. NBV approaches scale
better to complex real-world. Some of the NBV ap-
proaches in literature are of probabilistic characteristics
which provide estimation not fixed value.
One of the recent NBV CPP approaches is described
in [48] where an exploration algorithm is proposed to
collect water samples by building spatially varying phe-
nomenon over a specified region without prior knowl-
Fig. 6: The sample chromosome of a solution including
number of robots: 3 and number of allowed patterns: 2
is shown in (a). The chromosome decoding is shown in
(b) where three shapes ( diamond, circle, and square )
are presented and have two representations including (i)
the robot initial location if numbered 1, and (ii) shortest
path’s starting and ending points related to repeated
coverage motion if numbered 2 or more. Courtesy of
Springer [38]
edge about the spatial field. The built spatial varying
fields provide emphasis on locations that are good for
sampling. The candidate locations for the explorer are
generated using two techniques including fixed-window,
and Contour-based location selection methods as illus-
trated in Fig. 7. It includes two related subproblems
including exploration algorithm to generate the phe-
nomenon map, which concurrently facilitates collecting
actual physical samples using the sampling algorithm.
The exploration is performed using Gaussian Process
frontier-based approach by which it measures variables
to suggest sample utility. A look-back selective tech-
nique is used for sampling where the new candidates
are appended to a list which is used later by the robot
to look back for non-eligible candidates within the spec-
ified time period. The main aim of the proposed work
is to generate a good spatial phenomenon model and to
compute a path optimized for distance and time. An-
other approach is presented in [53] which builds prob-
abilistic decision maps to compute exploration paths
using a grid that represents the sum of expected scores
to pick and place static and dynamic objects. The main
A Survey on Multi-Robot Coverage Path Planning for Model Reconstruction and Mapping 11
aim of this work is to minimize the combined search and
action time with targets found in an environment us-
ing finite horizon plans. The Probability Density Maps
(PDM) is built using reward prediction function which
corresponds to the probability of finding new objects or
targets. The PDM is updated through a Bayesian fil-
tering procedure applied on the grid of the arena area
where each cell represents a state.
Fig. 7: The colormap shows the variance in the spatial
representation of the field. The potential candidate lo-
cations are represented by red circles. The contours are
shown in black lines.Two location selection approaches
are shown including Contour-based (a) and Fixed-
window locations selection. Courtesy of IEEE [48]
Another work following NBV approach is pre-
sented [54] which aims at increasing information about
uncertain areas while performing coverage. The en-
vironment is modeled using Voronoi partitioning ap-
proach where the Voronoi regions are assigned to vehi-
cles using a locational optimization approach. The pro-
posed distribution density function is formulated based
on some unknown targets’ positions which can be de-
tected using the appropriate sensors. The vehicles are
divided into two teams including service and search ve-
hicles. The search vehicles find the targets in order to
allow the service vehicles perform coverage efficiently.
The search vehicles find their path that maximizes the
gathered information individually using a look-ahead
dynamic programming algorithm. The objective of ser-
vice vehicles is to optimally cover the terrain by spread-
ing out over the environment of interest. Similarly, the
work in [65] proposed a two layered exploration ap-
proach including coarse exploration layer performed by
UGV, and fine mapping layer performed by UAVs in
GPS denied environment. The UGV starts exploring
the area of interest using 3D laser to create a rough 2.5D
volumetric map using SLAM, which is then used by the
UAV to update occupancy data of the gaps using 2D
tilting laser. A frontier coverage planning approach is
followed to generate a set of viewpoints which are then
used to compute the path using a Fixed Start Open
Traveling Salesman Problem (FSOTSP). The cost func-
tion used in FSOTSP is formulated using traveled dis-
tance and information gain. A volumetric motion plan-
ning interface is developed to support the navigation
of both UGV and UAV applying Batch Informed Trees
(BIT).
Vera et. al [52] proposed a variation of VRP with an
insertion heuristic and a negotiation mechanism for en-
ergy resources, and a heuristic for the continuous mon-
itoring problem with inter-depot routes and priorities
(CMPIDP) that uses all available information and is
fast enough to react to dynamic environmental changes.
3.2.5 Incremental Random Planners
The main two important sampling methods include
Rapidly exploring Random Trees (RRT) and Proba-
bilistic Road Map (PRM). The most used sampling
based method is the RRT method with all its varia-
tions. The work presented in [14] proposed a receding
horizon planning approach to compute an optimized ex-
ploration path. The method utilizes an occupancy map
to represent the environment where a finite iteration
random tree (RRT*) is grown in the free space part
of the map. The best branch of the random tree is se-
lected based on the amount of the unmapped space by
which the first edge is executed at every planning step
iteratively to complete the exploration.
The work in [79] described previously in section
3.1.2, utilized random planners as the second stage of
the planning framework which performs belief space
planning. This is performed for the vehicles individu-
ally using the Rapidly Exploring Random Belief Tree
(RRBT) algorithm. This method generates paths that
ensures having high image quality and improved vehi-
cle confidence by optimizing the path cost and reducing
the localization uncertainty.
This type of methods have limitations related to dy-
namic environments especially in multi-robot systems.
If one of the robots’ paths is blocked by an obstacle
or one of the robots failed, then a re-planning process
need to be preformed for the entire team which is of
high computational cost.
3.3 Communication and Task Allocation
Communication is a challenging aspect in multi-robot
systems. The common assumption in literature is that
communication is unlimited in range and bandwidth.
Nearly all centralized systems assume that the individ-
ual robots can communicate directly with the central
controller such as [70,45], and algorithms that create
12 Randa Almadhoun1et al.
maps assume global communication [55,58]. Since com-
munication systems are often down in the aftermath of
a disaster [52,62], achieving coverage with limited com-
munication must be a critical aspect that must to be
considered.
The actions of the robots in a decentralized system
are based on the collected information about the other
visible robots. These collected information include the
actions and locations of other visible robots, with re-
spect to their local coordinate system. Most robots
can provide some local communication means such as
wifi [24,48] or line-of-sight (LOS) [54,24,15] methods,
which can then be used to direct the team of robots and
avoid overlapping actions. Line-Of-Sight (LOS) com-
munication as described in [54] is based on categoriza-
tion scheme by which the team is divided into search
vehicles and service vehicles. The search vehicles maxi-
mize the amount of gathered information while the ser-
vice vehicles spread in the environment within a spec-
ified range to cover the entire area. The work in [15]
defines critical points for allocating cells based on sev-
eral factors, one of them is the LOS. Teams that are
involved in GPS denied or uncertain areas (disasters)
utilize these kind of communication technique.
Decentralized multi-robot systems are considered
fault tolerant, and scalable as the system is of dis-
tributed architecture and modular. Distributed mod-
ular systems allow detecting faults and replacing af-
fected robots without having global control which in-
creases the system robustness such as the work in [36].
Furthermore, using decentralized systems reduce the
network overhead introduced by information exchange
between robots. Several decentralized multi-robot sys-
tems have been proposed which shares information with
the nearest neighbor such as [63,53]. The work pre-
sented in [58] performed distributed Voronoi partition-
ing where each robot uses a relative configuration of
other robots within its sensor range. The distributed
partitioning is performed by incrementing pseudo sen-
sor range in steps while the range constrained Voronoi
cell gets contracted to obtain the desired region. The
work in [70] proposed a hybrid decentralized multi-
agent path finding framework. It combines Reinforce-
ment Learning (RL) for planning single agent paths and
Imitation learning from centralized path planner. In
this work, the agents select movements that will benefit
the whole team using the learned decentralized policy
where the agents perform implicit coordination during
the online path planning without having explicit com-
munication among the team.
Another decentralized approach for information
sharing between a team of UAVs is proposed in [42]
where the UAVs performs two types of map updates at
each time step: uncoordinated map update, and coordi-
nated map merging. The first type generates an uncoor-
dinated occupancy probability using local information
only, while the second update combines the local infor-
mation with the other UAVs robots in order to com-
pute the actual probability for the map. Three types of
coordinated map merging approaches are proposed in-
cluding belief update, average, and modified occupancy
grid map merging. The work in [36] proposed a one-to-
one decentralized coordination where each pair of UAVs
share their own area to survey. Taking into considera-
tion all the one-to-one coordinations between neighbor-
ing UAVs distributes the whole team efficiently. The
work in [65] performed decentralized exploration using
UGV and UAV platforms where each one of them up-
dates their map and navigation on their own then share
it with the other vehicle.
3.4 Mapping
Different kinds of applications require detailed 3D mod-
els and 2D maps such as: terrain surveillance, indoor
mapping, inspection, fire monitoring, and urban plan-
ning. In general, a model reconstruction and an envi-
ronment mapping process involves four steps including:
viewpoint planning, scanning, registering, and integra-
tion [73]. Robot localization is considered an impor-
tant requirement to perform accurate reconstruction
and mapping. It is performed usually as part of the
most used approach in mapping which is the Simulta-
neous Localization and Mapping (SLAM). Some work
in literature represents the workspace reconstruction as
an environment mapping application and others rep-
resents it as a model reconstruction of a structure of
interest.
Most of the reviewed work in literature performed
area type of coverage for various kinds of environments.
This type of environments include terrain, arena, in-
door environment, and forests. For example, The work
presented in [38,82,83,27] utilized topological maps in
the planning process. For providing the covered map,
some presented work performed monocular Structure
from Motion (SfM) and stereo Iterative Closest Point
(ICP) as in [49] generating 2D map. Other work per-
formed 3D reconstruction using Real Time Appear-
ance Based Mapping (RTABMap) [1] generating 3D
textured meshes and octomap [32] generating 3D oc-
cupancy maps as in [74,61]. Moreover, some of the re-
viewed work especially the ones performed in natural
environment overlay the path on the google map as il-
lustrated in [10,41]. Additionally, some work performs
meshing as to reconstruct 3D mesh structures utilizing
techniques like TSDF or triangulations as in [79,75].
A Survey on Multi-Robot Coverage Path Planning for Model Reconstruction and Mapping 13
Another aspect important in mapping and informa-
tion gathering is related whether the information is pro-
cessed online while the planning is performed or offline
after completing the coverage mission. Most of the work
presented follows an offline process as described in [11,
40,52]. Based on the type of processed information, per-
forming the mapping online as in [55,35,58] facilitates
monitoring and adapting to new changes but at the
same time is computational expensive and could affect
the planning process.
4 Discussion and Future Research Directions
Modeling and mapping environmental terrains and
complex structures is considered an essential process
for increasing the level of autonomy in wide range of
application domains. Generating an efficient coverage
route is a critical requirement in order to gather accu-
rate information through the viewpoints encapsulated
in the route in order to perform modeling and map-
ping. Utilizing a team of robots minimizes the effort
required by each member to achieve the coverage task.
The team of deployed robots in CPP could include var-
ious homogeneous and heterogeneous robotic platforms
which facilitates the coverage and mapping tasks.
Based on the work surveyed in the literature, vari-
ous research components have to be taken into consid-
eration while developing multi-robot CPP approaches,
such as: type of the environment, the dimensionality of
the CPP problem, the number of sensors and agents
type, and the coordination and communication tech-
nique. The CPP problem is divided into two processes,
viewpoints generation and path planning. The sequence
of using these processes is affected by the existence of
the environment (model-based/non-model based) and
the type of prior available information (2D map, 3D
mesh, 3D point cloud). Based on the dimensionality
of the problem, various types of information could be
used during mapping and reconstruction, and different
mapping techniques are utilized to generate the final
map or model. Various evaluation metrics were consid-
ered in literature but the main common metrics include
coverage completeness, path length, execution time, ro-
bustness, and energy consumption. The different parts
of performing CPP using multi-robot system are sum-
marized in Figure 8.
Most of the presented CPP approaches in litera-
ture implement model-based viewpoints generation al-
gorithms such as [48,60,15,57,27]. The majority of
these papers are generating grids consisting of cells, and
are applied on 2D type of environment. However, grid
based representations have limitations in handling par-
tially occluded cells or cover areas close to the bound-
aries in continuous spaces.
Another commonly used approach for multi-robot
CPP in 2D type of environments is sweep coverage as
described in [23,34,10]. In sweep coverage, the robots
move progressively through the environment and ensure
every point has been seen by at least one robot. While
sweep coverage requires only a single pass, it is also used
for repeated coverage as explained in [26,38,67]. This
type of coverage keeps the team small, hence easier to
deploy, but still provides complete coverage, as required
for search and rescue. Another common assumption is
that robots have a map to direct their movement. Usu-
ally, the overall region is divided to minimize the num-
ber of sweep turns which consumes significant amount
of time for the UAVs to stop, rotate and then accelerate
for the following sweep. Moreover, performing Cyclic
Coverage as described in [26] may not be an appropriate
approach compared with the Cluster-based algorithms
as in [11] especially in situations where the target area
include regions with different coverage priorities or the
robots have different speeds.
Heuristic based algorithms combine randomness
and heuristics to drive the exploration process as in [24,
60,22]. A good ratio between performance and cost
could be provided by these type of methods especially
that they do not consume much computational re-
sources and do not require expensive sensors. However,
parts of the area of interest may remain unexplored.
Most of the methods surveyed in this article either fol-
lows a grid based search approach or reward based ap-
proach, hence both dominates the other described ap-
proaches. Probabilistic and spanning tree (grid based
search) approaches are considered critical in perform-
ing coverage especially that it generates a continuous
path. Different spanning trees implementations in lit-
erature differ regarding aspects such as computational
complexity and quality of the generated model. Some of
the methods in literature performed trajectory planning
to generate a continuous smooth path for the robots as
in [82,24,54]. It is considered a good property for con-
tinuous surveillance and modeling operations.
Achieving coverage completeness is the main aim
of applying CPP. The work discussing multi-robot cov-
erage of 2D regions [13,38,82] dominates the research
on coverage of 3D structure [79,49]. Working with sim-
ple 2D regions allows achieving coverage completeness
faster. Some of the grid based search and reward based
approaches achieves the coverage completeness by par-
titioning the region of interest and allocating the re-
gions among the team members which is hard to be
done in 3D.
14 Randa Almadhoun1et al.
Fig. 8: The main components of the multi-robot CPP
Several assumptions were made in the different sur-
veyed work in literature, especially those targeting
multi-robot systems. Some previous work assumes that
the robots can create maps and merge them when they
regroup, but this is actually difficult to achieve in prac-
tice. Another common assumption is that communi-
cation is unlimited in range and bandwidth. Nearly
all centralized systems assume that the central con-
troller can communicate directly with all the individual
robots, and algorithms that create maps assume global
communication. The focus need to be on achieving com-
plete coverage with limited communication since com-
munication systems are often down in the aftermath of
a disaster [52,69,62].
The surveyed articles evaluated the performance of
their proposed work based on several metrics. The most
important metric which is common between all the pa-
pers is the path execution time, especially that it is
correlated to the path length and the energy level of
the robots [34,83,39]. The different surveyed CPP al-
gorithms showed that NBV and non-model based meth-
ods take much longer time than using other model-
A Survey on Multi-Robot Coverage Path Planning for Model Reconstruction and Mapping 15
based or classical approaches. The sensing technique is
considered another important aspect affecting the path
execution duration which is either performed discretely
where the data is gathered at each viewpoint or contin-
uously along the path. Discrete sensing is used in most
of the work presented in literature, since it allows the
robot to perform scanning without missing observations
by allowing it to stabilize at each viewpoint [54,48].
The main components of the CPP process used in
the recent surveyed work in this article are summa-
rized in Table 1. Table 1summarizes the recent sur-
veyed research on the multi-robot CPP. In this table,
each paper is summarized in a row, where the columns
list the type of environment, the algorithm processing
technique, viewpoints generation method, the coverage
path generation method, and the evaluations metrics of
the coverage method.
Based on the surveyed literature, there are numer-
ous challenging hindering the progress of an efficient
multi-robot cooperative CPP. These challenges could
be classified as follows:
1. Heterogeneity: heterogeneity in a coverage sce-
nario can be defined in different aspects, such as
different movement or sensing capabilities of the
robots or different platforms. Heterogeneity is crit-
ical in multi-robot systems to enhance the visibil-
ity in applications that require complete coverage
and accurate data. Using heterogeneous sensors pro-
vides different types of data that could be utilized
in providing a map of different information. One of
the recent used sensors in wide applications is the
event based camera which provides events instead
of frames where events are generated at local in-
tensity changes as independent asynchronous times-
tamped spikes. These kind of sensors can operate in
different light intensities, and provide the property
of low latency, and power consumption. An exten-
sive review of all the algorithms, applications, and
datasets is provided in [29]. Moreover, using hetero-
geneous platforms facilitates dividing complex tasks
in harsh environments and enhance flexibility and
mobility. For example, inspecting a bridge could be
performed using a collaborative system consisting of
AUV, UGV and UAV platforms providing data from
underwater and on the ground. Most of the work
presented in the literature utilize homogeneous type
of robots and sensors such as the work presented
in [83,30,69,42] which limits the coverage dimen-
sionality. Few papers utilized heterogeneous type of
robots and sensors as in [23,51,47,65].
2. Prioritization: In some applications, parts of the
target area should be visited or covered sooner than
others due to different priorities. Prioritization is of
great significance in large areas and big structures
especially for time critical tasks where it facilitates
the detection of fire, and danger. Some of the work
presented in this review utilized the priority in an
area coverage application as in [38] where the work
provided a priority index to the robots, selected pat-
terns in the grid. In another work presented in [31],
a prioritization is performed to the objectives of the
optimization function based on the allocated parti-
tioned areas. The work in [83] also prioritized the
robots to avoid planning conflicts.
3. Robustness: Robustness is another critical part in
multi-robot systems since it is related to handling
robot failure. There are different robustness criteria
that need to be considered in the real world, such
as message loss, robot action failure, and commu-
nication failure. Different robust techniques exists
in literature to detect robot failures and reallocate
tasks between the remaining robots including the
use of more precise GPS system like DGPS, and the
use of active sensing techniques to update the cover-
age path in real time. Robustness is considered one
of the challenging problems that need to be main-
tained in multi-robot systems to allow the rest of
the team adapt to the new changes that occur to
the system online. A lot of the work in literature
implemented robustness in various ways as in [40],
and [27] where robot failure is handled in different
ways.
4. Communication Modality: Most of the robotic
systems have limited range of communication. For
example, robots transmit messages to other robots
within a specific distance from it. Also, based on
the cooperation method used whether it’s a cen-
tralized, distributed or decentralized, the team of
robots need to maintain communication especially
if the team shares information. Most of the reviewed
work assumes perfect communication and utilizes
centralized type of cooperation as presented in [6,
22,70,45] which is subject to scalability, overhead
and single point of failure problems. Some of the
work presented a decentralized CPP approach for
multi-robot systems as presented in [54,53,11,70].
The type of cooperation and communication need
to be defined for the team of robots based on the
application and type of information that need to be
shared and processed.
5. Adaptability: One of the main properties that a
team of robots need to have is the ability to change
behavior over time and react to changes in the en-
vironment in order to prevent unnecessary degra-
dation in performance or improve the performance.
Dynamic environment characterized by the presence
16 Randa Almadhoun1et al.
Table 1: Review of multi-robot CPP approaches
Paper
Year
Application
Type of Environment Algorithm Processing Viewpoints Generation CPP Approach Evaluation
[38]
2012
area coverage -2D area coverage - offline
- centralized
-model-based
-The area of interest is modeled with disks
representing the range of sensing devices
distributed among robots considering their travel times
-reward based
-pattern based genetic algorithm
-Eight premeditated neighbor
disk prioritization patterns
-compared to hierarchical oriented genetic algorithm
-completion time
-probability of the mutation and crossover,
-generation
-layout settings
[34]
2013
cleaning areas
-2D area coverage -offline
-centralized -model-based
-BCD
-grid based search
-A robot covers each cell with
one of 12 templates consisting of
several back and forth motions.
-time efficiency
-coverage completeness
-number of turns
-robustness (changing the obstacles locations)
-compared with the single robot
[82]
2014
aerial operations:
search and rescue,
mapping,
and surveillance
-2D arbitrary environment -online
-centralized
-model-based
-extending BCD allowing certain cells
to be divided into half cells,
ensuring each cell
will not be covered twice
-grid based search
-Chinese postman problem to compute an
Eulerian circuit traversing through the cells
-concatenates per cell seed spreader motion
patterns into a complete coverage path
-analysis of completeness
-analysis of efficiency (coverage pattern,
traversal ordering)
-coverage optimality
-time
-distance
-number of turns vs orientation
of the direction
[39]
2017
terrain coverage
-2D arbitrary environment -offline
-centralized
-model-based
-divides the terrain into a number of equal areas
each corresponding to a specific robot
-discretize areas into finite set of equal cells
-grid based search
-generate MST for all the
unblocked nodes then Apply the ST
to the original terrain and circumnavigate
the robot around the area
-comparison with MFC
and Optimized MSTC algorithm
-maximum and min coverage time
-path length
-idealized coverage time [Ideal Max],
this value is simply calculated by
dividing the number of unoccupied
cells with the number of robots
[79]
2018
generating improved
maps for localization
-3D model
-partial known information
-online
-NBV centralized and
Rapidly Random Belief
Tree acts individually
on the vehicles
-non-model based (randomized)
-NBV and
random planner approach
-minimization problem for a set of vehicles over a space
of poses solved by CMA-ES algorithm
-heuristic include: Visibility, Span, Overlap, Basline,
Vergence angle, Collisions and occlusion:
-the map density
-mean squared error (MSE)
of point cloud matching
-localization accuracy
vs the generated maps of NBV
-distance
-Comparison of the paths
with covariance ellipsoids
[63]
2017
monitor wild fire
-2D area coverage
-online
-decentralized
(distributed manner)
-model-based
-a discrete grid-based
that include fire spread info
-grid based search
-potential field control
-The UAVs follow the border region
of the wildfire as it keeps expanding
, while maintaining coverage
of the entire wildfire area
-spreading of fire
[11]
2018
geophysical surveys
-2D area coverage -offline
-centralized
-model-based
-segments the environment into hexagonal cells
and allocates groups of robots to different
clusters of non-obstructed cells to acquire data.
-allocate them to robots based on the battery and location
-grid based search
-Cells can be covered by lawnmower,
square or centroid patterns
with specific configurations to address the
constraints of magneto-metric surveys
-Distance between parallel
coverage lines
-Number of robots
-cells quantityt vs length of the path
-Coverage angle optimization
-Hexagon size and battery
-compared to Voronoi cellular
decomposition (time analysis)
-Comparative sensing analysis.
(reconstruction quality)
[49]
2016
cooperative inspection
of complex structures
-3D structure -offline
-model-based
-horizontal planes translated vertically
along z axiz to check number
of loops and intersection points.
-using graph theory to check number of loops
-clustering to categorize the number
of points to each loop and use them
as waypoints
-geometric based
-branches (loops) are assigned by
dividing the loop according to num
of agents and yaw difference
-execution time
-yaw changes
-3D meshes and pointcloud
[62]
2016
search and rescue
missions (disaster
scenarios)
-2D area
-offline(CPP)
-online in the
recovery relocation)
-centralized
(user need to define
region of interest)
-model-based
-cell decomposition based on hexagons
-grid based search
-The lawnmower path angle
is modeled as graph based problem
-Subdivision of cells among agents
using K-means (clustering)
-TSP is used to generate path that connect
each cluster internal cells centroids separately
(minimum Euler path)
-The lawnmower pattern is used as the
basic coverage pattern for each hexagonal cell
-the final 3D route is generated adding z
(sum of minimum specified altitude
and current position in the elevation map)
-Survey time
-Recovery scheme
of fulfilling tasks
of damaged robots
[83]
2014
sensor based coverage
in narrow spaces
-2D area
(known
and Partially unknown
environment)
-online
-centralized for
prioritizing the robots
-model-based
-modeled the environment as GVD
-edges of voronoi diagram need
to be covered
-reward based
-The problem is capacitated ARP
which is solved by (UA)
-The approach modifies UA :
-to add energy demands (energy, coverage)
-to re-plan path utilizing the
remaining energy capacity
-analysis of tour length
per the number of agents
-tour length vs
energy consumption
-CPU time
[10]
2015
area coverage for
digital terrain map
and vegetation indexes
-2D area -offline
-model-based
-area is decomposed into sweeping rows
where each row represent an edge
in graph representation
-rotated polygon is used to define
the sweeping direction with low
turns and rows
-distance between rows is
based on the image footprint overlap
-reward based
-formulated as min max optimization problem
to minimize the maximum mission time
(minimize the number of UAVs required for coverage)
- VRP is used to formulate
the routing approach where the UAVs are
vehicles and row extreme points are the customers
-different constraints are added related
to setup time, individual UAV time,
time duration (battery), and visiting nodes once.
-battery duration and the mission time
-constraints effect on the resulting path
-number of rows in case of using
different number of UAVs
[27]
2010
area coverage
or border inspection
-2D area
(static obstacles) -offline
-model-based
-subsequent Trapezoidal decomposition (convex polygons)
of the 2D environment until one guard can
cover one convex polygon
-compute graph representation which is a reduced
constrained delaunay triangulation
-grid based search
-partitions the graph using Multi-Prim’s algorithm
into a forest consisting of partial trees .
-performs graph reduction
-Then using CST method, it build cycle on
each partial tree and assign each cycle to a robot.
-analyzed robustness
-overall computation complexity
-completeness
-worst case running time
[12]
2018
fire monitoring
and measurements
-2D area -online
-centralized
-model-based
-grid of cells holding information
about the fire propagation
-reward based
-VNS approach plans the trajectories
of the UAVs to observe the fire front
-fire front depends on the main propagation
direction and the rate of spread from
one cell to another calculated using
Rothermel’s method
-flight duration,
-number of UAVs to deploy
-take-off time
[47]
2018
surveillance and mapping
( persistent monitoring
of terrains)
-1.5D area -offline
-centralized
-model-based
-Visibility polygon and visibility region
corresponding to a point x on the terrain
-A terrain can be interpreted as a function
that returns an altitude value for every x.
-reward based
-VRP for planning
-the UAV must repeat a certain tour in
the environment.
-The cost function used for the ground
robot is asymmetric and dependent on
the slope of the terrain on
-computational time
[22]
2015
office like environment
symmetric hall coverage
-2D area -online
-distributed -non-model based (randomized)
-
-reward based
-frontier based and performing
rank based allocation
-compared to greedy, nearest
and rank based
-execution time
-explored area percentage
[53]
2018
MBZIRC
search, pick and place
-2D area
-area with moving objects -online
-decentralized
-randomized
-graph-like grid environment
with edge-connected cells
-NBV
-using PDM to plan exploratory paths
on a grid representing the sum of
expected scores to be found
-minimize the combined search and action time
with targets found in an environment
using finite-horizon plan
-reward prediction
-PDM changes
[48]
2018
sampling of water
for off-site analysis
-2D area -online
-model-based - consider locations on the outer-most
contour between a region with high
variance and a region with low variance
-or consider all the locations on a
fixed planning window centered on the
current position of the robot
-NBV
-an explorer that measures variables to suggest
sample utility (GP frontier-based exploration)
-a sampler that collects physical samples
(secretary hiring problem is used for the sampler),
-mean error in the interpolated data
as a function of distance traveled
-compare the GP-frontier based explorer
to two other exploration techniques:
global maximum variance search,
and lawnmower coverage
-sampling score
-compare the sampler with
submodular secretary algorithm
[60]
2018
coastal areas
coverage
-2D area
-online
-centralized (initial
partitioning process is
executed on the
ground station and
the cell decomposition
and coverage planning
are computed
on-board each UAV)
-model-based
-two steps:
1-a growing regions algorithm performs an
isotropic partitioning of the area based
on the initial locations of the UAVs
and their relative capabilities
2-then CDT is computed based on the
largest FoV among the availableUAVs
-reward based
-(AWP) isotropic cost attribution function
by a step transition algorithm, starting
from the initial position of each UAV,
propagating towards the other UAVs
or the borders of the area
-FOV projection size
-complexity
-average divergence vs number of robots
-altitude
-UAV capability vs area
[24]
2012
surveillance coverage
missions over a terrain
of arbitrary morphology
-3D environment -online
-centralized -non-model based (randomized)
-reward based
-two main steps
that can be expressed as follows:
1- The part of the terrain that is visible
2- The team members are arranged so that
for every point in the terrain the
closest robot is as close as possible to that point.
-convergence,
-scalability
-applicability to non-convex
3D environments
-cost function
-coverage %
of dynamic obstacles as in [83,53,68] or by the pos- sibility of change in shape or size as in [44,48,50,
A Survey on Multi-Robot Coverage Path Planning for Model Reconstruction and Mapping 17
19] should be reacted to accordingly in a manner
that still achieves the desired CPP performance.
This aspect is significant in applications that in-
volves changes in the natural environment such as
fire fighting more than a static structure.
6. Open Systems: The ability of the multi-robot sys-
tem to adapt to a new joining robot describes the
openness feature of the system. In most CPP ap-
plications, this aspect was not mentioned although
it is an important property in the multi-robot sys-
tems especially if complex large environments and
structures need coverage. Adding a new robot to the
team will support and enhance the efficiency of the
team.
7. Collective Intelligence: generating collaborative
policy for the team of robots to achieve one goal
could enhance the performance of the CPP. The
policy could be generated by training the agents on
a set of actions or types of data which will accel-
erate the CPP process. Collaborative manipulation
and multi-agent path planning are another two sub-
problems where machine learning could be used. Re-
cent work utilized different types of machine learn-
ing such as Reinforcement learning are presented
in [9,33,44,5,71,80].
5 Conclusion
In this paper, we surveyed the various literature work
related to the Coverage Path Planning (CPP) using
multi-robot systems. The major components of CPP
were identified and discussed. The viewpoints genera-
tion is classified based on the used exploration method
into model-based and non-model based approaches.
The model-based approach have deterministic charac-
teristics, by which it utilizes the reference model of
the structure or region of interest provided priori. This
knowledge is used in the coverage path planning to gen-
erate a path that encapsulates a set of viewpoints that
provide the maximum coverage for reconstruction and
mapping. The non-model based approach is considered
of randomized characteristics since it does not use pri-
ori knowledge about the structure or region and evalu-
ates a set of candidate views based on the coverage and
information gain. The non-model based approach per-
forms a cycle of online iterative steps including: candi-
date viewpoints generation, viewpoints evaluation and
path planning, path execution and scanning, and per-
forming reconstruction. The CPP algorithms are fur-
ther classified based on the path planning approach into
grid based search methods, reward based methods, ge-
ometric methods, NBV methods, and incremental ran-
dom planners. In general, performing CPP using multi-
robot systems requires defining the viewpoints genera-
tion approach, path planning method, communication
and task allocation method, and mapping technique.
The main aim of using multi-robot systems in CPP ap-
plications is to reduce the coverage effort and distribute
it among the team of robots in order to provide full high
quality reconstructions and maps with the minimum
energy and time.
Based on the surveyed papers, the main future ap-
plication domains of multi-robot CPP include search
and rescue, inspection, and agriculture. The main fu-
ture research trends in this topic include heterogeneity,
prioritization and robustness techniques, communica-
tion modality, system openness and adaptability, and
collective intelligence.
6 Conflict of Interest
On behalf of all authors, the corresponding author
states that there is no conflict of interest.
Acknowledgements This publication is based upon work
supported by the Khalifa University of Science and Technol-
ogy under Award No. RC1-2018-KUCARS
References
1. Real-time appearance-based mapping (2015). URL http:
//introlab.github.io/rtabmap/
2. Drones and robots are taking over industrial inspection
(2017). URL https://www.technologyreview.com/
s/608811/drones-and- robots-are- taking-over-
industrial-inspection/
3. Search and rescue robots current applica-
tions on land, sea, and air (2017). URL
https://www.techemergence.com/search-and- rescue-
robots-current- applications/
4. Adaldo, A., Mansouri, S.S., Kanellakis, C., Dimarogo-
nas, D.V., Johansson, K.H., Nikolakopoulos, G.: Coop-
erative coverage for surveillance of 3D structures. IEEE
International Conference on Intelligent Robots and Sys-
tems 2017-Septe, 1838–1845 (2017). DOI 10.1109/
IROS.2017.8205999
5. Adepegba, A., Miah, S., Spinello, D.: Multi-Agent Area
Coverage Control Using Reinforcement Learning. Florida
Artificial Intelligence Research Society Conference pp.
368–373 (2016)
6. Ahmad, W.: Distributed Navigation of Multi-Robot Sys-
tems for Sensing Coverage (2016)
7. Alitappeh, R.J., Pimenta, L.C.A.: Distributed safe de-
ployment of networked robots. In: N.Y. Chong, Y.J. Cho
(eds.) Distributed Autonomous Robotic Systems, pp. 65–
77. Springer Japan, Tokyo (2016)
8. Almadhoun, R., Taha, T., Seneviratne, L., Dias, J., Cai,
G.: A survey on inspecting structures using robotic sys-
tems. International Journal of Advanced Robotic Sys-
tems 13(6), 1729881416663664 (2016). DOI 10.1177/
1729881416663664. URL https://doi.org/10.1177/
1729881416663664
18 Randa Almadhoun1et al.
9. Andrew, W., Greatwood, C., Burghardt, T.: Deep Learn-
ing for Exploration and Recovery of Uncharted and
Dynamic Targets from UAV-like Vision pp. 1124–1131
(2018)
10. Avellar, G.S.C., Pereira, G.A.S., Pimenta, L.C.A., Iscold,
P.: Multi-UAV routing for area coverage and remote sens-
ing with minimum time. Sensors (Switzerland) 15(11),
27783–27803 (2015). DOI 10.3390/s151127783
11. Azp´urua, H., Freitas, G.M., Macharet, D.G., Cam-
pos, M.F.: Multi-robot coverage path planning us-
ing hexagonal segmentation for geophysical surveys.
Robotica 36(8), 1144–1166 (2018). DOI 10.1017/
S0263574718000292
12. Bailon-Ruiz, R., Bit-Monnot, A., Lacroix, S.: Planning
to monitor wildfires with a fleet of UAVs. Iros pp. 4729–
4734 (2018)
13. Balampanis, F., Maza, I., Ollero, A.: Area partition for
coastal regions with multiple uas. Journal of Intelligent
& Robotic Systems 88(2), 751–766 (2017). DOI 10.1007/
s10846-017-0559-9
14. Bircher, A., Kamel, M., Alexis, K., Oleynikova, H., Sieg-
wart, R.: Receding horizon ”next-best-view” planner for
3d exploration. In: 2016 IEEE International Conference
on Robotics and Automation (ICRA), pp. 1462–1468.
IEEE (2016)
15. Breitenmoser, A., Tˆache, F., Caprari, G., Siegwart, R.,
Moser, R.: MagneBike - Toward multi climbing robots
for power plant inspection. Proc. of 9th Int. Conf. on
Autonomous Agents and Multia- gent Systems (AAMAS
2010) (Aamas), 1713 – 1722 (2010)
16. Cabreira, T.M., Franco, C.D., Ferreira, P.R., But-
tazzo, G.C.: Energy-Aware Spiral Coverage Path Plan-
ning for UAV Photogrammetric Applications. IEEE
Robotics and Automation Letters 3(4), 3662–3668
(2018). DOI 10.1109/LRA.2018.2854967. URL https:
//ieeexplore.ieee.org/document/8411478/
17. Chan, S.K., New, A.P., Rekleitis, I.: Distributed cov-
erage with multi-robot system. Proceedings - IEEE
International Conference on Robotics and Automa-
tion 2006(May), 2423–2429 (2006). DOI 10.1109/
ROBOT.2006.1642065
18. Choset, H.: Coverage for robotics - A survey of recent
results. Annals of Mathematics and Artificial Intelligence
31(1-4), 113–126 (2001). DOI 10.1023/A:1016639210559
19. Cruz, H., Eckert, M., Meneses, J., Mart´ınez, J.F.: Effi-
cient forest fire detection index for application in Un-
manned Aerial Systems (UASs). Sensors (Switzerland)
16(6) (2016). DOI 10.3390/s16060893
20. Darrah, M., Trujillo, M.M., Speransky, K., Wathen, M.:
Optimized 3d mapping of a large area with structures us-
ing multiple multirotors. In: 2017 International Confer-
ence on Unmanned Aircraft Systems (ICUAS), pp. 716–
722 (2017). DOI 10.1109/ICUAS.2017.7991414
21. Dedousis, D., Kalogeraki, V.: Complete Coverage Path
Planning for arbitrary number of Unmanned Aerial
Vehicles. Proceedings of the 9th ACM International
Conference on PErvasive Technologies Related to As-
sistive Environments - PETRA ’16 pp. 1–4 (2016).
DOI 10.1145/2910674.2910719. URL http://dl.acm.org/
citation.cfm?doid=2910674.2910719
22. Dileep Muddu, R.S., Wu, D., Wu, L.: A frontier based
multi-robot approach for coverage of unknown environ-
ments. 2015 IEEE International Conference on Robotics
and Biomimetics, IEEE-ROBIO 2015 pp. 72–77 (2015).
DOI 10.1109/ROBIO.2015.7414626
23. Mart´ınez-de Dios, J.R., Merino, L., Ollero, A., Ribeiro,
L.M., Viegas, X.: Multi-UAV experiments: Application
to forest fires. Springer Tracts in Advanced Robotics 37,
207–228 (2007). DOI 10.1007/978-3-540-73958-6 8
24. Doitsidis, L., Weiss, S., Renzaglia, A., Achtelik, M.W.,
Kosmatopoulos, E., Siegwart, R., Scaramuzza, D.: Opti-
mal surveillance coverage for teams of micro aerial vehi-
cles in GPS-denied environments using onboard vision.
Autonomous Robots 33(1-2), 173–188 (2012). DOI
10.1007/s10514-012-9292-1
25. Ellefsen, K.O., Lepikson, H.A., Albiez, J.C.: Multiob-
jective coverage path planning: Enabling automated in-
spection of complex, real-world structures. Applied Soft
Computing Journal 61, 264–282 (2017). DOI 10.1016/
j.asoc.2017.07.051. URL http://dx.doi.org/10.1016/
j.asoc.2017.07.051
26. Fazli, P., Davoodi, A., MacKworth, A.K.: Multi-robot re-
peated area coverage. Autonomous Robots 34(4), 251–
276 (2013). DOI 10.1007/s10514-012-9319-7
27. Fazli, P., Davoodi, A., Pasquier, P., Mackworth, A.K.:
Complete and robust cooperative robot area coverage
with limited range. IEEE/RSJ 2010 International Con-
ference on Intelligent Robots and Systems, IROS 2010
- Conference Proceedings pp. 5577–5582 (2010). DOI
10.1109/IROS.2010.5651321
28. Galceran, E., Carreras, M.: A survey on coverage
path planning for robotics. Robotics and Autonomous
Systems 61(12), 1258–1276 (2013). DOI 10.1016/
j.robot.2013.09.004
29. Gallego, G., Delbruck, T., Orchard, G., Bartolozzi, C.,
Taba, B., Censi, A., Leutenegger, S., Davison, A., Con-
radt, J., Daniilidis, K., Scaramuzza, D.: Event-based Vi-
sion: A Survey pp. 1–25 (2019). URL http://arxiv.org/
abs/1904.08405
30. Gautam, A., Murthy, J.K., Kumar, G., Ram, S.P., Jha,
B., Mohan, S.: Cluster, Allocate, Cover: An Efficient Ap-
proach for Multi-robot Coverage. Proceedings - 2015
IEEE International Conference on Systems, Man, and
Cybernetics, SMC 2015 pp. 197–203 (2016). DOI
10.1109/SMC.2015.47
31. Hassan, M., Liu, D.: Simultaneous area partitioning and
allocation for complete coverage by multiple autonomous
industrial robots. Autonomous Robots 41(8), 1609–1628
(2017). DOI 10.1007/s10514-017-9631-3
32. Hornung, A., Wurm, K.M., Bennewitz, M., Stachniss,
C., Burgard, W.: OctoMap: An efficient probabilistic
3D mapping framework based on octrees. Autonomous
Robots 34(3), 189–206 (2013). DOI 10.1007/s10514-012-
9321-0
33. Hsu, S.h., Chan, S.h., Wu, P.t., Xiao, K., Fu, L.c.: Dis-
tributed Deep Reinforcement Learning based Indoor Vi-
sual Navigation pp. 2532–2537 (2018)
34. Janchiv, A., Batsaikhan, D., Kim, B., Lee, W.G., Lee,
S.G.: Time-efficient and complete coverage path plan-
ning based on flow networks for multi-robots. Inter-
national Journal of Control, Automation and Systems
11(2), 369–376 (2013). DOI 10.1007/s12555-011-0184-5.
URL https://doi.org/10.1007/s12555-011- 0184-5
35. Janchiv, A., Batsaikhan, D., hwan Kim, G., Lee, S.G.:
Complete coverage path planning for multi-robots based
on. Control, Automation and Systems (ICCAS), 2011
11th International Conference on pp. 824–827 (2011)
36. Joaquin, J., Bego˜na, A., Maza, I., Ollero, A.: Cooperative
Large Area Surveillance with a Team of Aerial Mobile
Robots for Long Endurance Missions pp. 329–345 (2013).
DOI 10.1007/s10846-012-9716-3
37. Kantaros, Y., Thanou, M., Tzes, A.: Distributed
coverage control for concave areas by a heteroge-
neous Robot-Swarm with visibility sensing constraints.
A Survey on Multi-Robot Coverage Path Planning for Model Reconstruction and Mapping 19
Automatica 53, 195–207 (2015). DOI 10.1016/
j.automatica.2014.12.034
38. Kapanoglu, M., Alikalfa, M., Ozkan, M., Yazıcı, A., Par-
laktuna, O.: A pattern-based genetic algorithm for multi-
robot coverage path planning minimizing completion
time. Journal of Intelligent Manufacturing 23(4), 1035–
1045 (2012). DOI 10.1007/s10845-010-0404-5. URL
https://doi.org/10.1007/s10845-010- 0404-5
39. Kapoutsis, A.C., Chatzichristofis, S.A., Kosmatopoulos,
E.B.: DARP: Divide Areas Algorithm for Optimal Multi-
Robot Coverage Path Planning. Journal of Intelligent
and Robotic Systems: Theory and Applications 86(3-4),
663–680 (2017). DOI 10.1007/s10846-016-0461-x
40. Karapetyan, N., Benson, K., McKinney, C., Taslakian, P.,
Rekleitis, I.: Efficient multi-robot coverage of a known en-
vironment. IEEE International Conference on Intelligent
Robots and Systems 2017-Septe, 1846–1852 (2017).
DOI 10.1109/IROS.2017.8206000
41. Karapetyan, N., Moulton, J., Lewis, J.S., Li, A.Q., Kane,
J.M.O., Rekleitis, I.: Multi-robot Dubins Coverage with
Autonomous Surface Vehicles. IEEE International Con-
ference on Robotics and Automation (ICRA’18), Bris-
bane, Australia pp. 2373–2379 (2018)
42. Khan, A., Yanmaz, E., Rinner, B.: Information Merging
in Multi-UAV Cooperative Search. 2014 IEEE Interna-
tional Conference on Robotics and Automation (ICRA)
pp. 3122–3129 (2014). DOI 10.1109/ICRA.2014.6907308
43. Kim, T.S., Lee, Y.H., Park, J.K., Kuc, T.Y., Ji, S.H.,
Lee, S.M., Moon, Y.S.: A time synchronized multi-
robot coverage algorithm for unstructured environ-
ment. International Conference on Control, Automa-
tion and Systems (Iccas), 503–508 (2014). DOI 10.1109/
ICCAS.2014.6988042
44. Lihua, Z., Miao, W., Hangyu, G.U.O., Tao, Y.: Multi-
platform Fire Control Strike Track Planning Method
Based on Deep Enhance Learning. IEEE computer soci-
ety pp. 747–750 (2017)
45. Luo, W., Member, S., Sycara, K.: Adaptive Sampling and
Online Learning in Multi-Robot Sensor Coverage with
Mixture of Gaussian Processes. IEEE International Con-
ference on Robotics and Automation (ICRA’18), Bris-
bane, Australia pp. 6359–6364 (2018)
46. M Debusk, W.: Unmanned aerial vehicle systems
for disaster relief: Tornado alley pp. In AIAA In-
fotech@ Aerospace Conference, AIAA–2010–3506, At-
lanta (2010). DOI 10.2514/6.2009- 2045
47. Maini, P., Yu, K., Sujit, P.B., Tokekar, P.: Persistent
Monitoring With Refueling On A Terrain Using A Team
Of Aerial And Ground Robots pp. 8493–8498 (2018)
48. Manjanna, S., Li, A.Q., Smith, R.N., Rekleitis, I., Dudek,
G.: Heterogeneous Multirobot System for Exploration
and Strategic Water Sampling. IEEE International Con-
ference on Robotics and Automation (ICRA’18), Bris-
bane, Australia pp. 4873–4880 (2018)
49. Mansouri, S.S., Kanellakis, C., Wuthier, D., Fresk, E.,
Nikolakopoulos, G.: Cooperative Coverage Path Planning
for Visual Inspection of Complex Infrastructures (2016)
50. Merino, L., Caballero, F., Martinez-de Dios, J.R., Maza,
I., Ollero, A.: An Unmanned Aerial System for Auto-
matic Forest Fire Monitoring and Measurement. Journal
of Intelligent and Robotic Systems In Press,, – (2011)
51. Merino, Luis Fernando Caballero, J. R. Mart´ınez-de Dios,
Joaqu´ın Ferruz, Ollero, A.: A Cooperative Perception
System for Multiple UAVs: Application to Automatic De-
tection of Forest Fires. J. Field Robotics 23, 245–267
(2006). DOI 10.1002/rob
52. Mersheeva, V., Friedrich, G.: Multi-UAV Monitoring
with Priorities and Limited Energy Resources. In-
ternational Conference on Automated Planning and
Scheduling -ICAPS15 (Layer 1), 347–355 (2015). URL
http://www.aaai.org/ocs/index.php/ICAPS/ICAPS15/
paper/view/10460
53. Miki, T., Popovi, M., Gawel, A., Hitz, G., Siegwart, R.:
Multi-agent Time-based Decision-making for the Search
and Action Problem. IEEE International Conference
on Robotics and Automation (ICRA’18), Brisbane, Aus-
tralia pp. 2365–2372 (2018)
54. Mirzaei, M., Sharifi, F., Gordon, B.W., Rabbath, C.A.,
Zhang, Y.M.: Cooperative multi-vehicle search and cov-
erage problem in uncertain environments. Proceedings of
the IEEE Conference on Decision and Control pp. 4140–
4145 (2011). DOI 10.1109/CDC.2011.6161448
55. Mirzaei, M., Sharifi, F., Gordon, B.W., Rabbath, C.A.,
Zhang, Y.M.: Cooperative multi-vehicle search and cov-
erage problem in uncertain environments. Proceed-
ings of the IEEE Conference on Decision and Con-
trol (August 2014), 4140–4145 (2011). DOI 10.1109/
CDC.2011.6161448
56. Modares, J., Ghanei, F., Mastronarde, N., Dantu, K.:
UB-ANC planner: Energy efficient coverage path plan-
ning with multiple drones. Proceedings - IEEE Interna-
tional Conference on Robotics and Automation pp. 6182–
6189 (2017). DOI 10.1109/ICRA.2017.7989732
57. Nedjati, A., Izbirak, G., Vizvari, B., Arkat, J.: Com-
plete Coverage Path Planning for a Multi-UAV Response
System in Post-Earthquake Assessment. Robotics 5(4),
26 (2016). DOI 10.3390/robotics5040026. URL http:
//www.mdpi.com/2218-6581/5/4/26
58. Palacios-Gasos, J.M., Talebpour, Z., Montijano, E.,
Sagues, C., Martinoli, A.: Optimal path planning and
coverage control for multi-robot persistent coverage in
environments with obstacles. Proceedings - IEEE In-
ternational Conference on Robotics and Automation pp.
1321–1327 (2017). DOI 10.1109/ICRA.2017.7989156
59. Palazzolo, E., Stachniss, C.: Effective Exploration for
MAVs Based on the Expected Information Gain. Drones
2(1), 9 (2018). DOI 10.3390/drones2010009. URL http:
//www.mdpi.com/2504-446X/2/1/9
60. Panerati, J., Gianoli, L., Pinciroli, C., Shabah, A., Nico-
lescu, G., Beltrame, G.: From Swarms to Stars : Task
Coverage in Robot Swarms with Connectivity Con-
straints. IEEE International Conference on Robotics and
Automation (ICRA’18), Brisbane, Australia pp. 7674–
7681 (2018)
61. Papachristos, C., Alexis, K., Carrillo, L.R.G., Tzes,
A.: Distributed infrastructure inspection path planning
for aerial robotics subject to time constraints. 2016
International Conference on Unmanned Aircraft Sys-
tems, ICUAS 2016 pp. 406–412 (2016). DOI 10.1109/
ICUAS.2016.7502523
62. Perez-Imaz, H.I.A., Rezeck, P.A.F., MacHaret, D.G.,
Campos, M.F.M.: Multi-robot 3D coverage path plan-
ning for First Responders teams. IEEE Interna-
tional Conference on Automation Science and Engineer-
ing 2016-Novem, 1374–1379 (2016). DOI 10.1109/
COASE.2016.7743569
63. Pham, H.X., La, H.M., Feil-Seifer, D., Deans, M.: A
distributed control framework for a team of unmanned
aerial vehicles for dynamic wildfire tracking. IEEE In-
ternational Conference on Intelligent Robots and Sys-
tems 2017-Septe, 6648–6653 (2017). DOI 10.1109/
IROS.2017.8206579
20 Randa Almadhoun1et al.
64. Phung, M.D., Quach, C.H., Dinh, T.H., Ha, Q.: En-
hanced discrete particle swarm optimization path plan-
ning for UAV vision-based surface inspection. Automa-
tion in Construction 81(April), 25–33 (2017). DOI
10.1016/j.autcon.2017.04.013. URL http://dx.doi.org/
10.1016/j.autcon.2017.04.013
65. Qin, H., Meng, Z., Meng, W., Chen, X., Sun, H.,
Lin, F., Ang, M.H.: Autonomous exploration and map-
ping system using heterogeneous uavs and ugvs in gps-
denied environments. IEEE Transactions on Vehicu-
lar Technology 68(2), 1339–1350 (2019). DOI 10.1109/
TVT.2018.2890416
66. Ranjbar-Sahraei, B., Weiss, G., Nakisaee, A.: A multi-
robot coverage approach based on stigmergic communi-
cation. In: I.J. Timm, C. Guttmann (eds.) Multiagent
System Technologies, pp. 126–138. Springer Berlin Hei-
delberg, Berlin, Heidelberg (2012)
67. Rekleitis, I., New, A.P., Rankin, E.S., Choset, H.: Effi-
cient boustrophedon multi-robot coverage: An algorith-
mic approach. Annals of Mathematics and Artificial In-
telligence 52(2-4), 109–142 (2008). DOI 10.1007/s10472-
009-9120-2
68. Sampedro, C., Bavle, H., Rodriguez-ramos, A., Puente,
P.D., Campoy, P.: Laser-Based Reactive Navigation for
Multirotor Aerial Robots using Deep Reinforcement
Learning pp. 1024–1031 (2018)
69. Santamaria, E., Segor, F., Tchouchenkov, I.: Rapid aerial
mapping with multiple heterogeneous unmanned vehi-
cles. ISCRAM 2013 Conference Proceedings - 10th In-
ternational Conference on Information Systems for Crisis
Response and Management 6(May), 592–596 (2013)
70. Sartoretti, G., Kerr, J., Shi, Y., Wagner, G.,
K. Satish Kumar, T., Koenig, S., Choset, H.: Pri-
mal: Pathfinding via reinforcement and imitation
multi-agent learning (2018)
71. Sartoretti, G., Paivine, W., Choset, H., Wu, Y.,
Satish Kumar, T.K., Koenig, S.: Distributed Rein-
forcement Learning for Multi-Robot Decentralized Col-
lective Construction (August) (2018). DOI 10.1103/
PhysRevLett.101.056802
72. Scott, W.R.: Model-based view planning. Machine Vi-
sion and Applications 20(1), 47–69 (2009). DOI 10.1007/
s00138-007-0110-2
73. Scott, W.R., Roth, G., Rivest, J.F.: View planning for au-
tomated three-dimensional object reconstruction and in-
spection. ACM Computing Surveys 35(1), 64–96 (2003).
DOI 10.1145/641865.641868
74. Song, S., Jo, S.: Online inspection path planning for au-
tonomous 3D modeling using a micro-aerial vehicle. Pro-
ceedings - IEEE International Conference on Robotics
and Automation pp. 6217–6224 (2017). DOI 10.1109/
ICRA.2017.7989737
75. Song, S., Jo, S.: Surface-based Exploration for Au-
tonomous 3D Modeling. IEEE International Conference
on Robotics and Automation (ICRA’18), Brisbane, Aus-
tralia pp. 4319–4326 (2018)
76. Tan, Y., Marchal, K., Van de Peer, Y.: Handbook of
research on design, control, and modeling of swarm
robotics. Handbook of research on design, control and
modeling of swarm robotics i, 68–82 (2016). DOI
10.4018/978-1-4666-9572-6
77. Valente, J., Barrientos, A., Cerro, J.D.: Coverage Path
Planning to Survey Large Outdoor Areas with Aerial
Robots: A Comprehensive Analysis. Introduction to
Modern Robotics (2011)
78. Vehicles, A.: Survey on Coverage Path Planning with
Unmanned Aerial Vehicles (2019). DOI 10.3390/
drones3010004
79. Vemprala, S., Saripalli, S.: Vision based Collaborative
Path Planning for Micro Aerial Vehicles pp. 3889–3895
(2018)
80. Wang, C., Wang, J., Shen, Y., Zhang, X.: Autonomous
navigation of uavs in large-scale complex environments:
A deep reinforcement learning approach. IEEE Transac-
tions on Vehicular Technology 68(3), 2124–2136 (2019).
DOI 10.1109/TVT.2018.2890773
81. Wei, M., Isler, V.: Coverage Path Planning under the
Energy Constraint. IEEE International Conference on
Robotics and Automation (ICRA’18), Brisbane, Aus-
tralia (2018)
82. Xu, A., Viriyasuthee, C., Rekleitis, I.: Efficient complete
coverage of a known arbitrary environment with appli-
cations to aerial operations. Autonomous Robots 36(4),
365–381 (2014). DOI 10.1007/s10514-013-9364-x. URL
https://doi.org/10.1007/s10514-013- 9364-x
83. Yazici, A., Kirlik, G., Parlaktuna, O., Sipahioglu, A.: A
dynamic path planning approach for multirobot sensor-
based coverage considering energy constraints. IEEE
Transactions on Cybernetics 44(3), 305–314 (2014). DOI
10.1109/TCYB.2013.2253605
84. Zheng, X., Koenig, S., Kempe, D., Jain, S.: Multirobot
forest coverage for weighted and unweighted terrain.
IEEE Transactions on Robotics 26(6), 1018–1031 (2010).
DOI 10.1109/TRO.2010.2072271
85. Zhou, X., Wang, H., Ding, B.: How many robots are
enough : A multi-objective genetic algorithm for the
single-objective time-limited complete coverage problem.
IEEE International Conference on Robotics and Au-
tomation (ICRA’18), Brisbane, Australia pp. 2380–2387
(2018)
86. Zkxw, O.M., Fq, H.G.X., Sodqqlqj, W., Hiihfwlyh, L.V.,
Hiilflhqw, D.Q.G., Vhjphqwdwlrq, D.: Multi-UAV Coop-
erative Coverage Path Planning in Plateau and Mounain
Environment. pp. 820–824 (2018)
... Most of these applications focus on obtaining a complete 2D or 3D coverage maps of areas, terrains or structures utilizing the well known process of Coverage Path Planning (CPP) [1]. These applications include area surveillance, industrial inspection, cultural heritage sites digitization, and forest fire monitoring [2]. CPP is the process of generating an optimized path that provides complete coverage of the structure or environment of interest, in order to gather highly accurate information. ...
... ing on the environment dimensionality, and the structure complexity [2]. Most of these approaches utilizes single robot or multi-robot systems for wide 2D areas coverage utilizing classical approaches [3]. ...
... CPP is usually performed to autonomously map and reconstruct models of the area/structure being covered. Depending on the application and the sensor suite, the generated maps or models can hold various levels of information relevant to the application, such as semantics [23], heat maps, occupancy, or signals strength [1], [2]. ...
Article
Coverage Path Planning (CPP) is an essential capability for autonomous robots operating in various critical applications such as fire fighting, and inspection. Performing autonomous coverage using a single robot system consumes time and energy. In particular, 3D large structures might contain some complex and occluded areas that shall be scanned rapidly in certain application domains. In this paper, a new Hybrid Coverage Path Planning (HCPP) approach is proposed to explore and cover unknown 3D large structures using a decentralized multi-robot system. The HCPP approach combines a guided Next Best View (NBV) approach with a developed Long Short Term Memory (LSTM) waypoint prediction approach to decrease the CPP exploration time at each iteration and simultaneously achieve high coverage. The hybrid approach is the new ML paradigm which fosters intelligence by balancing between data efficiency and generality allowing the exchange of some CPP parts with a learned model. The HCPP uses a stateful LSTM network architecture which is trained based on collected paths that cover different 3D structures to predict the next viewpoint. This architecture captures the dynamic dependencies of adjacent viewpoints in the long term sequences like the coverage paths. The HCPP switches between these methods triggered by either the number of iterations or an entropy threshold. In the decentralized multi-robot system, the proposed HCPP is embedded in each robot where each one of them shares its global 3D map ensuring robustness. The results performed in a realistic Gazebo robotic simulator confirmed the advantage of the proposed HCPP approach by achieving high coverage on different 3D unknown structures in a shorter time compared to conventional NBV.
... To formally define the coverage problem a grid discretisation of a space is generally first defined with a certain resolution. Grids can be of different shapes including cells with triangular, square, hexagonal or diamonds shapes [16]. Different definitions of coverage of the grid exist, including dynamic coverage and static coverage [17]. ...
... Algorithms using a single-agent may be optimal if sufficient time is available or if agents are expensive (e.g., planetary exploration missions). However, navigating and exploring a vast area using a single robot suffers from drawbacks, including long turnaround time, depletion of robot power resources, and poor quality and quantity of information collected [16]. In particular, single-agent systems may not be an ideal choice for surveillance or disaster response when mission time is critical. ...
... Performance degrades when the number of agents increases or when communication bandwidth is limited or intermittent. Thus the multi-agent coverage task is widely considered in a distributed fashion [16,[30][31][32]. Operating a distributed robotics framework considerably reduces the load of communication and enhances the resilience of system connectivity -important factors for large-scale networks [33]. ...
Preprint
In mobile robotics, area exploration and coverage are critical capabilities. In most of the available research, a common assumption is global, long-range communication and centralised cooperation. This paper proposes a novel swarm-based coverage control algorithm that relaxes these assumptions. The algorithm combines two elements: swarm rules and frontier search algorithms. Inspired by natural systems in which large numbers of simple agents (e.g., schooling fish, flocking birds, swarming insects) perform complicated collective behaviors, the first element uses three simple rules to maintain a swarm formation in a distributed manner. The second element provides means to select promising regions to explore (and cover) using the minimization of a cost function involving the agent's relative position to the frontier cells and the frontier's size. We tested our approach's performance on both heterogeneous and homogeneous groups of mobile robots in different environments. We measure both coverage performance and swarm formation statistics that permit the group to maintain communication. Through a series of comparison experiments, we demonstrate the proposed strategy has superior performance over recently presented map coverage methodologies and the conventional artificial potential field based on a percentage of cell-coverage, turnaround, and safe paths while maintaining a formation that permits short-range communication.
... However, researchers also aimed at other properties like semantic information, which could be extracted from the UAV data (Popovic et al., 2017;Stache et al., 2021;Valente et al., 2013). Apart from many other applications like precision agriculture (Basiri et al., 2022;Just et al., 2020), indoor exploration and reconstruction (Shen et al., 2012;Sun et al., 2021;Zhu et al., 2018) to name a few, we focus on viewpoint and path planning for 3D reconstruction of complex outdoor man-made objects like buildings and bridges which is a topic of interest for many applications (Almadhoun et al., 2019). A closely related topic is coverage path planning (Bircher et al., 2015;Galceran and Carreras, 2013;Kaba et al., 2017;Peng and Isler, 2020;Shang et al., 2020;Tan et al., 2021) which deals with optimizing the trajectories that cover a known 2D or 3D environment. ...
Preprint
Full-text available
Unmanned aerial vehicles (UAVs) are widely used platforms to carry data capturing sensors for various applications. The reason for this success can be found in many aspects: the high maneuverability of the UAVs, the capability of performing autonomous data acquisition, flying at different heights, and the possibility to reach almost any vantage point. The selection of appropriate viewpoints and planning the optimum trajectories of UAVs is an emerging topic that aims at increasing the automation, efficiency and reliability of the data capturing process to achieve a dataset with desired quality. On the other hand, 3D reconstruction using the data captured by UAVs is also attracting attention in research and industry. This review paper investigates a wide range of model-free and model-based algorithms for viewpoint and path planning for 3D reconstruction of large-scale objects. The analyzed approaches are limited to those that employ a single-UAV as a data capturing platform for outdoor 3D reconstruction purposes. In addition to discussing the evaluation strategies, this paper also highlights the innovations and limitations of the investigated approaches. It concludes with a critical analysis of the existing challenges and future research perspectives.
... CPP algorithms are often developed for the reconstruction of the environment and creation of the digital models. Categorization of approaches into modelbased and non-model-based is presented in [10]. Model-based approaches use the reference model of the area or structure of interest to generate a path that encapsulates a set of viewpoints providing the maximum coverage for reconstruction and mapping. ...
Preprint
Full-text available
This paper presents a new trajectory planning algorithm for 3D autonomous UAV volume coverage and visual inspection. The algorithm is an extension of a state-of-the-art Heat Equation Driven Area Coverage (HEDAC) multi-agent area coverage algorithm for 3D domains. With a given target exploration density field, the algorithm designs a potential field and directs UAVs to the regions of higher potential, i.e., higher values of remaining density. Collisions between the agents and agents with domain boundaries are prevented by implementing the distance field and correcting the agent's directional vector when the distance threshold is reached. A unit cube test case is considered to evaluate this trajectory planning strategy for volume coverage. For visual inspection applications, the algorithm is supplemented with camera direction control. A field containing the nearest distance from any point in the domain to the structure surface is designed. The gradient of this field is calculated to obtain the camera orientation throughout the trajectory. Three different test cases of varying complexities are considered to validate the proposed method for visual inspection. The simplest scenario is a synthetic portal-like structure inspected using three UAVs. The other two inspection scenarios are based on realistic structures where UAVs are commonly utilized: a wind turbine and a bridge. When deployed to a wind turbine inspection, two simulated UAVs traversing smooth spiral trajectories have successfully explored the entire turbine structure while cameras are directed to the curved surfaces of the turbine's blades. In the bridge test case an efficacious visual inspection of a complex structure is demonstrated by employing a single UAV and five UAVs. The proposed methodology is successful, flexible and applicable in real-world UAV inspection tasks.
... Cabreira et al. [30] Galceran and Carreras [20] Almandhoun et al. [31] Chen et al. [32] Our work Table 1 compares the present work to already existing surveys of CPP methods for robotics or UAVs. The present paper is focused not only on surveying the CPP methods for UAVs, but also on: (a) examining all the decomposition methods, (b) reviewing the multi-robot strategies, (c) the multi-UAV's and standalone UAV's CPP methods, (d) UAVs' Sensors 2022, 22, 1235 3 of 20 energy-saving CPP algorithms, and (e) the comparison of the energy-saving CPP methods. ...
Article
Full-text available
The coverage path planning (CPP) algorithms aim to cover the total area of interest with minimum overlapping. The goal of the CPP algorithms is to minimize the total covering path and execution time. Significant research has been done in robotics, particularly for multi-unmanned unmanned aerial vehicles (UAVs) cooperation and energy efficiency in CPP problems. This paper presents a review of the early-stage CPP methods in the robotics field. Furthermore, we discuss multi-UAV CPP strategies and focus on energy-saving CPP algorithms. Likewise, we aim to present a comparison between energy efficient CPP algorithms and directions for future research.
... A few CPP articles that feature approximate cellular decomposition techniques are: an approach for image mosaicing in precision agriculture with irregular-shaped areas [26], an optimal CPP algorithm where the area is represented by points of interest [27], another approach for precision agriculture that uses a team of heterogeneous quadcopters [28], a metaheuristic algorithm named Harmony Search (HS) [29], online non-uniform CPP using different grid sizes [30,31], high-resolution aerial sensing with multiple heterogeneous UAVs for nonconvex areas [32], pheromone-based methods [33], and a combination of digital pheromones and evolutionary strategies for coordination of multiple UAVs [34], among others. ...
Article
Full-text available
Coverage path planning (CPP) is a field of study which objective is to find a path that covers every point of a certain area of interest. Recently, the use of Unmanned Aerial Vehicles (UAVs) has become more proficient in various applications such as surveillance, terrain coverage, mapping, natural disaster tracking, transport, and others. The aim of this paper is to design efficient coverage path planning collision-avoidance capable algorithms for single or multi UAV systems in cluttered urban environments. Two algorithms are developed and explored: one of them plans paths to cover a target zone delimited by a given perimeter with predefined coverage height and bandwidth, using a boustrophedon flight pattern, while the other proposed algorithm follows a set of predefined viewpoints , calculating a smooth path that ensures that the UAVs pass over the objectives. Both algorithms have been developed for a scalable number of UAVs, which fly in a triangular deformable leader-follower formation with the leader at its front. In the case of an even number of UAVs, there is no leader at the front of the formation and a virtual leader is used to plan the paths of the followers. The presented algorithms also have collision avoidance capabilities, powered by the Fast Marching Square algorithm. These algorithms are tested in various simulated urban and cluttered environments, and they prove capable of providing safe and smooth paths for the UAV formation in urban environments.
Conference Paper
Full-text available
Conference Paper
View Video Presentation: https://doi.org/10.2514/6.2022-1414.vid Space exploration refers to constructing a map with the aid of sensor data. This exploration is achieved utilizing a group of robots in an obstacle cluttered environment and distributing tasks amongst these robot(s). The robotic configuration is equipped with sensors to acquire data from the surroundings and to ensure collision-free motion. This paper presents a framework for the design of a Hybrid Stochastic Optimizer (HSO) for multi-robot space exploration. The proposed algorithm augments deterministic Coordinated Multi-Robot Exploration (CME) and stochastic Arithmetic Optimization (AO) techniques for maximizing the utility. The framework initially utilizes deterministic CME to ascertain the cost and utility values of adjacent cells around robot(s). The overall solution accuracy is then improved utilizing the Arithmetic Optimization algorithm. The proposed utilization of hybrid is interpreted that the algorithm starts with deterministic technique and continues off with stochastic method until the required improved solution with the desired accuracy is achieved. The effectiveness of the proposed Hybrid Stochastic Optimizer is then ascertained by training the multi-robotic framework in different complexity environmental conditions. The results efficacy is demonstrated by comparing the results of the proposed methodology with those achieved from conventional CME methods. Results indicate that the proposed HSO algorithm greatly enhanced the space exploration process as it gives enhanced area coverage apart from reduced search time.
Article
Full-text available
In this work, we tackle the problem of multi-robot convex workspace division. We present an algorithm to split a convex area among several robots into the corresponding number of parts based on the area requirements for each part. The core idea of the algorithm is a sequence of divisions into pairs with the lowest possible perimeters. In this way, the compactness of the partitions obtained is maximized. The performance of the algorithm, as well as the quality of the obtained parts, are analyzed in comparison with two different algorithms. The presented approach yields better results in all metrics compared to other algorithms.
Article
Non-holonomic Dubins-car like robots with limited control ranges have to autonomously form the densest net across a corridor environment, then to keep this formation, and to sweep the corridor with a given speed. The robots do not communicate with anybody, cannot distinguish between the partners, cannot play distinct roles in the team, and are unaware of the team size or corridor width. In its local frame, every robot has access to the relative coordinates of the objects within a finite sensing range and measures its own speed. It is shown that even in the face of essential constraints on the motion and available sensory and computational resources, the mission can be fully solved by a lower-level, continuous-time, and computationally inexpensive distributed control law. The proposed controller is common for all robots, individually operates at any of them and is reactive, i.e., it directly converts the current sensory data into the current control in a reflex-like fashion via executing a few elementary operations. The performance of the navigation algorithms is justified via a mathematically rigorous non-local convergence result for the corridors with straight parallel walls and is demonstrated via computer simulation tests in more complicated cases.
Article
Full-text available
Micro aerial vehicles (MAVs) are an excellent platform for autonomous exploration. Most MAVs rely mainly on cameras for buliding a map of the 3D environment. Therefore, vision-based MAVs require an efficient exploration algorithm to select viewpoints that provide informative measurements. In this paper, we propose an exploration approach that selects in real time the next-best-view that maximizes the expected information gain of new measurements. In addition, we take into account the cost of reaching a new viewpoint in terms of distance and predictability of the flight path for a human observer. Finally, our approach selects a path that reduces the risk of crashes when the expected battery life comes to an end, while still maximizing the information gain in the process. We implemented and thoroughly tested our approach and the experiments show that it offers an improved performance compared to other state-of-the-art algorithms in terms of precision of the reconstruction, execution time, and smoothness of the path.
Conference Paper
Full-text available
Inspired by recent advances in single agent reinforcement learning, this paper extends the single-agent asynchronous advantage actor-critic (A3C) algorithm to enable multiple agents to learn a homogeneous, distributed policy, where agents work together toward a common goal without explicitly interacting. Our approach relies on centralized policy and critic learning, but decentralized policy execution, in a fully-observable system. We show that the sum of experience of all agents can be leveraged to quickly train a collaborative policy that naturally scales to smaller and larger swarms. We demonstrate the applicability of our method on a multi-robot construction problem, where agents need to arrange simple block elements to build a user-specified structure. We present simulation results where swarms of various sizes successfully construct different test structures without the need for additional training.
Article
Full-text available
Multi-agent path finding (MAPF) is an essential component of many large-scale, real-world robot deployments, from aerial swarms to warehouse automation. However, despite the community's continued efforts, most state-of-the-art MAPF planners still rely on centralized planning and scale poorly past a few hundred agents. Such planning approaches are maladapted to real-world deployments, where noise and uncertainty often require paths be recomputed online, which is impossible when planning times are in seconds to minutes. We present PRIMAL, a novel framework for MAPF that combines reinforcement and imitation learning to teach fully-decentralized policies, where agents reactively plan paths online in a partially-observable world while exhibiting implicit coordination. This framework extends our previous work on distributed learning of collaborative policies by introducing demonstrations of an expert MAPF planner during training, as well as careful reward shaping and environment sampling. Once learned, the resulting policy can be copied onto any number of agents and naturally scales to different team sizes and world dimensions. We present results on randomized worlds with up to 1024 agents and compare success rates against stateof- the-art MAPF planners. Finally, we experimentally validate the learned policies in a hybrid simulation of a factory mockup, involving both real-world and simulated robots.
Article
Full-text available
Coverage path planning consists of finding the route which covers every point of a certain area of interest. In recent times, Unmanned Aerial Vehicles (UAVs) have been employed in several application domains involving terrain coverage, such as surveillance, smart farming, photogrammetry, disaster management, civil security, and wildfire tracking, among others. This paper aims to explore and analyze the existing studies in the literature related to the different approaches employed in coverage path planning problems, especially those using UAVs. We address simple geometric flight patterns and more complex grid-based solutions considering full and partial information about the area of interest. The surveyed coverage approaches are classified according to a classical taxonomy, such as no decomposition, exact cellular decomposition, and approximate cellular decomposition. This review also contemplates different shapes of the area of interest, such as rectangular, concave and convex polygons. The performance metrics usually applied to evaluate the success of the coverage missions are also presented.
Article
Full-text available
In this paper, we propose a deep reinforcement learning (DRL)-based method that allows unmanned aerial vehicles (UAVs) to execute navigation tasks in large-scale complex environments. This technique is important for many applications such as good delivery and remote surveillance. The problem is formulated as a partially observable Markov decision process (POMDP) and solved by a novel online DRL algorithm designed based on two strictly proved policy gradient theorems within the actor-critic framework. In contrast to conventional simultaneous localization and mapping-based or sensing and avoidance-based approaches, our method directly maps UAVs' raw sensory measurements into control signals for navigation. Experiment results demonstrate that our method can enable UAVs to autonomously perform navigation in a virtual large-scale complex environment and can be generalized to more complex, larger-scale and three-dimensional environments. Besides, the proposed online DRL algorithm addressing POMDPs outperforms the state-of-the-art.
Article
Event cameras are bio-inspired sensors that work radically different from traditional cameras. Instead of capturing images at a fixed rate, they measure per-pixel brightness changes asynchronously. This results in a stream of events, which encode the time, location and sign of the brightness changes. Event cameras posses outstanding properties compared to traditional cameras: very high dynamic range (140 dB vs. 60 dB), high temporal resolution (in the order of microseconds), low power consumption, and do not suffer from motion blur. Hence, event cameras have a large potential for robotics and computer vision in challenging scenarios for traditional cameras, such as high speed and high dynamic range. However, novel methods are required to process the unconventional output of these sensors in order to unlock their potential. This paper provides a comprehensive overview of the emerging field of event-based vision, with a focus on the applications and the algorithms developed to unlock the outstanding properties of event cameras. We present event cameras from their working principle, the actual sensors that are available and the tasks that they have been used for, from low-level vision (feature detection and tracking, optic flow, etc.) to high-level vision (reconstruction, segmentation, recognition). We also discuss the techniques developed to process events, including learning-based techniques, as well as specialized processors for these novel sensors, such as spiking neural networks. Additionally, we highlight the challenges that remain to be tackled and the opportunities that lie ahead in the search for a more efficient, bio-inspired way for machines to perceive and interact with the world.