PIPELINE INSPECTION WITH AUTONOMOUS

SWARM ROBOTICS.

RICHARD MOLYNEUX

A thesis submitted in partial fulfilment of the requirements for the

degree of Master of Philosophy

The University of Sheffield

Faculties of Mechanical Engineering and Automatic Control and

Systems Engineering

Date: 14/01/2023.
1
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 3

Abstract

Underground water pipelines require frequent inspection to prevent de-

cay and avoid costly repairs. The use of robots to inspect pipelines is
well documented but the uniquely hostile environment of subterranean
water networks means pipes require continual inspection. The idea of
autonomous inspection robots that can provide continuous coverage
has shown promise, but existing methods do not proactively aim to
overcome a variety of diverse challenges. Specifically, extreme variabil-
ity in the pipe conditions and the dense surrounding earth limit the
communication capabilities of the robots, while dynamic water flows
and power issues have a detrimental effect on their movement. This
thesis presents the implementation of a range of path planning algo-
rithms of varying levels of autonomy as governing swarm behaviours,
each with a focus on overcoming some of the specific challenges in-
herent in underground water networks, with the goal of improving the
efficiency with which the swarm can inspect a network. The Greedy
Walk uses stochastic processes to plan a locally optimal path, the novel
Ad Hoc algorithm aims to provide cyclic coverage, with robots mov-
ing as a fluid net throughout the network, and the k-Chinese Postman
Problem solution explicitly plans optimal paths round subsections of
a network. The thesis examines the performance of these behaviours
against existing methods and anticipates obstacles in their real world
implementation. The thesis then presents tailored versions of the path
planning behaviours that include the introduction of more advanced
methods focused on circumventing these issues. Specifically, the al-
gorithms are developed to incorporate Gaussian Process Regression
models to analyse strong water flows and use the data to plan intelli-
gently, mitigating the detrimental effects of the flow. The flow analysis
also provides a platform from which a novel Simultaneous Localisation
and Mapping algorithm is presented, alongside a Multi-Objective Ge-
netic Algorithm with the focus of increasing inspection frequency and
conserving robot charge. The thesis shows evidence that an approach
to pipeline inspection with autonomous swarm robotics based in path
planning algorithms can help overcome the likely physical limitations
of real world implementation.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 5

Acknowledgments

First and foremost I wish to thank my supervisor, Kirill Horoshenkov,

for his continued patience and mentorship throughout my academic
journey. His knowledge and guidance have given me the freedom to
pursue my research path, and have undoubtedly helped shape the di-
rection of my thesis. Not only that, but his calm demeanour and pro-
fessionalism have helped keep me sane and on track through peculiar
and difficult times, and for that I am immensely grateful. Both in and
out of academia, I wish him nothing but the best.

Likewise I give my thanks to my secondary supervisor, Sean Anderson,

for his technical insight and willingness to help. There was never a call
where I did not come away with a new idea or direction. I wish him
luck in his future research.

A special thanks must go to Christopher Parrott for his tutelage and

friendship. Outside of the sheer volume of time he put into my devel-
opment, both as a coder and a researcher, I appreciate his help getting
me through the long days, from sharing his projects to our collective
ramblings. He has undoubtedly had the biggest impact on my research
experience and I wish him the best of luck in his new career.

Thanks go to all my friends who have helped make my time in Sheffield

special. In my eight years here I have been fortunate enough to meet
some wonderful people who have enriched my experience in so many
ways. To each of you I am grateful for the memories you have given
me.

In addition, I wish to thank my lifelong friend Jack Evans. His steadfast

friendship is rare and I appreciate every act of kindness he has shown
to me and those around him. His consistency has kept me grounded,
particularly during my research, and I will always welcome his advice.

To Antonia I give my greatest thanks. Her love and support have

helped me endure my hardest times and I will be forever grateful to
her for the patience she has shown through all my studies. I can only
look forward to our future together.
6 MOLYNEUX

Finally, I wish to thank my family for their love and unconditional

support. It is impossible to express my gratitude for everything they
have done for me — without them I simply would not be who I am
today. To them I dedicate this work.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 7

Contents

Abstract 3

Acknowledgments 5

1. Introduction 10

1.1. Motivation. 11

1.2. Problem Definition. 12

1.3. Aims and Objectives. 13

1.4. Publication. 13

1.5. Thesis Outline. 14

2. Literature Review. 15

2.1. Existing Inspection Methods. 16

2.2. Robotic Inspection Methods. 18

2.3. Autonomous Inspection Methods. 20

2.4. An Introduction to Swarm Intelligence. 23

2.5. An Introduction to Path Planning. 26

2.6. Simultaneous Localisation and Mapping Algorithms. 28

2.7. Applications of Deep Learning Algorithms. 30

3. Simulation Platform and Communication Methods. 32

3.1. Introduction. 32

3.2. Simulation. 33

3.3. Behaviours. 43
8 MOLYNEUX

3.4. Stigmergy. 46

3.5. Greedy Walk. 48

3.6. Ad Hoc. 49

3.7. Multiple Chinese Postmen Problem. 54

3.8. Simulations. 57

3.9. Implementing Communication Between Robots. 61

3.10. Discussion 70

4. Implementing a Responsive Flow Intelligence. 72

4.1. Introduction. 72

4.2. Introducing Dynamic Flow. 73

4.3. Localization. 75

4.4. SLAM simulations 81

4.5. A Reactive Flow Intelligence. 83

4.6. The Gaussian Process. 85

4.7. Simulation Implementation. 89

4.8. Flow Intelligence Simulations. 93

4.9. Flow SLAM Algorithm 96

4.10. Discussion. 97

5. Creating a Solution to Autonomous Power. 99

5.1. Implementing the Power System. 100

5.2. Creating a Multi-Objective Metric. 101

5.3. Creating the Genetic Algorithm. 102

 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 9

6. Conclusion. 104

6.1. Discussion 104

6.2. Conclusion 108

6.3. Future Work 109

References 110
10 MOLYNEUX

1. Introduction

Water networks are invaluable to humans. They are essential to sus-

taining modern life as we know it and feature in almost every aspect
of our lives somewhere.

Despite their necessity, underground pipeline networks are increasingly

in disrepair. Many of these pipes are old, built before sufficient doc-
umentation was proper practice. As such, not only are they in poor
condition, but often their supposed location is not precise and they
frequently overlap with other, newer pipes. The harsh environment
in which they are placed accelerates decay, and their inaccessibility
provides little opportunity for maintenance. It has become clear that
current inspection methods are at best ineffective, so water companies
are looking to transition to a more reliable, cost-effective solution.

Most water companies desire to increase their inspection frequency with

minimal change to infrastructure, providing a new area to which au-
tonomous sensing might be applicable. By introducing mobile inspec-
tion robots with the ability to navigate the networks, it is possible to
deploy the robots from a central hub into the pipes, to inspect the sur-
rounding area. This circumvents vast infrastructure changes and pro-
vides a non-destructive, proactive inspection process capable of finding
leaks and blockages before they become problematic. Given the typi-
cal size of UK water networks, it is likely that multiple robots will be
required to service a single area. With proper implementation, these
swarms have the advantage of communicating and working together
autonomously to increase the efficiency of the inspection process.

Autonomous swarms are extremely capable and easy to implement,

and have subsequently branched out into a vast variety of fields. Un-
fortunately, swarm behaviours typically require frequent, if not con-
stant, communication. Unlike most swarm settings, underground pipe
networks are a highly challenging communication environment with
extremely limited communication ranges. This is due to the extreme
variability in the pipe conditions and the relatively high attenuation
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 11

of acoustic and radio waves with which robots would normally com-
municate. This impacts the swarm’s ability to make quick, informed
decisions as each agent is working with partial, or out of date informa-
tion.

Additionally, water networks have the unique problem of dynamic wa-

ter flows. A high flow has the potential to slow or halt a robot, which
provides particular difficulties in ensuring a consistent inspection pro-
cess. As well as this, the disruption from strong flows limit the fre-
quency with which robots in the swarm pass one another, further com-
pounding the communication issues. In addition, to cover the large size
of the water networks with a continuous inspection process requires the
robots to be untethered. Given that the robots are continuously in-
specting, fighting flow and attempting to communicate to one another,
their power demands vastly outweigh their capabilities.

There are many approaches that can be taken to address these issues,
from actively governing the swarm to encouraging biologically inspired
simple behaviours. Each has the potential to counteract the problems
in different ways, with the ultimate goal of creating an efficient pipeline
inspection process to sustain the water networks we rely on.

1.1. Motivation. Given its recent emergence as a field, swarm robot-

ics is already relatively advanced due to its vast array of capabilities
with simple application. Typically, robots require little information to
influence the swarm and, similar to many natural systems, the group
is able to make complex decisions based on a multitude of smaller bits
of information collected by the swarm. The ability to formulate ef-
ficient solutions with minimal computational or sensing requirements
proves advantageous in many fields, from nanotechnology to search
and rescue robots. Unfortunately, underground pipeline networks are
a uniquely challenging environment with various stipulations that limit
the potential of typical swarm behaviours, which can be detrimental to
the efficiency of autonomous swarm inspection. The simple nature of
swarm algorithms means they struggle to make use of other potentially
beneficial, intelligent solutions that can be created from the data collec-
tion. Additionally, pipeline environments do not favour swarm systems
12 MOLYNEUX

as the efficiency of swarm behaviour is highly dependent on frequent,

instantaneous communication. As such, the opportunity presents it-
self to examine the rigour of other existing methods to determine both
their efficiency against swarm behaviours, and whether they can be tai-
lored to include more advanced techniques. Path planning algorithms
of varying degrees of autonomy are commonplace and are used in many
fields, from navigation to video games. They require more direction to
function efficiently than swarm algorithms however they are able to
take advantage of much more information than swarm algorithms to
create solutions to more complex problems. It is hoped that by in-
troducing path planning algorithms, methods specifically focused on
mitigating the effects of the environment can be implemented to help
improve the efficiency of the autonomous swarm inspection process.

1.2. Problem Definition. Inspection robots have gained traction as

viable solutions to pipe maintenance in recent times. Though the field
is new, there is a multitude of evidence that shows the potential of
robots in a pipeline environment. However, the step up to an au-
tonomous platform is enormous. Immediate considerations surround-
ing the physical limitations of an autonomous robot range from the
power requirements to the communication methods and processing
power of a necessarily small robot. To this end, what is needed is
an exploration of potential systems that are capable of providing an
efficient inspection process, whilst minimising the physical demands of
the robot. A well-documented field surrounding network navigation
is rooted in graph theory, specifically path planning algorithms. Path
planning algorithms have the ability to create intelligent routes for a
robot and can have varying degrees of autonomy. Their capacity to
consider information and plan ahead allows them to determine opti-
mal paths with multiple objectives. As the field progresses, and the
anticipated physical obstacles of autonomous inspection become more
apparent, an assortment of behaviours that can circumvent these is-
sues could aid in reducing certain hardware requirements. To be sure
of an optimal and realistic inspection process, an exploration into path
planning algorithms and their potential in subterranean water networks
must be considered.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 13

1.3. Aims and Objectives. The overall objective of this research

project is to develop a swarm behaviour based in path planning capa-
ble of handling the unique technical challenges posed to autonomous
pipeline inspection robots. The work of Parrott, C. et al. [44] lays the
foundation of a simulation platform and the stigmergy behaviour [16],
against which the efficiency of path planning algorithms can be bench-
marked. To counteract the evident detrimental effect of dynamic water
flows, the work aims to integrate Learning Algorithms, with which the
agents can model the water flow, with path planning algorithms to
navigate the network efficiently. This work is to build upon the ex-
isting simulation to introduce other mechanics that represent issues
in implementing autonomous swarms that have not been considered
previously, and build in adapted algorithms to achieve optimal solu-
tions in a more complex simulation. Contributions to the simulation
should consider realistic situations such as strong flows overwhelming
robots, pushing them back and the robots becoming ’lost’, or the ef-
fects of adding power limitations to the robots with limited charging
stations. This provides the opportunity for the algorithms to be de-
veloped further, in response. For example, the work aims to utilise
information gained from the flow analysis to improve a novel SLAM
implementation for lost agents. As these considerations increase, the
aim is to similarly adapt any algorithms or behaviours with additional
solutions to mitigate the effects of these changes and keep consistently
low Time Between Inspections. It is hoped that a variety of path plan-
ning algorithms will provide a strong basis to be adapted in response
to each new stipulation. At the conclusion of the thesis, the resulting
algorithms should incorporate many different solutions to exhibit an
adaptive approach to autonomous pipeline inspection robots that sus-
tains low Time Between Inspection values and high network coverage,
regardless of the conditions imposed.

1.4. Publication. This thesis represents the authors own work and
has led to the following peer-reviewed publication:

Molyneux, R. , Parrott, C. , Horoshenkov, K. (2019), ’An Applica-

tion of Path Planning Algorithms for Autonomous Inspection of Buried
14 MOLYNEUX

Pipes with Swarm Robots’, World Academy of Science, Engineering

and Technology, Open Science Index 153, International Journal of Me-
chanical and Materials Engineering, 13(9), 555 - 563.

1.5. Thesis Outline. The thesis is structured as follows:

• Chapter 2 explores the literature surrounding autonomous robotic

inspection and potential fields of application. Particularly, it ex-
plores biologically inspired swarm behaviours and path planning
algorithms and how they can be applied to pipeline inspection.
First and foremost, existing inspection methods and robotic
capabilities are explored to examine the areas of challenge in
implementing autonomous swarms. The literature examines
the strengths of current acoustic and ultrasonic communica-
tion techniques and investigates fields in artificial intelligence
and machine learning. An emphasis is placed on potential ap-
plications in path planning before concluding with analysis of
genetic algorithms and their applications.
• Chapter 3 begins by providing an in-depth detailing of the
simulation platform presented by Parrot, C. et al. [44]. This is
important as the work in this thesis is built off adaptations and
additional work that is built on the simulation platform. The
detail informs the reader of the workings of the original simula-
tion, and enables more coherent explanations of the changes as
they are added. The Chapter goes on to introduce three differ-
ent path planning algorithms that govern the decision making
process of the agents in a network, before demonstrating the
simulation output to analyse their inspection efficiency against
the existing stigmergy behaviour. The Chapter concludes by
adding acoustic and ultrasonic communication methods to the
simulation before comparing the effects on the behaviours.
• Chapter 4 uses the simulations existing ability to model dy-
namic flow patterns to analyse the effects of strong flows on the
path planning algorithms. It goes on to add a mechanic to the
simulation — the ability for an agent to become overwhelmed
by the flow and become lost. A simple but effective SLAM
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 15

algorithm capable of localising quickly with limited topologi-

cal information is then presented and the adapted simulation is
used to evidence it’s efficiency. The Chapter then presents an
approach for stochastic modelling of the flow patterns with the
intent to learn the cyclic flow of a network, to aid in intelligent
traversal. To this end, Gaussian Regression models and the
speed at which they can learn the network flow are illustrated.
The Chapter finishes by examining the effects of path planning
with knowledge of the flow and concludes with the idea for a
novel SLAM algorithm, capable of utilising the flow data to
improve the rate at which it is able to localise.
• Chapter 5 details additional changes to the simulation plat-
form surrounding charging and power. Specifically, agents are
given a battery percentage and two methods of charging, as well
as adding conditions on a powerless robot. The Chapter then
presents a Multi-Objective Genetic Algorithm with the goal of
finding Pareto optimal solutions that balance the power needed
to inspect a path, against the reduction in Time Between In-
spections.
• The final Chapter 6 discusses the works and contributions of
the thesis and concludes with a specific vision for future work
that extends from lessons learned during the research project.

2. Literature Review.

This Chapter will provide an analysis of existing literature relevant to

autonomous swarm inspection. The chapter begins with a brief history
of existing inspection methods and recent developments in inspection
robotics before presenting novel work surrounding biological swarm be-
haviours and autonomy. This is followed up by a literature review of
path-planning algorithms and their potential applications in subter-
ranean networks. In anticipation of the methods this thesis presents,
the review then extends to other areas of relevance, such as Simultane-
ous Localisation and Mapping algorithms, Deep Learning algorithms
and Multi Objective Algorithms to provide context as to their uses and
appropriateness to work in tandem with path-planning algorithms.
16 MOLYNEUX

2.1. Existing Inspection Methods. Underground water pipeline net-

works are uniquely difficult to inspect frequently — the subterranean
nature of the pipes alongside the size of the networks makes proactive
manual inspection near impossible. This is a problem as water pipeline
networks are increasingly in disrepair, and water companies have lim-
ited knowledge of the conditions of their pipes. As such, faults in the
system are only exposed after they have caused a leak, leading to con-
sumer disruption and costly repairs — an estimated 3,031 million litres
of water are leaked each day [15].

The most manual inspection process use human inspection to either

enter a pipe or surveying the conditions surrounding a pipe. For se-
verely damaged pipes, the deterioration can be indicated by examining
the properties of the surrounding soil [35]. The pH value of soil above a
pipeline can provide an insight into potential corrosion alongside other
indicators such as the moisture levels in the soil. Changes in the elec-
trolyte levels in the soil surrounding a pipe can be surveyed to ascertain
whether gaps exist in the coating of ferrous pipework. Though non-
invasive, this is a time consuming approach to examing the surface
conditions of pipework. In a similar manner, sediment sample analysis
can be implemented to assess their chemical constituents and risk to
water quality [26].

In spite of this, existing methods do exist that provide a measure of

assurance to water companies, though they are limited. Commercial
inspections using remote CCTV cameras that inspect the internal bore
are commonplace. The system comprises of a CCTV camera and light-
ing apparatus that is moved throughout the network [33], and can be
fitted with multiple cameras to achieve a 360 degree view of the pipe
[24]. The method of implementation is either to fit the CCTV under-
ground or insert it at the end of a rod, similar to an umbilical cord
system, to explore the surrounding area [25]. The rod insertion also
provides additional inspection opportunities by attaching hydrophones
in tandem with the cameras for accurate in-pipe leak detection. Un-
fortunately both elements become expensive when considering the size
of the networks typically seen in the UK. A fitted CCTV unit requires
invasive drilling and given the topography of pipeline networks, with
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 17

extremely long pipes and bends, is inefficient in a fixed position to

examine an entire network. Similarly, the range on the rod insertion
is around 200 meters. An average size water network is explored as
part of the research in this thesis, the size of which is detailed in the
Looping network in Table 1. The cumulative length of this network is
nearly 30,000 metres, with limited natural entry points. When consid-
ering the 348,723km of water pipework in the UK [15] it is extremely
difficult to consider the rod able to sufficiently inspect a network with-
out a high quantity of manual labour and the use of invasive drilling.
The requirement to be close to an entry point is a common obstacle in
ensuring coverage of a network and is a contributing factor to the push
for autonomous inspection.

Another well documented inspection method surrounds the use of acous-

tic detection in water. The wide applicability of acoustic detection has
allowed it to develop into many commercial products, with one that
leans heavily on the use of acoustic sensors being the LeakFinder sys-
tem [21], focused on pinpointing leaks in a length of pipe. The system
is comprised of transmitters and sensors that can sense leak-induced
vibrations and sound. A signal is transmitted through the pipe to two
receivers at different points of the pipe. The technology couples both
primary acoustic sensors, accelerometers and hydrophones, to calcu-
late the time lag between the two sensors [33] and derive an accurate
location of the leak. Though the inspection returns accurate results,
the system still requires an operator and is a reactive response to being
aware of the existence of a leak.

Another commercially viable acoustic technology is that of the Smart-

Ball [22]. The SmartBall is a free-swimming device acoustic device
designed for leak detection. What differentiates SmartBall is that it
passes through a water pipeline network with the flow of the network.
As it traverses the network, SmartBall makes use of multiple acoustics
sensors alongside ultrasonic transmitters and temperature sensors to
continuously record data and the position of the ball within a network.
By emitting acoustic pulses and examining the feedback, the sensors
are able to determine approximate locations of air pockets and leaks in
a non-invasive manner. However, as the SmartBall is propelled along
18 MOLYNEUX

by the flow of the network, the user has very little control over the
direction or speed the ball takes. As such, frequency analysis is often
required to properly determine the nature of an anomaly. Additionally,
the lack of directional control means that to gain a proper coverage of
the network, the ball must be placed at different entry points so as to
avoid simply repeating the same few paths.

2.2. Robotic Inspection Methods. Given the limited ranges that

can be achieved via current manual inspection processes, it is un-
surprising that vast leaps have been made in the areas of robotic
pipeline inspection. Robotic inspection has particular application in
Non-Destructive Testing and can be implemented to reduced work
time, reach areas of difficulty, or provide a safer inspection process.
The field is well documented, particularly in gas pipeline networks.
Typically, the inspection revolves around inserting a tethered robot
into a pipe, equipped with visual or acoustic sensors, and traversing
each direction from the entry point whilst continuously scanning [5].
The tethered nature of the robots removes any requirement for power
considerations, and provides a fast, reliable feedback of inspection data
to the receiver. However, the range that can be inspected is of course
limited to the length of the tether.

Perhaps the most advanced field when considering robotic pipeline in-
spection is centred around gas pipelines. Gas and oil leaks pose a
much more serious threat than water leaks, and as such extensive re-
search has been invested to ensure a thorough inspection process. Early
robotic systems have been presented that use articulated structures and
tether cables to develop snakelike robots capable of navigating complex
pipeline configurations [48]. Pipeline inspection gauges or gadgets,
commonly referred to as pigs, can be propelled by a pressure-driven
flow at speeds to gather fast inspection data [42] [13]. Unfortunately
this is not a complete method for all pipelines as obstacles and physical
barriers can disrupt this inspection method. As such, advances have
been made in crawler technology to create inspection robots of higher
maneuverability. Crawlers are typically mounted on wheels or tracks
and enable the user to assess a wide range of pipeline conditions due to
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 19

their independent movement and sensor functionality. They are typi-

cally equipped with cameras and lights, alongside acoustic or ultrasonic
sensors for leak detection [38]. By moving independently to the flow
they are able to move slower and steadier to generate a more accurate
depiction of the pipeline. Additionally, they can be built smaller than
the pigs, and as such provide an option to inspect smaller pipes that
were previously inaccessible.

There are many different designs for robotic inspection, typically in

response to the pipes in question [27]. Because of the strength of cer-
tain dynamic flows in water networks, water inspection robots tend to
follow a locomotion design that anchors them to the wall [41]. Usually
this is done by compressing the robots upon entry to the network, then
applying outward force until the wheels/tracts/movement mechanism
is snug to the pipe wall [47]. Therefore, one of the main considerations
in the design stage is the pipe diameter, as this provides the biggest ob-
stacle with respect to the physical size the robots can be, and therefore
their capabilities. Additionally, topological constraints such as sharp
corners or obstacles in the pipe are necessary to circumvent to gain
full use of the tethered range. As such, a favourable design in water
pipelines is that similar to that of a caterpillar. Elongating the ro-
bot enables the system to fit inside pipes of small diameters without
compromising on functionality, spreading out necessary components
such as motors and sensors. The conditions of many pipelines mean
pipes frequently have internal deposits lining the pipe wall that can in-
hibit movement methods of other robots — those relying on magnetic
attraction to a ferrous surface to grip the pipe walls can falter when
brought into contact with high quantities of sludge. The See Snake [45]
is a commercial adaptation for water pipelines focused on non-contact
Non-Destructive Testing. The See Snake aims to circumvent the effects
of the pipe wall lining by attaching wheels to multiple modular bodies
at different degrees and linking the bodies together in the caterpillar
design. Additionally, the modular split provides the robot with en-
hanced movement capabilities, and enables it to traverse sharp bends.
This is particularly useful in water networks where pipes can meet at
sharp junctions and unusual angles. Despite it’s maneuverability, the
20 MOLYNEUX

See Snake is again a tethered robot that requires manual instruction

to navigate a network.

Advances in robotic pipeline inspection have demonstrated that through

various innovative designs, robots are capable of maneuvering across
the network and around sharp bends. Additionally, they are loaded
with sensors that already collate valuable inspection data. This helps
support the proposal that robotic swarms can inspect autonomously.
However, the field has struggled to break away from the anchor of teth-
ered systems, with no system capable of continuously inspecting on it’s
own. Untethered robots lack the high charge capacity needed to sustain
inspection for long, and there is no mention of subterranean charging
capabilities.

2.3. Autonomous Inspection Methods. It has been established

that the concept of utilising robots to inspect pipelines is well doc-
umented. What differentiates the field of autonomous inspection is
the goal of creating a system that does not require an operator. Imple-
menting autonomous inspection circumvents many of the problems that
exist with pipeline inspection — continuous inspection can occur with-
out the need for invasive or destructive manual inspection techniques.
An autonomous robot that is continually inspecting would be able to
collate larger amounts of data, generating a much more accurate depic-
tion of pipeline networks with a much greater coverage. By extension,
introducing multiple robots into a network would inevitably increases
the frequency with which pipes can be inspected as this increases the
cumulative distance of pipe inspected by the swarm.

The literature in Section 2.2 evidences that aside from the charging
and power considerations, appropriate robotic platforms exist that can
navigate and inspect a network. A milestone in evidencing the feasi-
bility of autonomous inspection is the work of Parrot, C., Dodd, T.,
Boxall, J.B. and Horoshenkov, K. titled ‘Simulation of the Behaviour
of Biologically-Inspired Swarm Robots for the Autonomous Inspection
of Buried Pipes’ [44]. The work details a simulation platform that ex-
plores how a simple swarm intelligence can be applied to autonomous
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 21

robots to create a powerful inspection process based off of indirect coor-

dination. Specifically, a simulation environment was designed to model
the structure of real life water networks and place robots, or agents,
within the simulation capable of traversing the network. By assum-
ing continual inspection as the agents moved throughout the network,
the simulation was capable of determining an average Time Between
Inspections or TBI for each pipe and was able to evidence improved ef-
ficiency to existing methods. In addition to this, the simulation makes
use of EPANET software files to construct the corresponding water
flows within the network across a 24 hour period. By assuming that
strong water flows might inhibit the movement speed of robots in a
pipeline, the paper was able to illustrate the detrimental effects of ex-
isting flow cycles to both the coverage and frequency of inspections.

The work extends further, seeking to prove not only the effectiveness
of autonomous inspection, but to provide a governing intelligence for
the swarm of robots to improve the coordination of the swarm. To
govern the movement of the swarm, the work by Parrott, C. et al. [44]
digitised the Stigmergy [16] swarm behaviour to implement an implicit
form of cooperation between the agents, based solely off choosing a
pipe when the agent reaches a junction. Provided there was communi-
cation between the agents, the agents were able to share their ‘mem-
ories’ of the network — their individual inspection histories. In doing
so, two agents sharing their inspection history were able to update
their own awareness of pipes that had been inspected most recently
to inform decisions. Each pipe was therefore anointed with a virtual
pheromone that enabled agents at a junction to choose the pipe which
most required inspection. The paper goes on to illustrate the benefits
of increased communication ranges and coordination to the efficiency
of the Stigmergy implementation, particularly with respect to network
coverage.

The work of Parrott, C et al. [44] provides a fantastic foundation

for future research into autonomous inspection. The work proves the
feasibility of autonomous pipeline inspection using swarms by evidenc-
ing a behaviour capable of adapting to the expected physical obstacles
the might surround the implementation of real robots into a network.
22 MOLYNEUX

Building on this, it is possible to consider other aspects relevant to

autonomous inspection such as intelligent solutions that seek to proac-
tively mitigate the negative affects illustrated in the work. The simu-
lation results have shown both strong water flows and limited commu-
nication capabilities to be complications unique to underground water
pipeline inspection and as such, potential areas to be navigated to im-
prove the efficiency of the swarm. Additionally, it must be noted that
continuous inspection was assumed and the real life power stipulations
inherent in robotics were not modelled.

Other notable contributions to autonomous inspection includes the

work of Li, X., Yu, W., Lin, X. and Iyengar, S.S. titled ‘On Opti-
mizing Autonomous Pipeline Inspection’ [32]. The work proposes a
solution to gallery guarding, a problem aiming to find the smallest set
of points inside a region from which all the boundary points are visible.
In the application of pipeline networks, the work assumes the robots
are fitted with a camera and lighting and can inspect everything they
see. The work implements a hierarchical integer linear programming
(HILP) algorithm to find the optimal formulation of points on a net-
work from which coverage can be guaranteed. The work details a 3D
model evidencing leaks and blockages found during the simulation, and
provides a method with which an individual robot could ensure traver-
sal of the network and complete inspection — by simply moving from
point to point.

The application of the work of Li, X. et al. [32] into water networks
requires some consideration. The system is reliant on the initial geom-
etry of the network to be modelled before any computation can occur.
Water companies in the UK do not necessarily have the information re-
quired to build the 3D model — pipes have frequently been built on top
of other pipes and before proper documentation was implemented. As-
suming the information is available, the work details how the pipeline
wall needs to be marked in order for the robot to localise at the exact
location needed to ensure complete coverage. Applying the markings
to water networks is a task in itself; given the size of water networks in
the UK and the relative range a robot can assume to be able to inspect
from one spot, it is likely that an extremely high number of markings
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 23

would be required to run efficiently and is likely to be a manual task in

itself. Finally, water networks have the unique issue of dynamic flow
rates which are capable of blocking off large subsections of network at
a time. This would make it difficult to consistently reach the spots
required for optimal inspection without coordination.

Continuous autonomous inspection of water networks is undoubtedly

desirable but, due to hardware capabilities and power requirements, it
seems a long way off. However, if the ability to continuously inspect
can be assumed, it is now evident that robots could be applied for this
purpose. To that end, as with the work of Parrott, C. et al. [44],
through the medium of simulations it is possible to begin analysing
efficient ways to get the robots to work together and build and optimal
inspection process.

2.4. An Introduction to Swarm Intelligence. Swarm intelligence

can be considered as the emulation of some biological system to achieve
a collective goal. Bonabeau [4] offers a specific definition that swarm in-
telligence is ‘any attempt to design algorithms or distributed problem-
solving devices inspired by the collective behaviour of social insect
colonies and other animal societies’. The field is technically a sub-
set of Artificial Intelligence (AI) algorithms and has gained increasing
popularity in the last two decades due to it’s wide variety of applica-
tions. Swarm algorithms typically use small amounts of data to inform
a rudimentary choice that, in tandem with the rest of the swarm, often
contribute to offering a versatile solution. An example of their strength
lies in NP-hard problems — a classification of polynomial-time prob-
lems that are not solvable by a non-deterministic Turing machine in
polynomial time. They are particular difficult to solve, and as the
problem grows, so too does the solution — exponentially. As such it
is frequently difficult to solve them exactly in real time. Swarm intel-
ligence algorithms have been applied to find solutions sufficiently close
to the global optima in a much more viable time frame [9]. This helps
illustrate how powerful proper application of swarm intelligence can be.

Perhaps the best example of a simple swarm intelligence is stigmergy,

a form of indirect communication used by ant colonies to leave relevant
24 MOLYNEUX

information for other members of the colony in their environment [16].

The concept was first introduced in 1959 by Pierre Paul Grasse [55]
and linked together the behaviour of individual ants with that of the
swarm. In essence the worker ants would leave pheromone signatures
on their environment as they continue their work. For instance, when
traversing a food trail and discovering all the food has been harvested
and there is none left, a single ant would leave a pheromone signature
to let other workers know. As more and more ants begin to leave
similar signatures, the group eventually reaches a quorum, a state in
which they have sufficient information to agree that trail is no longer
a viable food root. This was shown to be used to share indicators
of danger or new resources. The evidence of what could be achieved
with such simple information and implementation is the basis for Ant
Colony Optimization algorithms.

It is unsurprising then that swarm intelligence has been applied in a

variety of fields with notable success. Biologically inspired algorithms
are not limited to insects, and range from glow worm to bat and lion
based algorithms [9]. They have huge applications and can simplify
complex problems into manageable operations. Of note are Particle
Swarm Optimization algorithms, built to mimic fish schooling and bird
flocking [11]. PSOs communicate not only their position but also their
velocity, enabling the members to manage their own positions and move
together intelligently. This has lead to applications in UAV swarms in
search and rescue operations [10], tremor analysis for the diagnosis of
Parkinson’s disease [19] and auto-tuning for telecommunication systems
[18].

The application of swarm intelligence in subterranean networks is lim-

ited. Despite their affluence above ground, swarm behaviours are highly
reliant on being able to communicate with one another, directly or in-
directly. As soon as a member becomes isolated, their efficiency and
contributions to the swarm objective will drop. Given the hostile com-
municative environment of underground networks, it is therefore un-
surprising that the implementation of swarm behaviours in pipeline
inspection is difficult. Despite this, the work of Parrott, C. et al. [44]
mentioned in Section 2.3 digitised the stigmergy swarm behaviour to
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 25

assign virtual pheromones to pipes in a network. This helped formulate

the governing behaviour of the swarm, informing decisions implicitly
to enforce an impressive indirect coordination method. The work did
explore the effect of different communication ranges and was more ef-
ficient when the agents were able to communicate more frequently.

Mandal, S. K., Chan, F. T. and Tiwari, M. K. [34] also found a way to

implement swarm intelligence in pipeline inspection by combining an
Artificial Bee Colony (ABC) algorithm with a Support Vector Machine
(SVM) algorithm to develop a pipeline leak detection system. ABC al-
gorithms are typically concerned with optimisation for problems that
are deterministic or stochastic in nature, and are inspired by the per-
ceived foraging behaviour of honey bees in their search for nectar [29].
SVMs are supervised learning algorithms centred on regression analy-
sis. The work uses the split of Employed Bees focused on investigating
food sources and updating their memory to complete their search and
share this information in the dance area. Subsequently, Onlooker Bees
decipher information from the dance area to choose a food source based
on the best estimated returns. In this instance, a Genetic Algorithm is
applied to select ’food sources’ for the Onlookers. Finally, the Scouts
are Employed Bees who have exhausted their food sources to randomly
search out another food source. Combining the two approaches results
in a behaviour where, during the search, employed bees and onlookers
modify their positions while the scouts exchange abandoned solutions
with more fruitful ones. The paper goes on to show the benefits of using
this search process in tandem with an SVM and the SVM’s improved
ability to obtain high detection rates of leaks.

Swarm intelligence is clearly a powerful and well established field. How-

ever, the underlying commonality between all swarm behaviour is a fast
rate of communication — it is essential to ensuring intelligent contri-
bution to a common goal. Without proactively seeking to increase the
rate of communication exchange, extrapolating the work to a subter-
ranean environment is unlikely to yield similarly brilliant results as it
does above ground.
26 MOLYNEUX

2.5. An Introduction to Path Planning. Path planning is a crucial

component to many robotic and automated systems. The field is based
in geometry and is an extension of graph theory in mathematics, with
the objective of calculating the best way to traverse a graph with one or
more secondary objectives. Gasparetto, A., et al. gives a straightfor-
ward definition: ‘find a collision-free motion between an initial (start)
and a final configuration (goal) within a specified environment’ [23].
Contrary to the indirect approach of swarm intelligence, path planning
algorithms aggressively seek optimal solutions but are typically reliant
on higher computer processing power to use data to inform their de-
cisions. Literature in the field is well documented as the goals of the
algorithm can vary drastically, from finding the minimum energy route
to an end point, to planning a path that visits allows traversal of an
edge only once. The latter is the famous Konigsberg bridge problem
which questions whether it is possible to walk over the seven bridges
of Konigsberg only once each and return to the same spot [58].

Dijkstra’s algorithm is a complete method for finding the shortest path

between two vertices on a graph. A graph is simply a mathematical
representation of relationships and is built with vertices and edges [6].
Modelling water networks in this manner allows for the introduction
of other path planning algorithms. It works by dividing all nodes (ver-
tices) into unvisited and visited sets and iteratively refining the sets,
moving nodes to the visited set once the shortest path to the visited
node has been found. Eventually, the destination vertex is visited and
the algorithm is complete. The algorithm is widely accepted as the de-
fault, optimal solution for shortest path search problems which has lead
to implementations in navigation systems such as Google Maps [31],
intelligent fire evacuation systems [61] and soft computing [14]. How-
ever, the method is computationally expensive and becomes unsuitable
as the size of the graph increases [39]. This lead to the development of
the A* algorithm, a heuristic approach to shortest path problems [54].

Another well documented example of a path planning algorithm is the

Chinese Postman Problem [20]. The Chinese Postman Problem is a
heavily studied topic within graph theory, and hence has a huge back
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 27

catalogue of relevant literature of adaptations, typically involving min-

max approximations [1]. The Chinese Postman Problem itself focuses
on finding the optimal route around every edge in a graph such that
every edge is passed at least once [43]. The problem originally revolved
around a single postman servicing a network of streets delivering post
whilst trying to minimise the total time it took to complete a route,
always returning to the depot node. This problem is highly relevant
to the problem of pipe inspection with an autonomous robot swarm
as the swarm network can be modelled as a graph, with junctions as
vertices and the pipe links as edges [56]. A CPP solution would provide
an optimal traversal route for the agents in the network. The CPP has
a counterpart routing problem, the Traveling Salesman Problem [17].
The TSP attempts to traverse a graph while touching every vertex at
least once instead. Both of these algorithms constitute path-planning
algorithms as they create a connected path between two points with
some optimization function.

The Vehicle Routing Problem (VRP) is an advanced research topic

within path planning and has an overlap with swarm intelligence [2].
It is an optimisation problem concerned with minimising or maximising
some goal but for multiple vehicles. The earliest example is the Truck
Dispatching Problem [12], which required that a fleet of oil trucks leav-
ing a central hub traverse the minimal amount of distance as a fleet,
whilst still fulfilling the demand for oil from customers diffused around
the network. This has been increasingly adapted to incorporate more
worldly considerations such as time windows for delivery slots and dy-
namic information [28]. Research is also driven forward in a solo ca-
pacity for vehicles — Global Positioning Systems (GPS) are inherent
in most vehicles and smartphones nowadays. To ensure the fastest
route, GPS routes take much more into account than the topography
of UK road networks. Traffic jams, roadworks and even user selected
options such as avoiding motorways are all necessary considerations to
advise on the best route. Maaref, M. and Kassas, Z. M. introduce sta-
tistical bounds that guarantee the probability of the absolute position
error exceeding a set threshold cannot be larger than the integrity risk
[36]. The work shows that the additional considerations created a route
outside of the shortest path that was able to outperform it.
28 MOLYNEUX

The k-CPP is an extension of the Chinese Postman Problem, intro-

ducing multiple postmen to service an area [43]. As opposed to swarm
behaviours, ideal solutions in this instance revolve around the post-
men actively planning their paths together. The most time consuming
method and inefficient method would have k-1 postmen simply remain
at the depot node, whilst the remaining postman serviced an area on
their own. On the other hand, an optimal algorithm would see the
workload distributed relatively evenly with the overall route time de-
creasing linearly as k, the number of working postmen, increases. The
process of finding CPP solutions revolves around ensuring that an the
graph is Eulerian — that is to say, no vertices in the network have an
odd degree. An elegant, albeit inefficient method for then extracting a
route around the graph is Fleury’s algorithm. Fleury’s algorithm works
by iteratively removing edges from the network who’s deletion does not
disconnect the graph [49]. The process repeats from connecting ver-
tices until the last edge has been removed, and the Eulerian cycle is
the order in which the edges were removed.

It is hoped that the literature presented can be adapted to offer an

alternative consideration for the governing intelligence of the swarms
in underground pipeline inspection. The network components in roads
which allow for the application of path planning algorithms correlate
with those in water networks — roads and junctions serve as appropri-
ate reflections of pipes and their intersections. The necessary step is to
present the network as a graph. Upon doing so, many applications of
path planning can be implemented, such as shortest path algorithms
or multiple agent planning algorithms such as the k-CPP and VRP
solutions.

2.6. Simultaneous Localisation and Mapping Algorithms. Si-

multaneous Localisation and Mapping algorithms, or SLAM algorithms,
are typically used in unknown environments to localise themselves
based on features in their surroundings, often landmark features [3].
The field is broad with many applications from large scale mapping of
urban structures [57] to autonomous mine mapping [40]. Solutions to
SLAM problems must be considered when attempting to make truly
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 29

autonomous robots. With respect to pipeline networks, it could be

desirable for robots to be able to map out the network. As afore-
mentioned, the information surrounding some networks is unreliable.
Previous methods of mapping out networks have included Branch and
Bound algorithms [7]. However, these techniques use brute force and
are exhaustive. They are therefore undesirable when considering the
size of UK water networks. The potential for robots to map their own
environment before traversing them provides a newer definition of au-
tonomy.

An example of gas pipeline is presented by Zhang, S. and Dubljevic,

S. [64]. The work details how a pipeline crawler can navigate small-
diameter gas pipelines whilst simultaneously mapping out the topology
of the network. The system uses an inertial measurement unit and
odometers to generate kinematic and measurement models that can
accurately relay pipeline features and a complete topology.

Once a system has been mapped, SLAM algorithms are typically em-
ployed into robotic systems, which provides them with the ability to
quickly localise themselves. Worley, R., Yu, Y. and Anderson, S. de-
scribe a method of acoustic echo-localisation for pipe inspection robots
[60]. Underground pipeline networks are extremely dark, and as such
other visual SLAM (vSLAM) methods lose their appeal [53]. Similarly,
pipelines are often devoid of landmark features and as such become a
difficult environment with which to truly be certain that a robot has
localised. The echo system presented detects distant features and uses
acoustic processing to make measurements of the robots position. The
acoustic data is refined before a state estimation is declared, leaving
the method with a high success rate.

With strong flows in the network it is necessary to consider the possi-

bility of robots getting pushed around. If a robot is pushed too hard
it might lose its bearings. In this instance, a Simultaneous Localisa-
tion And Mapping algorithm can assist in localising the robot again.
Relying on acoustic echo-localisation compliments the acoustic com-
munication and inspection methods as the robots will be necessarily
30 MOLYNEUX

small, so any equipment that can be shared will significantly lighten

the weight of the robot.

2.7. Applications of Deep Learning Algorithms. Machine learn-

ing is a subset of artificial intelligence focused on obtaining specific
information without being explicitly programmed to. Supervised learn-
ing requires human interaction for input data, and requires feedback
on the accuracy after the output [50]. On the contrast, unsupervised
learning algorithms require no input from the user. Though supervised
learning algorithms tend to be simpler than their counterparts, they
are more powerful in their ability to output data based on previous
experiences. However, the strength of a supervised model is entirely
reliant on the efficiency of the training data set. Incoherent data leads
to incoherent learning.

Deep Learning Algorithms are a subset of machine learning algorithms

that use multiple layers of neural networks to build computational mod-
els [63]. They are able to train deep neural networks and can be used in
diverse tasks such as image classification, object detection and image
retrieval [62]. Often it is the case that they have several algorithms ca-
pable of handling specific tasks better than others. The deep learning
refers to the use of layers in the network, and learning can be split into
supervised, semi-supervised or unsupervised.

Gaussian Process Regression models are a subset of deep learning ar-

tificial intelligence that are simple to implement, flexible, fully proba-
bilistic models [8]. They use a non-parametric, Bayesian approach to
regression that allows for powerful prediction analysis on datasets [51].
This means that instead of calculating a probability distribution using
data, the PDF is calculated using functions that fit the data. Gaussian
Processes are particularly adept at solving unknown function curves
that map inputs to outputs [52]. In contrast to supervised learning
methods, this allows the relationship between an input and output to
be solved as a function curve using very little data. As the data set
matures, the Gaussian Process refines the curve to better represent
the new distribution function. This is an extremely powerful method
for early analysis when little is known but the data set. This could
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 31

have applications in pipeline swarms, building up a database of flow

readings to fit a curve that accurately depicts flow cycles which could
inform decision making surrounding path planning.
32 MOLYNEUX

3. Simulation Platform and Communication Methods.

3.1. Introduction. In order to provide a competitive inspection pro-

cess, it is desirable that the autonomous robots are able to work ef-
fectively together as a swarm. Given pipe diameters can be as little
as 30cm tall, the robots are necessarily small and are therefore un-
likely to be computationally powerful. As such, the robots will likely
lack the means to think intelligently individually so must be able to
make complex decisions from limited information in order to achieve
optimal results independent of manual instruction. To make informed
decisions the robots must overcome an oppressive subterranean envi-
ronment which limits their communication and movement.

Underground networks provide some unique difficulties for autonomous

robots. Strong water flows inherent in pressurised water networks can
slow, or even halt a robot. Network size demands the robots be unteth-
ered which introduces power requirements that currently surpass their
capacity. Leaks and blockages can obstruct access to whole segments of
network, rendering them incapable of inspection. Autonomous inspec-
tion robots must take these considerations into account, which forces
both swarm and path planning algorithms to become more reliant on
a consistent feed of up-to-date information. Above ground these issues
can be mitigated but in an underground network they are compounded
by strenuous communication capabilities.

A strong communication method helps swarms share information about

their environment. Typically this is a simple process that can be done
with acoustic exchanges or pheromones, both of which can convey a
simple message quickly. Unfortunately, underground pipeline networks
provide a particular challenge in this regard. Because of the density of
the surrounding earth and the variability of the pipe conditions, typical
communication methods are rendered ineffective. Without the range
customarily afforded to swarms, the robots in the system are unable
to make use of what little information they can gather as they inspect
a network. By demonstrating an effective communication process, im-
provements in the robots decision making processes and ability to work
together are made evident.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 33

This chapter details the custom simulation platform presented by Par-

rott, C. et al. [44] for autonomous inspection of pipeline networks.
Three path planning algorithms are presented to provide a basis to ex-
plore efficiency solutions, and the effects of communication ranges are
explored in a manner consistent with the existing stigmergy results. Fi-
nally, realistic ultrasonic and acoustic communication types and ranges
are added to the simulation, to get a realistic representation of the fre-
quency of communication in larger networks.

3.2. Simulation. The nature of pipeline inspection creates difficulties

in exploring proposed solutions — the scale of the networks demands
high investment for calculable results and testing must be non-invasive.
Similarly, implementing robots in water pipelines requires hardware de-
velopment and potential infrastructure upgrades. Given the abstract
nature of the problem, simulation platforms are the only medium ca-
pable of demonstrating the feasibility of autonomous inspection. Mod-
elling the robots in a network as point agents removes the necessity for
immediate hardware expansion while an accurate network reconstruc-
tion allows for the exploration of solutions.

The simulation detailed here, in Section 3.2, continuing until Section

3.3, is an in depth description of the simulation described in the work
of Parrott, C., Dodd, T., Boxall, J.B. and Horoshenkov, K. (2020)
‘Simulation of the Behaviour of Biologically-Inspired Swarm Robots for
the Autonomous Inspection of Buried Pipes’ [44]. The work is detailed
here to inform the reader of the technicalities of the simulation as the
work is adapted and expanded throughout the remainder of the thesis
(Section 3.5 onwards). Additionally, it should support the reader in
understanding context surrounding the structure of the code, which
allows for separate behaviours to be kept in separate classes, how the
Time Between Inspection metric is generated and how movement, flow
and communication are modelled.

The presented simulation was built in C++, an extension of the C

programming language, and has been compiled using Microsoft Visual
Studio 2017. The advantage of C++ is that it supports object-oriented
programming, a programming paradigm that allows the manipulation
34 MOLYNEUX

of objects and classes. Not only does this enable the simulation to
model dynamic systems, but it ensures cohesiveness of results. The
simulation is simply compiled beforehand and parameters are chosen
via a user interface which allows core attributes of the simulation to
be kept separate from the selected parameters. In addition, the sim-
ulation uses EPANET, a water distribution modelling software. This
allows the simulation to read in .inp files which contain data from real
water networks for reconstruction. The reconstruction is rendered us-
ing the built-in Windows Application Programming Interface — Figure
1 illustrates the EPANET render alongside the simulation counterpart
with point agents for the Net2 network.

Figure 1. The EPANET render and simulation render

of the Net2 network respectively.

The network itself is reconstructed as an undirected graph wherein ver-

tices and edges represent junctions and pipes respectively. The data
provided by EPANET consists of network positions for each junction
alongside connectivity information which suffices to form the graph.
Additionally, the edges themselves have an ID alongside information
such as their length, diameter and roughness. The dynamic water flow
is represented by the demand of water at each junction, for which there
is sufficient data for a 24 hour cycle.

Table 1 details four presented networks and their number of junctions,

pipes and cumulative pipe length. The Net2 (Figure 1) and Net3
(Figure 2) networks are artificial whereas the Branching (Figure 3)
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 35

Table 1. A table detailing the topography of multiple

networks.

Network Junctions Pipes Total Length (m)

Net2 35 40 360.00
Net3 92 117 2157.12
Branching 168 175 2858.50
Looping 510 560 28765.51

and Looping (Figure 4) networks are real pipeline networks in England.

These networks are also present in the paper detailing the simulation
by Parrott, C. et al. [44]. This is because the networks were the only
ones available at the time of the research project.

Figure 2. The EPANET render for the Net3 network.

Figure 3. The EPANET render for the Branching net-

work.
36 MOLYNEUX

Figure 4. The EPANET render for the Looping net-

work.

After the simulation stores the network, the user interface allows for
parameter specification and creates the subsequent output files. Once
the parameters have been inputted, the simulation creates the agents
and their starting positions in the network. From there, the agents
traverse the network and are assumed to be continuously inspecting,
with the behaviour they are following making the decisions based on
accumulated information. A necessary distinction to make is though
the behaviour decides the optimal paths, it is the agents which actively
move through the network and follow the strict laws in place within
the system. For example, a behaviour might tell an agent to enter a
pipe for which the flow is too strong to overcome, in which case the
agent in question will forcibly stop moving and enter a rest state until
the flow allows traversal. If specified in the parameters, the simulation
can replicate dynamic water flows or impose charge requirements on
the agents. As the agents traverse the network, their communication
methods are simulated and each is provided with an individual belief
of the inspection history of the network which assists in the decision
making process. Once the simulation has concluded, the normalised
mean and standard deviation of the average Time Between Inspections
(TBI ) is outputted alongside any other specified metrics.

The Time Between Inspections is a natural measure of the efficiency

of the inspection process. For pipeline inspection the objective is to
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 37

provide global, continuous inspection. The TBI is defined as the av-

erage amount of time between two successive inspections of a specific
pipe. At the conclusion of the simulation, it is possible to find the
mean value and the standard deviation of the network’s TBI. The two
in tandem portray how efficient or biased a behaviour is — a low mean
demonstrates intelligent, continuous inspection, whereas a low stan-
dard deviation shows an unbiased, equal distribution of inspection.

3.2.1. Simulation Detail. The simulation has been designed with mul-
tiple parameters and experiments in mind. As such there are many
functions and applications that require further explanation to relate
meaningful results to real networks. Similarly, due to the experimental
nature of the simulation certain assumptions have been made which
require justification. It is therefore beneficial to provide key elements
of the simulation with more detail and address the reasoning behind
specific implementations.

The base simulation platform presented by Parrot, C. et al. [44] begins

by opening an interface and loading in the following parameters for the
simulation:

• Simulation Duration — the duration in milliseconds a single

simulation will run for. Typically set to 2,419,200,000 to repre-
sent a 28 day month.
• Beginning and End Number of Robots — a simulation set is
formed of a single simulation for each sequential number of
agents.
• Number of Iterations — the number of simulations sets in the
batch.
• Robot Driving Speed — the base speed at which the agents will
move.
• Communication Type — selects the communication method the
agents will use.
• Dynamic Flow System — turns the water flow in the network
on or off.
• Visual Render — determines whether the render will be dis-
played.
38 MOLYNEUX

As elements of the simulation are built on or adapted (Section 3.3

onwards), the following parameters are required to be set also.

• Overwhelming Flow — turns the ability to become lost on or

off.
• SLAM Algorithm — sets the SLAM algorithm to use if lost.
• Flow Intelligence — selects the level of flow intelligence the
agents implement.
• Power System — turns the power system on or off.
• Power Intelligence — selects the level of power intelligence the
agents implement.

Once verified the parameters fall within the bounds of the system, the
program creates directories and files for the results. From here, a sim-
ulation batch is created — a grouping of simulation sets and instances
depending on the iterations specified in the interface. To lower compu-
tation time, the program supports threading, a technique that allows
separate pieces of code to run concurrently instead of consequentially.
In this instance, multiple simulations are able to run in parallel on
individual cores and processors of the computer. As a core finishes a
simulation, it is assigned a new one from the remaining batch. Upon
conclusion of the batch the results are compiled and the normalised
values are printed on the results files for the reader.

A simulation itself begins by reconstructing the selected network from

the .inp file and storing the copy in a class. From here it is possible
to access the pipes and their intersections as links and junctions re-
spectively, with each link having a start and end junction, and each
junction having at least one connected link. If dynamic flow is spec-
ified, the flow information is stored in a separate class as the water
demand of a junction as a function of time.

After the network is stored, the starting positions for the agents are
randomised and the agents created. The agents are assigned an index
and a behaviour which will govern their decision making processes.
Whereas the assigned behaviour uses information to decide a course of
action, the agent class controls the traversal across the network, from
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 39

physically entering junctions and links to calculating movement speeds

in the presence of flow. In addition, the agent class contains functions
that enable the agents to broadcast and receive transmittables, as well
as an ability to measure the water flow of the pipe it is in.

At this point the simulation is executed and the agents placed in the
network. For each link in the network, the Time Since Last Inspection
(TSLI ) is initialised to zero. The Time Since Last Inspection value
for a link represents the time passed since it was last inspected by an
agent. As the agent traverses the network, the TSLI value for each
link is continuously updated and is only reset to zero upon inspection.
Figure 1 demonstrates how the render converts the TSLI value into a
link colour, to illustrate the current pipes most in need of inspection. A
recently inspected pipe is represented by a bright yellow which steadily
gets cooler the longer it goes without inspection.

As the agents traverse the network, a PipelineInspectionLogger class

responsible for recording inspections is continuously called for the du-
ration of the simulation. As an agent enters a pipe, the logger records
the time of the simulation and stores it for each pipe such that at the
culmination of every simulation, the time and location of every single
inspection has been recorded. This enables the TBI value to be cal-
culated at the end of the simulation, independent of changes that can
occur from fluctuating patterns in the behaviours.

To help the behaviours make informed decisions, each agent is given

an individual memory. This memory contains a list of every pipe and
it’s own perceived TSLI value of when that pipe was last inspected. At
the start of the simulation every pipe is given a TSLI of zero, which in-
creases linearly with the simulation time so long as it is not inspected.
On entry to a new pipe, the agent sets the TSLI for that pipe in it’s own
memory back to zero. This information allows the agents to inadver-
tently keep track of pipes they have inspected recently, while actively
seeking to inspect pipes in need of inspection. The agents are assumed
to be capable of communicating their memories to one another, from
which they update each pipes TSLI with the most recent of the two
memories. This helps the agents to provide equal inspections to pipes
40 MOLYNEUX

that would naturally not be frequented as much as others. Figure 5

illustrates the different beliefs a group of eight agents have for the TSLI
values in a network. Each cube segment in the memory bar represents
a pipe in the network and, as with the network render, the colour corre-
lates to the TSLI value. The Actual Time Since Last Inspection bar at
the top shows the cumulative memory for the network, and is a direct
translation of the true values represented on the network render.

Figure 5. The individual memories of a network of a

group of inspection robots.

Similarities in the agent memories point towards a recent exchange of

memory. In Figure 5, Agents 2 and 4 have a near identical memory,
whereas Agent 6 shares the pattern, albeit much cooler. This informs
the user that Agents 2 and 4 have recently passed one another and
communicated their respective beliefs of the network, and that at least
one of them communicated with Agent 6 a little further back. Though
the render is purely visual and has no bearing on the simulation, it
aptly illustrates how the agents are able to exchange information in
order to avoid inspecting recently traversed pipes.

The communication itself is controlled by a communication class with

a broadcast function that can be called as an agent leaves a junction
to enter the next link. It is done so in this manner as at this point,
the agent’s behaviour must have decided on the next link to follow,
and thus the broadcast includes the link about to be inspected in the
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 41

broadcast. Not only does this provide all recipients with the most up-
to-date information, but it inadvertently avoids detrimental behaviours
such as agents on the same link syncing up and following one another.
The broadcast is called when the behaviour signals to the agent that
there is a transmittable to share at the next junction. A transmit-
table is individual to each behaviour, and can be any information the
behaviour needs to communicate in order to inspect as a swarm effec-
tively. Typically this includes the aforementioned agent memories, but
can also include relevant path planning information.

The platform supports multiple communication types with different

properties and ranges. Depending on the communication type selected,
the agent is able to enter a transmission state, pausing the agent in
place for a prespecified amount of time while it communicates. The
transmission mechanic of the simulation is a separate addition to the
base simulation presented by Parrott, C. et al. [44] as this is introduced
in Section 3.9.1. In addition, the communication class is responsible
for calculating all agents in range of the broadcasting agent who can
therefore receive the broadcast. This calculation differs depending on
the communication type and range selected.

The platform can access the flow data from the .inp file to influence
the movement speed of the agents in a link. The flow data available
constitutes a 24 hour cycle that changes every half hour. The flow
in a link is calculated as a demand of the two junctions it connects
from which the flow impact on the movement speed is calculated. If
the impact is against the agent direction it will slow the agent, and if
it goes with the agent direction it will speed up. In addition, if the
flow impact is against the agent direction and exceeds the movement
speed, the agent enters a sleep behaviour. This behaviour stops the
agent moving for half an hour, before waking to check if the new flow
velocity is traversable. This behaviour is repeated until the flow ve-
locity subsides enough to move again, at which point the behaviour
returns to the original planning behaviour. The behaviour represents
a hardware assumption that robots are able to clamp themselves to a
pipe wall and rest, conserving energy instead of fighting the flow. It is
assumed that the robots can be equipped with flow sensing capabilities
42 MOLYNEUX

and as such, all behaviours are gifted the ability to measure the flow
velocities for the connected links at the agent’s current junction. This
enables the agents to avoid entering a link that immediately traps it in
a sleep cycle.

Figure 6. The simulation render of Net2 showing the

TSLI spectrum and the dynamic flow rates.

Figure 6 shows the simulation render of the TSLI map against the flow
velocity map. The links in the flow velocity render are small triangles,
with the thicker end representing the direction the flow is heading. The
thickness and colour of the triangle represents the flow strength. As is
evident from the TSLI map, the topmost section has gone the longest
without inspection as it has the coolest colours. The flow map advises
this is due to the strength of the flow in the connecting links slowing
agents reaching the subsection.

Once the simulation is complete the Time Between Inspection metric

is calculated. The time between each inspection event is calculated
from the PipelineInspectionLogger records before being averaged and
normalised with respect to the cumulative network length, L, and the
nominal driving speed of the agents, S. This allows for comparison
of TBI values regardless of the network in question. The normalised
Time Between Inspection value is therefore defined as
TBI × S
TN =
L
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 43

where TBI represents the raw TBI value and TN represents the nor-
malised value. Both the mean and standard deviation are calculated,
with low values indicating a continual inspection and an impartial in-
spection respectively.

Figure 7 depicts a state machine of the platform. A finite-state ma-

chine is a behaviour model which illustrates the transition between
building blocks of states and functions. In this instance it is used to
outline the raw design of the simulation process.

Figure 7. The simulation state machine.

3.3. Behaviours. In the context of the simulation, a behaviour is a

governing intelligence of some description that is designed to improve
the inspection process by calculating the best path an agent can take.
They can be very simple or complex, and will be adapted through the
44 MOLYNEUX

course of this thesis to make use of additional information that becomes

available to them in the hopes of becoming more efficient. There are
four presented behaviours, each with a unique focus or justification.
The first is Stigmergy, as presented in the work of Parrott, C. et al.
[44], a simple behaviour inspired by ant colonies that relies on indirect
coordination to lower inspection times. The second, Greedy Walk is a
computationally inexpensive and fluid stochastic process that plans and
broadcasts it’s path to encourage a more direct coordination. The Ad
Hoc behaviour is an extension of the Greedy Walk that uses graph the-
ory to plan routes that culminate in close proximity of one another to
circumvent communication issues. Finally, the multiple Chinese Post-
man Solution is a computationally expensive approach that finds opti-
mal routes around individual subsections of the network. This section
of the thesis details the behaviours further, alongside their advantages
and pseudocodes.

The Stigmergy indirect coordination method is surprisingly powerful

for such a simple method. The agents traverse a pipe, each with their
own memory of the network and the last time each individual pipe was
inspected. As an agent reaches a junction, it determines which pipes
have been inspected most recently, choosing the least recent one. Only
once a robot comes within communication range of another robot will
it exchange information, updating their maps with the others inspec-
tion route — there is no planned path to encounter other agents and
the interactions occur by chance. The process is repeated each time
another robot is encountered so that there is a continuous process of
information exchange between robots and corrections to their respec-
tive future routes. By following this rudimentary behaviour, unbiased
pipe coverage is achieved as the robot cares little about the topology
of the network and is solely interested in the immediate TSLI values.

The Greedy Walk behaviour aims to introduce forward planning whilst

still being computationally inexpensive. The agent simulates different
random walks that last a pre-specified time, before choosing the walk
which inspected the highest cumulative TSLI. In graph theory, a ran-
dom walk is a stochastic process that consists of successive random
steps, while a greedy strategy is focused on pursuing the highest value
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 45

at every possibility. The Greedy Walk therefore seeks to find the most
optimal solution it can without compromising it’s simplicity and lack
of bias. The random nature of the walks simulated ensure solutions do
not just tend to local optima, but global. As the agent follows the spec-
ified walk, it broadcasts not only the inspection memory, but the path
it is currently undertaking. This informs other agents of pipes that
have been allocated to be inspected in the near future, and therefore
aren’t in need of inspection.

The Ad Hoc behaviour constitutes multiple agents traversing the net-

work in close proximity to one another. An ad hoc network is a self-
sufficient network that does not require full knowledge of an infrastruc-
ture. The core concept of the ‘sweeping net’ is to plan paths for each
robot together such that they are within range of another robot at the
predetermined time of a communal communication exchange, at which
point the path for each is planned again. The robots work together to
ensure that they are not traversing similar trails around one another,
and the whole group tends towards a cyclic net sweeping around the
network. The behaviour uses Dijkstra’s algorithm to ensure cyclic cov-
erage and aggressively seek outlying junctions that might otherwise be
less frequently inspected. This implementation provides a rudimentary
solution to communication problems, whilst allowing complete auton-
omy and a generally even coverage. The efficiency of the algorithm is
highly dependent on the number of robots, increasing as the number
of robots does. In short, a bigger net covers more surface area.

The multiple Chinese Postman Problem is an ancient graph theory

problem wherein a postman wishes to find the quickest route to de-
liver post to every street on their route, which is solvable by making
the graph Eulerian. By splitting a network into regions using a simple
cluster analysis, it is possible to reduce the pipeline graph into a series
of distinct smaller sub-graphs, each with their own individual robot.
Once a robot has been allocated a sub-graph area, it will find a com-
plete solution traversing a much smaller area, reducing computational
complexity whilst completely bypassing communication issues. Despite
the computational complexity, the CPP ensures complete coverage of a
46 MOLYNEUX

network as each link is part of a cluster an individual robot is respon-

sible for, for which the solution dictates each link must be inspected at
least once in a cycle.

3.4. Stigmergy. The Stigmergy behaviour detailed presents the adap-

tations of biological Stigmergy from Bonabeau, E. et al. [4] as imple-
mented by Parrott, C. et al. [44] as a virtual swarm intelligence. At
this point the behaviour is unaltered from the presented work, and
is included as a benchmark with which subsequent algorithms will be
compared. Additionally, the pseudocode is presented.

As previously stated, the Stigmergy behaviour is considered the purest

of the swarm algorithms in this research. The algorithms strength lies
in its use of indirect coordination — the process of achieving sychro-
nized, efficient behaviour by sharing small bits of information. The
most common example of this is found within ant colonies. As ants
venture from their nests to forage for food, they leave small trace
pheromones along the paths they travel. Each distinct pheromone con-
veys a simple message, for example if a path has a bountiful supply of
food, or if there is an enemy colony nearby. Another ant sensing these
pheromones is able to make a quick decision, on its own, without the
help of the swarm. This process allows armies of ants to draw the same
conclusions and make seemingly simultaneous decisions.

To implement this efficiently, the algorithm utilises a ’virtual pheromone’

system, the Time Since Last Inspected values previously mentioned.
No agent relays instructions to another but by sharing one another
memories, the agents are able to achieve indirect coordination. Out-
side of the communication exchange, the agents follow a very simple
algorithm. They travel along a pipe and when they reach a junction
they check their memory. They eliminate the pipe they have just tra-
versed as an option, unless they are at a dead end, and choose the pipe
which has been inspected least recently. If two pipes are equally in
need of inspection, the agent chooses one of them at random. When
dynamic flow is introduced, the agents follow the same algorithm with
the distinct difference that they eliminate any path with too strong a
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 47

flow to traverse as an option. If no options are available at a junc-

tion, for example at a dead end where strong flow now keeps the agent
trapped, the agent will enter the sleep behaviour before waking and
trying again.

Algorithms 1 and 2 details the pseudocode used in the simulation for

the Stigmergy behaviour.

Algorithm 1 Stigmergy Algorithm

Require: CurrentJunction 6= null, DrivingSpeed > 0
1: LastLink ← null
2: loop
3: N extLink ← SelectNextLink
(CurrentJunction, DrivingSpeed, GetT ime(), LastLink)
4: if N extLink 6= null then
5: T imeSinceLastInspection of N extLink ← 0
6: Broadcast all known T imeSinceLastInspection values to
nearby agents
7: Travel along N extLink at DrivingSpeed
8: CurrentJunction ← opposite junction of N extLink
9: LastLink ← N extLink
10: if Only 1 link is connected to CurrentJunction then
11: Synchronise T imeSinceLastInspection values with
Node
12: end if
13: else
14: Sleep for 30 minutes
15: end if
16: end loop

The Stigmergy behaviour provides an unbiased approach to pipeline

inspection. By considering only the immediate links connected to the
junction it inadvertently achieves a high impartiality which, when cou-
pled with the indirect coordination, results in a highly robust and even
inspection process. The limitation of not thinking ahead means the per-
formance has the potential to suffer when flow is introduced, getting
stuck behind high flows it could have avoided. However, the behaviour
is resiliently fair and capable of achieving consistently low TBI values.
Often the behaviour becomes cyclic and segmented towards the end of
the simulation, as certain groups of agents disperse to certain areas.
48 MOLYNEUX

Algorithm 2 Stigmergy SelectNextLink Function

1: function SelectLink((Junction, DrivingSpeed, CurrentT ime, LastLink))
2: CandidateList ← empty
3: BackT rackingCandidate ← null
4: for all Links connected to Junction do
5: if DrivingSpeed > F lowV elocity of Link at CurrentT ime
then
6: if Link 6= LastLink then
7: Add Link to CandidateList
8: else
9: BackT rackingCandidate ← Link
10: end if
11: end if
12: end for
13: if CandidateList 6= empty then
14: SelectionList ← empty
15: M axT imeSinceLastInspection ← 0
16: for all Links in CandidateList do
17: if T imeSinceLastInspection of Link >
M axT imeSinceLastInspection then
18: M axT imeSinceLastInspection ←
T imeSinceLastInspection of Link
19: SelectionList ← empty
20: end if
21: Add Link to SelectionList
22: end for
23: return A random link from SelectionList
24: else if BackT rackingCandidate 6= null then
25: return BackT rackingCandidate
26: end if
27: return null
28: end function

This does not seem to hinder the performance of the agents as they
seem to settle in even areas with little overlap.

3.5. Greedy Walk. Random walks are a stochastic process of sequen-

tial random steps with the end goal of finding a solution within some
prespecified bounds. The Greedy Walk behaviour relies on simplicity
and forward planning in an attempt to find locally optimal paths. The
behaviour begins with multiple iterations of random walks in the sur-
rounding area and calculates the cumulative TSLI value they inspect.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 49

The iterations are random, with each walk a collection of random deci-
sions at each junction. However, the behaviour estimates the time each
link takes to inspect and the path concludes once the cumulative time
exceeds a prespecified value. By weighting each link with the time it
will take to inspect, the Greedy Walk is able to include small outlying
pipes or clusters of small pipes in the path as they take little time to
traverse, but provide a large TSLI gain.

In addition to planning ahead, the Greedy Walk behaviour is able to

broadcast the path it is following to surrounding agents. Though most
behaviours share the inspection memory in a similar manner to Stig-
mergy, the addition of including a future path in the broadcast enables
the agent to avoid future crossover as well. As such, agents in close
proximity either directly inspect together efficiently to clear a small sec-
tion, or one of the agents finds another route and they split to inspect
independent areas. This helps reduce the time between inspections as
the agents are able to plan around one another and prioritise pipes that
might otherwise have been left.

Algorithm 3 details the pseudocode used in the simulation for the

Greedy Walk behaviour.

Algorithm 3 Greedy Walk SelectNextLink Function

1: function Greedy Walk((CurrentJunction, LastLink))
Require: CurrentJunction 6= null
2: if P athList non-empty then
3: if DrivingSpeed < P athList Front (F lowV elocity) then
4: Clear P athList
5: end if
6: end if
7: if P athList empty then
8: P athP lan(CurrentJunction, LastLink)
9: end if
10: N extLink = P athList Front
11: P athList delete Front
12: return N extLink
13: end function

3.6. Ad Hoc. The Ad Hoc behaviour is inspired by the implementa-

tion of ad hoc networks in natural disasters and their ability to remain
50 MOLYNEUX

Algorithm 4 Greedy Walk PathPlan Function

1: function PathPlan((CurrentJunction, LastLink))
2: bestCumulativeT SLI ← 0
3: for all Iterations do
4: while P athT ime < M axW alkT ime
5: for all Links connected to CurrentJunction do
6: if DrivingSpeed > F lowV elocity of Link at
CurrentT ime then
7: if Link 6= LastLink then
8: Add Link to CandidateList
9: end if
10: end if
11: end for
12: if CandidateList non-empty then
13: N extLink ← random Candidate
14: Add N extLink to IterationsList
15: CurrentJunction ← N extLink end junction
16: CumulativeT SLI = CumulativeT SLI +
T imeSinceLastInspection of Link
17: end if
18: end while
19: if CumulativeT SLI > bestCumulativeT SLI then
20: bestCumulativeT SLI ← CumulativeT SLI
21: bestP athP lan ← IterationsList
22: end if
23: end for
24: return bestP athP lan
25: end function

cohesive with little to no infrastructure. The focus of this algorithm is

to have the agents in close proximity to one another at an appointed
time intervals where a communication exchange will occur. By over-
coming the communication dilemma the algorithm aims to actively
have agents work together, with each agent sharing their planned path.
The agents follow a hierachy, with the first planning a route and shar-
ing it’s chosen route with the next agent. This agent then calculates a
path that ends in communication range of the former, before passing
on both plans to the next agent. Each agent will subsequently choose
a path that compliments the swarms current plan. The end result is a
unique swarm behaviour that sees a net of agents initially expand as
they inspect the network before contracting in the concluding stages
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 51

of their path plans so as to be in communication range for the next

’meeting’.

The Ad Hoc behaviour relies heavily on Dijkstra’s algorithm — a

method of finding the shortest path on a graph between two vertices.
Using Dijkstra’s, the first agent is able to find all the junctions in
the network it can reach in time for the next communication ’ping’.
Though computationally more expensive, unlike the A* heuristic, Di-
jkstra’s is complete and will always find the shortest path. It works
by producing the shortest-path tree — fixing a source node, spreading
from this to calculate the shortest path to every node in the network
which ultimately includes the second node in question. Once the first
agent has found all junctions within range, it selects the junction with
the largest cumulative TSLI value. From here, the agent will calculate
which of the surrounding junctions are within communication range,
depending on the type and range selected. Once calculated, the agent
broadcasts a memory, a path it will follow and the candidate list of
junctions in range of it’s own new end point.

Upon receiving a broadcast, the following agent will examine the can-
didate list for any junctions it can reach in time. If this exists it will
follow the same process the first agent and add to the candidate list.
If it cannot reach one of the candidate junctions in range, it will in-
stead find the highest TSLI junction it can reach with Dijkstra’s. By
iteratively passing a junction candidate list through the agents and
implementing a greedy Dijkstra’s approach, the agents tend to clus-
ters of agents actively inspecting together with rare satellite agents
that aggressively seek outlying nodes. It should be noted, that whilst
it seems nonsensical to calculate the shortest paths using Dijkstra’s
to then choose the most expensive of those options, the alternative is
to use a much more computationally expensive longest-path algorithm
before selecting a path that can be completed in time, a shorter one.

Algorithms 5 and 6 detail the pseudocode used in the simulation for

the Ad Hoc behaviour and Dijkstra’s algorithm.

The efficiency of the Ad Hoc algorithm is reliant on the number of

robots in a network — too few and the swarm struggles to maintain
52 MOLYNEUX

Algorithm 5 Ad Hoc Algorithm

1: procedure Ad Hoc((CurrentJunction, N odesInRange, CommunicationT ime))
2: if N odesInRange empty then
3: for all Junctions do
4: ShortestT ravelT ime ←
Dijkstras(CurrentJunction, Junction)
5: if ShortestT ravelT ime ≤ CommunicationF requency
then
6: Add Junction to CandidateList
7: end if
8: end for
9: BestV alue ← 0
10: for all CandidateList do
11: if P athV alue ≥ BestV alue then
12: BestV alue ← P athV alue
13: end if
14: end for
15: T imeSinceLastInspection of P athLinks ← 0
16: for all Junctions do
17: if InCommunicationRange(EndN ode, Junction) = true
then . This function checks if a junction
is in communication range of the node at the next exchange. Add
Junction to N odesInRange
18: end if
19: end for
20: Broadcast all known T imeSinceLastInspection values and
N odesInRange to nearby agents
21: else
22: for all N odesInRange do
23: ShortestT ravelT ime ←
Dijkstras(CurrentJunction, Junction)
24: if ShortestT ravelT ime ≤ CommunicationT ime then
25: Add Junction to CandidateList
26: end if
27: end for
28: BestV alue ← 0
29: for all CandidateList do
30: if P athV alue ≥ BestV alue then
31: BestV alue ← P athV alue
32: end if
33: end for
34: T imeSinceLastInspection of P athLinks ← 0
35: for all Junctions excluding N odesInRange do
36: if InCommunicationRange(EndN ode, Junction) = true
then Add Junction to N odesInRange
37: end if
38: end for
39: Broadcast all known T imeSinceLastInspection values and
N odesInRange to nearby agents
40: end if
41: end procedure
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 53

Algorithm 6 Dijkstras Algorithm

1: function Dijkstras((StartJunction, EndJunction, Graph))
2: while Not at EndJunction
3: for all Links at CurrentJunction do
4: if OptimalDistance of CurrentJunction + LinkLength <
OptimalDistance of OtherJunction then
5: OptimalDistance of OtherJunction ←
OptimalDistance of CurrentJunction + LinkLength
6: end if
7: if LinkLength < ClosestN eighbour and U nvisited then
8: ClosestN eighbour = LinkLength
9: N extJunction ← CurrentJunction
10: end if
11: end for
12: if N extJunction index is invalid then
13: N extJunction ← V isitedSet Front
14: Pop V isitedSet Front to the back
15: else
16: Add CurrentJunction to V isitedSet
17: end if
18: Set CurrentJunction as U nvisited
19: CurrentJunction ← N extJunction
20: end while .
Now every Junction has an optimal distance from StartJunction
and Dijkstra’s follows this back from the EndJunction
21: while Not at StartJunction
22: for all Links do
23: if OptimalDistance of OtherJunction < ShortestP ath
then
24: ShortestP ath = OptimalDistance of OtherJunction
25: N extJunction ← OtherJunction
26: end if
27: end for
28: CurrentJunction ← N extJunction
29: end while
30: return OptimalDistance of EndJunction
31: end function

form, or does so with agents following similar paths. On the other

hand, too many robots causes the ’net’ to saturate, with those lower
in the hierachy seeing increasingly low TSLI returns on their paths.
A sweet spot exists, where the network is sufficiently saturated with
agents to maintain a cyclic net, but not so many agents that the latter
54 MOLYNEUX

agents are surplus. In spite of this, the goal of the Ad Hoc approach
is to limit the frequency of communication required, particularly as
transmission times are introduced into the network. Where other be-
haviours requirement for consistent transmission compounds the effect
of longer transmission times, the Ad Hoc approach suffers only a frac-
tion of the detriment. Similarly, because of the fluidity of the behaviour
— planning around only a sole end point, the behaviour is able to re-
tain consistent inspection times when flow and communication issues
are examined in tandem.

3.7. Multiple Chinese Postmen Problem. The multiple Chinese

Postmen Problem investigates the possibility of autonomous inspection
without communication. The concept has evolved around the Chinese
Postman Problem, a very old mathematical dilemma that attempts to
find the optimal path around a graph such that every edge is traversed
at least once. This can be extrapolated to pipeline inspection, creating
a cyclic path wherein every pipe is inspected by a robot at least once a
cycle. The solution revolves around making the graph Eulerian — that
is, every vertex must have an even degree. This is done by matching up
pairs of odd degree vertices and creating an additional artificial path
between them consisting of current real edges that can be traversed
twice. Once the network is Eulerian, it is possible to create a shortest
cycle that traverses each real and artificial path once using Fleury or
Hierholzer’s algorithms — algorithms for finding paths around Eulerian
networks.

The issue for the Chinese Postman Solution revolves around it’s com-
putational expense in larger graphs. As the graph size increases, so
too does the number of odd degree vertices. The combination of odd
degree vertices that gives the optimal pairings is iteratively calculated
by examining the lexicographic order — a technique for ordering and
swapping variables in a manner that ensures every possible combination
is examined. Therefore, the combination of odd degree vertex pairings
that finds the optimal articial path introductions increases exponen-
tially as the network grows. 12 odd degree vertices means there are
12! = 479, 001, 600 lexicographic combinations of odd pairings. Given
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 55

that the Looping network in Figure 4 has 296 odd degree vertices, a
smarter implementation is necessary.

Figure 8. An example of how a DMA split is intro-

duced to divide the Looping network with four agents.

Figure 8 illustrates how the network can be divided using District

Metred Area analysis. At the beginning of a simulation, each junction
and link in the network is assigned to a robot based off its proximity
to the robot’s starting position — it joins the cluster of the robot it is
closest to using Dijkstra’s. Once the clusters have been assigned, the
boundary links between different clusters are added to both clusters
they join by adding the other junction end to each cluster. This ensures
every link is accounted for, and every cluster is a connected graph. Not
only does this help to reduce the computational complexity of a network
but it ensures consistent, local coverage for each cluster.

In addition, a heuristic approach can make use of additional, advan-

tageous information provided by EPANET. Typically, the ID numbers
that junctions are assigned are ordered. That is, a junction who’s ID
is close to another junction’s ID is likely to be close to that junction
in the network. Extending this logic allows us to create the first lexi-
cographic list wherein every odd degree vertex ID is in size order. By
coupling off junction pairs with close IDs it is possible to assume that
the cumulative additional artificial paths is close to the optimal, as
the paths introduced will, in the majority, be between junctions close
together. By introducing this simple heuristic approach whenever the
56 MOLYNEUX

number of odd degree vertices in a cluster is deemed excessive, it is

possible to rapidly calculate CPP solutions that otherwise make this
method unusable.

Below the pseudocodes used in the simulation for the CPP behaviour
algorithm are detailed:

Algorithm 7 Multiple CPP SelectNextLink Function

1: function SelectLink((Junction, CurrentT ime, LastLink))
2: if P athList = empty then
3: P athList ← MCPP(CurrentJunction, ClusterAssignment)
4: while P athListF irstEntry 6= CurrentJunction do
5: Remove P athListF irstEntry from P athList and add it
to the back of P athList
6: end while
7: end if
8: N extLink ← P athListF irstEntry
9: Remove P athListF irstEntry from P athList and add it to the
back of P athList
10: return N extLink
11: end function

Algorithm 8 MCPP Function

1: procedure MCPP((CurrentJunction,ClusterAssignment))
2: P athP lanned ← f alse
3: while P athP lanned = f alse do
4: EulerianCluster ← Eulerian(ClusterAssignment) . This
function takes the cluster assignment and makes it Eulerian.
5: SolutionF ound ← f alse
6: while SolutionF ound = f alse do
7: P athList ← FleurysAlgorithm
(CurrentJunction, EulerianCluster)
8: end while
9: P athP lanned ← true
10: end while
11: return P athList
12: end procedure

This algorithm requires no communication to be effective, and the com-

putational costs of running once a path is found are minimal. Without
flow, the algorithm guarantees consistent, cyclic coverage of every pipe
in the system. However, the starting computations required to find the
paths can become extremely expensive without a heuristic approach
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 57

Algorithm 9 Eulerian Function

1: procedure Eulerian((ClusterAssignment))
2: for all Junctions do
3: if Junction degree is odd then
4: Add Junction to OddDegreeList
5: end if
6: end for
7: for all OddDegreeList do . Creates initial lexicographic list.
8: Arrange in size order depending on JunctionID
9: end for
10: Define InverseLexicographic
11: while Lexicographic 6= InverseLexicographic
12: for all LexicographicPairs do
13: AdditionalP aths = AdditionalP aths +
Dijkstras(LexicographicP air)
14: end for
15: if AdditionalP aths < bestCumulativeAdditionalP aths then
16: bestCumulativeAdditionalP aths = AdditionalP aths
17: bestArtif icialP athsAdded = Artif icialP athsAdded
18: end if
19: Lexicographic(LexicographicList)
20: end while
21: end procedure

and, given its rigid one-path solution, it struggles to adapt to flow

changes. Though realistic anyway, it becomes necessary to spread out
the agents at the beginning of the inspection process intelligently. This
improves the behaviours efficiency by ensuring that the region distri-
bution, and therefore the work distribution, are even.

3.8. Simulations. To analyse the efficiency of the proposed behaviours,

a set of base simulations were run to provide a benchmark compari-
son. This simulation was run for a 28 day period 20 times for 1 to 32
robots. The 28 day timestep is chosen to counter any bias from the
chosen starting positions, giving the agents time to break out of their
respective areas, whilst inspecting long enough to provide meaningful
results. The 20 iterations aims to eliminate any anomalies that can oc-
cur in single simulation runs. To determine how efficiently the agents
function as a swarm, the number of agents increases after each set of
simulations, from 1 to 32. This allows us find a critical point at which
58 MOLYNEUX

Algorithm 10 Lexicographic
1: function Lexicographic(((LexicographicList)))
2: for all LexicographicList going backwards do
3: if LexicographicsList[Index] <
LexicographicList[Index + 1] then
4: F irstCharacter ← LexicographicList[Index]
5: F irstIndex = Index
6: end if
7: end for
8: for all LexicographicList above Index do
9: if LexicographicsList[Index] < F irstCharacter then
10: CeilingCharacter ← LexicographicList[Index]
11: CeilingIndex = Index
12: break loop
13: end if
14: end for
15: Swap F irstCharacter and CeilingCharacter
16: for all LexicographicList above F irstIndex do
17: Sort into size order.
18: end for
19: end function
Algorithm 11 Fleury’s Algorithm
1: function FleurysAlgorithm((CurrentJunction, EulerianCluster))
2: CycleSize ← N umberLinksInEulerianCluster
3: for CycleSize do
4: for all ConnectedLinks do
5: if Link also has an Artif icialLink then
6: Add Link to V iableOptions
7: end if
8: end for
9: if V iableOptions is empty then
10: for all ConnectedLinks do
11: Add Link to V iableOptions
12: end for
13: end if
14: Choose random Link from V iableOptions
15: CurrentJunction ← LinkOtherJunction
16: P athList add Link
17: end for
18: return P athList
19: end function
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 59

the introduction of new agents provides little additional benefit as well

as determining which behaviours function better in sparser or denser
environment. For these simulations the driving speed has been kept
consistent so as not to interfere with the TBI.

The simulations were run without communication or dynamic flow in

order to ascertain a base level of inspection. The results have been
normalised and plotted on a log-log scale to illustrate the relationship
between the number of robots and TBI clearer.

Figure 9. A graph showing the relationship between

the Time Between Inspection and number of robots in
the Looping Network.

Initially it is obvious that the more agents in a system, the more fre-
quent the inspections. This is shown by the generally linear downhill
slope of mean and standard deviations in Figure 9. As the number
of agents in a network is increased, the frequency of inspections rises
as there are simply more agents traversing the network at one time.
Early additional agents provide a much higher return likely due to low
numbers of agents in a system being unable to cover the entire sys-
tem. Therefore additional agents provide a bigger benefit than when
the swarm is already large. As the number of agents in the network
reaches 16, the introduction of more agents starts to see diminishing
returns, as is evident by the congested mean values. The network has
become saturated as agents find themselves frequently in close prox-
imity to one another, inspecting pipes that have just been inspected.
The effect is mirrored in the standard deviation of the TBI, where
additional early agents provide a much more significant benefit than
those introduced in the latter stages. Again, this indicates a critical
point where the current swarm is able to inspect the entire network
sufficiently.
60 MOLYNEUX

The mean values for each behaviour are extremely close together, es-
pecially as the network saturates with agents. However, this is to be
expected when the simulation is run without a dynamic flow present.
Without flow, the agents in a network are continuously moving and
inspecting regardless of the paths they take. As such, the mean values
tend to be quite close together. For flow-less simulations, a better indi-
cator of performance is the standard deviation. The standard deviation
of the TBI represents the disparity between the frequency that individ-
ual pipes are inspected. A low standard deviation indicates that most
pipes in the network are inspected with the same frequency, whereas
a high standard deviation highlights a discrepancy in the inspection
process.

With regards to the behaviours themselves the multiple CPP solution

seems to perform the best. This is unsurprising — the algorithm specif-
ically constructs an optimum path around a cluster, which in itself
ensures consistent coverage. As the network is split into clusters, the
mCPP can ensure consistent, cyclic coverage. In addition, the agents
can only overlap at the cluster borders which helps contribute to the low
mean TBI. However, the mCPP is the most rigid of all the behaviours
and therefore most susceptible to flow and other considerations.

The Stigmergy and Greedy Walk behaviour are extremely consistent,

as demonstrated by the perfectly linear lines in both the mean and
standard deviation plots. Though they have higher mean values than
the mCPP, the behaviour outputs competitively low standard deviation
values. Without communication, the agents have only their own TSLI
memory to utilise. Therefore, the agents inspect the areas and pipes
they personally have not inspected yet and as such, tend toward a cycle
around the network. This means the introduction of new agents has no
weight on the decision process of other agents, and subsequently each
follows their own cyclic route of the network. This helps account for
the extremely linear relationship between the TBI and the number of
robots as introducing a new agent is simply introducing a new cycle of
coverage into the system.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 61

Due to issues with the code during the research project, the Ad Hoc
behaviour has not been included in the simulations, and is instead left
as Future Work in Section 6.3.

3.9. Implementing Communication Between Robots. The abil-

ity to communicate is a necessary and powerful tool in biological sys-
tems. Sharing information about the surrounding habitat, usually re-
garding food or danger, is beneficial to each member of the society. An
ideal swarm environment relies on this, with each individual fulfilling
a task and communicating to the swarm both to convey and receive
information as necessary. Unfortunately, due to the extreme variabil-
ity of the pipe conditions and the density of the surrounding earth,
an underground pipeline network attenuates common communication
waves robots can communicate with. Creating a swarm behaviour with
limited communication capabilities is a challenge for the field as ulti-
mately the robots are forced to compromise the routes they can take
to exchange the necessary information to achieve optimal performance.

It is possible to demonstrate the detrimental effect of a lack of com-

munication with a simple addition to the simulation. As aforemen-
tioned, the communication is controlled by a communication class with
a broadcast function that is called. It is therefore easy to add a basic
communication process capable of exchanging the relevant information
agents need to make informed decisions. For the purpose of this specific
simulation, the broadcasting and receiving of a message are considered
instantaneous by any robot within range. Though subsection 3.9.1
introduces a transmission delay to reflect the time needed to commu-
nicate, this simulation is concerned only with the effect of the range.

In this instance, the range constitutes the sum of the length of the
pipes a signal can travel. That is, two agents are deemed in range of
one another if the length of the shortest path between them does not
exceed the range. The communication class calls a basic exploratory
Branch and Bound algorithm, Algorithm 12 to ascertain all junctions
and links within range. Though the simplest experiment would have
detailed the range by a linear radius with the broadcasting agent at
its epicentre, the environment of the network dictates the signals must
62 MOLYNEUX

also traverse the network, as they cannot pass through the pipe walls
or surrounding earth.
Algorithm 12 Communication Branch and Bound Algorithm
1: function Branch and Bound((Junction, InitialBroadcastRange))
2: if At a Junction then
3: Add (Junction, InitialBroadcastRange) to
BroadcastJunctions
4: else
5: DistanceF romLastJunction = LinkT ravel
6: DistanceT oN extJunction = LinkLength − LinkT ravel
7: for all Agents in Link do
8: if Agent in broadcast range then
9: Add Agent to AgentsT oBroadcastT o
10: end if
11: end for
12: Add (StartJunction, InitialBroadcastRange −
DistanceF romStart) and (EndJunction, InitialBroadcastRange−
DistanceF romEnd) to BroadcastJunctions
13: end if
14: while BroadcastJunctions non empty
15: Broadcast ← BroadcastJunctions Front
16: BroadcastJunctions remove Front
17: if Broadcast has not been broadcast to then
18: Broadcast set to broadcast to
19: for all BroadcastConnectedLinks do
20: if Agents in Link in range then
21: Add Agent to AgentsT oBroadcastT o
22: end if
23: BroadcastRemainingRange = Broadcast range
−LinkLength
24: if BroadcastRemainingRange > 0 then
25: Create Iterator until BroadcastRemainingRange
expires
26: Add Agents to AgentsT oBroadcastT o
27: end if
28: end for
29: end if
30: for all AgentsT oBroadcastT o do
31: Agent receive transmittable
32: end for
33: end function

In order to study the effects of the communication range of the robots,

each simulation was run with four different communication ranges as
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 63

a proportion of the cumulative length of the network. The frequency

of a communication exchange in a larger network is much lower than a
smaller network with the same communication range so by implement-
ing the range as a percentage of the network length, the TBI values
are normalised for every network. The ranges used are No Commu-
nication (0%), 1%, 10% and Infinite (100%). Figures 10 and 11
demonstrate the effect that different communication lengths have on
the Time Between Inspection for behaviours in the Looping network as
an increasing number of robots are introduced to the network. As the
multiple Chinese Postman solution was specifically chosen as a com-
parison because of it’s indifference to communication, the effects can
only be analysed on the Stigmergy and Greedy Walk behaviours. It
must be noted that though both Figure 10 and Figure 11 have been
run on the same simulation version for parity, the Stigmergy behaviour
and base simulation remain the same as those presented by Parrott, C.
et al. [44] and as such, Figure 10 will match these results closely.

Figure 10. The effects different communication ranges

have on the Stigmergy Time Between Inspection in the
Looping network.

Figure 11. The effects different communication ranges

have on the Greedy Walk Time Between Inspection in
the Looping network.
64 MOLYNEUX

The most important thing to note when examining the communication

results is the distinction between the mean and standard deviation
TBI results. Whereas the mean is concerned with the overall average
time it took to inspect the network, the standard deviation describes
the consistency of coverage throughout the network. At this stage
it is unsurprising that the mean TBIs remain consistent throughout
— the simulation is not considering flow and therefore the agents are
constantly moving at the same arbitrary speed. This leads to identical
mean TBI results as the mean is concerned with the frequency at which
pipes are inspected, and the agents are constantly inspecting regardless
of their communication ranges.

To investigate the effects of a simple communication process it is there-

fore much more appropriate to examine the standard deviation of the
TBIs. This illustrates how effectively agents are working together to
attain consistent coverage. A high standard deviation is undesirable as
it indicates a discrepancy between the frequency at which individual
pipes are inspected.

Examining the standard deviation values in Figure 10, it is apparent

that Stigmergy’s coverage remains relatively consistent. As expected,
the behaviour follows the trend that the longer the communication
range, the more effective the behaviour. This is because a larger com-
munication range is directly proportional to a more frequent exchange
of communication. As such, the agents are able to make better in-
formed decisions as to the state of the network. Interestingly the 10%
range is comparable to the infinite communication. Due to the size of
the network (Table 1), the Stigmergy agents are just as affected by the
10% range as the infinite. The 10% range is more than sufficient for
them to make locally optimal decisions, and by the time the network
saturates the network is sufficiently covered by the 10% range regard-
less. The consistency stems from Stigmergy’s relatively simple decision
making process — only considering the immediate surrounding pipes at
a junction unconsciously provides Stigmergy with an unbias approach
to inspection. As communication is introduced, before the agent can
traverse the whole network, it is communicated that a specific region
has already been done and as such the agents implicitly tend toward a
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 65

network partition, with each group providing an even coverage of their

area.

A similar scenario is evident in the Greedy Walk standard deviation in

Figure 11. Instead however there is the unique case where a 1% com-
munication range performs worse than the No Communication range.
This problem is specific to the Greedy Walk behaviour, and is a reflec-
tion on it’s desire to inspect locally optimal routes. A 1% communi-
cation range encourages an exchange of information which statistically
occurs mid-plan. As such, at the conclusion of the current path, two
agents again create a greedy path, that inevitably concludes with them
each inspecting the same paths from their shared memory. As the
communication range becomes sufficiently large for the agents to com-
municate without overlap, the performance increases again.

Figures 10 and 11 illustrate the necessity for an effective communica-

tion system. An irresponsible communication process can inadvertently
reduce the inspection efficiency, whilst a suitable one will inevitably
improve the consistency of inspections. As additional factors such as
water flow or robot power are introduced it is imperative that robots
are able to communicate effectively to negate the additional detrimen-
tal effects.

3.9.1. Acoustic and Ultrasonic Communication. It is evident that an

optimal inspection process requires sufficiently frequent communica-
tion for robots to work together effectively. The main difficulty the
communication process must overcome is the variability of the pipes,
from material to condition of the pipe. The attenuation of radio waves
with which robots might communicate is too high, struggling to con-
vey a signal that details large volumes of information inspection history
and planning information. Other obstacles include multi-path propa-
gation and limited available bandwidth. In contrast, there is a vast
array of literature detailing the advancements in underwater acoustic
communication.

Acoustic waves have an extended range that can travel through water,
but also propagate round corners well despite their slow propagation
66 MOLYNEUX

speed. This is advantageous as the robots do not necessarily have to be

in sight of one another to communicate, meaning robots can communi-
cate from further distances around problematic topological features. In
addition, acoustic methods are well documented and adaptable, with
hydrophones small and cheap.

As well as acoustic sensing, ultrasonic methods show surprising promise

for underwater communication. By operating in the smaller wave-
lengths of ultrasonic frequencies, transmitters and receivers operate
on a lower energy consumption because of their size. This allows for
frequent communication exchange, albeit it at a significantly smaller
range. Ultrasonic waves struggle to propagate round corners and for
the purposes of pipeline communication, require a line of sight to the
receiver.

The benefits of each of these types of communication is that they coin-

cide with common inspection techniques and provide relatively cheap
and small additions to an inspection robot. Similarly, both have doc-
umented applications in pipes of a range of typical materials, from
cast iron to PVC and concrete. The simulation will implement both
methods of communication as viable techniques as each behaviour may
benefit from a differing communication type.

3.9.2. Acoustic and Ultrasonic Implementation. When implementing

communication in the simulation it is important to keep the behaviour
and communication separate. By keeping the communication in a sep-
arate class, the behaviour is able to choose when to call the commu-
nication type to share information, as opposed to being at the mercy
of when the communication class might choose to communicate. This
provides a more intelligent approach as each behaviour shares informa-
tion a different way. The simulation is hard-coded with transmission
times for both communication types, calculated by the amount of in-
formation to be exchange over the speed at which the communication
type can exchange information. As is likely to be the case in a real net-
work, the agents pause the calculated amount of time to communicate
and receive this information.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 67

For the size of the information to be exchanged, it is assumed the In-

spection Memory is the only transmittable passed in. It is also assumed
that the transmittable is communicated via binary code as this is the
quickest and most reliable method. Therefore, to convey time as a func-
tion of days, hours, minutes and seconds requires 28, 24, 60, 60 binary
options or 11100, 11000, 111100, 111100. As such, for each link in the
network, the information to be transmitted is of the size 5+5+6+6 = 22
bits.

For acoustic communication, the aim is to provide a stream of reli-

able data at a long range. Chirp Linear Frequency Modulation is the
acoustic method that provides the longest range in a pipe at 50 me-
tres. However, the communication exchange is slow, with a rate of
approximately 7 bits per second. In addition, the method requires a 10
second preamble before the message pad to inform the receivers they
are about to receive a broadcast, as well as a 10 second post-amble to
inform them the broadcast is over. Examining Table 1, we see that
were the agents to communicate an entire network’s inspection history
in every broadcast, the transmission time might take just under half an
hour. However, another acoustic method, Amplitude Shift Keying pro-
vides a much more competitive bit transmission speed of 6350 bits per
second. This greatly reduces the transmission time to approximately
2 seconds, though it still requires the pre and post amble. As such,
the acoustic communication is modelled with a 22 second transmission
time, a 50 metre range and the ability to go around corners. The range
given corners is again calculated with the Branch and Bound algorithm
— Algorithm 12.

On the contrary, ultrasonic provides a much quicker transmission at the

expense of a competitive communication range. Because of the speed
and size of ultrasonic waves, the range struggles to exceed 10 metres
in a pipe and does not attenuate well round corners. However, the
speed at which information is communicated is rapid even compared
to the Amplitude Shift Keying acoustic method. As such the ultrasonic
method is modelled with negligible transmission time, but with a 10
metre range and an inability to go round corners.
68 MOLYNEUX

3.9.3. Acoustic and Ultrasonic Simulations. To analyse the effective-

ness of acoustic and ultrasonic waves requires a slightly different ap-
proach to the normalised communication ranges. Because the ranges of
acoustic and ultrasonic waves are finite, it is not realistic to model them
as a normalise value dependent on the cumulative size of a network.
In addition, the size of the network is an important factor — the same
number of robots in a large network are likely to be spread further out
than those in a small network. As such, because the communication
range is specified, by definition the frequency with which the robots
will communicate will be higher in a smaller network.

To compare the two ranges, the simulations have been run on the Stig-
mergy behaviour in both the Looping and Net2 networks. The Stig-
mergy behaviour has been selected because it is the behaviour that
communicates most frequently, at the exit of every junction. There-
fore the differences in the communication type, including transmission
time, will be highlighted the most. The Looping and Net2 networks
were selected for these simulations as they are the largest and smallest
networks respectively.

Figure 12. The effects of Acoustic and Ultrasonic

Communication for the Stigmergy behaviour on the
Looping network.

Immediately it is obvious that there is little difference between the two

ranges. As the behaviours are the same, the mean values being iden-
tical is predictable. However, the similarities between the standard
deviations is telling. Specifically, at every intersection the acoustic
communication is waiting for 22 seconds while the ultrasonic transmis-
sion time is negligible. The fact the standard deviations are so similar
means that the acoustic range is supplementing the deficit of the trans-
mission time with a more intelligent swarm inspection. On the other
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 69

Figure 13. The effects of Acoustic and Ultrasonic

Communication for the Stigmergy behaviour on the Net2
network.

hand, the ultrasonic method allows the agents to be continuously mov-

ing to help overcome the disadvantage of not signals not attenuating
around corners.

The standard deviation in Figures 12 and 13 illustrate the effects of

network size with relation to both ranges. As expected, the smaller
network has the lower standard deviation out of the two as a whole.
This is due to the range sizes physically filling a larger portion of the
network to allow more frequent communication. Interestingly in the
larger network both signals are closer to the No Communication seen
in Figure 10 than the infinite communication range, which is a strong
indication of the insufficient range size appropriate for a network that
large. As suspected, in this instance the acoustic method seems to be
better by the smallest of margins, as the reach outweighs the additional
distance the ultrasonic agents are required to cover to find one another
in more sparsely populated network. Similarly, the Looping network
has more corners round which the ultrasonic waves cannot attenuate.
Because of the scale of Net2, ultrasonic waves can reach other junctions
which aids in increasing communications frequency. For a network the
size of the Looping network, it as apparent that ultrasonic agents have
to be in almost the exact same place to communicate effectively.

A final comment must address the infinite communication crossover in

the Net2 network. Despite having no transmission time, the infinite
behaviour is overtaken by the acoustic and ultrasonic methods as the
network saturates. This illustrates the necessity of a disparity between
70 MOLYNEUX

the individual beliefs of agents. In a larger network infinite communi-

cation is good as the agents are spread out and almost inspect their
own areas, or follow a similar cycle out of sync for optimal effective-
ness. However, in a small network the agents are condensed and each
of them having the exact same memory can be detrimental as they
become increasingly aware every pipe is being inspected at all times.
The behaviour can then turn to random selection or agents following
one another and making similar decisions. The effect of having at
least some different beliefs is that not every agent will then make the
same decision, and at least some semblance of the planned behaviour
remains.

3.10. Discussion. The focus of this chapter was to detail the simu-
lation platform presented by Parrot, C. et al. [44] and why it is an
appropriate mechanism to model autonomous pipeline inspection. By
creating a system that can take in real networks and generate point
agents with limited hardware assumptions it has been able to give a
low-cost insight into the challenges of subterranean inspection, particu-
larly highlighting the communication difficulties faced in this particular
environment. In addition, the platform has laid the foundation from
which the rest of the thesis is built by creating a simulation capable
of easy manipulation without interfering with the core mechanics of
the system. This was exhibited in the adaptation of the simulation to
include ultrasonic and acoustic communication methods, the transmis-
sion state and three new planning algorithms.

The chapter has been able to provide a preliminary investigation into

the aptitude of multiple algorithms in a network. The behaviours pre-
sented each aimed to circumvent some pipeline specific issue, but most
importantly provide different truly autonomous methods. Importantly,
the scale of the number of robots gives an insight into network satura-
tion and can hint toward an appropriate number of robots required for
a specified level of inspection. One of the main outcomes of this thesis
is to prove the feasibility of autonomous inspection and the simulation
has shown that continuous inspection is, at least in theory, a possible
and efficient solution.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 71

With regards to the behaviours themselves, it was apparent that the

benchmark implementation of the behaviours gave favour to the mul-
tiple CPP solution. However, as additional considerations were taken
into account, in this instance communication, the other algorithms had
opportunity to improve. Unfortunately the CPP is a rigid solution and
fares increasingly worse the more considerations it must make. It is evi-
dent that an efficient communication method is beneficial to inspection
performance and contributes to creating a robust swarm intelligence, as
demonstrated by the normalised communication ranges in Figures 10
and 11. Similarly, the simulations in Figure 12 and 13 illustrate how
a real application of signal processing is still able to achieve consistent
inspection levels. This is important because, as the thesis progresses,
it is apparent that autonomous inspection becomes more reliant on
gathering and exchanging additional information to remain effective.
72 MOLYNEUX

4. Implementing a Responsive Flow Intelligence.

4.1. Introduction. Water networks provide a unique obstacle to in-

spection robots, with dynamic flows hindering the movements of the
robots. Not only does this physically limit the speed at which inspec-
tions can be completed but, given the already limited communication
ranges between robots, reduces their ability to cooperate effectively.
This results in higher times between inspection for pipes, increasing the
windows in which cracks and blockages can form. A solution to this is
to develop a responsive artificial intelligence, capable of analysing flow
patterns in order to reduce the impact of strong flows on the robots.
Additionally, strong flows have the potential to overcome an agent,
pushing it back down pipes with the strong flow. At this point an
agent can become lost, ceasing to make optimal decisions as it cannot
localise itself to the memory of the network.

This chapter begins by illustrating the detrimental effects dynamic flow

has on the planning algorithms, using the simulation. The chapter then
adapts the simulation to enable strong flows to overwhelm an agent,
putting them in a ‘lost’ state. Subsequently, a simple Simultaneous
Localisation And Mapping algorithm is introduced that the agents can
implement as they become lost to localise quickly. The algorithm re-
quires only the knowledge of how many connected links there are at a
junction to be effective.

This chapter goes on to present a learning algorithm based off of Gauss-

ian Processes for Regression problems to predict flow rates. Incorpo-
rating a measure of flow intelligence allows the behaviours to mitigate
the detrimental effects of flow, and the use of a deep-learning Bayesian
method improves this process as time goes on. It is common for water
companies to have a rough knowledge of the network flows, but there
are often cases where flows are unknown or are not accurate enough
to be meaningful. The two cases are considered — where robots have
prior knowledge of the networks flow, and when they do not. Imple-
menting a supervised learning algorithm helps circumvent these issues
to provide a solution to flow.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 73

The chapter concludes by briefly outlining a novel Flow-Based SLAM

algorithm that uses additional information from the Gaussian Process
to improve the rate at which the existing SLAM algorithm localises.

4.2. Introducing Dynamic Flow. Dynamic flow is undoubtedly the

biggest obstacle autonomous robots face in the inspection process. To
analyse their impact, the individual network’s dynamic data can be
implemented in a 24 hour flow cycle in the simulation. The required
data is available in the EPANET export of a network that is read into
the simulation, but unless specified in the user interface, has no influ-
ence on the agents. With the implementation of flow, the agents follow
the same decision making algorithms as before; the simple difference
is that the robot driving speed is now proportionally influenced by the
flow.

Unfortunately the flow data itself is discrete and changes hourly, which
can cause instant fluctuations as an agent is traversing a pipe. How-
ever, though not a discrete change, it is common for flow to suddenly
fluctuate with the demands of human routine. For example, the collec-
tive early morning shower can cause a spike in flow demand, not too
distant from a discrete leap. Similarly, though the flow data available
in these simulations is repeated every 24 hours due to the availabil-
ity of the data, this fortunately mirrors the monotony of the working
week. In short, the flow data provided for these simulations is real and,
though not ideal, a fair representation of flow demands and properties.

To explore the effects of dynamic flow, Figure 14 shows the Greedy

Walk behaviour run on the Looping network both with and without
flow, for a variety of communication ranges. The communication ranges
have been set to percentages of the cumulative network size again as
this allows the detrimental effects of flow to be isolated from the com-
munication issues it compounds.

Figure 14 demonstrates a clear loss of performance when flow is in-

troduced. The mean represents the overall speed at which agents are
inspecting the network and in each instance the introduction of flow
74 MOLYNEUX

Figure 14. The effects of introducing flow to the

Greedy Walk behaviour for the Looping network.

Figure 15. The effects of introducing flow to the CPP

behaviour for the Looping network.

greatly reduces how mobile the agents are. The mean is increased sim-
ply because the agents are slower at traversing the network as a whole,
and therefore reach new pipes for inspection much less frequently. In
addition, Figure 14 highlights how the Greedy Walk behaviour specifi-
cally could be improved with the addition of flow; as the algorithm uses
forward planning, so long as the flow is traversable, it will continue to
push on despite the slow rate. It is conceivable that introducing a break
function when the flow exceeds a certain value to reanalyse the future
path and instead choose a path of lesser resistance could help improve
the behaviour. Despite this, the behaviour is more fluid than the CPP
solution, which is shown to struggle further with flow in Figure 15.

The CPP algorithm struggles so much when flow is introduced due to

the algorithms stubborn nature. By concerning itself with a topologi-
cally optimal path, the algorithm fails to take precautions for dynamic
flows and their effects on the agents. By splitting up the network, the
algorithm is sure to eventually inspect all of its cluster at some point.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 75

However, a cluster who’s junctions are surrounded by strong flows will

inevitably waste vast amounts of time getting to those junctions. In
addition, the standard deviation sees a vast increase as areas with low
flow are still able to function somewhat effectively. The contrast to
those in high flow areas helps formulate this discrepancy.

In addition, the flow has a huge impact on the standard deviation of the
inspection for the Greedy Walk. This is to be expected, as particular
pipes can have consistently high flows, with extremely infrequent lulls
in which an agent can traverse it. Without accounting for the flow it
is difficult to then achieve similarly low standard deviation values, as
the small windows in which an agent can enter a pipe are not being
prioritised. As such, whole areas of a network can be cut off save for the
rare moment an agent happened to being inspecting nearby during a
lull. The Looping network particularly has topological obstacles, such
as bottlenecks to large segments of the network that are compounded
by an unusually high flow. Without any flow intelligence, a behaviours
ability to navigate flow in a network is dependent on the topology of
the network. Similarly, areas with low flow are likely to be inspected
every time they are passed, which helps cement the disparity between
the inspection frequencies.

The introduction of flow reaffirms the necessity for a behaviour to be

adaptive — the Greedy Walk behaviour suffers but not in the devas-
tating manner of the CPP algorithm. This can be done indirectly, with
behaviours that frequently re-plan their paths, moving in tandem with
the flow. For path planning algorithms, a more beneficial response is
to actively assimilate flow data in the hopes of using it to plan around
the flow.

4.3. Localization. Introducing flow also introduces another issue for

robots in a network. As previously stated, in the instance that the flow
becomes too powerful for an agent, overcoming a realistic moving speed,
the agent will enter the sleep state, becoming immobile and conserving
‘charge’ before waking every 30 minutes to see if the flow has subsided
enough for it to move again. However, it is not realistic to assume
that the clamping mechanism can withstand any determinable force
76 MOLYNEUX

imaginable. There is a likely scenario wherein the flow can become too
powerful, overwhelming the clamped robot and pushing it back down
the pipes with the flow. When this happens, it is important to assume
the agent is aware of the strength of the flow, and the possibility it has
been moved. In this instance, when the agent awakes it will enter a lost
state, traversing aimlessly — that is, with no governing behaviour, until
it enters the communication range of another agent. At this point the
agent receives a broadcast and position of the broadcasting agent, and
is assumed to be able to localise from that, returning to the previous
behaviour.

Without a flow intelligence it is all too frequent that an agent enters

a long pipe as the flow is traversable, only to be rendered immobile or
pushed back by a discrete switch mid-way through the pipe. This is
detrimental to the swarm inspection process, as once a robot becomes
lost, not only will that single robot be unknowingly inspecting a less
than optimal path, but it could potentially be communicating incorrect
inspection and flow data to the rest of the swarm. To avoid this it is
necessary to use the agents flow sensor to ascertain when the flow is
too high, and will have pushed it back. For the purposes of the simula-
tion, the agent enters a sleep behaviour when the flow velocity against
it is equal to it’s driving speed. Therefore the lost state is triggered
when the flow velocity is over double the driving speed of the agents.
Triggering the lost state prevents the agent from broadcasting infor-
mation or inspecting efficiently until it has been ’found’. Therefore it
is necessary to consider a localisation process that can quickly evaluate
the surrounding area and inform the robot of it’s location in order to
quickly return to function.

Localisation algorithms are a subset of Simultaneous Localisation And

Mapping algorithms and typically revolve around ascertaining a precise
location given surrounding landmarks and features by increasing the
certainty of a position as additional information is gathered. Given
the monotony of water pipe networks, it is difficult to to distinguish
distinct features to provide clarity. Though it is possible to analyse
the cracks, leaks or wear of pipes to identify them, this requires high
levels of hardware and an intelligent analyses system. Fortunately,
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 77

most water networks have a unique and random topological map. This
makes topological localisation well suited to inspection robots, as a
quick localisation process can be achieved simply by considering the
topology of the network — in this case, the junctions of a network.
By simply analysing the number of pipes at a junction and comparing
them to the topology of the network, an efficient localisation algorithm
can be implemented that uses little information and allows the robot
to simultaneously inspect as it finds itself.

4.3.1. No SLAM. We first consider the case where there is no proper

localisation algorithm as a benchmark. When an agent becomes lost,
they must still be able to traverse the network in search of an agent.
If the flow were to push an agent to an outlying pipe it could be a
long time before another agent passes within communication range.
In the meantime, the agent would be providing no discernible benefit
to the swarm and would simply be prolonging the time before it can
return to an inspection process. By continuously moving the agent
improves it’s chances of statistically finding another agent. However,
when lost, an agent has no local network knowledge. Therefore, it must
rely solely on what it can see which in this instance, is either the pipe it
is moving forwards in, or the intersection it reaches at the end. Without
a localisation algorithm, the lost behaviour defaults to an agent picking
links at random when faced with a choice at a junction.

Though not specifically a behaviour, the simulation requires a gov-

erning behaviour for the agent to move around a map. As such, the
NoSLAM SelectN extLink function necessary for movement simply re-
turns a random LinkIndex of those links connected to the current
junction. The concept is rudimentary and serves only to fill the ab-
sence of a proper behaviour.

4.3.2. SLAM Algorithm. Underground pipeline networks give the agents

a big advantage when it comes to localisation — their topology. Topol-
ogy based SLAM algorithms can face challenges when they encounter
symmetrical or cyclic problems as repeated identical topological fea-
tures provide no real information that can be used to localise. Because
78 MOLYNEUX

pipeline networks are typically designed in response to demand, partic-

ularly surrounding housing expansion, there is little natural symmetry.
Similarly, there is though cycles exist, they are rarely confined and ex-
ist within surrounding clusters with their own features. As such it is
possible to view pipeline networks as graphs, with their topologies and
accompanying sub-structures appearing uniquely complex. Without
symmetry or cycles, chaining subsequent vertex degrees begins to form
a unique topological landmark, much like a bar code. This provides
opportunities for reliable localisation with the only assumption being
that at an intersection, a robot is able to count the number of pipes
connected to it.

The SLAM algorithm itself is inspired by the work of Worley, R. and

Anderson, S.[59] which presents a Hidden Markov Model localisation
model that uses the number of links at a junction to influence state
transitions and likelihood models to achieve localisation. Because of
the information available to the agents in the simulation environment,
it is beneficial to make use of the junction information in a similar
manner, except in a deterministic fashion. Dealing only in absolutes is
sufficient in this instance for the agents to localise quicker and helps
reduce the computational power of the simulations.

The algorithm itself utilises the stored map and junction information
to correlate to the current junction it is at. As the agent continues
through the SLAM process, it can store the index of each link it chooses
for future reference. An important distinction to be made is though
the simulation assigns a connected link an index with relation to the
junction, this is unavailable to the lost agents. Because they are lost,
they do not know the direction they arrive at the junction from, so
cannot localise the ’correct’ index assigned by the simulation. The
index is a matter of perspective, however, and by storing the index of
each link from the direction the agent arrived at the junction, upon
conclusion of the SLAM algorithm it is possible to reconstruct the
path taken from when the agent initially became lost. Not only is a
perspective based index important for the SLAM behaviour, it also
allows the agent to continue inspecting as it progresses through the
SLAM process. As the agents can reconstruct their entire path, simply
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 79

storing the local TSLI values allows them to update their memory upon
conclusion of the algorithm. Though while actively localising the agent
is unable to broadcast reliable TSLI values, upon completion it can still
contribute to the swarm memory of the system.

Though recording the index entries remains important for reconstruct-

ing the SLAM path at the end, the algorithm itself is also concerned
with the degree of the junctions it passes. Once lost, the algorithm be-
gins by having the agent count the degree of the junction it is initially
at, and removing any junctions who’s degree does not match from a
potential starting junctions list. It then chooses a link who’s flow it
can traverse and follows that path until it reaches the next junction.
The agent counts again, and removes any junction from the starting
list who is not connected to a junction of a matching degree. This con-
cludes the initial step — once the agent reaches the second junction, it
has a bearing and begins the second phase.

Having a bearing means the agent is able to concretely assign a per-

spective index to each link as it arrives at a junction based off the
link it has arrived from. For example, upon arrival to the junction it
would assign the leftmost link the index 0. This allows the algorithm
to store a temporary local map of the link taken at a junction. The
process is iterated, with the the link index and junction degree con-
tinuously recorded. At the arrival of each junction, the agent is then
able to examine the initial starting positions list and ask which are
connected to links which, when they follow the same index path, tick
off the same vertex degrees along the way. This rapidly reduces the
list of potential starting positions as, though the very first link index
is unknown, the connected junctions have a strict path and therefore
criteria to follow. Once the starting positions list has reached one, the
process is complete. The agent is able to work out the missing link —
the second junction, and from there reconstruct the entire path from
the only starting position back to the current position.

Algorithm 13 details the pseudocode for the SLAM algorithm:

80 MOLYNEUX

Algorithm 13 SLAM Algorithm

Require: CurrentJunction 6= null
1: for all doJunctions
2: if Junction degree = CurrentJunction degree then
3: P otentialStartingJunctions add Junction
4: end if
5: end for
6: Choose random Link
7: for all P otentialStartingJunctions do
8: JunctionM atch ← f alse
9: for all P otentialStartingJunctions connected Junctions do
10: if ConnectedJunction degree = CurrentJunction degree
then
11: JunctionM atch ← true
12: end if
13: end for
14: if JunctionM atch = f alse then
15: Remove from P otentialStartingJunctions
16: end if
17: end for . End of Initial Phase
18: while P otentialStartingJunctions > 1
19: Choose random Link
20: IndexList add IndexOf Link
21: At N extJunction add N extJunction degree to
JunctionDegreeList
22: for all P otentialStartingJunctions do
23: JunctionM atch ← true
24: for all ConnectedJunctions do
25: for IndexList do size
26: Iterate through IndexList
27: if IterationJunction degree 6= JunctionDegreeList
then
28: JunctionM atch ← f alse
29: end if
30: end for
31: end for
32: if JunctionM atch ← f alse then
33: Remove from P otentialStartingJunctions
34: end if
35: end for
36: end while
37: for all OnlyStartingJunction ConnectedJunctions do
38: JunctionM atch ← true
39: Iterate through IndexList
40: if IterationJunction degree 6= JunctionDegreeList then
41: JunctionM atch ← f alse
42: end if
43: if JunctionM atch = true then
44: JunctionT wo = ConnectedJunction
45: end if
46: end for
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 81

4.4. SLAM simulations. In order to examine the effectiveness of the

SLAM algorithm it has been implemented in two behaviours. The
first, Stigmergy, is demonstrated in Figures 16 and 17 with ranges
constituting 1 and 10% of the network respectively. Additionally the
algorithm has been added to the mCPP solution in Figure 18. The two
behaviours represent opposite sides of the autonomous spectrum, with
Stigmergy effective inspecting anywhere and the CPP solution strictly
designated to one cluster. As such, the CPP requires an additional cou-
ple of tweaks for the condition in which it localises outside of it’s own
cluster. Given the path it plans is dependent on the cluster passed in, it
is necessary to return it to it’s own. As such, it simply implements Di-
jkstra’s algorithm to work out which of it’s cluster junctions is closest,
before returning and beginning it’s cycle again. In a similar fashion,
if the CPP agent wanders into another cluster, either due to the flow
or the localisation algorithm, it is necessary to be able to localise from
the communication of other agents. Otherwise, in the No SLAM case,
there would be no means by which the behaviour could localise. As
such the CPP agents are able to communicate their locations alone, in
the range of 1% of the network.

The Looping network has been selected to demonstrate the efficiency

of the SLAM algorithm as, numerous junctions of different degrees and
complex topological map, it would be the most difficult to localise in.

Figure 16. The effects of implementing a SLAM algo-

rithm with Stigmergy and a 1% communication range in
the Looping network.

When analysing the efficiency of the SLAM algorithm, the most no-
table feature of the simulation runs is the SLAM application in Figure
18. While the Stigmergy’s TBI values drastically drop when using the
82 MOLYNEUX

Figure 17. The effects of implementing a SLAM algo-

rithm with Stigmergy and a 10% communication range
in the Looping network.

Figure 18. The effects of implementing a SLAM algo-

rithm with the mCPP in the Looping network.

SLAM algorithm over random behaviour, the CPP actually finds ben-
efit in not localising at all. As the SLAM behaviour inevitably localises
the agent quicker than a No SLAM approach, the only conclusion that
can be drawn is when flow is introduced, the random No SLAM be-
haviour is actually a more adequate inspection process than the CPP
solution. This is unsurprising when Figure 15 is recalled — the ability
to move fluidly around the network, only piecing together the TSLI
values once localised, is still a better approach than localising quickly
to return to stagnating in a strong flow.

The Stigmergy then provides a better insight into the efficiency of the
SLAM algorithm. As is evident, for both communication ranges, the
SLAM algorithm returns a better TBI. This is because of the speed at
which it can localise to return to the Stigmergy behaviour — a func-
tionally better method of inspection. Interestingly there is very little
difference between the two communication ranges and their SLAM line.
As part of the SLAM algorithm is localising once within communication
range of another agent, it would be conventional to see a larger devia-
tion based on the communication range. This indicates that the SLAM
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 83

algorithm frequently self-localises without the assistance of other agents

in the system.

The topology based SLAM algorithm must therefore be considered an

effective approach to solving the problem of localisation in underground
networks. Not only is it able to make use of the simplest available
information to localise effectively, but it does so quickly and in a man-
ner where the agent is able to store TSLI values so inspection time
is not wasted. Given that in large networks the acoustic and ultra-
sonic communication struggles more, it is also beneficial that there is
little reliance on communication range to be effective. In tandem with
an appropriate flow intelligence, the algorithm can help minimise the
wasted inspection time introduced by dynamic flows.

4.5. A Reactive Flow Intelligence. Before developing a flow in-

telligence it is necessary to examine the information available to the
robots, from which they can build a portrayal of the network. Col-
lating live data as the robots naturally traverse the network can be
used to provide a clearer indication of flow patterns and is particularly
useful for unknown networks. Working with the flow will help navigate
the softer flows, moving quicker and lowering the mean TBI, and is
crucial in keeping the standard deviation TBI low in outlying areas.
Additionally, as with the simulation data itself, many water companies
have accessible, if not complete, flow data, which readings can reaffirm
or contradict depending on its validity. It is therefore desirable that
the flow intelligence be able to build a clear understanding of the flow
of the network regardless of the data it begins with.

In a similar sense, it is apparent the robots must have a basic under-

standing of the topology of the network. Specifically, if a robot is to
make use of the flow patterns of the network, it must be able to differ-
entiate between pipes. Not only does this enable the robot to traverse
the network more efficiently, but it allows the robot to take accurate
readings to further improve the data. It is sufficient to assume the
robots have knowledge of the degree of each junction in a network and
the respective connections to other junctions. This enables the robots
84 MOLYNEUX

to distinguish between individual pipes and consequently, the flow data

specific to that pipe.

The presented flow intelligence revolves around modelling individual

flow patterns for each pipe in a network. Once a robot is able to differ-
entiate between pipes, it can begin to learn the individual flow patterns,
using this information to negate the impact of the flow. The most com-
mon flow fluctuations that must be navigated mirror the demand for
water — for example, a sharp increase in the flow strengths in the early
mornings as people shower before work. Similarly, there is a long lull
through the night as people sleep, resulting in a low flow. The most
appropriate interpretation of the flow data then is to model it over a
24 hour cycle as this best represents the demands on the network, and
will thus prove more beneficiary as the robots learn the flow patterns.

The proposed flow model uses Gaussian Process (GP) methods to de-
fine the relationship between the flow strength and the time of day.
Gaussian regression methods are a stochastic process used in super-
vised learning which focus on generalizing probability distributions to
a function. Using training sets it is possible to fit a continuous model
to discrete data. The principle behind the presented flow intelligence
system is to partition the network into individual pipes and fit a regres-
sion model to each, representing the flow cycle throughout a 24 hour
day. As the robots traverse the network, they can use the regression
model to predict the flow in each pipe, whilst simultaneously collecting
training data that can improve the model. The most important prop-
erty of the process is that, by implementing a regression model, the
Gaussian process can model the flow as a continuous function. This
is an important stipulation as it contradicts the discrete data provided
by water companies for this thesis. The data provided is an hourly ap-
proximation of the flow in individual pipes and is mirrored as such in
the simulation. In real life however, spikes in flow are not instantaneous
and on the hour but rather fluctuate gradually. As robots traverse the
network, a continuous models gradient will better describe the strength
of the flow than a discrete jump.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 85

4.6. The Gaussian Process. Gaussian processes are a non-parametric,

Bayesian approach to regression that assume that every finite collection
of random variables in a stochastic process have a multivariate normal
distribution. The process makes use of prior distribution, typically
assumed to be Gaussian, in tandem with a training set to calculate
a posterior distribution using Bayes theorem. From this, a predictive
distribution can be calculated, which can be used to obtain a point pre-
diction for unknown data. This gives the added benefit that each pre-
diction carries a calculable uncertainty measurement, as the prediction
is calculated using the mean and variance of a Gaussian distribution.

Figure 19. A comparison of the predictive distribution

obtained via Gaussian processes and the distribution for
the training data.

Figure 19 illustrates an application of Gaussian processes. The red dot-

ted line represents the distribution with which the 20 points of training
data was created, g(x) = xsin(x), and is therefore the target replica
function. The green line represents the predictive distribution, with the
grey shaded region representing the expected variance. By inferring a
86 MOLYNEUX

probability distribution, the predictive distribution is able to get re-

markably close to the target function for such a small sample size, with
suspect areas of few data points represented by an increased variance
region.

Removing the requirement to learn exact data as other machine learn-

ing algorithms do makes the process efficient with small data sets, as
well as reducing the computational power required for inference. This
is particularly beneficial to pipeline inspection when there is limited
prior information about the flow of the network as it allows the robots
to quickly learn a fundamental pattern, where other methods would
take longer to provide a beneficial intelligence. In addition, pipes with
strong flows have infrequent opportunities to be inspected, so making
use of limited data sets is essential in modelling the harsher flows too.

4.6.1. Creating the Predictive Distribution. Figure 19 details two simi-

lar predictive distributions, one created with MATLAB’s in-built GPR
function and a simpler, version which makes various noise assumptions
for computational ease. Though MATLAB offers an in-built GPR func-
tion, it is desirable that the custom simulation be self-contained to bet-
ter represent the computational challenges a robot in a network would
face. It is unlikely that having a robot repeatedly return to one of the
limited contact points with the surface every time it was necessary to
update the GPR prediction would be very efficient. Figure 19 demon-
strates that with relatively few data points it is possible to recreate an
extremely close GPR match by eliminating some noise requirements
and using ’basic’ maths. Implementing this in the simulation allows
the agents to sufficiently predict the flow of the network to make near-
optimal decisions, without having to compromise the routes they take
to outsource the computations.

The equations detailed below are presented by Rasmussen, C. E. and

Williams, C. K. I in their work ‘Gaussian Processes for Regression’ [46].

To create a predictive distribution it is intuitive to take the weight-

space view, a Bayesian treatment of the linear model. The development
of a standard regression model in tandem with Bayes rule and some
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 87

Gaussian noise assumptions is sufficient to create an effective predictive

model. The following demonstrates how a predictive distribution is
built.

The Gaussian process begins with a Bayesian analysis of the standard

linear regression model;

(1) f (x) = x| w, y = f (x) + 

where x is the input vector, w is the vector of parameters of the linear

model, f and y are the function value and observed value respectively.
The difference in y and the function values f (x) differ by additive noise
which is assumed an independent, identically distributed Gaussian dis-
tribution with zero mean and variance σn2 :

(2)  ∼ N (0, σn2 ).

To calculate the posterior distribution, we must employ Bayes Rule:

likelihood × prior p(y|X, w)p(w)
(3) posterior = , p(w|y, X) =
marginallikelihood p(y|X)
where X is the matrix of training inputs — the design matrix. The
prior must express the beliefs about the parameters before the obser-
vations are examined, and as such a simple zero mean Gaussian prior
with covariance matrix Σp will suffice:

(4) w ∼ N (0, Σp ).

The likelihood is derived directly from the noise assumption and can
be factored over cases in the training set to reduce the model to

(5) p(y|X, w) = N (X | w, σn2 I).

The normalizing constant, or marginal likelihood, is independent of the

weights and is given by
Z
(6) p(y|X) = p(y|X, w)p(w)dw.
88 MOLYNEUX

Completing the square we recognise that the posterior distribution is

a Gaussian distribution with mean w̃ and covariance matrix A−1 :
1
(7) p(w|X, y) ∼ N (w̃ = 2 A−1 Xy, A−1 )
σn
where A = σn−2 XX | + Σ−1
p .

To make predictions using the posterior probability, all possible pa-

rameter values are averaged before being weighted by their posterior
probability. Therefore, the final predictive distribution is given by:
Z
p(f∗ |x∗ , X, y) = p(f∗ |x∗ , w)p(w|X, y)dw
(8)
1
= N ( 2 x|∗ A−1 Xy, x|∗ A−1 x∗ ).
σn

4.6.2. Function Space and the Predictive Function. Though the weight-
space view allows a direct derivation for the predictive distribution,
the function-space view considers inference in function space to predict
expected values. The calculations for mean, covariance and expected
values required for the robots to understand the flow intelligently are
much more computationally efficient due to their reliance on the kernel
function. The function-space view works under the definition that a
Gaussian process is a collection of random variables, any finite number
of which have a joint distribution. This means a process is completely
specified by its mean function and covariance function and therefore

(9) f (x) ∼ GP(m(x), k(x, x0 ))

where m(x) = E[f (x)] and k(x, x0 ) = E[(f (x)−m(x))(f (x0 )−m(x0 ))].
Specifically, the covariance function used is the squared exponential
covariance function, defined as follows:
1
(10) k(xp , xq ) = exp(− |xp − xq |2 ).
2

In this particular instance, it is possible to assume that the observa-

tions obtained by the simulation will be noise free — they are in fact
simulated values, representing discrete time and flow. However, in real
life the robots are likely to have noise surrounding their observations,
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 89

 which we assume is independently, identically distributed Gaussian

with variance σn2 . As such, the prior on noisy observations is repre-
sented by
cov(yp , yq ) = k(xp , xq ) + σn2 δpq or
(11)
cov(y) = K(X, X) + σn2 I
where K(X, X∗ ) denotes the n × n∗ matrix of covariances evaluated at
all pairs of training and test points. With the noise term it is now pos-
sible to examine the joint distribution of the target values and function
values under the prior and, in deriving the conditional distribution, the
predictive equations can finally be simplified to:

(12) f˜∗ = k|∗ (K + σn2 I)−1 y

(13) V ar[f∗ ] = k(x∗ , x∗ ) − k|∗ (K + σn2 I)−1 k∗

The above equations give the predicted mean and variance for an un-
known point under the predictive function. By consistently updating
the training data with flow data and remodeling the predictive dis-
tribution, robots are able to make more accurate predictions as the
variance drops. This pinching effect is illustrated where the variance
surrounding training data has reduced, though it remains large in ar-
eas with little training data. As the Gaussian process manipulates a
distribution as opposed to exact data, the computational complexity
to calculate the mean and variance values is low, even as the volume
of test data increases. Therefore the model simply improves as time
progresses, with little cost to the robots.

4.7. Simulation Implementation. As aforementioned, the flow in

the network is represented by dividing the network into individual pipes
and using a Gaussian process to define the pipes 24 hour flow cycle.
From this, the robots are better able to calculate the time it takes
to traverse a pipe, and can subsequently plan the most efficient paths
to reduce the average Time Between Inspection. The network of the
flow can be considered as n separate Gaussian processes, as illustrated
90 MOLYNEUX

in Figure 21 which shows the EPAnet representation of Net2’s upper

pipes in Figure 20.

Figure 20. The upper part of Net2.

Figure 21. The flow of the upper part of Net2 repre-

sented as a group of Gaussian priors.

Figure 21 was created with three months worth of inspection data from
the simulation. Obtaining an appropriate amount of training data is
fundamental to the efficiency of the model and repeated recomputa-
tions of the model are likely to slow the robots considerably. It is
already assumed that robots in the network have a means of sharing
their inspection data, either by passing a communication point in the
network and uploading to a cloud or by exchanging as they pass one
another. It is therefore realistic to consider the robots also sharing
the flow data they have measured. Upon doing so, the models can be
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 91

updated with the new training data, before the new predictive distri-
bution is relayed for consideration. The mean and variance flows are
then equally calculable from Equations 12 and 13 respectively.

The flow intelligence implementation in the simulation follows a sim-

ilar format to that of the TSLI memory. Upon entry to a pipe, an
agent will record the time and the flow of the pipe. The time can be
reversed engineered to its 24 hour time using a 24 hour modulo op-
eration. Throughout the simulation an agent will continuously store
the data point and upon completion will ’upload’ this batch of new
training data to ’the cloud’. At this point, the data is added to the
existing pool of training data and the model recalculated. The next
simulation is able to work off the improved model, as it itself collects
new data and as such, each simulation sequentially improves, lowering
the average TBI values.

Given the typically short simulation time (28 days), for convenience
the simulation updates the training data at the end of a simulation, as
this allows a comparison of complete simulations with better or worse
training sets. Alternatively, it is possible to introduce a class that can
export the training data for a MATLAB reconstruction of the predictive
distribution, though due to its continuous nature, the calculations of
expected point values must be done within the simulation; the class
must contain the necessary equations to calculate the expected value
of the flow, which the agents call as they calculate their paths.

Figures 22, 23 and 24 illustrate the development of the Gaussian

model of Pipe 38 in Net2 over three months of training data. Each
data point represents a the time in a 24 hour cycle it was recorded
and the flow of the link at that time. Given each month has 50 data
points, it is remarkable how similar to the third month the first month
is able to get. This illustrates the efficiency of the model in working
with small datasets. Similarly, we see the training data mature in the
second month, as it becomes almost indistinguishable from the third
months data, despite a 50 point separation.
92 MOLYNEUX

Figure 22. The GPR model obtained within the sim-

ulation for pipe 38 using one months worth of collected
data.

Figure 23. The GPR model obtained within the sim-

ulation for pipe 38 using two months worth of collected
data.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 93

Figure 24. The GPR model obtained within the simu-

lation for pipe 38 using three months worth of collected
data.

4.8. Flow Intelligence Simulations. Real life implementation of flow

intelligence demands a Gaussian Process Regression due to its low com-
putational cost and ability to interpret small datasets so accurately. In
addition, GPs have the ability to model continuous data, which flow
readings outside of the simulation will form. However, the simula-
tion data is discrete and cyclic. Additionally the simulation platform
can be adapted to provide the agents with perfect data which would
sufficiently replicate mature training data. The purpose of these simu-
lations are to ascertain how flow knowledge can influence the inspection
process. As such, for a simulation based approach, it is beneficial to
work with a different model.

Given that the flow data is a repeated 24 hour cycle, the flow intelli-
gence records the flow in the same manner a Gaussian Process would.
The time is recorded in milliseconds mod 24 hours and each pipe is still
represented by an individual flow memory of these recordings. How-
ever, when it comes to estimating a future flow value, if an agent has
already inspected that pipe at the specified time, it is aware of the flow
rate with certainty.
94 MOLYNEUX

The model creates a list of each pipe and a 24 hour window. As it

inspects the network, the model fills with certain flow data for future
reference. As the agent plans paths and wishes to predict a future flow
rate, it can cross check the pipes record. If a value is present at the
necessary time, it is assigned as the predicted future flow. However, if
no value is present, the agent seeks the nearest values either side of the
time in question. If they are present, the agent will create a weighted
average based on the time it is closest to, and use this to calculate a
weighted difference between the pair. From here, an expected value is
assigned.

This method allows the simulation to record data in the same manner
as a Gaussian Process would. By iteratively filling out the timesteps
of each pipe, it is able to assimilate to perfect knowledge in a similar
manner to the GP model. The advantage remains that once a perfect
copy of the flow has been established, longer simulations can be carried
out that are similar to a Gaussian model with mature datasets. By
using this model, the flow intelligence simulations in Figure 25 were
run for the Net2 network spanning a month, three months, six months
and a year.

The simulations were run using the Greedy Walk behaviour. With
the exclusion of the Ad Hoc behaviour, the Greedy Walk is the only
behaviour that can make use of the additional flow intelligence. The
Stigmergy behaviour is solely concerned with the immediate flow which
it already has access to, and as such has no need for the additional
intelligence. In addition, the CPP is incapable of calculating a path
outside of the topological optimum. As such, the intelligence was added
to the Greedy Walk and ran for a year to demonstrate the effects of
accumulating sufficient data. Net2 was chosen due to its volatile flow.

The simulation results in Figure 25 are initially extremely surprising.

The No Intelligence case appears to closely resemble the flow intelli-
gence when it has had One Months worth of training data, both in
terms of mean and standard deviation. This is to be expected — Fig-
ure 22 has already demonstrated that the training data after a month
has not yet ’matured’, despite forming an effective model. Therefore
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 95

Figure 25. The effects of flow intelligence in the

Greedy Walk behaviour over the course of a year for
Net2.

it is unsurprising that the No Intelligence and One Month case should

match as the model is still adjusting. However, the flow intelligence re-
sults then seem to follow a trend that indicates a relationship between
increased training set size and worse means and standard deviation
values. As the duration of the simulation increases, the mean TBI val-
ues slowly rise. In addition there is a drastic surge in the standard
deviations of the TBI values.

Contrary to what these results initially indicate, it is apparent that

the flow intelligence is effective. This is indicated by the discrepancy
between the rise in mean values and the astronomical rise in standard
deviation values. The standard deviation surge demonstrates that the
agents have learnt the flow of the network, and use it to continuously
traverse the path of least resistance. This explains why the mean value
also rises — from a mathematical standpoint, continuously inspecting
the same pipes does less to reduce the mean TBI than the aging pipes
do to raise it. As such, there is a slow rise in mean value the longer
the simulation progresses. It is likely the case that proportionally the
12 month intelligence inspects more pipes a month than the single
month one, they just happen to be the same set of pipes along an
easily traversable flow path. This compounds the considerable rise in
standard deviation, as not only does the behaviour have a tendency to
avoid in-need pipes, but it then actively seeks easy pipes to inspect.

The flow intelligence has been undeniably effective, as the standard

deviations in Figure 25 are testament to. What is abundantly clear
96 MOLYNEUX

however, is that a different application is necessary. Though the Greedy

Walk has some merit in adapting to flow and communication types, it
clearly is not a behaviour designed to use the flow. In learning the
network the behaviour is able to exploit it’s own greedy nature, to the
detriment of high flow pipes. It was hoped that by implementing the
Greedy Walk with a time based planning duration that additional flow
knowledge would simply improve the amount of links it could reach
within it’s allotted time, by extension more frequently inspecting in-
need pipes.

As a flow modelling technique, modelling the simulation flow in a dis-

crete, learning manner has been effective. Similarly, a real world imple-
mentation of the Gaussian Process Regression model is likely to yield
increasingly similar results with continuous data as the training set ma-
tures. As the models grow in depth it would benefit from a behaviour
that is able to utilise the flow more intelligently, to ensure a fairer distri-
bution of inspection. An alternative solution is proposed in Chapter 5,
wherein the estimated flow velocity helps form part of Multi-Objective
Genetic Algorithm metric that is weighted with different combinations
to ensure an impartial and efficient inspection.

4.9. Flow SLAM Algorithm. The Flow SLAM algorithm is a poten-

tial extension of the SLAM algorithm detailed in Section 4.3 that uses
the acquired flow data from the inspection process to localise quicker.
Though the SLAM algorithm was efficient in the Looping network,
there are real and artificial networks in which it would undoubtedly
perform worse. As aforementioned, symmetric and cyclic topologies
can introduce loops to the algorithm. To help counteract this, it is
proposed that once the training data has matured, the expected flow
of a pipe can be used to discount it in the localisation process. Using the
information acquired from an existing metric would allow the SLAM
algorithm to more quickly distinguish between pipes, as a further vari-
able separates the pipes. Similarly, any computational expense has
already been calculated in the construction of the initial GPR model.

The agents in a network have internal clocks to record accurate inspec-

tion times and flow readings. Therefore, as an agent is localising, it is
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 97

possible to determine the current flow of the link it is in and use that
as an identifying trait, much like the number of links at a junction. If
an agent measures the flow at it’s current link, it can cross reference
it with any link with sufficient training data in the network for the
specific 24 hour time. If the current flow matches with a links expected
flow, or within the realms of the Gaussian Process standard deviation,
it can be considered. However, if the current flow is far outside the
realms of the expected flow of the link in question, it can be safely
discarded.

Though this is a theoretical algorithm with no basis for discussion, it is

considered for any Future Work extending from this thesis in Section
6.3. If the SLAM algorithm and a Gaussian Process are implemented
in a real life application of the system, it stands to reason that an
additional step introducing readily available information might increase
the efficiency of the SLAM algorithm. Given the speed with which the
training data was collated in Figure 25, it is likely the data would
quickly assimilate to an accurate model to assist with SLAM decisions
with certainty.

4.10. Discussion. This Chapter aimed to demonstrate the unique is-

sues that dynamic flows create when introduced into pipeline networks.
The simple process of slowing an agent is sufficient enough to both
lower the frequency of inspection, but also actively prevent specific
pipes from being able to be inspected. In addition, the flow creates
numerous issues for planning algorithms specifically, as seen with the
huge detrimental impact on the CPP solution. The CPP serves as an
example of a theoretically optimal approach unable to adapt to the
sheer displacement of dynamic flows.

The introduction of overwhelming flow and a lost state also highlighted

the necessity for a localisation algorithm in real applications. The
presented SLAM algorithm is able to dampen the impact of strong
flow by creating a quick localisation process. In addition, the algorithm
relies solely on the topology of the network to localise, which aides in
reducing the reliance on a strong communication range for this purpose.
In spite of this, the performance of the SLAM algorithm does little to
98 MOLYNEUX

increase the performance of the actual behaviours and an active flow

intelligence is required.

The active intelligence is designed based on Gaussian Process Regres-

sion models and their ability to form complex models off little data.
Given the low frequency with which certain pipes in the network can be
inspected, this makes GPRs an ideal modelling platform for dynamic
flows. Specifically, by dividing the network into individual pipes with
their own flow history, a flow model of the entire network was con-
structed with the focus of recording and communicating this data. It
is hoped that by providing agents with an advanced knowledge of the
flow data, it might be possible to circumvent the effects of the flow and
increase the raw frequency of inspections in a network. Though the
models were efficient at creating a precise representation of the pipes,
the Greedy Walk behaviour was unable to take advantage of the in-
creased level of knowledge in the desired manner. A behaviour should
be considered that could make active use of the flow data to seek out
in need pipes to decrease the inspection discrepancies.

Finally, the presentation of a flow based SLAM algorithm seeks to capi-

talise on the success of both the SLAM and Gaussian Process efficiency.
It is hoped that the two together can improve the speed at which an
agent can localise by creating an additional variable with which other
pipes can be distinguished.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 99

5. Creating a Solution to Autonomous Power.

To continuously inspect underground networks of the size of water net-

work, it is necessary for the robots to be untethered — that is, not
connected to a power source. Not only does this allow the robots to
travel past the range a tether might hold them to, but it eliminates
the need for manual interaction. Additionally, the physical pipeline
environment varies, and can often be rugged and sharp on the ground.
Severing a power cord renders the agent stuck and unable to be re-
trieved non-invasively. It is therefore ideal that the robots in a pipeline
network are untethered, which itself introduces issues.

Given the necessarily small size of the robots, fitting in pipes as small
as 30cm, they are unable to hold a large charge. In addition, the agents
are continuously inspecting, often against strong flows that hinder their
movement, and are operating sensing and communication equipment.
As such, the robots batteries quickly drain, rendering them unable to
continue inspecting until they have recharged.

The Chapter begins by detailing the implementation of the power sys-

tem and two charge states into the simulation. The Hydrant charge
state makes use of specified junction ends in a network that are assigned
as hydrants for a quick influx of charge. On the other hand, the Tur-
bine state is based on the assumption that the robots in a network can
clamp themselves to the wall still, this time with a small hydro-electric
turbine that converts strong flow into charge. The process takes longer
than the Hydrant charge, but is able to clamp anywhere.

This Chapter goes on to present a Multi-Objective Genetic Algorithm

that aims to find a balance between the robots desire to continuously
be inspecting, and the necessary requirement of frequently charging.
Multi-Objective algorithms aim to find a mutually optimal weighting
assigned to the metrics in consideration that together return the highest
Pareto optimal value. A Pareto optimal set is a set of weightings where
no weighting can be improved without worsening the performance of
another metric[30]. By assigning weightings to the TSLI gain of a
100 MOLYNEUX

decision, and the expected charge that decision will expend, the Genetic
Algorithm component is able to find a Pareto optimal balance.

5.1. Implementing the Power System. In order to implement the

power system, the agent class in the simulation was given an additional
variable — charge. Without flow, the charge decreases at a steady rate
as the agent traverses the network. Each agents power is initialised to
100% which has enough charge for eight hours of movement. When
flow is introduced, the power decrease is calculated in another manner.
The additional power required to push through a flow is calculated as
the percentage of the flow against the driving speed. Implementing the
power in this manner means traversing a link with a spike in flow is
proportionally detrimental to the difficulty to overcome it.

In the instance that an agent’s charge expires, it enters the Turbine

state. Here the agent is assumed to have clamped itself in a similar
manner to the Sleep state. The Turbine behaviour recharges the agent’s
battery at a rate of 100% per hour as opposed to calculating the charge
by converting the links flow velocity into power as it is assumed that the
flow is always sufficiently high to charge the battery at its maximum
rate. However, when the Turbine state is forcibly entered, it remains
immobile for the whole hour, representing a system reboot due to the
shut down. On the contrary, if the agent chooses to enter the Turbine
state, it is able to specify the time it wishes to remain in the state to
gain sufficient charge.

Additionally, the agents are equipped with a Hydrant charging be-

haviour. Hydrants are a natural entry and exit point into subterranean
networks because of their non-invasive access. As such it is assumed
that agents at hydrant are accessible to the surface, in this instance
for a quick-charge. At the start of the simulation various junctions in
the network are assigned as hydrants. Specifically, the starting junc-
tions are always assigned as hydrants as per simulation they represent
the entry points to the network. Additionally multiple junctions who
have only one connected link can serve as hydrant nodes if specified.
When an agent enters a hydrant junction it charges to 100% in 10 min-
utes before it is placed back into the nework. As opposed to actively
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 101

removing the point agent from the system, it is modelled again as a

sleeping behaviour, with no communication or movement capabilities.
However, if the Hydrant system is selected, it will automatically set a
charge percentage at which the agent exits the current behaviour and
begins to search for a Hydrant charging point. This is a necessary
mechanic to introduce as pipeline networks can be vast and it is not
always a guarantee that a hydrant will be passed in one charge cycle.
The behaviour searches for the nearest hydrant based off it’s current
junction and using Dijkstra’s algorithm.

The simulation interface gives the user the option to specify which, or
both, power systems it would like to implement.

5.2. Creating a Multi-Objective Metric. Once the power system

has been specified, the Multi-Objective metric is created. Multi-Objective
Genetic Algorithms (MOGAs) are optimization problems, aiming to
find the best potential weightings to optimize the shared fitness value.
By iteratively changing the weightings in different simulation runs and
examining pairs that produced a low TBI value, the MOGA can hone
in on a Pareto optimal pairing. In a simple two dimension Genetic
Algorithm, a simple shared fitness function can be defined as:

(14) f (x, y) = f1 (x).w1 + f1 (y).w2

The shared fitness value, f (x, y) is the function to be optimized and

determines the strength of the weighting. The f1 (x) and f2 (y) func-
tions represent the functions in contention, while the w1 and w2 values
represent the assigned weightings.

The MOGA is concerned with increasing the TSLI gain while simulta-
neously reducing the power expenditure. This translates to the follow-
ing shared fitness function:

(15) f (x, y) = (T SLIGained).w1 − (ChargeExpended).w2

In order to have the shared fitness function influence behaviours, the

TSLI metric from which they usually make their decisions is overriden
102 MOLYNEUX

by the shared fitness function. For example, the Greedy Walk be-
haviour would iterate through its paths, accumulating the TSLI gains
and the expected charge it would expend to inspect those pipes. De-
pending on the weightings assigned to the MOGA, the Greedy Walk
chooses the path with the highest shared fitness function.

Once the shared fitness has been created, the initial weighting is as-
signed.

5.3. Creating the Genetic Algorithm. Genetic Algorithms are an

evolutionary algorithm inspired by the natural selection process. They
rely on the philosophy of survival of the fittest to keep parents that
provide good offspring. In this instance, the parents are the weight-
ings assigned to the shared fitness function, while the offspring is the
normalised TBI value at the end of the simulation. If the TBI value is
strong, the parents will be considered higher class than those weight-
ings that caused a low TBI.

The essence of Genetic Algorithms is based on giving stronger parents a

better chance of passing on their qualities. This is done with a range of
methods, but most typically focus around crossover and mutation. At
the end of a pairing when the offspring has been evaluated, the Genetic
Algorithm must choose a new pair of weightings for the subsequent
simulation. However, the algorithm will first skew the likelihood of the
parents being chosen again, depending on the quality of their offspring.

This is enforced mainly through the crossover function. The crossover

operator combines two parent weightings of good quality to form new
offspring. The parents are selected from the population and their off-
spring evaluated. Evaluated offspring have a chance to be added to the
parent population if their evaluated value is deemed high enough. By
iterating through good combinations of parents and implementing a
system where only acceptable pairings are reintroduced to the system,
the population tends towards more optimal solutions.

In addition, each offspring value has a pre-specified chance to mutate to

a different value. Therefore less than optimal values still have a chance
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 103

to be introduced to the population. Though this seems non-sensical

it is extremely important as it gives the algorithm the opportunity to
escape from tending towards a local optima.

Algorithm 14 is a pseudocode presented by Konak, A. Coit, D.W.

and Smith, A.E in a MOGA guide titled ‘Multi-Objective optimiza-
tion using genetic algorithms: A tutorial’ [30]. This illustrates the
methodology clearly.

Algorithm 14 Genetic Algorithm

1: Set t = 1. Randomly generate N solutions to form the first popu-
lation, P1 . Evaluate the fitness of solutions in P1 .
2: Crossover : Generate an offspring population Qt as follows:
3: Choose two solutions x and y from P1 based on the fitness values.
4: Using the crossover operator, generate offspring and add them to
Qt .
5: Mutation: Mutate each solution x ∈ Qt with a predefined mutation
rate.
6: Fitness Assignment: Evaluate and assign a fitness value to each
solution x ∈ Qt based on its objective function value and infeasi-
bility.
7: Selection: Select N solutions from Qt based on their fitness and
copy them to Pt+1 .
8: If the criterion is satisfied, terminate the search and return to the
current population.

With regards to the simulation, the population values are calculated

before the first simulation and updated at the end of a run. Because of
the difficulties presented by concurrently updating the same population
pool, the simulations run consequentially and avoid the threading and
batching processes until the MOGA ends. Only then are the final
weightings passed in to a usual batch of 20 runs to analyse the efficiency
of the MOGAs final weightings.

Due to the extensive simulation time required to run a full simulation

based MOGA cycle of sufficient size to be meaningful, in tandem with
the complications of the COVID-19 pandemic, the power simulations
have not been run to completion. This has been left as Future Work
in Section 6.3.
104 MOLYNEUX

6. Conclusion.

The success of the research project can be examined by considering the

achievements of the thesis against the aims and objectives. The aims
and objectives of this thesis were to anticipate issues in autonomous
pipeline inspection for underground water networks and develop an
advanced swarm behaviour based in path planning. Specifically, the
final behaviour should proactively seek out solutions to the problems
so as to keep high levels of inspection performance. Examining the
contributions of the work presents an opportunity to explore discuss
any achievements relating to the objectives. The Chapter presents a
discussion into the success and contributions of the research project,
potential areas of future work and concluding comments.

6.1. Discussion. Exploring the literature it quickly becomes appar-

ent that the idea of autonomous robotic swarm inspection is a rela-
tively new field and there are no documented uses of the technology
in underground pipeline networks. As such, the technical difficulties
of implementing autonomous inspection robots are unknown. To gain
an insight into the expected challenges the literature reviewed existing
inspection methods and robotic capabilities in pipeline networks. The
main obstacles to autonomous inspection robots were shown to be a
hostile communication environment, strong dynamic water flows and a
lack of existing autonomous systems capable of powering themselves to
sufficiently inspect a network. Early examinations of swarms of inspec-
tion robots in simulations were able to provide an effective inspection
process through swarm intelligence. However, the behaviour was reliant
on good communication, suffered in performance with the introduction
of dynamic flows and did not consider power and charge requirements.
Given the unique combination of issues and their detrimental results,
the research project aimed to develop a governing intelligence for the
swarms that considered all major obstacles and could mitigate the ef-
fects. The scope shifted to examine path planning algorithms in an
attempt to create a more intelligent swarm behaviour. It is these con-
siderations that any contributions can be attributed to.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 105

The contributions of the thesis are discussed below.

• Path Planning Algorithms: The focus of the thesis was to ex-

plore options in path planning for potential solutions to the in-
spection issues. To that end, the thesis presented three path
planning algorithms of varying levels of autonomy.The CPP
had early success in the simulation run without dynamic flow.
In splitting up the network areas to assign robots individual
subsections to inspect, the typical computational constraints of
complete solutions was removed. As such, the CPP was simply
implementing an optimal solution in a graph. The CPP’s fall
from grace proved the necessity for versatile behaviours as it was
unable to adapt to the introduction of dynamic flows. On the
other hand, the Greedy Walk behaviour proved to be robust in
responding to dynamic flow. In the early stages of the research,
it was a competitive inspection model with similar means and
standard deviations to stigmergy and the ability to plan ahead.
However the failings of the behaviour were illustrated when it
was provided with detailed flow information, which it exploited
to frequently inspect easy pipes. It did however serve to illus-
trate how a system could utilise perfect flow data to navigate a
network in a more intelligent manner. At the very least, the two
behaviours provided a comparison to the Stigmergy behaviour,
further validating its use as a swarm intelligence. The Ad Hoc
system was designed to circumvent the communication issues
in underground networks and would in theory be able to utilise
the flow data in a more effective way than the Greedy Walk.
Unfortunately, issues with the code during the research project
means the behaviour cannot be considered a contribution, and
is instead left to future work.
• Simulation Adaptations: Adapting the simulation presented by
Parrott, C. et al. [44] was necessary to better reflect the hard-
ships of underground pipeline inspection. Implementing new
mechanics tests the rigour of existing systems and provides
insight into potential failings. Specific additional mechanics
that were added to the simulation are the transmission time
106 MOLYNEUX

of acoustic communication, the potential to be overwhelmed

by flow and adding a power and charge system to the agents
and the network. Though the changes can be considered sim-
ple or rudimentary, they open the door for intelligent response.
Overwhelming flow provided the opportunity to model SLAM
algorithms, and continuous autonomous inspection will not oc-
cur until a solution to the power issues is found. The simulation
can now be passed forward with the capacity to explore other
ideas in these fields.
• Ultrasonic and Acoustic Communication: The thesis went on
the analyse the effects of communication ranges in pipeline
networks, comparing both normalised and realistic methods.
Where typically a larger communication range was beneficial to
the inspection process, the unrealistic ranges were highlighted
by the implementation of Acoustic and Ultrasonic signalling.
This was necessary to prove realistic methods of communica-
tion were still sufficient to govern an efficient swarm behaviour.
The results made it apparent that the behaviours were still able
to inspect efficiently by sharing their TSLI memories on the rare
occasions they communicated.
• SLAM Implementation: A topology base SLAM algorithm was
presented to relocalise agents that had become lost. The strength
of the algorithm lay in how little information was required for
it to localise quickly — the only real landmarks in underground
pipework are the junctions. The algorithm solves a real world
issue and improved the efficiency of the swarm by returning to
inspection behaviours quickly. Additionally, a novel SLAM al-
gorithm was presented that utilises collated flow cycle data and
the surrounding flow to further narrow down the potential pipe
location, in an attempt to localise quicker. Again, the strength
of the proposed approach lies in it’s simplicity — once sufficient
flow data has been collected, and the models created, this would
be as simple as checking the expected flows at the current time
and cross referencing this with the strength of the flow on the
sensor.
PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 107

• Flow Intelligence: The disruptive nature of dynamic flow causes

every single behaviour becoming less efficient, regardless of com-
munication range or prior proficiency. To counteract this, the
thesis presented a novel flow intelligence method that revolved
around modelling each individual pipe in a network as a flow
pattern. The theory surrounding this relied heavily on Gauss-
ian Process Regression models to prove that the data could
collate quickly into a useable and accurate model. Fortunately
due to the flow data required for the simulation it was possible
to reconstruct a simple flow intelligence that tended towards a
mature, ’perfect’ data set quicker than a GPR might. It was
hoped that providing a behaviour with near perfect flow in-
formation would allow the swarm to navigate tricky flows and
keep a low TBI value. Though the Greedy Walk was unable
to use the flow models to improve the inspection process, the
method illustrated how the information could generated quickly
and inform the behaviours.
• Multi-Objective Genetic Algorithm: Given the charge required
for movement and sensors, the power capacity of autonomous
inspection robots is limited. In order to find the correct bal-
ance between continuous inspection and losing charge, a Multi-
Objective Genetic Algorithm was presented to help the be-
haviours make decisions that would optimise the TSLI they
gained, while conserving power. The algorithm is designed to
work in tandem with the flow data collected to accurately ap-
proximate the energy expended in a path plan. By finding a
Pareto optimal solution between the charge expenditure and
the TSLI gain, an equilibrium could be gained that improves
the efficiency with which the agents traverse the network, de-
laying the need to charge. Unfortunately, the processing power
and simulation time required to find an optimal solution com-
pounded the COVID-19 pandemic and forced the work into
Future Research.
108 MOLYNEUX

6.2. Conclusion. To conclude, this thesis presented many ideas and

solutions designed to improve the effectiveness of path planning algo-
rithms in autonomous robotic swarm inspection. Though the research
vision of finding a behaviour capable of navigating all the difficulties in
pipeline networks has not been realised, progress has been made in cre-
ating solutions that are applicable to elements of water pipeline inspec-
tion. The presentation of an effective SLAM algorithm based solely on
counting the number of junctions now allows robots to localise quickly.
Additionally, the use of Gaussian Processes to model the flow of in-
dividual pipes provides the opportunity for intelligent navigation of a
network in dynamic flow. The path planning algorithms presented in
this thesis provided a good basis with which to explore solutions and
adaptations to the simulation will enable further research into these
areas.

Additionally, it is hoped that the Future Work presented is able to

provide applications in autonomous swarm inspection. The Multi-
Objective Genetic Algorithm provides a solid foundation from which
the power constraints and feasibility can be explored. Similarly, it is
hoped that the Ad Hoc behaviour and other, forward planning systems
are be explored as these are the behaviours most likely to make use of
the flow models and other data picked up in the inspection process.

Ultimately, though the adaptability of the stigmergy swarm behaviour

provides a thorough and even inspection process, it is hoped that this
thesis has illustrated that path planning algorithms are also viable in-
spection processes, and their capacity to integrate with other intelligent
systems might elevate their efficiency further.
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 109

6.3. Future Work. Complications in the research project have left

many areas for considerations of Future Work:

• Simulations should be run with the Ad Hoc behaviour intro-

duced in Section 3.6. The algorithm was primarily designed
to circumvent the communication issues in the network whilst
remaining autonomous. Additionally, it’s path planning func-
tionality in small areas made it the preferred candidate to make
use of the Gaussian flow intelligence capabilities.
• It is recommended that the Multi-Objective Genetic Algorithm
in Section 5 should also be run, as it is possible to track the effi-
ciency of the solution as the genetic algorithm progresses. The
MOGA was expected to bring the work together, using path
planning algorithms capable of adapting to both flow and power
intelligently. The Pareto optimal weighting will undoubtedly
provide the most efficient solution to the power stipulations.
• Additionally, Chapter 4.9 details an extension of the SLAM
algorithm that uses gathered data to aid in the localisation
process. Though this addendum was inspired by the availability
of the Gaussian Process data, any appropriate model of the
flow data is likely to improve the speed at which the algorithm
localises. The additional variable will help distinguish between
otherwise identical pipes and the check can be completed in the
same manner as the number of junction links.
• Similarly, Chapter 4.8 demonstrates the need for a behaviour
that can actively work with flow to ensure an even inspection
process. Given the vast availability of flow data and the speed
with which it can be collated and analysed, the opportunity to
implement intelligent systems is difficult to ignore. It is rea-
sonable to assume the behaviour will tend to having perfect
knowledge of the network with the goal of circumventing the
effects of the dynamic flow, providing a much more efficient
inspection process.
110 MOLYNEUX

References

[1] Ahr, D., Reinelt, G., (2006) ’A tabu search algorithm for the min-max k-
Chinese Postman Problem’, Computers and Operations Research 33, 3403 -
3422
[2] Aleman, R. E., Zhang, X., Hill, R. R., (2008). ’An adaptive memory algorithm
for the split delivery vehicle routing problem’. Springer Science and Business
Media – Heuristics. 16, 441 – 473.
[3] T. Bailey and H. Durrant-Whyte, (2006) ’Simultaneous localization and map-
ping (SLAM): part II,’ Robotics and Automation Magazine, 13(3), 108-117
[4] Bonabeau, E., M. Dorigo, G. Theraulaz, ’Swarm Intelligence: From Natural
to Artificial Systems’, New York, NY: Oxford University Press, 1999.
[5] Breivoll EU. (2023). Breivoll Diagnostics - Pipeline Inspection. Available
at https://breivoll.eu/diagnostics/pipeline-inspection/ (Accessed: January
2023).
[6] Bondy, J.A. and Murty, U.S.R., 1976. Graph theory with applications (Vol.
290). London: Macmillan.
[7] Boyd, S. and Mattingley, J., 2007. Branch and bound methods. Notes for
EE364b, Stanford University, 2006, p.07.
[8] Quinonero-Candela, J., (2005), ’A Unifying View of Sparse Approximate
Gaussian Process Regression’, Journal of Machine Learning Research, 6, 1939-
1959
[9] Chakraborty, A. and Kar, A.K., 2017. Swarm intelligence: A review of algo-
rithms. Nature-inspired computing and optimization, pp.475-494.
[10] Couceiro, M.S., 2017. An overview of swarm robotics for search and rescue
applications. Artificial Intelligence: Concepts, Methodologies, Tools, and Ap-
plications, pp.1522-1561.
[11] Cui, Z. and Gao, X., 2012. Theory and applications of swarm intelligence.
Neural Computing and Applications, 21(2), pp.205-206.
[12] Dantzig, G.B. and Ramser, J.H., 1959. The truck dispatching problem. Man-
agement science, 6(1), pp.80-91.
[13] Davidson, R., 2002. An Introduction to Pipeline Pigging, Pigging Products and
Services Association, 9.
[14] Deng, Y., Chen, Y., Zhang, Y. and Mahadevan, S., 2012. Fuzzy Dijkstra al-
gorithm for shortest path problem under uncertain environment. Applied Soft
Computing, 12(3), pp.1231-1237.
[15] Discover Water. (March 2022). Discover Water - Leaking Pipes. Available at
https://www.discoverwater.co.uk/leaking-pipes (Accessed: January 2023).
[16] Dorigo, M., Bonabeau, E. and Theraulaz, G., 2000. Ant algorithms and stig-
mergy. Future generation computer systems, 16(8), pp.851-871.
[17] Dorigo, M., Gambardella, L.M., (1997), ’Ant Colonies for the Travelling Sales-
man Problem’ Biosystems, 43(2), 73-81,ISSN 0303-2647
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 111

[18] H. Dubreil, Z. Altman, V. Diascorn, J.-M. Picard, and M. Clerc. Particle

swarm optimization of fuzzy logic controller for high quality rrm auto-tuning of
umts networks. In Vehicular Technology Conference, 2005. VTC 2005-Spring.
2005 IEEE 61st, pages 1865 – 1869 Vol. 3, 2005.
[19] Eberhart, R.C and Xiaohui, H., Human tremor analysis using particle swarm
optimization. In Evolutionary Computation, 1999. CEC 99. Proceedings of the
1999 Congress on, 1999.
[20] Eiselt, H. A., Gendreau, M., Laporte, G., (1995). Arc Routing Problems, Part
I: The Chinese Postman Problem. Operations Research. 43(2), 231 – 242.
[21] Echologics. (2023). Echologics - LeakFinder. Available at
https://www.echologics.com/products/leakfinderst/ (Accessed: January
2023).
[22] Fletcher, R., Chandrasekaran, M., SmartBall: A New Approach in Pipeline
Leak Detection, American Society of Mechanical Engineers, IPC2008-64064,
117-133.
[23] Gasparetto, A., Boscariol, P., Lanzutti, A. and Vidoni, R., 2015. Path planning
and trajectory planning algorithms: A general overview. Motion and operation
planning of robotic systems, pp.3-27.
[24] Hydromax. (2006). Set Inspection: Side Scanning Evaluation Technology.
Available at http://www.hydromaxusa.com/SSET.htm (Accessed June 2021).
[25] Hydrosave. (2023). Hydrosave Pipeline Camera Inspections. Avail-
able at https://www.hydrosave.co.uk/services/pipeline-condition-
inspections/pipeline-camera-inspections/ (Accessed: January 2023).
[26] Hydrosave. (2023). Hydrosave Sampling Services. Available
at https://www.hydrosave.co.uk/services/pipeline-condition-
inspections/sampling-services/ (Accessed: January 2023).
[27] Iszmir Nazmi, I., et al., Development of In-Pipe Inspection Robot: A Review,
IEEE Conference on Sustainable Utilization and Development in Engineering
and Technology, 2012.
[28] De Jaegere, N., Defraeye, M., Van Nieuwenhuyse, I. (2023). The Vehicle Rout-
ing Problem: State of the Art Classification and Review. Faculty of Economics
and Business, KU Leuven, Belgium
[29] Karaboga, D., Gorkemli, B., Ozturk, C. and Karaboga, N., 2014. A comprehen-
sive survey: artificial bee colony (ABC) algorithm and applications. Artificial
Intelligence Review, 42(1), pp.21-57.
[30] Konak, A. , Coit, D.W. , Smith, A. E (2006), ’Multi-Objective optimization us-
ing genetic algorithms: A tutorial ’, Reliability Engineering and System Safety,
91, 922-1007.
[31] Lanning, D.R., Harrell, G.K. and Wang, J., 2014, March. Dijkstra’s algorithm
and Google maps. In Proceedings of the 2014 ACM Southeast Regional Con-
ference (pp. 1-3).
112 MOLYNEUX

[32] Li, X., Yu, W., Lin, X. and Iyengar, S.S., 2011. On optimizing autonomous
pipeline inspection. IEEE Transactions on Robotics, 28(1), pp.223-233.
[33] Liu, Z., Kleiner, Y., State of the art review of inspection technologies for condi-
tion assessment of water pipes, Elsevier, Measurement, vol. 46, Issue 1, January
2013.
[34] Mandal, S.K., Chan, F.T. and Tiwari, M.K., 2012. Leak detection of pipeline:
An integrated approach of rough set theory and artificial bee colony trained
SVM. Expert Systems with Applications, 39(3), pp.3071-3080.
[35] Marlow, D., et al., Condition Assessment Strategies and Protocols for Wa-
ter and Wastewater Utility Assets, Tech. Rep., Water Environment Research
Foundation, 2007.
[36] Maaref, M. and Kassas, Z.M., 2020, April. Optimal GPS integrity-constrained
path planning for ground vehicles. In 2020 IEEE/ION Position, Location and
Navigation Symposium (PLANS) (pp. 655-660). IEEE.
[37] Molyneux, R. , Parrott, C. , Horoshenkov, K. (2019), ’An Application of Path
Planning Algorithms for Autonomous Inspection of Buried Pipes with Swarm
Robots’, World Academy of Science, Engineering and Technology, Open Science
Index 153, International Journal of Mechanical and Materials Engineering,
13(9), 555 - 563.
[38] Nestleroth, J.B., 2004. Implementing current in-line inspection technologies
on crawler systems. Technology Status Report, http://www. netl. doe. gov,
pp.1-14.
[39] Noto, M. and Sato, H., 2000, October. A method for the shortest path search
by extended Dijkstra algorithm. In Smc 2000 conference proceedings. 2000 ieee
international conference on systems, man and cybernetics.’cybernetics evolving
to systems, humans, organizations, and their complex interactions’(cat. no. 0
(Vol. 3, pp. 2316-2320). IEEE.
[40] Nuchter, A., Surmann, H., Lingemann, K., Hertzberg, J. and Thrun, S., 2004,
April. 6D SLAM with an application in autonomous mine mapping. In IEEE
International Conference on Robotics and Automation, 2004. Proceedings.
ICRA’04. 2004 (Vol. 2, pp. 1998-2003). IEEE.
[41] Nur Afiqah Binti Haji, Y. et al., Development and Adaptability of In-Pipe
Inspection Robots, IOSR Journal of Mechanical and Civil Engineering, vol. 11,
Issue 4, Aug 2014.
[42] Okamoto Jr, J., Adamowski, J.C., Tsuzuki, M.S., Buiochi, F. and Camerini,
C.S., 1999. Autonomous system for oil pipelines inspection. Mechatronics, 9(7),
pp.731-743
[43] Osterhause, A., Mariak, F., (2005). ’On variants of the k-Chinese Postman
Problem’. Operations Research and Wirtschaftsinformatik, 30.
[44] Parrott, C., Dodd, T., Boxall, J.B. and Horoshenkov, K. (2020) ‘Simulation
of the Behaviour of Biologically-Inspired Swarm Robots for the Autonomous
 PIPELINE INSPECTION WITH AUTONOMOUS SWARM ROBOTICS. 113

Inspection of Buried Pipes’ Tunnelling and Underground Space Technology.

Vol. 101 doi.org/10.1016/j.tust.2020.103356
[45] Picacorp. (2023). Pipeline Inspection and Condition Anal-
ysis Corporation - Water Main Inspection. Available at
https://www.picacorp.com/Services/Water-Main-Inspection (Accessed:
January 2023).
[46] Rasmussen, C.E., Williams, C.K.I., (1996) ’Gaussian Processes for Regres-
sion’, Advances In Neural Information Processing Systems, (8), Chapter 2,
Regression.
[47] Roh, S. G. et al., Actively Steerable Inpipe Inspection Robots for Underground
Urban Gas Pipelines, IEEE International Conference on Robotics and Automa-
tion, 2001.
[48] Ryew, S. M. et al., Inpipe Inspection Robot System with Active Steering Mecha-
nism, IEEE International Conference on Intelligent Robots and Systems, 2000.
[49] Sanchez-Torrubia, G., Torres-Blanc, C. and Gimenez-Martinez, V., 2008. An
eMathTeacher tool for active learning Fleury’s algorithm.
[50] Saravanan, R. and Sujatha, P., 2018, June. A state of art techniques on ma-
chine learning algorithms: a perspective of supervised learning approaches
in data classification. In 2018 Second International Conference on Intelligent
Computing and Control Systems (ICICCS) (pp. 945-949). IEEE.
[51] Sit, H., (2019). Quick Start to Gaussian Process Regression, Towards Data
Science. Available at https://towardsdatascience.com/quick-start-to-gaussian-
process-regression-36d838810319. (Accessed: January 2023)
[52] Schulz, E., Speekenbrink, M. and Krause, A., 2018. A tutorial on Gaussian
process regression: Modelling, exploring, and exploiting functions. Journal of
Mathematical Psychology, 85, pp.1-16.
[53] Taketomi, T., Uchiyama, H. and Ikeda, S., 2017. Visual SLAM algorithms: A
survey from 2010 to 2016. IPSJ Transactions on Computer Vision and Appli-
cations, 9(1), pp.1-11.
[54] Tang, G., Tang, C., Claramunt, C., Hu, X. and Zhou, P., 2021. Geometric
A-star algorithm: an improved A-star algorithm for AGV path planning in a
port environment. IEEE Access, 9, pp.59196-59210.
[55] Theraulaz, G. and Bonabeau, E., 1999. A brief history of stigmergy. Artificial
life, 5(2), pp.97-116.
[56] Thibleby, H., (2003). ’The directed Chinese Postman Problem’, Software -
Practice and Experience, 33: 1081-1096
[57] Thrun, S. and Montemerlo, M., 2006. The graph SLAM algorithm with appli-
cations to large-scale mapping of urban structures. The International Journal
of Robotics Research, 25(5-6), pp.403-429.
[58] Weisstein, Eric W. (2023). Königsberg Bridge Prob-
lem. MathWorld–A Wolfram Web Resource. Available at
114 MOLYNEUX

https://mathworld.wolfram.com/KoenigsbergBridgeProblem.html (Accessed:
January 2023)
[59] Worley, R. and Anderson, S. (2020) ’Topological Robot Localization in a Large-
Scale Water Pipe Network ’, Towards Autonomous Robotic Systems. Springer
International Publishing, 77 - 89.
[60] R. Worley, Y. Yu and S. Anderson,(2020) ’Acoustic Echo-Localization for Pipe
Inspection Robots,’ International Conference on Multisensor Fusion and Inte-
gration for Intelligent Systems 160-165
[61] Xu, Y., Wang, Z., Zheng, Q. and Han, Z., 2012, August. The application of
Dijkstra’s algorithm in the intelligent fire evacuation system. In 2012 4th Inter-
national Conference on Intelligent Human-Machine Systems and Cybernetics
(Vol. 1, pp. 3-6). IEEE.
[62] Yanming, G. et al., (2016), ’Deep Learning for Visual Understanding: A re-
view ’, Neurocomputing, 187, 27-48
[63] Yilin, Y., et al., (2019), ’A Survey on Deep Learning: Algorithms, Techniques
and Applications’, ACM Computing Surveys, 51(5), 1-36
[64] Zhang, S., Dubljevic, S., Pipeline crawler development for mapping gas pipeline
topology, American Control Conference, 2021.