Non-functional requirements and their

impact on AGV based systems

A mixed-method study on the impact of Non-functional re-

quirements on AGV based systems

Master’s thesis in Computer Science and Engineering

(Software Engineering and Technology)

Victor Svensson
Mahan Vahid Roudsari

Department of Computer Science and Engineering

C HALMERS U NIVERSITY OF T ECHNOLOGY
U NIVERSITY OF G OTHENBURG
Gothenburg, Sweden 2023
 Master’s thesis 2023

Non-functional requirements and their

impact on AGV based systems

A mixed-method study on the impact of Non-functional

requirements on AGV based systems

Victor Svensson
Mahan Vahid Roudsari

Department of Computer Science and Engineering

Chalmers University of Technology
University of Gothenburg
Gothenburg, Sweden 2023
Non-functional requirements and their impact on AGV based systems
A mixed-method study on the impact of Non-functional requirements on AGV based
systems
VICTOR SVENSSON
MAHAN VAHID ROUDSARI

© VICTOR SVENSSON MAHAN VAHID ROUDSARI, 2023.

Supervisor: Hans-Martin Heyn, Computer Science and Engineering

Supervisors company: Magnus Andersson Martin Lundh, Consat Engineering AB
Examiner: Gregory Gay, Computer Science and Engineering

Master’s Thesis 2023

Department of Computer Science and Engineering
Chalmers University of Technology and University of Gothenburg
SE-412 96 Gothenburg
Telephone +46 31 772 1000

Typeset in LATEX
Gothenburg, Sweden 2023

iv
Non-functional requirements and their impact on AGV based systems
A mixed-method study on the impact of Non-functional requirements on AGV based
systems
VICTOR SVENSSON
MAHAN VAHID ROUDSARI
Department of Computer Science and Engineering
Chalmers University of Technology and University of Gothenburg

Abstract
The use of Automated Guided Vehicles (AGVs) has grown rapidly in recent years.
The global market was valued at 3.81 billion USD in 2021 and is expected to grow
at a compound annual rate of 10.2% between 2022 and 2030. The efficiency of these
systems in warehouses and production facilities are impacted by the choice of non-
functional requirements (NFRs).

This study is based on a mixed-method approach to assess the impact that these
NFRs have on system throughput, and whether the results differentiate based on
the layout and the fleet size of the system. An interview study and a literature
review were conducted to ascertain what the most essential requirements are for
those systems. The results show that requirements regarding safety and robustness
are the most imperative for the industry.

A simulation tool was created during the study which utilizes a modification of the
A∗ algorithm, based on both space and time. The tool was used to create two
different layouts, used to represent different warehouse scenarios. A smaller layout
(75m x 75m) and a larger layout (150m x 150m) was created. The identified NFRs
were then independently varied and compared to a base case.

The results show that the identified NFRs have a profound effect on system through-
put. It was established that the impact of the different NFRs generally are amplified
in the smaller layout. Decreasing the fleet size leads to a larger, relative decrease
in throughput for the smaller layout due to increased congestion. Increasing the
fleet size has diminishing returns until the system is at max capacity and a fur-
ther increase in the number of agents only leads to more congestion, lowering the
throughput.

Keywords: Requirements Engineering, Non-functional requirements, Automated

Guided Vehicle, AGV, Safety, Robustness

v
Acknowledgements
Firstly, we would like to express our gratitude to Hans-Martin Heyn, our academic
supervisor. We would also like to thank our industrial supervisors at Consat AB,
Magnus Andersson and Martin Lundh as well as the participants in our interview
study. Lastly, we would like to thank our family and friends for their support during
the thesis process.

Victor Svensson
Mahan Vahid
Gothenburg, June 2023

vii
 Contents

List of Figures xiii

List of Tables xv

1 Introduction 1
1.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Significance of the study . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.5 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Background 5
2.1 AGV systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Means of navigation . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Control systems . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.3 Core AGV tasks . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Path-finding algorithms for AGV systems . . . . . . . . . . . . . . . . 8
2.2.1 Dijkstra’s algorithm . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.2 Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.3 The A* algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Non-functional requirements . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.1 Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.1.1 Communication . . . . . . . . . . . . . . . . . . . . . 10
2.3.1.2 Computational time . . . . . . . . . . . . . . . . . . 11
2.3.1.3 Memory allocation . . . . . . . . . . . . . . . . . . . 11
2.3.2 Reliability and Availability . . . . . . . . . . . . . . . . . . . . 12
2.3.2.1 Reliability in AGV systems . . . . . . . . . . . . . . 12
2.3.2.2 Maintenance . . . . . . . . . . . . . . . . . . . . . . 13
2.3.3 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.3.1 Single point of failure . . . . . . . . . . . . . . . . . 14
2.3.3.2 Deadlocks in AGV systems . . . . . . . . . . . . . . 14
2.3.4 Safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.4.1 Safety Standards . . . . . . . . . . . . . . . . . . . . 15
2.3.4.2 The Machinery Directive . . . . . . . . . . . . . . . . 15
2.3.4.3 EN 1525:1997 . . . . . . . . . . . . . . . . . . . . . . 16
2.3.4.4 ISO 3691-4:2020 . . . . . . . . . . . . . . . . . . . . 17

ix
Contents

2.3.5 Scalability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.5.1 Scalability in AGV Systems . . . . . . . . . . . . . . 18
2.3.6 Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.6.1 Security in AGV systems . . . . . . . . . . . . . . . 19
2.3.7 Usability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.7.1 Usability in AGV systems . . . . . . . . . . . . . . . 20
2.3.8 State of the art . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3 Literature Review 23
3.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1.1 Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1.2 Reliability and availability . . . . . . . . . . . . . . . . . . . . 24
3.1.3 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.4 Safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.5 Scalability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1.6 Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1.7 Usability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.1.8 Existing gaps in knowledge . . . . . . . . . . . . . . . . . . . . 28

4 Methodology 31
4.1 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.1.1 Search strategy and analysis . . . . . . . . . . . . . . . . . . . 33
4.2 Interview study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2.1 Data collection . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2.2 Data analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3 Simulation study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

5 Results 41
5.1 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2 Interview study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.3 RQ1: Essential requirements . . . . . . . . . . . . . . . . . . . . . . 48
5.4 Results from the simulations . . . . . . . . . . . . . . . . . . . . . . . 55
5.4.1 Layout 1: 150m x 150m . . . . . . . . . . . . . . . . . . . . . 55
5.4.2 Layout 2: 75m x 75m . . . . . . . . . . . . . . . . . . . . . . . 57
5.5 Statistical significance . . . . . . . . . . . . . . . . . . . . . . . . . . 60

6 Discussion 63
6.1 Research question 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.2 Research question 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.3 Research question 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.4 Generalizable experiences . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.5 Threats to validity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.5.1 Conclusion validity . . . . . . . . . . . . . . . . . . . . . . . . 68
6.5.2 Internal validity . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.5.3 Construct validity . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.5.4 External validity . . . . . . . . . . . . . . . . . . . . . . . . . 70

x
 Contents

7 Conclusion 73
7.1 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

Bibliography 75

A Interview Guide I

xi
Contents

xii
 List of Figures

1.1 Example AGV system in a warehouse setting. . . . . . . . . . . . . . 2

2.1 Example of a centralized AGV system. . . . . . . . . . . . . . . . . . 7

2.2 The Machinery Directive . . . . . . . . . . . . . . . . . . . . . . . . . 16

4.1 Overview of the research process . . . . . . . . . . . . . . . . . . . . . 31

4.2 An illustration of the 3D-A* where z-axis represents the time. . . . . 38
4.3 Overview of the final parameters of the simulation divided into five
categories. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

5.1 Overview of the thematic coding board from Miro . . . . . . . . . . . 45

5.2 Visualization of the affinity diagram. Link . . . . . . . . . . . . . . . 45
5.3 Total system throughput deviation from the base case given each
parameter (Layout 1 - 150m x 150m). . . . . . . . . . . . . . . . . . 57
5.4 Total system throughput deviation from the base case given each
parameter (Layout 2 - 75m x 75m). . . . . . . . . . . . . . . . . . . 59

xiii
List of Figures

xiv
 List of Tables

2.1 Calculated frequencies of AGV subsystem failures according to [1]. . . 13

2.2 Throughput of the system using corrective or corrective and predictive
maintenance [2]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 ISO 3691-4:2020, Examples of requirements for speed control. . . . . 18

4.1 Search terms and related sources . . . . . . . . . . . . . . . . . . . . 34

4.2 Roles of interview subjects . . . . . . . . . . . . . . . . . . . . . . . . 35

5.1 Analysis table for literature review. Continued on the next page . . . 41
5.2 Planguage table for frequency of loading failures . . . . . . . . . . . . 49
5.3 Planguage table for recovery time after loading failures . . . . . . . . 50
5.4 Planguage table for frequency of obstacles in vehicle path . . . . . . . 50
5.5 Planguage table for manual recovery after obstacle in vehicle path . . 51
5.6 Example of planguage requirement for safety . . . . . . . . . . . . . . 51
5.7 Requirement regarding speed control for agent. Distance to object(s)
is 2 tiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.8 Requirement regarding speed control for agent. Distance to object(s)
is 3 tiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.9 Requirement regarding speed control for agent. Distance to object(s)
is 1 tile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.10 Requirement regarding the rated speed for an agent. The rated speed
is set at 2 m/s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.11 Requirement regarding the rated speed for an agent. The rated speed
is set at 1 m/s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.12 Requirement regarding the rated speed for an agent. The rated speed
is set at 3 m/s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.13 Simulation results based on each factor, Layout 1 (150m x 150m) . . 56
5.14 Simulation results based on each factor, Layout 2 (75m x 75m) . . . 58
5.15 Shapiro-Wilk test results per parameter after 5 iterations. (p >
α(0.05) -> Assumes normal distribution) . . . . . . . . . . . . . . . . 60
5.16 Paired t-test results per parameter after 5 iterations. (p < α(0.05)
-> Statistically significant) . . . . . . . . . . . . . . . . . . . . . . . 61

xv
List of Tables

xvi
 1
Introduction

The following chapter is divided into five sections and serves to introduce the re-
search. The first section aims to provide the reader with a problem description for
the research. The second section seeks to explain the purpose of the study followed
by the third section containing the research questions that the thesis aims to explore.
The fourth section aims to investigate the significance of the study and finally, the
fifth section is a thesis outline.

1.1 Problem Description

Automated Guided Vehicles (AGVs) in the context of this thesis are autonomous
robots capable of transporting goods or material, commonly in a production fa-
cility or a warehouse. They navigate with the help of sensors, lasers, and either
reflexive coils or tracks in the floor of such facilities [3]. AGVs form a large and
important part of modern logistics and transport systems [4]. Their use has grown
rapidly in recent years, and the market is expected to be worth around 9.38 billion
USD by 2030 [5]. Core tasks for AGV systems include task allocation, localization,
path-planning, motion-planning, and vehicle management [4]. Depending on the
allocation of intelligence in the system, they are either centralized (controlled by a
central controller) or decentralized (each vehicle makes decisions independently) [4].
According to Ryck et al. [4] most modern systems are centralized. There are several
ways to measure the effectiveness of such a system. Among the most widely used
measurements, is the system throughput [4, 6, 7]. This is a metric that refers to the
number of units delivered within a specified time period.

The throughput of an AGV system is impacted by various factors, such as; the choice
of algorithms for path- and motion-planning and their computational effort, the size
of the AGV fleet and warehouse layout, etc. These factors themselves are also im-
pacted by the choice of non-functional requirements (NFRs). These are requirements
pertaining how well a system must perform it’s functions. They could take the form
of properties or characteristics that the system must exhibit or constraints to which
it must comply. For example, safety directives issued by the European Union are
mandatory to follow if products are to be sold within the union. This entails safety
NFRs regarding collision avoidance, speed limits, etc. Requirements regarding the
reliability of the system could entail NFRs regarding the uptime and mean time be-
tween failures for the vehicles. All of these NFRs ultimately impact the throughput
of the system. However, the field of AGVs has yet to discuss the impact of NFRs

1
1. Introduction

and be studied in the context of requirements engineering. This has been done more
thoroughly in adjoining fields, such as Automated Vehicles (AVs). For example,
Ryck et al [4] mentions issues regarding scalability, robustness, and flexibility for
centralized AGV systems, but does not measure their impact nor investigate these
factors from a requirements engineering standpoint. Research conducted by Javed
et al [8] investigated security aspects, but once again from a more general perspec-
tive.

As mentioned above, the throughput of an AGV system is impacted by factors such

as warehouse layout and fleet size. Therefore, determining that impact in combina-
tion with a set of NFRs is of interest due to the fact that the choice of NFRs and
the properties of the system will be impacted differently in differentiating scenarios.
This thesis will simulate different scenarios and present them in order to fulfill the
upcoming objectives.

The objectives of this thesis are as follows:

1. Explore essential non-functional requirements for AGV systems in a warehouse
setting
2. Evaluate the impact that the identified non-functional requirements have on
the throughput of a system.
3. Evaluate the impact that the different identified non-functional requirements
have in specific, real-world alike scenarios through simulation.

Through this research, we hope to contribute to the understanding of the impact

of non-functional requirements in AGV systems and provide practical insights and
recommendations for optimizing the usage of these algorithms in real-world scenar-
ios.

Figure 1.1: Example AGV system in a warehouse setting.

2
 1. Introduction

1.2 Purpose
The purpose of this study is to examine the necessary requirements for an Au-
tomated Guided Vehicle (AGV) system in a warehouse environment and how they
influence the throughput of the system in different scenarios. The study will identify
the essential non-functional requirements for AGV systems, and assess the impact
of these requirements on the throughput. Additionally, we will validate our findings
by demonstrating the impact of a set of real-world scenarios given a set of different
non-functional requirements. The research aims to enhance our understanding of
AGV systems design and optimization in warehouse settings.

Overall, the purpose of this study is to enhance our knowledge of AGV systems
within a warehouse setting in regard to NFRs and their impact on the system
throughput. The target audience for the thesis is industrial practitioners and re-
searchers within the general field of AGVs (Not only warehouse-related).

1.3 Research Questions

RQ1: Identify essential non-functional requirements for AGV systems within a
warehouse setting. What are the essential requirements needed to ensure the func-
tionality of the system according to the stakeholders and literature?

RQ2: Evaluate the impact of the non-functional requirements found in RQ1 on the
system throughput. How does each non-functional requirement from RQ1 affect the
system throughput?

RQ3 : Evaluate the impact of the non-functional requirements found in RQ1 on

the system throughput given a set of specific, real-world scenarios. Given the results
from RQ1, do differences in warehouse layout and fleet size impact the throughput?

1.4 Significance of the study

The growing importance of AGVs in warehouses and production facilities establishes
a need to assess how different NFRs impact their effectiveness (throughput). Little
research has been done on this subject. The objectives of the study are to explore
the essential NFRs for AGV systems and evaluate their impact on the throughput on
different warehouse layouts and fleet sizes. By doing this, the research will provide
practical insights and recommendations that are relevant to logistics companies, as
well as researchers and practitioners in the field of robotics and automation. The
study’s result could also have broader implications by contributing to improving the
efficiency and effectiveness of such systems in general.

3
1. Introduction

1.5 Thesis outline

The thesis is structured as follows:

1. Introduction: This chapter aims to introduce the research by providing a

problem description, a purpose, research questions, the significance of the
study, and lastly a thesis outline.

2. Background: This chapter aims to introduce relevant background informa-

tion for the research. Sequentially it will introduce the topics of AGVs, path-
finding algorithms, and NFRs.

3. Methodology: This chapter explains the methodology used to conduct the

research. It starts with explaining the overall methodology. After that, it
defines how the literature review, interview study, and the simulation model
were created.

4. Literature review: This chapter contains literature that was deemed to be

of interest to the study.

5. Results: This chapter will present the result of the interview study and the
literature review used to answer RQ1. Sequentially, it will present the NFRs
that will be used for RQ2 and RQ3. After that, results for RQ2 and RQ3
will be presented together with some tests for the statistical power of the re-
sults.

6. Discussion: In this chapter we will discuss the findings of the study and
common validity threats.

7. Conclusion: This chapter contains the authors conclusions regarding the re-
search.

4
 2
Background

The following chapter is divided into three sections. The first section aims to give
the reader a basic understanding of how centralized AGV systems work in general.
The second section aims to explain the underlying theory behind the path-finding
algorithm used for the simulation tool. The third section aims to give the reader
insights into the topic of non-functional requirements (NFRs) and their role in the
context of AGV systems.

2.1 AGV systems

AGVs are automated, driverless vehicles used in industry and intralogistics [4, 3].
AGVs perform a range of tasks ranging from processes for handling goods and ma-
terials in warehouses, to the use as means of transportation[3]. The global market
for AGVs was valued at 3.81 billion USD in 2021 and is expected to expand at a
compound annual growth rate of 10.2% between 2022 and 2030 [5].

2.1.1 Means of navigation

In order to navigate through their surroundings, AGVs use different navigational
procedures [3]. There are AGVs that navigate using physical guidelines, anchoring
points on the floor, and laser navigation. There are several common versions, among
them:
• Inductive guidance tracks, in which a current bearing conductor is set to the
floor. The layout of the warehouse system is divided into sections that can
be switched on and off. Two coils are mounted under the vehicle, inducing a
flowing current. The differential in the current between the coils is a measure of
the deviation from the guidance track, which guides the motor using negative
feedback [3].
• Passively inductive guidance tracks, in which a 5-10 cm wide metal strip is
affixed to the floor. A sensor on the vehicle detects the metal strip and uses
this to guide the motor [3].
• Optical guide tracks, in which a colored strip, distinctive from the floor is
painted. A camera on the vehicle uses edge detection algorithms to calculate
guidance signals to the motor [3].
• Anchoring points in the floor are commonly used as representatives of free
navigation spaces. The route is not physically fixed but realized virtually in a

5
2. Background

computer. To navigate, it relies on dead reckoning as well as using the anchor

points, usually magnets or quasi-active transponders, to take bearings [3].
• Laser navigation is the most prominent representative of free navigation.
Retro-reflecting foil is mounted on walls and pillars and can be read pre-
cisely and at great distances using a rotating laser scanner on the vehicle. The
scanner continuously scans its surroundings and measures the positions of the
fixed reflectors [3] which in turn helps calculate its own position accurately.

2.1.2 Control systems

AGV systems can be centralized or decentralized, depending on the allocation of
intelligence in the system [4]. A centralized system generally consists of three mod-
ules: a warehouse management system, a central controller, and a communication
server [9]. The warehouse management system provides information regarding the
goods, the central controller supervises the AGV by giving high-level instructions
[9], and the communication server receives and transfers information between the
central controller and the AGV [9]. Each AGV is allocated with a vehicle guidance
control system [3]. This could be a single- or multi-board computer, or an indi-
vidually designed computer based on micro-controllers [3]. The central controller
receives and prioritizes tasks or orders from the warehouse management system [10].
Information regarding the location and status of the vehicle is sent from the vehicle
guidance control system to the communication server which provides this to the
central controller. The central controller then facilitates the selection of for example
a vehicle to a certain task or a path which the AGV is to follow [10]. In a decen-
tralized system, there is no central controller that coordinates the fleet of AGVs
[11], instead, each AGV makes individual decisions based on information available
to them [11].

2.1.3 Core AGV tasks

A complete AGV system consists of five core tasks to be able to operate in its en-
vironment [4, 11]. These can be seen below in Figure 2.1

The first core task is task allocation. A set of tasks or orders need to be optimally
distributed to a set of AGVs. Task allocation is a constrained NP-hard problem, in
which the total cost needs to be as low as possible. The problem can be solved in
some cases by meta-heuristic optimization and other heuristics [4].

The second core task is localization. In this task, information about the environment
is communicated to the central controller from individual AGVs. This task is often
already decentralized. Information regarding the position of the AGV can be com-
municated on an individual level to the central controller or to neighboring AGVs [4].

The third core task, path-planning, is used to find the shortest path to a destina-
tion [4]. Path-planning or routing computes a basic, collision-free path using known
information. It involves the representation of the free configuration space and a

6
 2. Background

graph-search algorithm to generate the shortest path, minimizing factors such as

fuel consumption, travel time, and travel distance [4, 12].

Obstacles encountered during operation may require modification of the path using
information about positions, goals, and static paths to avoid collisions and dead-
locks, which is handled by the fourth core task, motion planning [4]. Algorithms
for collisions, deadlock avoidance, and zone control are needed which depend on the
usage of a centralized or a decentralized system [4].

Parallel to the tasks mentioned above, the fifth task, vehicle management controls
and monitors the status of the AGVs. This could include management issues such
as battery lifetime, maintenance requirements, or error status [4].

Figure 2.1: Example of a centralized AGV system.

7
2. Background

2.2 Path-finding algorithms for AGV systems

Path-finding for AGVs refers to the index of a certain parameter (shortest path,
shortest operation time), and the selection of a suboptimal or optimal path that
can be connected from the location of the vehicle to the goal destination [13]. Fur-
thermore, path-finding in an AGV system is a multi-agent path-finding problem
(MAPF) [14]. The problem is such that each entity in the system should reach its
goal whilst avoiding collisions [14]. A common requirement for such problems is
to minimize the sum of costs of the agent’s plans, which is considered an NP-hard
problem. MAPF is described by [14] as:

A MAPF problem with n agents is defined by a tuple (G, S, G) where G = (V, E)

is an undirected graph and S : [1, , n] → V maps an agent to a start vertex and
G : [1, , n] → V maps an agent to a goal vertex.

Time is discretized into time steps. For each step, t, each agent occupies one of the
graph vertices. An action, a in a classical MAPF is a function a: V → V such that
a(v) = v ′ means the agent’s current location is v and the next time step is v. There
are two types of actions, wait and move. A sequence of actions πi = (a1 , , an ) is a
single-agent plan. The number of actions in the plan defines its cost C(πi ).
∑
The total cost for all agents is ni=1 C(πi ). A solution to a classical MAPF problem
is a set of n single agent plans, which cant conflict, one for each agent, such that
Π = (π1 , , πn ). These could then be, for example, minimized minπ∈Π C(π).

2.2.1 Dijkstra’s algorithm

There are several existing algorithms that solve the problem of path-finding for AGV
systems. One of the fundamental algorithms, of which many others are based, is
Djikstra’s algorithm. This algorithm solves the problem of finding the shortest path
from a starting point in a graph to a destination [15]. It is described as the following
by Dijkstra [16]:

Suppose we want to find the minimum total length between two given nodes, P and
Q. In the course of the solution, the nodes are divided into three sets. A the nodes
for which the minimum length from P is known. Nodes will be added to this set in
order of increasing the minimum path from P. B is the node from which the node
to be added to set A will be selected. Does not belong to A themselves. C is the
remaining node.

The branches are also subdivided into three sets. I is the branches occuring in the
minimal paths from P to the nodes in A. II is the branches from which the next
branch is to be placed in set I will be selected. III is the remaining branches.

The solution for the minimum total length can be found by iterating the following
steps, first transferring node P to the set A:

8
 2. Background

1. Consider all branches r connecting the the node we just transferred to A with
nodes R in sets B or C. If R belongs to B, we investigate whether the use of
r gives us a shorter path than the known path in II. If not, it is rejected. If
it is, we replace the corresponding branch in II. If R belongs to C, it is added
to set B and branch r is added to set II.
2. Every node in B can be connected to P in only one way if we restrict ourselves
to branches from I and one from II. Each node in B has a distance from P:
the node with the minimum distance is transferred from B to A, and the
corresponding branch is transferred from II to I. After this, we return to step
1 and repeat until the goal node Q is in A, giving us the solution.

2.2.2 Heuristics
The heuristic function h(v) directs the search in path-finding. It’s a function that
estimates the distance from a given node v to the destination node [17]. Heuristics
that does not overestimate the true distance to that goal node is called admissible
heuristics [17]. There are several well-known heuristic functions, among them Man-
hattan, Euclidean, and Chebyshev distance.

The Manhattan distance calculates the true distance to the goal node, given that
there are no obstacles [17]. It is a geometric term used in geometric space to indicate
the absolute wheelbase on the standard coordinate system between two points [18].
Given two coordinates, (x1 , y1 ) and (x2 , y2 ). It is described by Li et al. [18] as:
h(n) = |x1 − x2 | + |y1 − y2 |
The Euclidean distance refers to the real distance between to points in the m-
dimensional space, or the natural length of the vector [18]. It is the square sum of
the difference between two points according to Li et al [18]:
√
h(n) = (x1 − x2 )2 + (y1 − y2 )2
The Chebyshev distance is a measure in the vector space [18]. The definition of
the distance is the maximum of two coordinates (x1 , y1 ) and (x2 , y2 ) of the absolute
difference of each according to Li et al. [18]:
h(n) = max(|x1 − x2 | + |y1 − y2 |)

2.2.3 The A* algorithm

One of the best, known path-finding algorithms, that could be seen as an extension
of Dijkstra’s is the A∗ search algorithm [19, 20]. Dijkstra’s obtains the shortest path
by traversing all nodes [19]. The success rate for obtaining the optimal paths is
high, but it has a very low computational efficiency [19]. The A∗ algorithm uses
a combination of heuristic searching and the shortest path [20]. It has good real-
time performance when the environment is not too complex [19]. Each cell in the
configuration space is evaluated by the value:
f (v) = h(v) + g(v)

9
2. Background

where h(v) is the heuristic distance of the cell to the goal state and g(v) is the length
of the path from the initial state to the goal state, as calculated by Dijkstra’s [20].
Each adjacent cell of the reached cell is evaluated by the value of f (v) and the cell
with the lowest value is chosen. An advantage of the algorithm is the wide range of
modifications that can be done to f (v) [20]. It could be modified to include different
criterion’s such as safety or energy consumption, thus allowing for some flexibility
[20].

2.3 Non-functional requirements

Non-functional, or quality requirements (NFRs) specify how well a system must
perform its functions [21]. It’s a description of a property or characteristic that a
system must exhibit or a constraint that it must respect [22]. The importance of
this type of requirement grows as markets develop [21].

Wiegers and Beatty [22] ascertain that NFRs for embedded systems are especially
critical, as the environment in which they exist is naturally more complex than
pure software applications. The quality attributes that Wiegers and Beatty deem
most important for such systems include: efficiency, reliability, robustness, safety,
security, and usability. Additionally, NFRs regarding scalability are important in the
context of centralized AGV systems due to issues with scaling for example computing
power as the number of agents in the system increases.

2.3.1 Efficiency
Efficiency requirements stipulate how efficiently a system uses computer resources.
Lauesen [21] defines efficiency as how fast the system responds, how much resources
it uses, and how accurately it computes values. Lausen also [21] deems it to be the
same as performance, whilst Wiegers and Beatty [22] agree that it’s closely related,
it refers to internal quality attributes rather than external.

2.3.1.1 Communication
Wireless-based solutions have become increasingly popular in the industry for con-
trolling AGVs and managing fleets [23]. There is a necessity that the solutions
provide flexibility and convenience, allowing for real-time communication between
vehicles and control centers.

To ensure effective communication between AGVs and control centers, there are sev-
eral standards in place, one of which is the IEEE 802.11p-2010 standard. IEE Stan-
dard for Information Technology - Local and metropolitan area networks - Specific
requirements - Part 11: Wireless LAN Medium Access Control (MAC) and Physical
Layer (PHY) Specifications Amendment 6: Wireless Access in Vehicular Environ-
ments [24]. This standard specifies the requirements for wireless LAN medium access
control and physical layer specifications, specifically for Wireless Access in Vehicular

10
 2. Background

Environments (WAVE).

The WAVE standard is designed for devices with rapidly changing physical layers
and short-duration communication exchanges. WAVE applications that adhere to
the standard operating in the 5.9 GHz frequency range and use seven channels with
10 MHz bandwidth each [24, 25].

To meet the requirements for WAVE applications, latency is an important consid-

eration. Latency refers to the time delay between the initiation of a message and
the receipt of its response. In the context of WAVE applications, latency is critical
for ensuring the safety and efficiency of AGV operations. To address this concern,
the IEEE 802.11p-2010 standard specifies maximum latency requirements for both
safety-related and non-safety-related messages. Safety-related messages, which are
critical for ensuring the safety of AGV operations, must have a maximum latency of
100 ms [25]. Non-safety-related messages, which are typically used for monitoring
and reporting purposes, must have a maximum latency of 300 ms [24].

2.3.1.2 Computational time

The scheduling and routing tasks for AGV systems are time critical, as they are
needed to work in real-time [26]. The computational speed of these tasks is depen-
dent on the choice of algorithms to carry them out. Furthermore, any requirement
on the latency sets an upper boundary for the time allowed for computing, as any
computing or algorithmic processing must be completed within the time specified
by the latency requirement. An AGV system adhering to for example the IEEE
802.11p-2010 standard needs to respond within the 100 ms threshold [24]. Thus,
any system adhering to the standard would also need to have a computing time
lower than 100 ms. Centralized solutions are often preferred for the guarantees of
finding an optimal solution [27]. This guarantee vanishes with this upper boundary,
as the algorithm might not be able to search the entire solution space before running
out of time, leading to the choice of sub-optimal paths to the goal destination.
.

2.3.1.3 Memory allocation

All embedded systems are constrained by their memory limitations. For an AGV
system, this limitation is set by the hardware that is chosen for the central controller.
However, the amount of memory that is consumed by the system is impacted by the
choice of algorithm for path-finding, the complexity of the represented production
facility, and the number of AGVs present in the system.

The space complexity is a measurement for the total usage of space by an algorithm.
For the A∗ algorithm, the space complexity is linear O(n) as it remembers the
current shortest path and costs for all visited nodes [28]. This is supported by the
results in [29]. The researchers studied the performance of the A* algorithm, as
well as an improvement with an additional heuristic, in a single robot system. The

11
2. Background

performance was measured in terms of time and space complexity as well as the
completeness and optimality of the solution. The maximum number of nodes and
paths the algorithm considered during the iteration of a correct path was used to
determine the space complexity. The results showed that these grew linearly as the
distance to the goal was increased.

2.3.2 Reliability and Availability

Reliability requirements stipulate how long the system can run before experiencing
a failure [22] while availability requirements stipulate the extent to which the users
of a system’s services are available when and where they are needed [22]. Reliability
problems occur as a result of improper inputs, errors, hardware failures, or compo-
nents not being available when need [22]. Availability is described by Wiegers and
Beatty [22] as the ratio of uptime to the sum of uptime and downtime, we will treat
it as a product of reliability, and therefore address both in the same chapter.

2.3.2.1 Reliability in AGV systems

There are different ways to assess and quantify reliability. A common measurement
used in reliability and risk assessment is the failure rate (λ) [30]. Another metric,
used to represent an index for repairable units, is the mean time between failures
M T BF [30].
Operating time
MTBF(θ) =
N umber of F ailures
N umber of F ailures
Failure rate(λ) =
Operating time
Reliability is a measure of the probability of successful performance of the system
over a period of time [30]. The reliability of most component families follows the
"reliability bathtub" curve [30]. This is a curve that can be divided into three differ-
ent phases. Firstly, the failure rate is reduced as weak components are eliminated.
In the second phase, the failure rate will remain approximately constant. In the last
phase, the failure rate increases as components wear out. Assuming a component
is in the second phase, with approximately constant failure rate, the reliability R(t)
at time t can be represented by the following equation [30]:

R(t) = e−λt
If we integrate this, it will give us reliability for a specified time period
∫ t
R(t) = e−λt dt
0
The effectiveness of an AGV system relies heavily on its reliability and availability
[2]. Yan et al. [1] studied reliability modeling of a single AGV system using a com-
bined fault tree and Petri net approach. The failure rates for different components
in an AGV system can be seen in table 2.1 below. The results suggest that the
accumulated failure rate for all subsystems of an AGV during an entire year would
be 10.45.

12
 2. Background

Table 2.1: Calculated frequencies of AGV subsystem failures according to [1].

Subsystem Failure rate (frequency/year)

Drive unit 1.50
AGV software control system 6.00
Laser navigation system 0.88
Safety Systems 0.38
Attachments 1.25
Brake system 0.20
Steering system 0.25

2.3.2.2 Maintenance
Maintenance of the AGVs is a key factor that affects the reliability of the system.
There are various maintenance strategies available, such as corrective and preventive
maintenance [2]. Corrective maintenance involves repairing an AGV only when it has
malfunctioned or stopped working [2]. On the other hand, preventive maintenance
refers to scheduled maintenance that is carried out regularly, regardless of the current
condition of the AGVs. In the study conducted by Yan et al. [2] the repair time for
a vehicle during corrective maintenance was described by a normal distribution. If
there was another vehicle being worked on, it would wait in a queue. Preventative
maintenance was described as a vehicle health check for a predefined amount of
time. The result from the study conducted by Yan et al. [2] can be seen in Table
2.2; T denotes the time interval for periodic maintenance; P the percentage of AGVs
failed within the time interval; N1 the number of missions completed per year with
periodic but without corrective maintenance; N2 the number of missions completed
per year with both.

Table 2.2: Throughput of the system using corrective or corrective and predictive
maintenance [2].

T P N1 N2
7 days 0,03 11518 11840
20 days 1,10 13213 14709
1 month 3,93 12840 15264
2 months 18,06 11028 15792
3 months 36,62 9372 15972
4 months 53,37 7983 16059
6 months 77,34 6084 16142
12 months 98,06 3282 16234

2.3.3 Robustness
Robustness requirements stipulate how well a system responds to unexpected op-
erating conditions [22]. Robustness is the degree to which the system continues
to function properly when confronted with invalid inputs, defects in hardware or

13
2. Background

software components, or unexpected operating conditions [22]. A closely associated

term with robustness is fault tolerance.

2.3.3.1 Single point of failure

The design and control of AGVs involve many potential technical issues which need
to be addressed, among them path-planning [31].

A problem related to robustness in centralized AGV systems is the fact that the
central control unit acts as a single point of failure [4]. The central control unit
requires centralized, global information, shared memory, and synchronization mech-
anisms. It has high communication demands and a low level of tolerance towards
fault conditions [32]. If there is any failure or malfunction in the central control unit
it can cause the entire system to shut down, leading to delays. The need for constant
communication between the central control unit and individual AGVs poses a threat
to the robustness of the system if it is experiencing technical issues or network prob-
lems. The result of this could be that the AGVs stop or deviate from their intended
path.

2.3.3.2 Deadlocks in AGV systems

An unexpected operating condition that impacts the robustness of AGV systems is
deadlock situations [4]. A deadlock is a situation where an AGV has no possible
action to take [4]. Deadlocks are caused by AGVs competing for or sharing limited
resources in a system [33]. When they occur, material(such as pallets) and vehicle
flows are blocked until a recovery procedure can be performed. This may result in
low throughput, loss of control, congestion, or force a shutdown of the entire system
[33].

The deadlock situations are handled by motion-planning algorithms [4]. When such
situations occur, the motion planner usually handles this by modifying and rerout-
ing from the path that was determined by the path-finding algorithm. All central
optimizing controllers are time-consuming and lack robustness and flexibility. This
is due to the fact that they only detect deadlocks moments before they occur, after
which they attempt to resolve it [4].

Another tool for deadlock prevention is more proactive. Deadlocks can be avoided
by limiting the number of vehicles in a particular area, called zone control [4]. How
this is implemented varies [33]. The general rule, however, is to break up the pro-
duction layout into zones. Only one vehicle may occupy a particular zone at one
time [33]. If a vehicle approaches an occupied zone, it will stay in its current zone
until the next zone is unoccupied. If such a rule is implemented, the throughput
rate of the system depends on the size of each zone, with a bigger zone lowering the
rate [34]. Typically, the minimum size of a zone is equal to the time required to
stop an AGV from its top speed through the use of a controlled braking mechanism.
Deadlocks in AGV systems are mostly the result of the zone-partitioning strategy

14
 2. Background

used within the system [34].

2.3.4 Safety
Safety requirements deal with the need to prevent a system of any kind from do-
ing injury to people or property [22]. Safety requirements are often dictated by
government regulations, and legal or certification issues are important to consider
[22].

2.3.4.1 Safety Standards

Standards are a set of guidelines, procedures, and specifications that have been
established to provide a framework for consistent and reliable performance in vari-
ous industries. They are used for different purposes in requirements, ranging from
mandatory to voluntary adoption, and they play a crucial role in ensuring that
products, services, and systems meet the required quality, safety, and environmen-
tal standards.

In some cases, standards are mandatory and act as constraints for the system [21].
In such instances, adherence to these standards is essential, and any deviation from
them could have adverse effects on the system’s performance, reliability, and safety.
It is therefore imperative that such mandatory standards are not omitted when
considering requirements. In contrast, voluntary standards are followed at the dis-
cretion of organizations or individuals. While they are not legally binding, they
provide a framework for best practices [21].

Technically, standards can be proposed by any organization, but internationally,

the largest organization for standardization is the International Organization for
Standardization (ISO). ISO develops and publishes international standards that are
recognized globally, covering various industries and sectors. In Europe, the main
body for standardization is the European Committee for Standardization (CEN),
which works closely with ISO to develop and promote standards that are specific to
the European Union (EU) and its member states.

CEN produces European Standards (EN). They come in two different types, non-
harmonized ENs, and harmonized ENs. The former refers to voluntary standards
while the latter refers to a specific category of standards that are used to prove that
products comply with the technical requirements of EU law [35]. Whilst the har-
monized standards could be considered voluntary, the technical requirements that
they comply with are mandatory [35].

2.3.4.2 The Machinery Directive

The Machinery Directive, which governs machinery, interchangeable equipment, and
safety components among other products is the underlying EU directive relevant to

15
2. Background

AGVs [36]. This directive makes up the technical requirements that are mandatory
to follow. In order to demonstrate compliance with the directive in the context
of AGVs, there is a harmonized EN. It’s called EN 1525:1997, Safety of industrial
trucks - Driverless trucks and their systems [37]. There is also an ISO standard,
ISO 3691-4:2020, which is slated to replace EN 1525:1997 but has not yet been har-
monized.

Figure 2.2: The Machinery Directive

2.3.4.3 EN 1525:1997
EN 1525:1997 proposes requirements regarding the safety of AGVs [37]. The stan-
dard encompasses a wide range of subjects, some of which has an impact on the
path-finding algorithms of an AGV system and some which lies beyond that scope.
The standard requires that AGVs are designed and operated to ensure safe move-
ment, taking into account the specific application and operating conditions [35].
The AGVs must be equipped with appropriate sensors to avoid collisions [37]. It
requires a risk assessment to be made in order to identify potential hazards for the
specific system [37]. This assessment should consider factors such as the speed of
the AGV, the layout of the facility and the presence of other vehicles or personnel
operating within the environment [37]. The responsibility to implement this lies on

16
 2. Background

the system designer and operators [37].

2.3.4.4 ISO 3691-4:2020

ISO 3691-4:2020 or also Industrial trucks - Safety requirements and verification -
Part 4: Driverless industrial trucks and their systems has been established by ISO
with the intention to align with major legislative regulations in the EU, Japan, North
America and Australia [38]. It is a C-type standard, which means it concerns safety
standards detailing safety requirements for a group of machines [39]. In the context
of path-finding algorithms, requirements regarding speed control and braking is of
interest. All systems adhering to the standard must have a personnel detection sys-
tem, i.e sensors that can detect objects or personnel within a certain distance from
the truck. This distance may vary depending on the specific application and type
of system that’s being used [38].

The standard stipulates that the AGV has to stop in the range of personnel detec-
tion means for the worst condition specified by the manufacturer, as a function of
speed, friction, gradient, and rated load. Furthermore, it should be able to stop be-
fore contact is made between the rigid parts of the AGV or its load and a stationary
person. In the case that the AGV has stopped due to an object within its path, it
can be restarted after a minimum delay of two seconds [38].

The speed control of the system should follow the requirements specified in Table
2.3, among others [38]. The requirements differ depending on whether the personnel
detection sensors are activated or muted and the distance to objects within the
AGV’s surroundings. The standard does not explicitly stipulate a maximum speed
limit. Any system complying with the standard, however, should have a speed that
is lower than the rated speed for the system [38]. This is the travel speed as defined
by the manufacturer [38]. As an example, Toyota’s current AGV solutions have a
maximum speed ranging between 2,0 - 2,2 m/s [40].

17
2. Background

Table 2.3: ISO 3691-4:2020, Examples of requirements for speed control.

Clearance on Clearance on Clearance from object Personal Detection

Max speed
one side, C1 other side, C2 in direction of travel Means
>500 mm >500 mm >500 mm Active Rated speed
>500 mm >500 mm >500 mm Muted 0,3 m/s
>500 mm >500 mm <500 mm Active 0,7 m/s
>500 mm <500 mm <500 mm Muted 0,3 m/s
>500 mm <500 mm >500 mm Active 1,2 m/s
>500 mm <500 mm >500 mm Muted 0,3 m/s
>500 mm <500 mm <500 mm Active 0,7 m/s
>500 mm <500 mm <500 mm Muted 0,3 m/s
<500 mm <500 mm >500 mm Active 1,2 m/s
<500 mm <500 mm >500 mm Muted 0,3 m/s
<500 mm <500 mm >500 mm Active Rated speed
<500 mm <500 mm <500 mm Active 0,3 m/s
<500 mm <500 mm <500 mm Muted 0,3 m/s

2.3.5 Scalability
Refers to requirements regarding how easily the system can grow to handle more
users, transactions, servers, or other extensions. Scalability has both hardware and
software implications. Scaling up a system could mean acquiring faster computers,
increasing network capacity, data compression, or increasing the number of produc-
ing units in a system [22].

2.3.5.1 Scalability in AGV Systems

The problem of path planning in a multi-agent environment refers to the challenge
of determining the best routes for multiple autonomous agents, such as robots or
vehicles, to move from their current locations to their desired destinations. This
problem is notoriously difficult, and it has been proven to be NP-hard, which means
that finding an optimal solution to this problem becomes increasingly challenging
as the number of agents, in this case, AGVs and tasks in the system grows [41].

The number of AGVs, n, that is required for a system can roughly be estimated as
the following, where TL = total loaded travel time, TE = total empty travel time,
TW = total waiting time and TA = time an AGV is available [42].
∑ TA +TE +TW
n= TA

In current industrial AGV systems, a centralized approach is often used, where a

18
 2. Background

central unit is responsible for planning the trajectories for all AGVs simultaneously.
The central unit has information about the positions, goals, and static paths that
the AGVs must follow [4]. It uses this information to search the solution space for
an optimal solution while ensuring that the planned trajectories are collision-free
and deadlock-free. However, this centralized approach becomes less scalable as the
number of AGVs and the environment size increases, as the computational require-
ments increase exponentially [4, 43]. This can lead to increased response times and
decreased system performance, limiting the number of AGVs that can be used in
the system.

Fransen et al. [44] ascertains that a system is deemed scalable if it can hold a one-
to-five zone and AGV ratio. A zone is represented by a cell in a grid-based layout.
This layout is not supposed to represent a physical system layout, but rather a vir-
tual grid representation of the layout to control the AGVs. Only one AGV can be
present in each zone at any given time. This implies that the scalability also relies
on the layout of the production facility. A higher amount of grids would imply the
opportunity to increase the amount of AGVs in the system to maintain the ratio.
Traditional bottlenecks and traffic rules within the facility can also impact the scal-
ability, for example, if the system allows for uni-directional or bi-directional paths
or the amount of input or output stations [42].

2.3.6 Security
Security requirements deal with unauthorized access to the system [22]. The re-
quirements could refer to physical, data, or software security. They are often based
on business rules or corporate policies [22].

2.3.6.1 Security in AGV systems

When considering the security of any embedded system, it’s important to consider
the vulnerabilities both of hardware and software components, as they all can be
compromised with respect to their confidentiality and integrity. Below are a few
considerations from different standards, stipulating requirements both for the AGV
itself, as well as the software controlling it.

ISO 3691-4:2020 specifies several security requirements for AGV systems, including
the use of personnel detection devices, such as scanners [38]. These scanners are
designed to detect people or objects, and they should be able to detect objects of
different sizes and shaped [38]. Interlocks or other means should also be used in
accordance with the standard to prevent unauthorized access to the AGV’s power
source or control systems to prevent hazards [38]. The standard also emphasizes
that AGVs should be operated only by authorized personnel who have received ad-
equate training on the safe use of the AGV [38].

ISO/IEC 27001:2013, Information technology Security techniques Information se-

curity, management systems Requirements is an information security management

19
2. Background

standard that provides a comprehensive framework for managing and protecting sen-
sitive information, not specifically tailored towards AGV systems [45]. It includes
requirements and guidelines that can be applied to AGV systems, however, organi-
zations should conduct a risk assessment to identify potential security threats and
vulnerabilities and develop a plan to mitigate them [45]. Organizations can also im-
plement access controls such as password authentication, biometric authentication,
or the use of access cards, which can help prevent unauthorized access to the AGV
system [45]. Measures can be taken to protect sensitive data from unauthorized
access or modification [45]. Regular security audits can be conducted to identify
potential security vulnerabilities and areas for improvement, and a plan can be in
place to respond to security incidents [45].

WAVE applications adhering to the IEEE 802.11p-2010 standard also should adhere
to requirements in regard to the security of the data [24]. The standard specifies
the use of a shared secret key for authentication, which is used to authenticate each
communication between devices [24]. This ensures that only authorized devices are
able to communicate over the network and that data transmitted over the network
cannot be intercepted by unauthorized parties. It includes measures to protect
against packet replay attacks, which involve capturing and re-transmitting packets
to disrupt the communication between devices [24]. The standard specifies the use of
a sequence number for each packet transmitted, which is used to detect and prevent
packet replay attacks [24]. It also includes measures to protect against denial-of-
service (DoS) attacks, which involve disrupting the communication between devices
by overwhelming the network with traffic. The standard includes mechanisms for
managing and prioritizing network traffic, which can help prevent DoS attacks by
ensuring that critical network traffic is given priority over less important traffic [24].

2.3.7 Usability
Usability requirements deal with ease of learning, ease of use, error avoidance and
recovery, the efficiency of interactions, and accessibility [22]. Usability deals with
the efficiency of the system in regards to the users: how easy should it be to learn,
how efficient should it be in daily use, etc [21].

2.3.7.1 Usability in AGV systems

Since AGVs are embedded systems, both the hardware and software parts need to
be considered for usability [46]. There are several standards suggesting requirements
regarding the usability of products, but there are no specific standards for AGVs.
One such standard is ISO 9241-11:2018 Ergonomics of human-system interaction
Part 11: Usability: Definitions and concepts [47]. The standard outlines the ab-
sence of negative consequences, satisfaction, and wide applicability as important for
the usability of a product [47].

The absence of negative consequences means that a system, such as an AGV, should
be designed in such a way that it does not cause any adverse effects or outcomes
to the user, equipment, or surrounding environment [47]. An AGV adhering to ISO

20
 2. Background

3691-4:2020 suggests requirements regarding personnel detection means, braking,

zones, and speed control [38]. These requirements are to be followed to ensure that
the system is safe to operate and that there are no hazards to individuals or equip-
ment, thus limiting negative consequences [38]. In an AGV system, economic harm
can be caused by for example deadlock situations, as they impact the throughput
of the system [33]. For larger AGV systems, the choice of path-finding algorithm
and heuristics can cause sub-optimal paths when the entire solution space can’t be
searched [4]. Accumulated over time, this could also result in economic harm.

Satisfaction is an essential component of usability, which means that a system must

be designed to ensure user satisfaction [47]. In the case of AGV systems, satisfac-
tion can be achieved by ensuring that the system meets the user’s expectations and
provides a positive experience. To achieve satisfaction in AGV systems, the system
must be designed to meet the needs of the user. This includes providing an interface
that is easy to use and understand and allowing for easy and efficient control of
the AGVs. The AGV should also be designed to perform its tasks smoothly and
effectively, without causing any disruptions or delays. Furthermore, the AGV sys-
tem should be designed to meet the specific needs and requirements of the user.
For example, the AGV should be configured to handle the specific materials, load
capacity, and travel distance required by the user’s operation which would ensure
that the AGV is optimized for the user’s needs and provides maximum satisfaction.

Wide applicability means that a system must be designed to be applicable to a wide

range of users and contexts [47]. To achieve wide applicability in AGV systems,
the system must be designed to be flexible and adaptable to different environments
and situations. For example, the system should be able to handle different types
of loads, travel distances, and work environments. Additionally, the system should
be designed to be compatible with different hardware and software configurations,
making it easier to integrate with other systems.

2.3.8 State of the art

Below we will present the current state-of-the-art, regarding AGVs in the context of
different NFRs.

• The current state-of-the-art safety requirements for AGVs is the ISO-3691-

4:2020 standard. This standard is to replace EN 1525:1997.
• ISO-3691-4:2020 also stipulates requirements regarding the security of AGVs.
This can be complemented with ISO/IEC 27001:2013 which is a general infor-
mation security standard.
• The current state-of-the-art AGV systems are centralized, as a consequence
of this they can have issues with scalability and robustness. The scalability of
a centralized system is limited due to increasing computational requirements
when the system expands. The centralized system also acts as a single point
of failure and malfunctions that can lead to system shutdowns. Furthermore,
the need for constant communication between the central unit and the AGVs

21
2. Background

impacts the robustness of the system as a whole.

• Reliability studies of modern AGVs show that they have a failure rate of
roughly 10,45 sub-component failures per year, which is impacted by the choice
of maintenance strategy.
• There is no AGV standard for communication. The IEEE 802.11p-2010 is
state-of-the-art in adjoining fields and could be applied. Depending on the
level of parallelism in the system, this could set an upper limit on the compu-
tational time of any algorithms used by the system.
• There is no AGV standard for usability. However, ISO 9241-11:2018 outlines
requirements for the usability of products. Usability comes according to the
standard as a result of the absence of negative consequences, satisfaction with
the product, and wide applicability.

22
 3
Literature Review

The objective of the upcoming chapter is to present related works for the research.
Articles from the field of AGVs will be included when such have been identified.
Otherwise, it will encompass relevant academic papers from the fields of automated
vehicles (AVs) and more general literature from the field of software engineering.
The chapter will end with a section regarding the identified gaps in knowledge in
the current literature.

3.1 Related Work

The following section will discuss different published articles in the context of AGVs
when possible or AVs or other industries where generalization to AGV systems
would be possible and studies in AGVs are scarce. The articles will concern the
non-functional requirements regarded by Weigers and Beatty [22] to be essential for
embedded systems. These include safety, scalability, robustness, reliability, avail-
ability, security, usability, and efficiency.

3.1.1 Efficiency
Tsang et al. [48] studied the problem of warehouse multi-robot automation in
discrete-time and space configuration with the focus on task allocation and path-
planning. A comparison between the computation times of a decentralized system
using a RERAPF algorithm for path-planning and a centralized system using the
conventional A∗ algorithm was made, for a varying amount of tasks and robots.
With the number of robots varying between 1 and 100, the results showed that the
computational time for the A∗ algorithm increased almost logarithmic from 0 ms to
340 ms as the number of vehicles in the system increased.

Setiawan et al. [49] conducted a comparative analysis of the A∗ and D ∗ Lite path-
planning algorithms for a single-AGV system was made. The resulting simulation
showed that the computing time, given that there were no unknown obstacles in
the path, was 0.3 ms for the A∗ algorithm and 5.2 ms for the D ∗ Lite algorithm.
However, in the case where the path had to be re-planned due to unknown obstacles,
the D ∗ Lite algorithm outperforms the A∗ algorithm given that the state space is
large enough.

Whilst the studies conducted by Tsang et al. and Setiawan et al. [48, 49] compare

23
3. Literature Review

different algorithms and their computational effort, no studies have been made that
take different NFRs into account. Taking different NFRs into account could give
the industry further perspective on what type of algorithms are most suitable given
different contexts.

3.1.2 Reliability and availability

Sarker et al. [50] conducted a review of sensing, communication, human factors,

and controller aspects for connected and automated vehicles (CAVs). The authors
deemed that the human factor was important for the mass adaptation of CAVs,
which largely depends on the accuracy and reliability of the vehicles. Among the
critical factors regarding the reliability of the vehicles are: the level of accuracy of
the vehicles’ routes, availability of current roadway information, level of training
and prior learning of a vehicle, system failure detection, and accuracy of the route
selection.

Schöner [51] studied the challenges and approaches for testing of highly automated
vehicles (HAVs). The pretext of the study was that this type of system requires high
levels of availability and effectiveness. The authors deemed that systems with high-
reliability requirements are designed with a functional analysis approach. Amongst
the goals of this methodology is that it allows for calculating reliability on a system
level based on properties in the sub-components.

Huang et al. [52] conducted a review on human-machine shared control systems for
AVs. Amongst the conclusions that the authors made was that software failure de-
tection, including component and communication failures must be addressed. They
deemed it necessary to develop a control system that can detect and implement any
potential fault, as the reliability of the system highly influences the driver’s trust in
the technology.

Yan et al. [2] investigated the impact that different maintenance strategies has on
the throughput of AGV systems was investigated. More specifically, they studied
the impact of corrective and predictive maintenance.

Availability and reliability in the context of AVs have different focal points than
AGVs. The reliability-related issues mentioned by Sarker et al., Schöner and Huang
et al .[50, 51, 52] are mostly regarding the safety of the vehicles. In the context
of AGVs, the availability is a function of their reliability instead. Whilst reliability
issues are connected with safety, the main concern is the impact that availability
has on the throughput of the system. This issue has not been studied thoroughly,
although Yan et al. [1] studied it within the context of maintenance strategies.

24
 3. Literature Review

3.1.3 Robustness
Orf et al. [53] presents a novel method for probabilistically deriving the function-
ality of a localization system, a part of the AV that is deemed safety-critical. The
authors deem that the high demands on safety for AVs make failure-free operation
and malfunction detection necessary.

Martinez and Francesc [54] provides an overview of the current state-of-the-art in

key aspects of autonomous driving. The authors ascertain the need for these systems
to continue working after a failure and suggest that the support of cloud computing
could help in this regard. Further, they establish the need for a cooperative environ-
ment regarding communication with other agents in the environment. In order to
accomplish this they ascertain the need for a robust communication network being
able to handle and transmit huge amounts of data in all possible conditions.

Li et al. [55] investigated platooning of connected and automated vehicles. Amongst

the conclusions the authors made was that communication issues, quantization er-
rors, and packet loss posed a significant challenge for the technology.

Gruyer et al. [56] studied the impact of Cooperative, Connected, and Automated
Mobility (CCAM). The researchers ascertain that the capability of the technology
is strongly related to the robustness of its components. Furthermore, they ascertain
that for automation beyond what is available today, onboard sensors are not enough
to guarantee the high requirements for robustness and reliability.

Whilst neither of the studies is researching robustness in the context of AGVs, some
of the issues are similar. The importance for the system to continue working after
a failure, as ascertained by Orf et al. and Martinez and Francesc [53, 54] is also
important in the context of AGV systems. Current AGV systems suffer from a lack
of robustness, impacting the system by 1) creating potential deadlock scenarios,
2) forcing re-planning of routes, and 3) lowering the throughput. Just as for the
CCAM systems mentioned by Gruyer et al. [56], the capability or capacity of an
AGV system is greatly dependent on its robustness. Evaluating robustness in the
context of AGVs could help the industry by providing insights regarding its impact,
and create incentives to come up with better solutions.

3.1.4 Safety
Marvel and Bostelman [57] presented an overview of safety standards for AGVs,
and describe how those related to other adjoining fields in modern manufacturing.
Among the safety standards discussed were the Machinery Directive and EN1525.
The authors further discussed the future of such standards, and speculated that
they should include criteria such as; measurement of dynamic obstacles, three-
dimensional imaging to detect overhanging obstacles, detection of humans located
near the AGV, and more.

Tiusanen et al [58] presented an overview of current safety requirements for AGVs

25
3. Literature Review

and other autonomous machines. The authors ascertain that there is a gap be-
tween the safety requirements set in standards and the current state-of-the-art in
the industry. Furthermore, they ascertain that one of the largest challenges for the
industry, in regard to safety standards, is the requirements for the sensor systems
used to detect objects.

Javed et al [59] discusses safety assurance in the context of Industry 4.0. They ascer-
tain that production in Industry 4.0 is to be deemed as safety critical. For example,
due to the simultaneous presence of AGVs and human workers in the same site.
They establish that safety requirements are derived either from a hazard analysis or
extracted from a set of safety standards.

Menzel et al. [60] analyzed a scenario-based approach for the design of vehicles us-
ing the ISO 26262 standard, which represents the state-of-the-art for safety-critical
electric vehicle systems. The authors deemed that automation beyond conditional
automation has a large challenge in ensuring that the vehicles behave in a safe way.

Riedmaier et al. [61] conducted research resulting in a survey on scenario-based

safety assessment for AVs. They consider two aspects when assessing the safety of
a vehicle. Firstly, they ascertain that safe functionality must be ensured. Secondly,
that the intended function, if proven safe, cannot introduce hazards caused by tech-
nical failures in the system’s software or hardware. They deemed that the market
of AVs will not grow unless the challenge of proving the safety of these systems to
legislators and society has been solved.

Stolte et al. [62] derived safety goals and safety requirements for actuation sys-
tems of AVs in accordance with ISO 26262. They ascertained that causal factors
for unsafe control actions for AVs include, among other aspects: inadequate control
algorithms, inconsistent or incomplete models, missing or delayed feedback from sen-
sors, component failures, and missing inputs. Their findings implied that measures
to go beyond the current state-of-the-art will be needed to ensure the functional
safety of AVs.

The studies by Marvel and Bostelman [57], Tiusanen et al. [58] and Javed et al.
[59] ascertain the use of safety standards as a way to elicit safety requirements for
AGV systems. Research such as the study conducted by Stolte et al. [62] would be
insightful for the industry, where safety requirements are derived from a standard.
This is due to the fact that ISO 3691-4:2020 has similar, demanding requirements for
safety. Just like AVs, AGVs operate in environments in close proximity to humans.
Therefore, as mentioned in Riedmaier et al. [61] regarding the market for AVs, AGVs
have to follow these strict safety standards in order to be deemed as an option to
conventional forklifts for the public.

26
 3. Literature Review

3.1.5 Scalability
Ryck et al. [4] presents state-of-the-art control algorithms and techniques for AGV
systems. The researchers look at decentralized versus centralized solutions. The
authors ascertain that current, centralized solutions suffer from a lack of scalability.
This indicates a need for an optimized algorithm and a low-latency decentralized
solution.

Fransen et al. [44] presented a real-time path planning approach for dense, grid-
based AGV systems. The authors ascertained that throughput is the key perfor-
mance index for such a system. Furthermore, they ascertained the key features that
an AGV system should possess. It should be flexible, applicable to different layouts,
robust, and able to cope with the inaccuracies and unexpected conditions that im-
pact the throughput. It should also be scalable, and applicable to systems with a
zone-to-AGV ratio of five to one.

Duboc et al. [63] presented a goal-oriented approach for eliciting, modeling, and rea-
soning about scalability requirements, for software systems in general. The authors
regard this type of NFRs to be one of the most important for any software-intensive
system. Furthermore, they ascertain the need to define the scalability goals of a sys-
tem through an a priori analysis. The authors deem that the context is important
when assessing the scalability of a system. Among things to consider; are systems
quality goals, the characteristics of the application domain and its variation, and
the acceptable levels of quality goal satisfaction under these variations.

As mentioned by Ryck et al. [4], centralized AGV systems suffer from a lack of
scalability. Among the problems identified is that the computational time increases
exponentially with the number of AGVs and the size of the environment. The
importance of scalability is ascertained by [63]. As this issue has been ascertained,
studying it further and simulating its impact would be of value to the industry.

3.1.6 Security
Chowdhury et al. [64] conducted research regarding cyberattack surfaces and se-
curity vulnerabilities for connected and automated vehicles (CAVs). The authors
recommended the following practices based on ISO 26262. Among the different
vulnerabilities identified were in-vehicle systems such as sensors and software, v2x
communication networks (vehicle-to-x), and attacks on the supporting digital in-
frastructure.

Javed et al. [65] studied the security and safety of platooning AGVs in production
facilities was investigated. The authors deemed that a safety-critical system, such
as AGVs, can only be considered safe if they are also secure. In light of this, they
conducted two different analyses to derive safety and security requirements and sim-
ulated their results using a digital twin.

The study conducted by Javed et al. [65] is recent, exhaustive, and covers most

27
3. Literature Review

possible security threats. The study, even though the context is platooning AGVs
in a quarry, ought to be generalizable to contexts such as warehouse settings as the
threats they identified are very similar.

3.1.7 Usability
Schömig et al. [66] created a standardized test procedure for usability evaluation
of external human-machine interfaces (eHMIs) in the context of automated vehicles
was created. In the study, they defined procedures for creating usability require-
ments. The authors proposed following the usability definition provided by ISO
9241-11, that the system must be effective, efficient, and satisfying the user’s needs.

Pauzie and Orfila [67] studied human factors that impact opinions regarding AVs.
The authors ascertained the importance of a human-centered approach to the design
of transport and automotive systems. They identified issues such as technology ac-
ceptability, situation awareness, and mental workload to be critical for the usability
of AVs.

Younessi and Lauesen [68] investigated usability requirements in general. The re-
searchers ascertained that usability can be divided into five factors; ease of learning,
task efficiency, ease of remembering, understandability, and subjective satisfaction.
The necessity of each factor depends on the context of the system under study.

Similar to Younessi and Lauesen [68], Bevan [69] investigated usability requirements
in a more general manner. The paper focuses on "quality of use" which the authors
deem to be the end goal of usability. Furthermore, the authors ascertain that us-
ability requirements should be stated in terms of the effectiveness, efficiency, and
satisfaction required in different contexts.

No studies have been made of the impact that usability requirements have on AGV
systems. The studies conducted by Schömig et al. and Pauzie and Orfila [66, 67] are
within the context of AVs, whilst Younessi and Lauesen and Bevan [68, 69] studied
usability for software in general. However, following the standard mentioned by
Schömig et al. [66] and the impact that it has on AGV systems would provide the
industry with insights regarding how important such factors are. However, given
the nature of the research at hand, simulations might not be the most appropriate
way to investigate such factors.

3.1.8 Existing gaps in knowledge

The literature review proves that there are multiple significant gaps in current aca-
demic literature regarding the use of NFRs in the context of AGV systems, which
establishes a need for such a study. Below follows a summary of the position of the
existing literature in the field.

• Studies regarding NFRs in the context of AGVs are lacking. Studies exist that

28
 3. Literature Review

consider different quality attributes, such as reliability, safety, and security.

The studies that do exist on these subjects are however not considering these
quality attributes from a requirements standpoint.
• No articles were identified that study several types of NFRs simultaneously in
the context of AGV systems and their impact on the system throughput.
• Articles discussing safety in the context of AGVs were identified, however,
they had no focus on requirements. Articles that studied safety as an NFR,
in the context of AVs were identified. Two of the identified studies used the
ISO 26262 standard as a baseline.
• Research in the field has concluded that centralized AGV systems suffer from
a lack of scalability. However, no in-depth studies using scalability as an NFR
were identified by the researchers.
• Studies have been conducted in adjoining fields regarding the robustness of
CCAMs and AVs. Whilst many of the more important points made are similar,
no studies have been conducted in the context of AGVs and requirements
engineering.
• Reliability-related issues have been studied both in the context of AGVs and
AVs. However, this has not been done from a requirements engineering stand-
point.
• Studies regarding the efficiency of different path-finding algorithms used in
AGV systems are common. However, no studies were identified taking different
NFRs into account.
• Studies have been conducted regarding the security of platooning AGVs in a
production facility. As the threats are similar to those for AGVs operating in
a warehouse environment, the results ought to be generalizable.
• No studies were identified regarding the impact of usability on AGV systems.
The studies that have been conducted are in the context of AVs or on a more
general level.

The thesis addresses several of the gaps mentioned above. It is a study conducted
from a requirements engineering standpoint on the subject of AGVs. Requirement
engineering studies in this context is rare and the articles that does exist on certain
subjects (safety, reliability for example) are discussed in a more general context.
The study also considers several different types of NFRs simultaneously, rather than
discussing them in isolation. Furthermore, it investigates the relationship between
these NFRs and throughput (efficiency). During the course of the research only one
such study was identified, which had a focus on reliability.

29
3. Literature Review

30
 4
Methodology

In this section, we will provide an overview of the methodology we used to address

the research questions posed in this study. Our approach was a mixed-methodology
one, which involves the use of both qualitative and quantitative methods of data
collection [70]. Specifically, we employed a non-systematic literature review, ex-
ploratory semi-structured interviews, and computer simulations to collect data.

Figure 4.1: Overview of the research process

The use of qualitative research methods is particularly valuable when investigating

human behavior and interactions in natural settings and data usually takes the form
of text or pictures [71]. The strength of qualitative methodologies is that they force
the researcher to delve into the complexity of the problem at hand, rather than ab-
stract it away [71]. This makes the potential results of the research richer and more

31
4. Methodology

informative [71]. The use of qualitative research methodologies has been shown to
be useful in the field of software engineering [72].

Quantitative methods are useful when describing anything other than human behav-
ior, using statistics or other numerical data and usually take the form of numbers
or classes [72]. Quantitative data often provides a better understanding of the stud-
ied phenomenon and has been the primary data type used in software engineering
research historically [72].

By employing a mixed-methods approach, we were able to take advantage of the

strengths of both qualitative and quantitative methods and gain a more compre-
hensive understanding of the research questions at hand [70, 72, 73]. Furthermore,
it made it possible to employ methodological triangulation to confirm the validity
of our findings. Methodological triangulation involves the use of multiple methods
to investigate the same research question, which can help to reduce the risk of bias
and increase the reliability of the results [72].

4.1 Literature review

All literature reviews are more or less structured ways of collecting and synthesizing
available data in regard to a specific topic [74]. Amongst the purposes for utilizing
a literature review as a tool in research are identifying the gaps in knowledge and
helping identify general patterns within the same area [75]. Literature reviews in
general are an appropriate methodological tool when the researcher wants to evalu-
ate theory or evidence in a certain area [74]. This could be done broadly, such as in
exploring the collective evidence within the topic of interest [74].

The conducted literature review is a targeted, semi-systematic review. Being tar-

geted implies that it will focus on specific areas of interest, rather than attempting
to cover all available literature on a topic [76]. A semi-systematic literature review
implies that it is not fully systematic, nor fully integrated [74] with the purpose
to gain an oversight of the research area [74]. Appropriate research questions are
often broad, similar to this thesis [74]. Samples from semi-structured reviews are
mostly in the form of research articles and the search strategy can be more or less
structured and the analysis and evaluation both qualitative and quantitative [74].
As the study is of the mixed-methodology approach, utilizing both quantitative and
qualitative methods of data collection, a semi-systematic approach is appropriate.
Furthermore, semi-systematic reviews are appropriate when studying a broad topic,
which has been conceptualized within several disciplines, all of which might not be
possible to study [74]. This is true for this thesis, as AGV systems have been stud-
ied in a wide range of disciplines from the field of robotics to logistics. The study
followed the general guidelines for semi-systematic reviews as proposed by Snyder
[74], and the resulting search strategy will be discussed in Section 4.1.1.

32
 4. Methodology

4.1.1 Search strategy and analysis

Snyder [74] ascertains that when designing a literature study, one must consider the
search strategy. This implies deciding and recording search terms and databases used
and that the researchers decide on inclusion and exclusion criteria for the review.
This is important for transparency, and maintaining a proper thread of evidence
for the reader. Furthermore, Snyder [74] ascertains that one shouldn’t have strict
inclusion or exclusion criteria, as this can result in missing important information
on the subject.

In Table 4.1 the search terms used for finding the sources for the study are visual-
ized, along with the articles they resulted in. The articles were found using Google
Scholar and come from databases such as Elsevier, IEEExplore or Taylor & Francis.
The inclusion criteria were not limited to articles within a certain year of publication
but were to be written in English. Worth mentioning is that previous research on the
specific topic of non-functional requirements for AGV systems is almost non-existent
and the researchers had to use alternative terms to identify relevant literature. For
each search term, roughly the top twenty results appearing in Google Scholar was
considered. In many cases the relevancy was deemed to be minimal, and articles
were omitted. When the articles was deemed to be pertaining to the subject that
was to be investigated, their abstracts were read. If they still was deemed to be of
interest, they were included in the study. A visualization of the search strategy can
be seen in Table 4.1.

The articles that were found during the process were used to:
1. Justify the research. The identified literature was used to identify gaps in the
existing research on the field of NFRs and AGV systems. For example, whilst
one article studied the impact of maintenance strategies on AGV systems,
the impact of reliability on the throughput of an AGV system has not been
researched.
2. It helped establish a theoretical framework by which the researchers could
extrapolate requirements to answer RQ1. For example, different standards
regarding safety were identified, which could serve as a baseline for any safety
NFRs used within the research.

The analysis of the literature review followed the guidelines provided by [77]. The
guideline differentiates between research questions with more subjective goals (expe-
riences, attitudes, or perceptions) and more objective goals (effect, risk, association,
or prevalence). The former will be followed in this case which implies identifying
findings that help answer the research question, condensing and summarizing those,
and lastly grouping them thematically, preferably in a table.

33
4. Methodology

Table 4.1: Search terms and related sources

Search term or source Search Engine Identified source(s) Database(s) Total number of results
"AGV efficiency path-planning" Google Scholar [48] IEEExplore 21
"Automated vehicles safety requirements" Google Scholar [61], [60], [62] IEEExplore, Elsevier, IEEExplore 30800
"Automated vehicles scalability requirements" Google Scholar [78] IEEExplore 1280
"Centralized AGV systems" Google Scholar [4] Elsevier 7510
"Comparison of a* and d* lite" Google Scholar [49] Springer 438000
"Reliability AGV" Google Scholar [2] Elsevier 24000
"Reliability AV" Google Scholar [51] Springer 3350000
"Reliability requirements automated vehicles" Google Scholar [50], [52] IEEExplore, IEEExplore 16300
"Robustness automated vehicles" Google Scholar [53], [54] Elsevier, Elsevier 173000
"Robustness AV" Google Scholar [56] - 738000
"Safety AGVs" Google Scholar [57], [58], [59] IEEExplore, -, Elsevier 27400
"Scalability automated vehicles" Google Scholar [79] IEEExplore 94300
"Scalability requirements" Google Scholar [63] Elsevier 638000
"Scalability requirements AGV" Google Scholar [44] Elsevier 353
"Security requirements" Google Scholar [65] Elsevier 4750000
"Security requirements automated vehicles" Google Scholar [64] - 449000
"Usability requirements" Google Scholar [68], [69] -, Elsevier 1340000
"Usability requirements automated vehicle" Google Scholar [67], [66] Elsevier, - 3280

4.2 Interview study

According to Building theories from multiple evidence sources [71], interviews can
be used for gathering requirements and evaluating them. The aim of interviews
is to gather general information, including opinions regarding processes, products,
and personal knowledge. The qualitative data in this study were collected from
semi-structured interviews with individuals domain-knowledgeable in the context of
AVG systems. The reasoning for the choice of semi-structured interviews is that
they allow for exploratory probing on issues deemed important for the study [80].
Furthermore, semi-structured interviews have the benefit of not limiting the scope
of the answers provided by the interviewers and letting information or questions
arise, that the interviewers might not first have considered [80]. They are designed
to not only provide the information that has been foreseen but also the unexpected.
Semi-structured interviews furthermore allow the gathering of detailed data and an
in-depth understanding of the topic at hand by allowing follow-up questions if re-
quired [81]. The general principle that our interview guide followed can be compared
with the funnel model as described by Runeson and Höst [72]. For each topic, we
started off with open questions, to then narrow them down to more specific ones.
The interview guide can be found in Appendix A for further details.

4.2.1 Data collection

In order to ensure that the data collected through the interviews are valuable and
appropriate, it is crucial to gain access to the right individuals [82]. This was
ascertained by the choice of interview subjects, which had differing but appropriate
roles within their respective organizations to provide a broader perspective on the
different topics. Differing roles among the interview subjects are to be preferred

34
 4. Methodology

according to [72]. The roles of the interview subjects can be seen in Table 3.2. The
sampling method used for the interview study was convenience sampling, i.e. the
subjects were chosen based on their accessibility for the study [83, 84] since these
subjects are often highly paid and hard to attain. Ideally, the study would have
included more subjects. However, the reason for the low number of subjects is the
same as stated above: the subjects are highly paid and hard to attain.

Table 4.2: Roles of interview subjects

ID Role Experience
A Solution Architect > 5 years
B Director of Technology > 20 years
C Senior Advisor AGV solutions > 20 years
D Manager of Automation & Vision Services < 5 years
E Solution Architect & Business Analyst > 10 years

As mentioned earlier, it is important to consider triangulation when doing research,

as it increases its precision [72]. Apart from the methodological triangulation that
was achieved by combining qualitative and quantitative methodologies, observer tri-
angulation was also realized during the interview sessions. According to Runeson
and Höst [72] this is achieved when more than one observer takes part in the study.
During the course of the interview study, both of the researchers participated in all
of the interviews.

A serious consideration when using interview studies is confidentiality, consent, and

the handling of sensitive results [72]. This was achieved by informing the subjects
during the process of procuring the interviews. Subjects were ensured that neither
their name nor the organization that they works for was to be disclosed and that
any material that was to be collected would be coded in the study. Furthermore,
subjects were given the names and contact information of the researchers, as well as
information regarding the research and the procedures that were to be used in the
study. Since all interviews were recorded for transcription, the subjects were asked
to give consent before recording each session.

Each interview started with a short explanation of the study and its purpose, in ac-
cordance to Case study research in software engineering: Guidelines and examples
[80] which is important, as it might affect the level of participation of the subjects.
After that, the focus of the first segment of the interview was the subject’s experi-
ence within the context of AGV systems. This is in accordance with Galletta [81],
which ascertains the importance of doing this at the start of the interview. Each
interview ended with a question regarding whether the subject perceived that there
was something that was missed, or if they had anything additional that they would
like to add.

35
4. Methodology

It is also essential that a full record of the interview is taken [75]. This was achieved
via the built-in Teams recording function, which facilitated a video recording of the
interviews for transcription and analysis in the later phases.

4.2.2 Data analysis

According to Runeson and Höst [72], the primary objective of analyzing qualitative
data is to draw conclusions while maintaining a clear chain of evidence. This re-
quires that the reader can trace how the conclusions and results have been derived
from the data. Furthermore, it is imperative to conduct analysis concurrently with
data collection, as this is the phase where new insights can be discovered and acted
upon [72, 73]. Since the interviews in this study were spaced over several days,
each interview was analyzed separately to gain a deeper understanding before the
following interviews.

The approach used to analyze the obtained data is a thematic coding approach, or
constant comparison method, as described by Seaman in Guidelines for conducting
and reporting case study in software engineering [73]. It is a generic approach to
the analysis of qualitative data and is described by Robson and McCartan [75]. Al-
though this approach has the benefit of being flexible and generic, it is often limited
to describing or exploring data [75]. However, this was not an issue in this case,
as the interview study was meant to be exploratory. According to Robson and Mc-
Cartan [75], coding plays a crucial role in this approach, involving identifying and
recording data items that exemplify the same theoretical or descriptive idea. Each
code is then broken down into smaller themes that relate to the research questions.
These groupings of themes are then examined for underlying themes or explanations
of phenomena [73].

The thematic coding analysis in this study was conducted in accordance with the
general guidelines provided by Robson and McCartan [75]. The first step was to
transcribe and re-read the data to note down initial ideas. Once that was done,
initial codes and themes were generated. These steps were conducted by both of
the researchers using a Miro board and simultaneously working on the same inter-
view transcription. The approach of coding after the interview sessions is defined
as post-forming [73]. An affinity diagram was created based on those codes and
themes, which were then interpreted. Through providing these codes, themes, and
an eventual affinity diagram, the study ought to provide a clear chain of evidence
for the reader. After the interview study was completed, the themes and codes were
discussed with subject E to ascertain reliable results. Worth mentioning is that the
researchers believe that there was a certain degree of saturation in the results de-
spite the number of interviews. After the third interview, a general pattern started
to emerge regarding which requirements the subjects believed to be important and
which were deemed less important.

36
 4. Methodology

4.3 Simulation study

In order to answer RQ2 and RQ3, there is a need to benchmark and assess the
effect of the found non-functional requirements on a system’s throughput and com-
putational complexity. Initially, the idea was to use a simulator provided by Toyota,
but due to the unavailability of their proprietary source code, it was decided to
create a new simulation system that met the required specifications and provided
greater control for this thesis.

The primary goal of simulating AGV trucks is to create a path planner which plans
routes for each agent in an AGV-based system (A warehouse). The algorithm should
take into account the movements of both moving and stationary objects to ensure
efficient and safe route planning. The starting point for the path planner is to create
a variation of the A-star algorithm and make it based on both time and space. For
simplicity in the following paragraphs, this will be referred to as 3D-A*.To achieve
this, the algorithm must take time into consideration, which would enable planning
the route for each AGV and avoiding collisions based on their planned paths.

In order to have a path planner to find the best routes there was a need to create a
layout for the subject warehouse. The original layout for the simulation system was
based on nodes and vertices in a weighted directed graph which allowed the creation
of complex warehouse layouts. However, it was later changed to a grid system to
enhance compatibility with the new algorithm and ease the testing and creating a
proof of concept phase.

In the updated grid system, each AGV can be allocated specific grids at any given
time (Depending on size, safety areas around the AGV, turn radius, etc), which the
3D-A* algorithm can then take into consideration for future planning. Once the
algorithm was functional and routes could be dynamically planned it was time to
move to the next step. In this step, there is a need to simulate a series of tasks
that can be executed by any given AGV. These tasks will have a start location,
a destination, and a time when the task can be started. In order to have these
predetermined tasks, two main area types were defined. The first area type act as
an inbound area for the warehouse similar to a conveyor belt, we will refer to them
as "Loading areas". The other area type will be similarly referred to as "output
stations" where the AGVs would finish their task.

At this point, the path planner needs to be able to process these incoming tasks in
a queue-based manner. For the purpose of this study, a FIFO queue was created
which could process AGVs once they were assigned a new task and required path
planning. In order to use this simulation tool for this study and simulate different
requirements and settings, there was a need to have control over different charac-
teristics and parameters of the simulation. For this, a number of parameters were
created and assigned as the input to the simulation tool that controlled the behavior
of the path planner, the AGVs themselves, the warehouse layout, and other miscel-
laneous settings.

37
4. Methodology

Figure 4.2: An illustration of the 3D-A* where z-axis represents the time.

These parameters were divided into five different categories; Environment settings,
AGV settings, Path planner settings, External factors, and debugging. The en-
vironment settings were used together with loading areas and output stations
to create desired warehouse layouts. The AGV settings are mainly directed for
AGVs but since speed and safety marginal are taken into consideration when path
planning, these settings also affect the behavior and results of the path planner. The
third category, Path planner settings was used for fine-tuning the heuristic and
the behavior of the path planner based on the different layouts. Another important
group of parameters, the External factors were used to create distributions to the
system which often occur in the real world. These parameters were used to simulate
problems during different stages of each task such as the loading/offloading stages
but also during transport. Finally, some Debugging parameters were used for test-
ing, debugging, and understanding the path planner’s behavior.

38
 4. Methodology

Figure 4.3: Overview of the final parameters of the simulation divided into five
categories.

39
4. Methodology

After the simulation system was operational, it was used to visualize the system’s
functionality and assess whether the algorithm is functioning as intended by check-
ing for collisions between the agents and assessing their movements. Using the
simulations, it is possible to simulate various scenarios and benchmark the system’s
throughput based on the parameterized NFRs. This together with different ware-
house layouts provide valuable insights into how these NFRs affect the AGV system’s
total throughput.

Using the simulation it was possible to assess the system’s performance (Total system
throughput) based on the given NFRs. For each NFR multiple simulations were per-
formed in 3 different value ranges which will be discussed later in the thesis. These
values reflected the Must, Plan, Wish ambitions of each of the parameterized
NFRs. In order to answer RQ3 multiple warehouse layouts were analyzed and two
layouts were created based on the main characteristics of these layouts. The main
characteristics of these layouts could be interpreted as loading areas being placed
around one or two edges of the warehouse while the output stations were clustered
together towards the center of the warehouse. These results were then used to an-
swer RQ2 and RQ3, which will be presented in Section 5.4

In order to detect any meaningful effects from the simulations and ensure that accu-
rate conclusions are made a series of statistical tests was performed. The data from
the simulations was used in a Shapiro-Wilks test to ascertain normality. The results
are normally distributed if p > α(0.05) and vice versa. Once the Shapiro-Wilks test
was completed, a paired t-test was conducted to test for statistical significance be-
tween the base case and the result of a simulated scenario. The paired t-test in this
study is done to identify which parameters lead to a statistically significant result
compared to the base case. The threshold for statistical significance is p > α(0.05).
The results for the statistical tests can be seen in Section 5.5

40
 5
Results

In this section, we will provide the results obtained during this study. It will start
with the interview study and literature review. From this a set of requirements will
be created, answering RQ1. After that, the results from simulating RQ2 and RQ3
will be presented along with a short statistical analysis of the results.

5.1 Literature review

In Table 5.1 below the findings of the literature review are presented. Each row
represents an article, which has a finding relevant to answering RQ1, a code posi-
tioning the article to a relevant NFR, and a subcode depending on the field of use
for each article.

Article Outline Relevant findings Code Subcode

[60] Scenario-based approach for the design Safety is critical for AV systems Safety Safety (AV)
of AVs.
Market will not grow unless
[61] Survey on scenario-based safety assess- Safety Safety (AV)
ment for AVs. safety is ensured to stakeholders

Derived safety goals and safety Safety standards used to derive NFRs.
[62] requirements for AV actuation systems A need to go beyond current state-of- Safety Safety (AV)
using the ISO 26262 standard. the-art.
An article discussing the current Current systems lack scalability, Scalability Scalability (AGV)
[4] state-of-the art for AGV robustness, flexibility and are comp- Robustness Robustness (AGV)
systems. utionally demanding. Efficiency Efficiency (AGV)
Provides a method
for probabilistically deriving the
[53] Failure free operation is deemed neces- Robustness Robustness
functionality of a localization sary for safe operation (AV)
system for AVs.
The authors ascertain the need
Current state-of-the-art in Robustness Robustness (AV)
[54] for the systems to keep
key aspects of autonomous driving. Safety Safety (AV)
working after a failure.
Communication issues, quantization
Robustness Robustness (AV)
[55] Investigation of platooning for AVs. errors and packet loss poses
Efficiency Efficiency (AV)
a significant threat to the technology.
Scalability is heavily dependant on the
[63] Goal-oriented approach for eliciting, Safety Safety (General)
modeling and reasoning regarding scal- context of the application domain
ability requirements

Table 5.1: Analysis table for literature review. Continued on the next page

41
5. Results

Article Outline Relevant findings Code Subcode

The capability of the technology is
[56] Studies the impact of the CCAM tech- strongly related to the robustness of Robustness Robustness
nology. (AV)
its components.
A review of sensing, communication The adaptation of the technology
[50] and human factors and controller depends largely on the reliability of Reliability Reliability (AV)
aspects for AVs. the vehicles
The reliability on a system level is largely
[51] Studied challenges and approaches for dependant on the properties of its sub- Reliability Reliability (AV)
testing of AVs.
components
A review on human-machine shared System reliability has a large influence
[52] Reliability Reliability (AV)
control systems for AVs. on the trust in the technology
A combination of corrective and
preventative maintenance strategies has
[2] Studied maintenance strategies for Reliability Reliability
AGV systems a positive impact on system (AGV)
throughput
Studied warehouse multi-robot
automation in discrete time- The computational time for the
[48] and space configuration, A∗ algorithm increased logarithmically Efficiency Efficiency
(AGV)
with a focus on path-planning with the amount of agents.
and task allocation
Comparative analysis of A∗ and The D ∗ Lite algorithm outperforms
[49] D ∗ Lite algorithms in a the A∗ algorithm when the path Efficiency Efficiency
(AGV)
single-AGV system. has to be re-planned often.
Outline of cyber-attack surfaces and The authors recommended following
[64] Security Security (AV)
security vulnerabilities for AV systems. the ISO 26262 standard.
Studied the security and safety of
AGV systems can only be deemed
[65] platooning AGVs in a production Security Security (AGV)
safe it is also secure.
facility.
Created a standardized test procedure
The authors propose using the
[66] for usability evaluation of Usability Usability (AV)
ISO 9241-11 to evaluate usability.
human-machine interfaces.
Ascertained the importance of a
[67] A study of human factors that impact human-centered approach for the Usability Usability (AV)
opinions regarding AVs
design of automotive systems.
Usability can be divided into five factors;
ease of learning, task efficiency, ease
[68] Studied usability requirements in gen- Usability Usability (Gen-
eral. of remembering, understandability and eral)
subjective satisfaction
Requirements (for usability) should
be stated i terms of the effectiveness,
[69] Usability requirements in general, focus Usability Usability (Gen-
on "ease-of-use". efficiency and satisfaction required eral)
in the context

42
 5. Results

Article Outline Relevant findings Code Subcode

Ascertains that the key performance

index is throughput. There
is a high desirability to have systems
[44] Presented a dynamic approach for Robustness, Robustness
path-planning in dense, grid-based that are capable of coping with Scalability (AGV), Scala-
AGV systems unexpected interruptions, and that are bility (AGV)

scalable.
Ascertains the use of safety standards
[57] Discusses safety standards for AGVs as a basis for safety requirements Safety Safety (AGV)
and adjoining fields
for AGV systems.
Believes that there is a gap
[58] Discusses safety standards for AGVs between existing safety standards Safety Safety (AGV)
and other autonomous machines
and the state-of-the art for the industry.
Ascertains that the production in
industry 4.0 is to be deemed as
[59] Discusses safety assurance in the con- safety-critical. Establishes that safety Safety Safety (AGV)
text of industry 4.0 and autonomous
machines such as AGVs requirements can be derived from
standards such as ISO 3691-4:2020.

As mentioned previously, articles regarding the use of requirements in AGV systems

are lacking. Therefore, an attempt was made to include articles from adjoining
fields, such as AVs. Whilst articles regarding requirements for the field were iden-
tified, they were not plentiful. A significant quantity of the articles discusses for
example "usability" from a quality perspective, rather than a requirements engineer-
ing perspective. For this reason, the results for each code will be discussed below,
taking into account the generalizability of the findings toward AGV systems.

The first identified code is Safety. There are articles from the field of AVs and AGVs.
The relevant findings from the former are that safety is critical for such systems and
that the market will not grow unless the technology satisfies this need. This is a
finding that ought to be generalizable across fields to some extent, even if the safety
aspect probably is more important for AVs than AGVs. The articles identified in the
latter ascertain that AGV systems are to be deemed as safety-critical. Furthermore,
they establish the utilization of standards such as ISO 3691-4:2020 as a basis for
deriving safety requirements.

The second identified code is Robustness. One relevant finding from the subcode
Robustness (AGV) is the lack of robustness in current AGV systems. This is due
to the centralized nature of the system, which has the limitations of not being very
proactive in avoiding unforeseen operating conditions. The findings from the sub-
code Robustness (AV) is not deemed very generalizable. This is due to the fact that
the issues for such systems, in terms of robustness, are very different from AGV
systems. The focus is on how robustness impacts the safety of the vehicle, rather
than the impact it has on some measure of efficiency, which is the focal point for
robustness in the context of AGV systems.

43
5. Results

The third identified code regards scalability. Relevant findings, within the context of
AGV systems, suggest that scalability is an issue. The literature suggests that this
is due to the fact that the computational effort required increases logarithmically
with the number of AGVs in the system.

The fourth identified code is Usability where no studies were found in the context of
AGVs. The chosen studies have been categorized either as Usability (AV) or Usabil-
ity (General). The latter has a more general focus on usability. The generalizability
of the findings for those can be considered as high, as they are not pertaining to
any particular type of system, but act more as general guidelines. The findings from
Usability (AV) suggest that it is appropriate to use the ISO 9241-11 standards to
evaluate usability for AV systems. This should be applicable to AGVs as well.

The fifth identified code is Reliability. The articles discussing it in the context of
AGVs mainly concern maintenance strategies and their impact on system through-
put. Articles in the subcode Reliability (AV) ascertains that this factor has a large
impact on the trust and adaptation of the technology. Whilst this might be true to
some extent in the case of AGVs, reliability is once again discussed as a factor in
system safety. This is important for AGVs as well, but most likely to an even higher
extent for an AV system.

The sixth code pertains to articles regarding Security. The articles from the subcode
Security AV recommended following the ISO 26262 standard to mitigate security
vulnerabilities for AV systems. Following standards to mitigate risks is something
that can be considered generalizable across fields. Research conducted in the field
of AGV systems ascertained the need for security as a necessity in order to achieve
a safe system.

The seventh and final code pertains to articles regarding Efficiency. Research con-
ducted in the field of AGVs considers this to be an issue, as the computational effort
for running these systems is high, increasing with layout and fleet size.

5.2 Interview study

The results of the interviews were analyzed using a thematic coding approach, as
mentioned in Section 4.2.2. The thematic analysis was performed with the help of
two tools, Atlas.ti and Miro. These tools helped with uncovering themes and codes
from the raw interview data. The generative AI from Atlas, which utilizes OpenAI’s
GPT model to generate codes, was used on the data first. This resulted in a set of
quotes and codes. These codes were not utilized for the final coding. The reason
for this was that the codes where deemed to narrow by the researchers. However,
it served to get some initial ideas regarding what codes could be used. Instead, the
thematic coding process was continued by a manual read-through of the transcripts.
Themes started to emerge, which where stored on a board in the Miro tool. The
data was grouped per interview and each code labeled with a different color. The

44
 5. Results

results are presented below in Figure 5.1.

Figure 5.1: Overview of the thematic coding board from Miro

In order to make some sense of the data, the thematic analysis was complemented
with an affinity diagram, which allows for themes and codes to be visually structured
and hierarchically organized. The affinity diagram was also created through the use
of Miro. Figure 5.2 below shows the result of this process.

Figure 5.2: Visualization of the affinity diagram. Link

Safety was the first emerging theme. During the course of the interviews, subjects
were asked to give their opinion regarding this NFR. This was generally perceived
as an important attribute. Several of the subjects explicitly named this as one of
the most important NFRs.

45
5. Results

"Having CE-marked products is a prerequisite to selling products within

certain markets" (E)

"All of our solutions follow the required standards ... The current
solutions are safe for humans, which is a part of the CE-certification.
This is very important for us" (B)

"Of the NFRs you’ve mentioned, I deem safety, robustness against

unexpected situations, and availability to be amongst the most
important" (E)

"There are several considerations on the hardware level ... There are
standards that companies are enforced to comply with, for example, the
Machinery Directive and the adjoining ISO and EN standards" (A)

"I consider that scalability, as the number of agents, robustness, and

safety is the most important factors" (D)

The second emerging theme was robustness. This was the NFR that was considered
the most important by the largest amount of subjects. Why this is can be sum-
marized with a quote from subject C, "The overall goal of an AGV system is to
execute all orders as fast as possible given the set of resources that you got". Many
of the subjects deemed that issues regarding the robustness of the systems were the
single factor that impacted this the most. This is due to factors such as obstacles,
unpredicted traffic, and failures during the loading function which causes delays.

"The hardware itself is very robust. However, the system’s robustness is

impacted by three types of factors. Firstly, different types of obstacles
in the vehicle path. Secondly, unpredicted traffic. Thirdly, failure
during the loading function" (C)

"I believe it’s very rare for this type of system to be robust. Ive only
seen a few systems that could be considered robust and this takes
constant surveillance because situations occur and multiply. Ive seen
systems that can only handle a few hours without surveillance without
something going wrong" (B)

"Its the issues regarding solving the problem with the robustness that is
important" (C)

46
 5. Results

"AGV systems are not very tolerant of external factors such as humans,
goods that are placed incorrectly on the pallets and so forth" (E)

The third emerging theme was reliability and availability. Most of the subjects be-
lieved that the availability of an AGV system should be between 98.5 and 99.99%.
Only one of the subjects mentioned this as one of the more important NFRs.

"The reliability for a single AGV should be above 99%. There should
also be zero downtime as a result of deadlocks and issues with software"
(A)

"On a single-AGV level, it should be available when it’s needed. The

overall availability is impacted by scheduled maintenance, service, and
so on. An adequate level is above 99%." (E)

"98-99% uptime, considering subsystem breakdowns and systematic

errors is to be considered adequate." (D)

The fourth emerging theme was scalability. There were some discrepancies regard-
ing the importance of this NFR. Two of the subjects named this amongst the most
important, whilst two of the others completely wrote it off.

"Scalability is not as important as some might consider. Most have

roughly the same performance in systems that dont exceed 20 AGVs.
Its not very often that customers decide to scale up the system without
changing/expanding the production facility. The times that this
happens its often negligible" (A)

"I only remember one project where this turned out to be an issue,
which we later solved. I have helped deliver systems with more than 50
AGVs. This is however not very common. I dont see scalability as a
big issue" (E)

"The solutions using MAPF dont have as many rules as say Djiktras.
In these old systems, everything was hardcoded and CPU power wasnt
a problem. For the MAPF however, a lot of CPU power is needed. It
increases almost logarithmically with the number of AGVs in the
system. This can become a bottleneck for systems with around 20-25

47
5. Results

agents or even as low as 10" (B)

The fifth emerging theme was efficiency. None of the subjects deemed this to be
amongst the more important NFRs, whilst one of them completely wrote it off.
However, there was close to consensus regarding the factors that impacted this par-
ticular NFR. The algorithmic complexity, the complexity of the layout, and the
quality of the input data. Neither of the subjects believed that communication has
a large impact on the system.

"I dont believe that computational efficiency is interesting in this

regard. Its the issues regarding solving the problem with robustness that
is important. If you have issues with computational efficiency, you are
using bad algorithms to solve the problem " (C)

"The communication isnt that limiting. It’s enough, depending on the

speed of the vehicle, to have a reactivity time of roughly 1 second" (C)

The sixth and final emerging theme was others. This regards issues with usability
and security, the final two NFRs considered in the study. Only (B) considered
security to be important, but still considered that scalability and robustness are more
important. (A) and (C) considered that usability could be considered as important
as some of the other NFRs. (A) even considered it to be the most important.

5.3 RQ1: Essential requirements

The findings from the literature review give little indication as to what NFRs are
to be considered the most important, as they are often discussed in isolation. On
the other hand, the results from the interview study ascertain that NFRs regarding
safety and robustness are the most important. An argument could have been made
to include requirements regarding scalability as well, which also was mentioned as an
issue in the literature. The number of subjects that thought they are of the highest
importance is the same as for safety. However, two of the subjects also explicitly
stated that scalability shouldn’t be an issue at all. On top of that, all subjects im-
plicitly stated the importance of safety requirements by ascertaining that a product
without a CE-certificate can’t be sold in major markets. Furthermore, the literature
suggests that the systems are safety-critical. Robustness on the other hand was ex-
plicitly stated as the most important NFR type by a majority of the subjects and
should therefore be included. The issue is furthermore discussed in the literature,
for example, the lack of robustness for deadlock-avoidance algorithms in centralized
systems.

All of the elicited requirements were discussed with subject E in order to ascer-
tain their correctness. The first set of requirements are stated in the planguage

48
 5. Results

format. This means having a field explaining the ambition, scale, meter and a set
of thresholds (must, plan, wish). The "ambition" states the purpose or objective of
the requirement. The "scale" defines the unit of measurement and "meter" how to
make those measurements. The "must" threshold establishes a minimally acceptable
result. The "plan" threshold defines the expected result. The "wish" threshold es-
tablishes a better than expected result. For the safety requirements another format
as chosen. The "description" is a description of what the requirement should do.
The "purpose" is the reasoning of why it is to be implemented. The "trigger" refers
to when the requirement occurs. The "pre-condition" refers to a set of conditions
that needs to be satisfied before the requirement can be triggered. The "frequency"
specifies how often the situation occurs in which the requirement is triggered.

The first stipulated requirement is RO1. This requirement emulates the issue ascer-
tained by subject (C) regarding the loading function. In practice, this is a situation
that occurs when an agent attempts to load a pallet from a rack and fails due
to the sensors identifying for example a faulty pallet or detecting plastic hanging
down from the packaging. This causes a fault that is non-recoverable and has to
be handled manually by an operator. The rate at which this occurs could not be
ascertained by the interviews or the literature study at a precise rate. However,
as ascertained by (B) this is something that occurs quite frequently and can be
calculated in percentages rather than per mille occurrences.

Tag: RO1 Robustness.LoadingFunction.Failures

Ambition Limit the downtime of the system by reducing the amount
of non-recoverable failures due to failures during the load-
ing function
Scale Percentage of locations blocked
Meter N/A
Must 7%
Plan 5%
Wish 3%

Table 5.2: Planguage table for frequency of loading failures

The second requirement, RO2 is also related to issues during the loading function.
As mentioned in the paragraph above, this is a non-recoverable fault that has to be
manually handled by an operator. The time this takes is highly dependent on the
environment (size of the layout, congestion) and the number of operators that are
available.

The third requirement, RO3 is regarding issues mentioned by all of the subjects.
Namely, obstacles blocking the path of the agent. This can be in the form of mis-
placed pallets, humans interacting with the system, and so forth. The issue is

49
5. Results

Tag: RO2 Robustness.LoadingFunction.ResponseTime

Ambition Fast manual recovery from failure during loading function
to limit system downtime
Scale Seconds from agent failing in loading function to manual
recovery by operator
Meter N/A
Must 60s
Plan 45s
Wish 30s

Table 5.3: Planguage table for recovery time after loading failures

touched upon in the literature, which mentions the lack of robustness for deadlock-
and collision-avoidance algorithms. They don’t work proactively but rely on the
sensors of the AGV. This is a recoverable fault. The occurrence of such faults is
once more highly dependent on the context. As mentioned by (B), this fault might
be negligible in completely isolated systems. However, these are not representative
of the industry as a whole. The factor increases with the number of humans inter-
acting with the system.

Tag: RO3 Robustness.Obstacles.Frequency

Ambition Limit the downtime of the system by reducing the amount
of recoverable failures due to obstacles in the vehicle path
Scale Percentage of tiles blocked in the layout
Meter N/A
Must 7%
Plan 5%
Wish 3%

Table 5.4: Planguage table for frequency of obstacles in vehicle path

The fourth requirement RO4 stipulates the recovery time after an object has been
in the vehicle path. The ISO 3691:4 standard stipulates that the AGV is allowed
to resume (at the earliest) two seconds after the object is removed from the sensor
range.

The safety requirements stipulated will attempt to emulate the ISO 3691-4:2020
standard, as this represents the state-of-the-art safety regulations according to the
literature. The use of such standards to identify requirements and conduct research

50
 5. Results

Tag: RO4 Robustness.Obstacles.ResponseTime

Ambition Fast automated recovery after vehicle detects an obstacle
in its path
Scale Seconds from agent detecting obstacle to automated recov-
ery
Meter N/A
Must 15s
Plan 10s
Wish 5s

Table 5.5: Planguage table for manual recovery after obstacle in vehicle path

in the field of requirements engineering has been documented. Furthermore, all

of the subjects mentioned the importance of standards within the industry. The
requirements will not be stipulated in the form of planguage as this would defeat
the purpose of analyzing their impact on the throughput. For example, a valid safety
planguage requirement for collisions could look like Table 5.6 below.

Tag: Example Safety.Obstacles.Collisions

Ambition A safe system with few occurrences of collisions between
an agent and an object
Scale Amount of collisions divided by amount of collisions
avoided
Meter N/A
Must 10−7
Plan 10−9
Wish 10−12

Table 5.6: Example of planguage requirement for safety

The example provided in Table 5.6 poses two different issues. Firstly, the purpose of
the study is to examine the impact on throughput. This requirement would measure
how safe the system is. Furthermore, if this type of requirement would have been
included, the impact it would have on the throughput of the system is infinitesimal.
This is due to very small likelihood of collision occurring. Furthermore, if it occurs it
doesn’t necessarily entail a lengthy stop. A collision would be defined as any contact
being made by the vehicle whilst it is in motion. For example, the vehicle could
make contact with a pallet at very low speed. This would force a manual restart,
which would not take very long. This combined with the low probability of such a
collision entails that the impact that this would have on system throughput is almost

51
5. Results

negligible. Secondly, the algorithm used within the system is collision-free. In order
to introduce collisions into the system, you would have to hard code parameters.
Their impact would be directly proportional to the values chosen. Because of these
reasons, the safety requirements will be stated as constraints to the system rather
than in planguage. The particulars of this type of requirements where explained in
the second paragraph in this section.

The first safety requirement, SAF1 concerns the speed control of the agent, in other
words, what speed is allowed within certain distances to objects. The ISO 3691:4
standard is specific in this case and the distances and velocities can be found in table
2.3. The second and third safety requirements, SAF2, SAF3 are alternations of
SAF1. SAF2 will emulate a system with stricter speed control, where the distances
to the objects are increased to three tiles. SAF3 will emulate a system with less
strict speed control, where the distances are lowered to one tile.

Tag: SAF1 Safety.SpeedControl1

Description The system shall adhere to the limitations of speed control
as stipulated in Table 2.3. The distance is set to two tiles
in the simulation environment
Purpose Compliance with Machinery Directive allowing for safe op-
eration of the agent.
Trigger Agent within specified distance to object(s)
Pre-condition Agent is in automatic mode with functioning personal de-
tection means
Frequency Any time agent identifies object(s) within the defined dis-
tances

Table 5.7: Requirement regarding speed control for agent. Distance to object(s) is
2 tiles

52
 5. Results

Tag: SAF1 Safety.SpeedControl2

Description The system shall adhere to the limitations of speed control
as stipulated in Table 2.3. The distance to an object is set
to three tiles in the simulation environment
Purpose Operation of agents with a larger marginal to the closest
object(s) allowing for a safer system.
Trigger Agent within specified distance to object(s)
Pre-condition Agent is in automatic mode with functioning personal de-
tection means
Frequency Any time agent identifies object(s) within the defined dis-
tances

Table 5.8: Requirement regarding speed control for agent. Distance to object(s) is
3 tiles

Tag: SAF1 Safety.SpeedControl3

Description The system shall adhere to the limitations of speed control
as stipulated in Table 2.3. The distance to an object is set
to one tile in the simulation environment
Purpose Operation of agents with a smaller marginal to the closest
object making a trade-off between efficiency and safety.
Trigger Agent within specified distance to object(s)
Pre-condition Agent is in automatic mode with functioning personal de-
tection means
Frequency Any time agent identifies object(s) within the defined dis-
tances

Table 5.9: Requirement regarding speed control for agent. Distance to object(s) is
1 tile

53
5. Results

The next set of requirements (SAF4, SAF5 and SAF6) will concern the rated
speed of the agents. The rated speed is the travel speed as defined by the man-
ufacturer according to ISO 3691-4:2020. This is of course something that varies
depending on the application and context in which the agents are used. Available
solutions from Toyota have a rated speed ranging between 2 - 2.2 m/s. SAF4 will
emulate such a solution, with a rated speed of 2 m/s. SAF5 and SAF6 will repre-
sent solutions with a lower and higher rated speed respectively.

Tag: SAF4 Safety.RatedSpeed1

Description The system shall have a maximum velocity of 2 m/s
Purpose Establish max speed for agent to ensure safe operation.
Trigger Agent within specified distance to object(s) allowing max-
imum rated speed.
Pre-condition Agent is in automatic mode with functioning personal de-
tection means.
Frequency Any time agent is within a permissible distance to an object
allowing max rated speed according to SAF1, SAF2 or
SAF3.

Table 5.10: Requirement regarding the rated speed for an agent. The rated speed
is set at 2 m/s

Tag: SAF5 Safety.RatedSpeed2

Description The system shall have a maximum velocity of 1.5 m/s
Purpose Establish max speed for agent to ensure safe operation.
Trigger Agent within specified distance to object(s) allowing max-
imum rated speed.
Pre-condition Agent is in automatic mode with functioning personal de-
tection means.
Frequency Any time agent is within a permissible distance to an object
allowing max rated speed according to SAF1, SAF2 or
SAF3.

Table 5.11: Requirement regarding the rated speed for an agent. The rated speed
is set at 1 m/s

54
 5. Results

Tag: SAF6 Safety.RatedSpeed3

Description The system shall have a maximum velocity of 2.5 m/s
Purpose Establish max speed for agent to ensure safe operation.
Trigger Agent within specified distance to object(s) allowing max-
imum rated speed.
Pre-condition Agent is in automatic mode with functioning personal de-
tection means.
Frequency Any time agent is within a permissible distance to an object
allowing max rated speed according to SAF1, SAF2 or
SAF3.

Table 5.12: Requirement regarding the rated speed for an agent. The rated speed
is set at 3 m/s

5.4 Results from the simulations

In this section we provide the results from our computer simulation, answering RQ2
and RQ3. Tables 5.13 and 5.14 represent the two layouts and show the different
simulation cases and their respective parameters, the total throughput, and their
respective deviation from the base case. The throughput is the median value ob-
tained after five iterations. The low number of iterations is due to the complexity
of the simulations. One iteration of each simulation case took roughly five minutes
(depending on the layout size and amount of AGVs). Figure 5.3 and Figure 5.4 visu-
alize the throughput deviation from the base case for each simulation configuration
in a histogram.

5.4.1 Layout 1: 150m x 150m

Table 5.13 summarizes the results for the larger layout (150m x 150m) given five
iterations of the simulation for each case. The values for the throughput is the me-
dian of these five iterations. For SAF4-6 we can see that the throughput decreases
significantly (-17) when lowering the rated speed. Increasing the rated speed had a
positive effect, albeit not as large (+11). Interestingly, both lowering and increasing
the amount of AGVs in the system decreased the output (-8 and -1 respectively).
Regarding the effects of speed control, SAF1-3, lowering the allowed distance had
a small positive effect (+2). Increasing the distance on the other hand had a sig-
nificant negative effect (-16). Achieving the "wish" threshold for problems during
the loading function, RO1, had a small, positive effect on the throughput (+2).
Decreasing it to the "must" level had a small negative effect (-3). Interestingly,
increasing the probability (R02) had no impact on the throughput. The random
stops also had a small impact on the "wish" and "must" levels (2, -2 respectively).
The same goes for increasing the probability (R03). The "must" level implied a

55
5. Results

decrease of -1 whilst the "wish" level implied a increase of 2. In Figure 5.3 below
the deviations from the base case for each of the simulations cases are shown.

)
)

03
01

(R
(R

se
3)

on
es

ca
1-

su

ti
F

se
ra
is
A

ba
)

)
du
s

t
(S

on

02

04
V

pu
6)

om
G

(R

R
ti
ol

ps
4-

gh
A

nc

(
tr

to

fr
F

ou
y

y
of

on

fu
A

lit

lit
s

on
hr
(S

g-

m
C
r

bi

bi
be

ti
lt
in

do
ba

ba
d

ia
um

ad
ee

ee

ta
an

ev
ro

ro

To
Sp

Sp

Lo
Simulation
N

D
R
P

P
Base case 5 2 2 45 0.05 10 0.05 36 0
#Layout 1: 01 3 2 2 45 0.05 10 0.05 28 -8
#Layout 1: 02 10 2 2 45 0.05 10 0.05 35 -1
#Layout 1: 03 5 1 2 45 0.05 10 0.05 19 -17
#Layout 1: 04 5 3 2 45 0.05 10 0.05 47 11
#Layout 1: 05 5 2 1 45 0.05 10 0.05 38 2
#Layout 1: 06 5 2 4 45 0.05 10 0.05 20 -16
#Layout 1: 07 5 2 2 30 0.05 10 0.05 38 2
#Layout 1: 08 5 2 2 60 0.05 10 0.05 33 -3
#Layout 1: 09 5 2 2 45 0.03 10 0.05 36 0
#Layout 1: 10 5 2 2 45 0.07 10 0.05 36 0
#Layout 1: 11 5 2 2 45 0.05 5 0.05 39 3
#Layout 1: 12 5 2 2 45 0.05 15 0.05 34 -2
#Layout 1: 13 5 2 2 45 0.05 10 0.03 38 2
#Layout 1: 14 5 2 2 45 0.05 10 0.07 35 -1

Table 5.13: Simulation results based on each factor, Layout 1 (150m x 150m)

56
 5. Results

Figure 5.3: Total system throughput deviation from the base case given each
parameter (Layout 1 - 150m x 150m).

5.4.2 Layout 2: 75m x 75m

Table 5.14 summarizes the results for the smaller layout (75m x 75m) given five
iterations of the simulation for each case. The values for the throughput are the
median of these five iterations. When decreasing the amount of AGVs, the through-
put decreased of the system deviated -30 from the base case. When increasing the
amount of AGVs, the throughput decreased by -3. Regarding the rated speed re-
quirements (SAF4-6), decreasing the speed had a substantial effect, lowering the
throughput by 50. Increasing the speed had the opposite, but smaller effect (+38).
Implementing the speed control requirements (SAF1-3) resulted in an increase of
18 and a decrease of 64 when decreasing and increasing the distance respectively.
Requirements regarding the loading function (RO1) had the effect of increasing the
throughput by 2 when the loading time was on the "wish" level. Increasing the time
for issues during the loading function to the "must" level resulted in a -3 decrease
in throughput. The results for R02 were unexpected. Both the higher and lower
probability resulted in a decrease from the base case (-29, -13 respectively). The
duration for random stops (RO3) had a positive effect on the "wish" level (+7) and
a negative effect on the "must" level (-3). Similar to the results for R02, the results
for R04 were unexpected. The increase in probability yielded a decrease of -15,
whilst the decreased probability lowered the throughput by -7. Figure 5.4 visualizes
the deviations from the base case for each of the simulation cases for layout two.

57
5. Results

)
)

03
01

(R
(R

se
3)

n
es

ca
io
1-

su

at

se
F

is

ur
A

ba
)

)
s

t
on

02

04
V

(S

pu
d
6)

om
G

(R

(R
ti

ps
4-

gh
l
A

ro

nc

to

fr
F

ou
y

y
nt
of

fu
A

lit

lit
s

on
hr
(S

co

om
r

ng

bi

bi
be

ti
lt
ba

ba
d

ia
i

d
um

ad
ee

ee

ta
an

ev
ro

ro

To
Sp

Sp

Lo
Simulation
N

D
R
P

P
Base case 5 2 2 45 0.05 10 0.05 103 0
#Layout 2: 01 3 2 2 45 0.05 10 0.05 71 -30
#Layout 2: 02 10 2 2 45 0.05 10 0.05 103 -3
#Layout 2: 03 5 1 2 45 0.05 10 0.05 55 -50
#Layout 2: 04 5 3 2 45 0.05 10 0.05 139 38
#Layout 2: 05 5 2 1 45 0.05 10 0.05 109 9
#Layout 2: 06 5 2 3 45 0.05 10 0.05 36 -64
#Layout 2: 07 5 2 2 30 0.05 10 0.05 103 2
#Layout 2: 08 5 2 2 60 0.05 10 0.05 99 -3
#Layout 2: 09 5 2 2 45 0.03 10 0.05 68 -29
#Layout 2: 10 5 2 2 45 0.07 10 0.05 84 -13
#Layout 2: 11 5 2 2 45 0.05 5 0.05 110 7
#Layout 2: 12 5 2 2 45 0.05 15 0.05 97 -3
#Layout 2: 13 5 2 2 45 0.05 10 0.03 88 -15
#Layout 2: 14 5 2 2 45 0.05 10 0.07 87 -7

Table 5.14: Simulation results based on each factor, Layout 2 (75m x 75m)

58
 5. Results

Figure 5.4: Total system throughput deviation from the base case given each
parameter (Layout 2 - 75m x 75m).

59
5. Results

5.5 Statistical significance

Once the different simulation configurations were simulated, it was decided to verify
that the different simulation runs for each configuration were normally distributed.
Normal distribution is not a must in this case since there can be some variations
between each simulation but the data range should be reasonable.

The Shapiro-Wilks test was utilized to analyze if the simulations were normally
distributed or not. The null hypothesis for the test is that the simulation data is
normally distributed. The p-value (the threshold for significance level annotated by
α) was decided to be 0.05 which is generally considered a conventional threshold.
The table below represents the p-value for each simulation configuration and whether
or not it exceeded the threshold. A p-value > α indicates a normal distribution while
a p-value < α suggests a significant deviation from a normal distribution. As seen
in the table below most results were normally distributed but since the sample size
is very small (5 simulations/configuration) the results are not definitive and should
be interpreted with caution.

Requirement Layout 1 p-value Layout 2 p-value

BASE 0.618 > α 0.350 > α
No. AGVs: 3 0.404 > α 0.920 > α
No. AGVs: 10 0.657 > α 0.287 > α
Speed: 1m/s 0.993 > α 0.486 > α
Speed: 3m/s 1.000 > α 0.209 > α
Marginal: 1m 0.373 > α 0.999 > α
Marginal: 4m 0.998 > α 0.906 > α
H. Duration: 30s 0.965 > α 0.519 > α
H. Duration: 60s 0.997 > α 0.075 > α
H. Probability: 3% 0.0187 0.310 > α
H. Probability: 7% 0.997 > α 0.920 > α
T. Duration: 30s 0.047 0.620 > α
T. Duration: 60s 0.722 > α 1.000 > α
T. Probability: 3% 0.959 > α 0.820 > α
T. Probability: 7% 0.190 > α 0.804 > α

Table 5.15: Shapiro-Wilk test results per parameter after 5 iterations. (p > α(0.05)
-> Assumes normal distribution)

60
 5. Results

Once the data was deemed to be rational and within the acceptance level there was
a need to identify which configurations significantly deviated from the base case.
For this, the Paired t-test was utilized which is used to compare the means of two
different data sets and identify a statistically significant difference. Here the null
hypothesis was declared as not significant and once again a threshold value of 0.05
was used as the α. The results for the paired t-test are visualized in the table below
where a p-value < α indicates a statistically significant deviation from the base case.

Requirement Layout 1 p-value Layout 2 p-value

No. AGVs: 3 0.051 0.001 < α
No. AGVs: 10 0.583 0.198
Speed: 1m/s 0.001 < α 0.000 < α
Speed: 3m/s 0.009 < α 0.0002 < α
Marginal: 1m 0.378 0.037 < α
Marginal: 3m 0.005 < α 0.000 < α
H. Duration: 30s 0.436 0.513
H. Duration: 60s 0.833 0.454
H. Probability: 3% 0.454 0.000 < α
H. Probability: 7% 0.811 0.003 < α
T. Duration: 30s 0.154 0.090
T. Duration: 60s 0.683 0.495
T. Probability: 3% 0.030 < α 0.006 < α
T. Probability: 7% 0.796 0.199

Table 5.16: Paired t-test results per parameter after 5 iterations. (p < α(0.05) ->
Statistically significant)

61
5. Results

62
 6
Discussion

In this chapter we will discuss the findings of the study and common threats to
validity

6.1 Research question 1

"Identify essential non-functional requirements for AGV systems

within a warehouse setting"

Based on the results from the literature review and the interview study, NFRs re-
garding robustness and safety were to be considered the "essential" requirements.
This doesn’t mean that the other requirements presented in the thesis aren’t im-
portant - it serves as a way to get a reasonable scope. Out of the remaining NFRs,
reliability/availability wasn’t really seen as an issue for AGV systems. Efficiency
was not deemed important by any of the subjects, and one of them wrote it off com-
pletely. Scalability could have been included since some subjects referred to it as a
non-issue, whilst others deemed that it was a problem. Its impact is however studied
to some extent with the increase/decrease in the number of AGVs as a case during
the simulation. However, what would have been interesting in that case would be to
get measurements regarding the computational effort, as this seems to be the most
limiting factor rather than throughput. Due to limitations in time during the inter-
views, there was never a deep-dive into the subjects of security and usability. The
subjects were asked after having discussed the other NFRs whether those would be
considered more important. This could be seen as somewhat leading, and perhaps
the answers would have been different if the questions were framed in some other way.

The combination of a literature review and an interview study was conducted in

order to achieve triangulation and avoid mono-method biases. However, the results
from the literature review were not very helpful in answering what type of require-
ments would be used for RQ1. This is mostly due to the fact that most of the
articles aren’t discussing "safety" or "scalability" from a requirements engineering
standpoint, but rather a quality perspective. Furthermore, they are most often dis-
cussed in isolation, so to gauge which ones are the most important given the articles
in the literature review is difficult. They did give insight into how the requirements
could look, however, for example using existing standards for safety requirements.

63
6. Discussion

The interview study was very insightful and helped answer RQ1. If the researchers
were to conduct a similar study again, this would be the focal point of that study.
By perhaps doubling the number of interviews and having a few follow-ups regarding
specific requirements, the answers for RQ1 could have been more reliable.

The requirements that were chosen for the study were based both on the answers
from the interview study (RO1-4) and findings in the literature (SAF1-6). For the
former, it was a bit difficult to establish the exact properties of such requirements.
The question was asked in the latter interviews, and the most straightforward an-
swer was that these types of unexpected operating conditions would be occurring in
percentages. It was a difficult question to answer, as it’s heavily context-dependent.
A system in full isolation might not suffer from these conditions at all, whilst a sys-
tem that is interacting with humans and manual forklifts might have large issues.
For the safety requirements (SAF1-6), a choice was made to make them more as
constraints to the system. This choice was made due to the fact that a normal plan-
guage NFR would be derived from a set of functional requirements. The planguage
requirement would be stipulated in terms of for example the risk of collision. This
risk would be very low, say between 10−9 and 10−12 . The impact that this would
have on the throughput is negligible. On the other hand, the system constraint that
this NFR is derived from will have an impact on the throughput, and therefore those
constraints were chosen to represent safety NFRs instead.

Summary
The types of NFRs that were deemed most important by the domain experts
were robustness and safety. Findings from the literature review and back-
ground helped shape the safety requirements (SAF1 - SAF6). Findings from
the interview study helped shape the robustness requirements (R01 - R04).

6.2 Research question 2

"Evaluate the impact of the non-functional requirements found in

RQ1 on the system throughput"

This pertains to cases 03-14 in Tables 5.13 and 5.14. Regarding the impact of dif-
ferent distances for speed control (SAF1 - SAF3), the results were as expected.
An increase in the distance for speed control led to a decrease in the throughput for
both layouts. This is rational, as a longer distance for speed control means longer
travel paths for the AGVs, and increased queue and waiting times at for example the
output stations. On the other end, decreasing the distances means decreased travel
path lengths and decreased queue and waiting times, yielding a positive impact on
the throughput.

The impact of the rated speed for the vehicles (SAF4 - SAF6) were also expected.

64
 6. Discussion

Decreasing the rated speed led to a negative impact on system throughput. A re-
verse, but smaller effect, was identified when the rated speed was increased. This
is most likely due to the fact that increasing the speed also increases congestion in
the system, which has a negative effect on the throughput.

Regarding the impact of R01 and R03 the results were as expected. In both lay-
outs, the deviation from the base case had a negligible effect on the throughput
when the time was decreased. One might wonder why the impact of the two was
very similar when the loading function errors were considerably longer. This is due
to the fact that there are more tiles in the layout where the random stops could
occur, relative to the amount of tiles where loading function errors could occur.

The results for R02 and R04 were as expected for layout one. They had a slightly
positive impact when increased and a slight negative impact when decreased. How-
ever, for the smaller layout (75m x 75m) the throughput decreased both when in-
creasing and decreasing the likelihood of loading time errors. The rational result
would be for the throughput to increase when the probability decreases, such as it
did in the first layout. The results were analyzed and inspected visually to give some
idea of why this was the case. The rational conclusion was that the waiting times
for the output station were a bottleneck in the smaller layout. When the agents
stopped in the middle of an aisle, they were less likely to be in the path of another
agent, and they decreased the time spent by the output stations. Thus decreasing
the probability of these random failures, led to the agents having to take routes
around each other instead of taking the shortest path leading to a negative impact
on the total system throughput. This is not something that is to be considered usual
for these types of systems. Rather, it is an edge case caused by the particulars in
the layout.

To relate back to the literature, a relevant finding was that there is a high desirabil-
ity to have systems that are capable of coping with unexpected operation conditions.
The results that has been obtained through this research further proves this point,
and provides an example to what degree different robustness issues that current
solutions are experiencing impact the throughput. As the results clearly show that
the impact of such issues are negative, it ought to provide incentives for industrial
practitioners and researchers to solve these issues. Regarding the safety require-
ments, it is of course interesting for practitioners to gain insight to how for example
standards impact system throughput. However, these standards exist for a reason
and sacrificing system safety for increased throughput is not advisable. Instead, the
industry should be working towards even more safe systems. Doing so could increase
the trust in the technology, leading to further possibilities for sales and investments
and a heightened growth rate for the industry as a whole.

65
6. Discussion

Summary
The results for SAF1 - SAF6 was as expected. Increasing the rated speed,
and decreasing the distance for speed control had a positive impact on the
throughput, and vice versa. The simulations yielded unexpected results in
layout 2 for R02 and R04. It was ascertained that this represents an edge
case for the system. The results for R01 and R03 were expected. There was
a slight increase in throughput when the time for loading function failures and
random stops was decreased, and vice versa.

6.3 Research question 3

"Evaluate the impact of the non-functional requirements found in

RQ1 on the system throughput given a set of specific, real-world
scenarios"

This pertains to the differences between the results found in the two layouts, as well
as the increase and decrease in the number of agents in the system (cases 01 and 02).

Regarding the former, decreasing the layout size had the impact of intensifying the
results from RQ2. This is most likely a result of shorter travel times, which leads
to the different requirements having an overall larger effect for each case. Regarding
the rated speed SAF4-6, the overall impact was similar for both layouts. When the
speed was decreased, the percentage loss of throughput was similar (47.2% for the
larger and 48.5% for the larger). Similar results were identified for increasing the
rated speed, which yielded an increased throughput of 36.8% for the smaller layout
and 30.5% for the larger layout. The differences in effect size could be explained by
a reduced relative impact on the travel time for the larger system. Increasing the
speed does reduce the travel time in the larger system, but the relative impact would
be smaller than for a system where the distances between objectives are shorter.

Regarding the speed control requirements (SAF1-3) the results were amplified in
the smaller system. Decreasing the distance had a relative impact of +5.6% for the
larger system and +8.7% for the smaller system. The discrepancies in the relative
impact could be explained by the fact that the smaller system is already congested,
and there is less empty space for the agents to operate in. The relative impact of
increasing the distance for the speed control requirements was -44.4% for the larger
layout and -62.1% for the smaller layout. This can be explained with the same
reasoning, the smaller system is already more congested.

Concerning the probabilities of loading function errors and random stops (R02 and
R04), the relative impact when increasing the probabilities in the larger layout was
measured at +2.7% for both NFRs. For the smaller layout, the relative impact
of R02 and R04 was measured at -28.3% and -14.5% respectively. As mentioned
in the previous section, there were discrepancies regarding the results between the

66
 6. Discussion

different layouts. This can be explained by system bottlenecks in the second layout.
As the travel paths are considerably shortened, the agents will spend more time at
the output stations, which becomes the system’s bottleneck. When an agent stops
inside an aisle due to a random stop or issues during the loading function, they
inadvertently impact the system throughput in a positive way as the queue and
waiting times for agents at the output stations are lowered.

The relative impacts of R01 and R03 (time for loading function errors, random
stops) on the different layouts were similar. When decreasing the time for R01, the
relative impact on the system throughput for the larger layout was 5.6% and 6.7%
for the smaller layout. When increasing the time, the relative impact on system
throughput was -8.3% for the larger layout and -2.9% for the smaller layout. When
the time for R03 was increased in the larger layout, the relative impact was -8.3%
while for the smaller layout, it was measured at -2.9%.

The impact of decreasing the number of agents in the system was as expected. For
the larger layout, the relative impact was measured at -22.2%. For the smaller lay-
out, the relative impact was measured at -29.1%. The reason why the relation is
not linear (which would entail -40% in throughput for both systems) is most likely
due to the decreased congestion in the system. This could also explain why the
impact is larger in the smaller layout, where congestion has a bigger impact. When
the amount of AGVs was increased, the relative impact for the larger layout was
-2.7%. The relative impact for the smaller layout was measured at -2.9%. These
results might seem to be unexpected, but there is an explanation. Rationally, one
would expect diminishing returns, where the throughput is increased and the effi-
ciency of each agent is decreased due to congestion. However, the system has in
this case reached a point where (in both layouts) the output stations are the system
bottleneck. Increasing the number of agents in such a system would not increase
the amount of tasks it could handle, but only add to the congestion. This would
lead to a small decrease in throughput, as can be seen for both layouts.

Summary
The impact that the different NFRs have on the system’s throughput is gener-
ally amplified in the smaller layout. Decreasing the fleet size leads to a larger,
relative decrease in throughput for the smaller layout due to increased conges-
tion compared to the larger layout. Increasing the fleet size have diminishing
returns until the system is at max capacity and an increase in agents mostly
leads to more congestion, lowering the throughput.

6.4 Generalizable experiences

The set of requirements that was studied is highly specific. In order to generalize
our findings beyond this specific context, we will offer advice to researchers aiming
to conduct similar research.

67
6. Discussion

The conducted study utilized an extensive background and literature review, as well
as an interview study to gain insights regarding the most important NFRs for this
type of system. If the researchers were to conduct a similar study again, the inter-
view study and obtaining more participants would be given higher focus in order to
answer this question. A focus group for the initial stages, combined with follow-up
interviews regarding specific requirements, would have been ideal.

In our case, a simulation tool was created to assess the impact of the NFRs. How-
ever, it is not necessary to use a simulation tool to assess the impact of NFRs on
AGV systems. It could, for example, be replaced with a mathematical formula to
estimate system throughput. Whilst this would abstract away some of the real-world
complexities, it would also save a substantial amount of time. This saved time could
have been used to gain more in-depth knowledge regarding the specific requirements
or conduct more interviews. It would also have mitigated the issue of the complexity
of the simulation tool, which led the researchers to use only five iterations for each
base case.

Whilst the interview subjects were certain about the type of issues AGV systems
face, they were hesitant to provide exact values to determine the frequency of these
problems. A time study at an actual warehouse or production facility that uses
AGVs could ascertain those values.

6.5 Threats to validity

In this section, we will discuss the impact of different validity threats. Sequen-
tially, we will discuss the conclusion validity, internal validity, construct validity,
and external validity of the research.

6.5.1 Conclusion validity

Conclusion validity pertains to the relationship being tested and refers to whether
it is reasonable to presume covariation. There are several major threats, for ex-
ample; low statistical power, random irrelevancies in the experimental setting, and
the random heterogeneity of possible respondents. Random irrelevancies refer to
noise or extraneous factors that are irrelevant to the research questions but still
might impact the results (such as randomness introduced by the simulation in this
study). The random heterogeneity of the respondents refers to the natural variation
of characteristics among the individuals that partake in the interview study. Tests
regarding the statistical significance was discussed in Section 4.3 and the results are
shown in Section 5.5.

Worth mentioning regarding the statistical significance tests however, are the fun-
damental limitations of the tests that have been used. Not all of the values in Table
5.15 are normally distributed. This departure from normality can affect the validity
of the results. Whilst t-tests can be used to detect significance for smaller sample

68
 6. Discussion

size (five iterations in this case), there is an argument to be made for using a higher
p value to detect significance with a smaller sample size. Furthermore, the use of
paired t-tests could have impacted the generalizability of the findings, as they per-
tain specifically to the data being compared.

Regarding random irrelevancies in the experimental setting, this could impact the
results for RQ2 and RQ3. However, both of the questions have been answered with
the help of a computer simulation. In other words, it takes place in a non-empirical
setting, where all the rules are predefined. For that reason, there shouldn’t be any
statistical noise or extraneous factors to consider. As a result, this factor was deemed
to not have impacted the results in any major way.

The random heterogeneity of respondents is a threat to the results obtained from

the interview study, which has been used to answer RQ1. The results from this
section may introduce variability that is not accounted for in the study, and the
generalizability of any findings could have been impacted.

6.5.2 Internal validity

Internal validity pertains to the causality between potential relationships in the
study and speaks for the validity of the research itself. Common threats to internal
validity include; selection bias, history, maturity and testing. Selection bias occurs
when participants of a study are not selected randomly. History and maturity refer
to external events that occur during the course of a study and changes that occur
naturally over time, respectively. Finally, testing refers to effects that occur when
participants repeatedly are tested for some response, which might influence their
behavior.

Regarding selection bias, the participants of the study were selected based on avail-
ability for the interviews. This might have impacted the causality of relationships
identified in the interview study and as a result the answers to RQ1. However, the
respondents were deemed to be representative of the industry as they include a range
of individuals with different experience levels and roles within different companies.

Concerning the history and maturity, it refers to external events that occur during
the course of a study and changes that occur naturally over time, respectively. Nei-
ther is deemed to have an impact on the research at hand.

Testing should not be an issue for any of the research questions. RQ2 and RQ3
are answered with a computer simulation and RQ1 does not involve any repeated
interviews or tests with the same subjects.

69
6. Discussion

6.5.3 Construct validity

Construct validity pertains to the transformation of an idea, concept, or behavior
into a functioning and operating reality. It refers to whether specific measurements
are model-independent and dependent on variables by which the hypothesized theory
is constructed. Common threats include; construct under-representation, irrelevant
variance in the construct, mono-operation bias, and mono-method bias. Construct
under-representation occurs when the model doesn’t adequately represent the re-
ality it is trying to emulate. As a result, measurements attained from said model
might be erroneous. Mono-operation bias occurs when only a single measure or
operationalization is used to represent reality. Mono-method bias occurs when only
one method is used to collect the data for the model.

The simulation model used to answer RQ2 and RQ3 will in part be suffering from
construct under-representation, as many of the complexities of a real-world ware-
house environment has to be abstracted away in the model. Irrelevant variances
in the construct are less of a problem for the model. This is due to the fact that
the variables will be controlled and pre-programmed and their values will not be
influenced by external factors. However, the variance may still occur due to the fact
that the simulation is affected by for example the processing speed of the computer
and also the fact that each simulation has some randomness to it.

Mono-operation bias could affect the results for RQ2 and RQ3. This is due to the
fact that the model might fail to capture the complexities of a real-world warehouse
setting.

The researchers tried to mitigate any mono-method bias by collecting data to an-
swer RQ1 using a combined literature and interview study. However, finding an
answer to this question in the literature was difficult. The results heavily relied on
the interview study as a result - which might induce bias. This bias could have been
propagated, as the results from RQ1 were used to simulate the results for RQ2 and
RQ3.

6.5.4 External validity

External validity pertains to the generalization of relationships across settings, ac-
tors, and time. In other words, external validity is concerned with the extent to
which the results of a study are relevant and applicable in the real world. The
choice of settings and actors might impact the generalizability. Furthermore, the
lack of a representative sample and the choice of methodology can impact the gen-
eralizability of any results.

The study’s ability to be generalized is limited. This means that the findings of the
study may not be applicable or relevant to other populations, settings, or conditions
beyond the specific sample or context in which the study was conducted. This lim-
itation is acknowledged in the literature, which suggests that computer simulations

70
 6. Discussion

are inherently limited in their ability to be generalized. This is because computer

simulations aim to model specific types of real-world systems, which may not ac-
curately represent all possible scenarios or conditions. As a result, the findings for
RQ2 and RQ3 in the study may be affected by this limitation. It is important
to note, however, that the limitations of generalizability do not necessarily detract
from the value or importance of the study’s findings within the specific context in
which it was conducted. Moreover, the generalization of findings is not the goal of
all studies. In this case, the aim is to gain an understanding of the subject rather
than generalize it across actors, contexts, and time.

The lack of a representative sample might have impacted the study results. As
mentioned previously, the utilized sampling method for the interview study was
convenience sampling which as mentioned earlier might have impacted the general-
izability of the results for RQ1.

71
6. Discussion

72
 7
Conclusion

The purpose of this study was to evaluate the impact that essential NFRs have on
system throughput and how it is impacted by the choice of layout and fleet size.
This was achieved by a mixed-methodology approach. The essential requirements
were identified with the help of an interview study and a literature review. The im-
pact that these requirements had on system throughput was then simulated using a
tool developed during the study.

The simulation tool was developed with the goal of emulating AGVs in a warehouse
environment. It utilizes a variation of the A∗ algorithm, based on both time and
space. Two different layouts were created for the simulation, representing different
warehouse environments. A smaller layout (75m x 75m) and a larger layout (150m
x 150m).

The interview study ascertained that requirements regarding safety and robustness
were to be considered the most important. Findings from the literature review and
background helped shape a set of safety requirements regarding speed control and
velocity. Findings from the interview study helped shape four robustness require-
ments regarding loading function issues and random stops along with their respective
probabilities.

The requirements were used as input to the simulation model, allowing for bench-
marking of the system throughput. For each layout, each of the requirements (and
thresholds) was varied independently and compared to a base case. The results as-
certain that the effect that the NFRs have on system throughput is profound. The
researchers established that the impact of the different NFRs generally are amplified
in the smaller layout. Decreasing the fleet size leads to a larger, relative decrease in
throughput for the smaller layout due to increased congestion. Increasing the fleet
size has diminishing returns until the system is at max capacity and the increase in
agents only leads to more congestion, lowering the throughput.

The thesis was conducted in collaboration with Consat AB as an industrial partner.

Consat delivers services in various fields, such as industrial technologies, IT and
automation. The goal of the thesis is to assess the impact that NFRs have on
AGV systems and give Consat further insights on the subject. The study also
aims to contribute to global research in the fields of automation and requirements
engineering.

73
7. Conclusion

7.1 Future work

This study has shown the impact that a certain set of NFRs regarding robustness
and safety has on the system throughput, and how it varies based on differentiating
layouts and fleet size. Future work could continue to study the impact of similar
requirements on more advanced simulation tools. The set of NFRs chosen for the
research is by no means exhaustive, and work could continue on the same line as
the research conducted.

Due to the limitations in the scope of the research, NFRs regarding reliability, avail-
ability, scalability, security and usability was omitted. It could be interesting for
future work to encompass such requirements for example the impact that reliability
and availability have on the system throughput. It would also be interesting to see
the impact that scalability has on for example the computational complexity and
effort for AGV systems. To expand upon the usability, future research could encom-
pass what impact the flexibility and adaptability of the system have on its efficiency.
Research regarding how usability requirements can impact operator training and re-
tention would also be of interest.

For this research, the A∗ algorithm was used as the path-finding algorithm to direct
the agents during the simulation. It would be interesting to see the impact of
different NFRs given other kinds of pathfinding algorithms, utilizing, for example,
D ∗Lite algorithm, the ant colony optimization algorithm, or the genetic algorithm.

74
 Bibliography

[1] R. Yan, L. M. Jackson, and S. J. Dunnett, “Automated guided vehicle mission

reliability modelling using a combined fault tree and petri net approach,” The
International Journal of Advanced Manufacturing Technology, vol. 92, no. 5-8,
pp. 1825–1837, 2017.
[2] R. Yan, S. Dunnett, and L. Jackson, “Novel methodology for optimising
the design, operation and maintenance of a multi-agv system,” Reliability
Engineering System Safety, vol. 178, pp. 130–139, 2018. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S0951832017311559
[3] G. Ullrich, Automated Guided Vehicle Systems A Primer with Practical Appli-
cations, 2015.
[4] M. De Ryck, M. Versteyhe, and F. Debrouwere, “Automated guided vehicle
systems, state-of-the-art control algorithms and techniques,” Journal of Man-
ufacturing Systems, vol. 54, pp. 152–173, 2020.
[5] G. V. Research, “Automated guided vehicle market size, share trends
analysis report by vehicle type, by navigation technology, by application, by
end use industry, by component, by battery type, and segment forecasts,
2022 - 2030,” 2021. [Online]. Available: https://www.grandviewresearch.com/
industry-analysis/automated-guided-vehicle-agv-market
[6] N. R. Mohamad, M. Fauadi, and A. Noor, “Optimization of material trans-
portation assignment for automated guided vehicle (agv) system,” International
Journal of Engineering & Technology, vol. 7, no. 3.20, pp. 334–338, 2018.
[7] D. Sinriech and J. Tanchoco, “An economic model for determining agv fleet
size,” International Journal of Production Research, vol. 30, no. 6, pp. 1255–
1268, 1992.
[8] M. A. Javed, F. U. Muram, S. Punnekkat, and H. Hansson, “Safe and secure
platooning of automated guided vehicles in industry 4.0,” Journal of systems
architecture, vol. 121, p. 102309, 2021.
[9] M. Qi, X. Li, X. Yan, and C. Zhang, “On the evaluation of agvs-based warehouse
operation performance,” Simulation Modelling Practice and Theory, vol. 87, pp.
379–394, 2018.
[10] C. G. Co and J. M. A. Tanchoco, “A review of research on agvs vehicle manage-
ment,” Engineering Costs and Production Economics, vol. 21, no. 1, pp. 35–42,
1991.
[11] M. De Ryck, D. Pissoort, T. Holvoet, and E. Demeester, “Decentral task allo-
cation for industrial agv-systems with resource constraints,” Journal of Manu-
facturing Systems, vol. 59, pp. 310–319, 2021.

75
Bibliography

[12] L. Qiu, W.-J. Hsu, S.-Y. Huang, and H. Wang, “Scheduling and routing al-
gorithms for agvs: A survey,” International Journal of Production Research,
vol. 40, no. 3, pp. 745–760, 2002.
[13] C. Wang and J. Mao, “Summary of agv path planning,” in 2019 3rd Interna-
tional Conference on Electronic Information Technology and Computer Engi-
neering (EITCE). IEEE, 2019, pp. 332–335.
[14] A. Andreychuk, K. Yakovlev, P. Surynek, D. Atzmon, and R. Stern,
“Multi-agent pathfinding with continuous time,” Artificial Intelligence, vol.
305, p. 103662, 2022. [Online]. Available: https://www.sciencedirect.com/
science/article/pii/S0004370222000029
[15] M. A. Javaid, “Understanding djikstra’s algorithm,” 2013.
[16] E. W. Dijkstra, “A note on two problems in connexion with graphs,” in Edsger
Wybe Dijkstra: His Life, Work, and Legacy, 2022, pp. 287–290.
[17] A. Botea, B. Bouzy, M. Buro, C. Bauckhage, and D. Nau, “Pathfinding
in Games,” in Artificial and Computational Intelligence in Games, ser.
Dagstuhl Follow-Ups, S. M. Lucas, M. Mateas, M. Preuss, P. Spronck,
and J. Togelius, Eds. Dagstuhl, Germany: Schloss Dagstuhl–Leibniz-
Zentrum fuer Informatik, 2013, vol. 6, pp. 21–31. [Online]. Available:
http://drops.dagstuhl.de/opus/volltexte/2013/4333
[18] B. Li, C. Dong, Q. Chen, Y. Mu, Z. Fan, Q. Wang, and X. Chen, “Path planning
of mobile robots based on an improved a* algorithm,” in Proceedings of the 2020
4th High Performance Computing and Cluster Technologies Conference & 2020
3rd International Conference on Big Data and Artificial Intelligence, 2020, pp.
49–53.
[19] J. Yao, C. Lin, X. Xie, A. J. Wang, and C.-C. Hung, “Path planning for virtual
human motion using improved a* star algorithm,” in 2010 Seventh international
conference on information technology: new generations. IEEE, 2010, pp. 1154–
1158.
[20] F. Ducho, A. Babinec, M. Kajan, P. Beo, M. Florek, T. Fico,
and L. Juriica, “Path planning with modified a star algorithm for
a mobile robot,” Procedia Engineering, vol. 96, pp. 59–69, 2014,
modelling of Mechanical and Mechatronic Systems. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S187770581403149X
[21] S. Lauesen, Software requirements: styles and techniques. Pearson Education,
2002.
[22] K. Wiegers and J. Beatty, Software requirements. Pearson Education, 2013.
[23] E. A. Oyekanlu, A. C. Smith, W. P. Thomas, G. Mulroy, D. Hitesh, M. Ramsey,
D. J. Kuhn, J. D. Mcghinnis, S. C. Buonavita, N. A. Looper et al., “A review
of recent advances in automated guided vehicle technologies: Integration chal-
lenges and research areas for 5g-based smart manufacturing applications,” IEEE
access, vol. 8, pp. 202 312–202 353, 2020.
[24] I. C. Society, IEE Standard for Information Technology - Local and metropolitan
area networks - Specfic requirements - Part 11: Wireless LAN Medium Access
Control (MAC) and Physical Layer (PHY) Specfications, Amendment 6: Wire-
less Access in Vehicular Environments, IEEE Std 802.11p-2010 ed. New York
City, United States: IEEE Computer Society, 2010.

76
 Bibliography

[25] S. Eichler, “Performance evaluation of the ieee 802.11 p wave communication

standard,” in 2007 IEEE 66th Vehicular Technology Conference. IEEE, 2007,
pp. 2199–2203.
[26] L. Qiu, W.-J. Hsu, S.-Y. Huang, and H. Wang, “Scheduling and routing al-
gorithms for agvs: a survey,” International Journal of Production Research,
vol. 40, no. 3, pp. 745–760, 2002.
[27] V. Digani, M. A. Hsieh, L. Sabattini, and C. Secchi, “Coordination of multiple
agvs: a quadratic optimization method,” Autonomous Robots, vol. 43, pp. 539–
555, 2019.
[28] M. Nosrati, R. Karimi, and H. A. Hasanvand, “Investigation of the*(star) search
algorithms: Characteristics, methods and approaches,” World Applied Program-
ming, vol. 2, no. 4, pp. 251–256, 2012.
[29] A. Veeraswamy and B. Amavasai, “Optimal path planning using an improved
a* algorithm for homeland security applications,” in Proceedings of the 24th
IASTED international Conference on Artificial intelligence and Applications,
2006, pp. 50–55.
[30] J. J. Andrews and M. Bob, Reliability and Risk Assessment. Professional
Engineering Publishing, 1993.
[31] Q. Sun, H. Liu, Q. Yang, and W. Yan, “On the design for agvs: Modeling,
path planning and localization,” in 2011 IEEE International Conference on
Mechatronics and Automation, 2011, pp. 1515–1520.
[32] I. Draganjac, D. Mikli, Z. Kovai, G. Vasiljevi, and S. Bogdan, “Decentralized
control of multi-agv systems in autonomous warehousing applications,” IEEE
Transactions on Automation Science and Engineering, vol. 13, no. 4, pp. 1433–
1447, 2016.
[33] M.-S. Yeh and W.-C. Yeh, “Deadlock prediction and avoidance for zone-control
agvs,” International Journal of Production Research, vol. 36, no. 10, pp. 2879–
2889, 1998.
[34] R. Lochana Moorthy, W. Hock-Guan, N. Wing-Cheong, and T. Chung-Piaw,
“Cyclic deadlock prediction and avoidance for zone-controlled agv system,”
International Journal of Production Economics, vol. 83, no. 3, pp. 309–324,
2003. [Online]. Available: https://www.sciencedirect.com/science/article/pii/
S0925527302003705
[35] E. Union, “Standards in europe,” https://europa.eu/youreurope/business/
product-requirements/standards/standards-in-europe/index_en.htm, 2023.
[36] E. A. for Safety and H. at Work, “Directive 2006/42/ec - new ma-
chinery directive,” https://osha.europa.eu/en/legislation/directives/
directive-2006-42-ec-of-the-european-parliament-and-of-the-council, 2021.
[37] E. C. for Standardization, Safety of industrial trucks - Driverless trucks and
their systems, EN 1525:1997 ed. Brussels, Belgium: European Committee for
Standardization, 1997.
[38] I. O. for Standardization, Industrial trucks - Safety requirements and veri-
fication - Part 4: Driverless industrial trucks and their systems, ISO 3681-
4:2020 ed. Vernier, Geneva, Switzerland: International Organization for Stan-
dardization, 2020.

77
Bibliography

[39] ISO, Safety of Machinery - General principles for design - Risk assessment and
risk reduction, ISO 12100:2010 ed. Vernier, Geneva, Switzerland: International
Organization for Standardization, 2010.
[40] T. M. Handling, “Automation of repetitive pallet handling,” 2023.
[41] J. Yu, “Intractability of optimal multirobot path planning on planar graphs,”
IEEE Robotics and Automation Letters, vol. 1, no. 1, pp. 33–40, 2016.
[42] R. J. Mantel and H. R. Landeweerd, “Design and operational control of
an agv system,” International Journal of Production Economics, vol. 41,
no. 1, pp. 257–266, 1995, proceedings of the 12th International Conference
on Production Research. [Online]. Available: https://www.sciencedirect.com/
science/article/pii/0925527395000186
[43] I. Draganjac, T. Petrovi, D. Mikli, Z. Kovai, and J. Oruli, “Highly-
scalable traffic management of autonomous industrial transportation systems,”
Robotics and Computer-Integrated Manufacturing, vol. 63, p. 101915,
2020. [Online]. Available: https://www.sciencedirect.com/science/article/pii/
S0736584518303843
[44] K. Fransen, J. Van Eekelen, A. Pogromsky, M. A. Boon, and I. J. Adan, “A
dynamic path planning approach for dense, large, grid-based automated guided
vehicle systems,” Computers & Operations Research, vol. 123, p. 105046, 2020.
[45] I. O. for Standardization, Information technology Security techniques Infor-
mation security, management systems Requirements, ISO/IEC 27001:2013 ed.
Vernier, Geneva, Switzerland: International Organization for Standardization,
2013.
[46] M. G. Morris and A. P. Dillon, “The importance of usability in the establish-
ment of organizational software standards for end user computing,” Interna-
tional Journal of Human-Computer Studies, vol. 45, no. 2, pp. 243–258, 1996.
[47] I. O. for Standardization, Ergonomics of human-system interaction Part 11:
Usability: Definitions and concepts, ISO 9241-11:2018 ed. Vernier, Geneva,
Switzerland: International Organization for Standardization, 2018.
[48] K. F. E. Tsang, Y. Ni, C. F. R. Wong, and L. Shi, “A novel warehouse
multi-robot automation system with semi-complete and computationally ef-
ficient path planning and adaptive genetic task allocation algorithms,” in 2018
15th International Conference on Control, Automation, Robotics and Vision
(ICARCV). IEEE, 2018, pp. 1671–1676.
[49] Y. D. Setiawan, P. S. Pratama, S. K. Jeong, V. H. Duy, and S. B. Kim, “Exper-
imental comparison of a* and d* lite path planning algorithms for differential
drive automated guided vehicle,” in AETA 2013: Recent Advances in Electrical
Engineering and Related Sciences. Springer, 2014, pp. 555–564.
[50] A. Sarker, H. Shen, M. Rahman, M. Chowdhury, K. Dey, F. Li, Y. Wang, and
H. S. Narman, “A review of sensing and communication, human factors, and
controller aspects for information-aware connected and automated vehicles,”
IEEE transactions on intelligent transportation systems, vol. 21, no. 1, pp. 7–
29, 2019.
[51] H.-P. Schöner, “Challenges and approaches for testing of highly automated
vehicles,” in Energy Consumption and Autonomous Driving: Proceedings of the

78
 Bibliography

3rd CESA Automotive Electronics Congress, Paris, 2014. Springer, 2016, pp.
101–109.
[52] C. Huang, F. Naghdy, H. Du, and H. Huang, “Review on human-machine shared
control system of automated vehicles,” in 2019 3rd International Symposium
on Autonomous Systems (ISAS). IEEE, 2019, pp. 47–51.
[53] S. Orf, N. Lambing, S. Ochs, M. R. Zofka, and J. M. Zöllner, “Modeling lo-
calization uncertainty for enhanced robustness of automated vehicles,” in 2022
IEEE 18th International Conference on Intelligent Computer Communication
and Processing (ICCP). IEEE, 2022, pp. 175–182.
[54] M. Martínez-Díaz and F. Soriguera, “Autonomous vehicles: theoretical and
practical challenges,” Transportation Research Procedia, vol. 33, pp. 275–282,
2018.
[55] S. E. Li, Y. Zheng, K. Li, Y. Wu, J. K. Hedrick, F. Gao, and H. Zhang, “Dy-
namical modeling and distributed control of connected and automated vehicles:
Challenges and opportunities,” IEEE Intelligent Transportation Systems Mag-
azine, vol. 9, no. 3, pp. 46–58, 2017.
[56] D. Gruyer, O. Orfila, S. Glaser, A. Hedhli, N. Hautière, and A. Rakotonirainy,
“Are connected and automated vehicles the silver bullet for future transporta-
tion challenges? benefits and weaknesses on safety, consumption, and traffic
congestion,” Frontiers in sustainable cities, vol. 2, p. 607054, 2021.
[57] J. Marvel and R. Bostelman, “Towards mobile manipulator safety standards,”
in 2013 IEEE International Symposium on Robotic and Sensors Environments
(ROSE). IEEE, 2013, pp. 31–36.
[58] R. Tiusanen, T. Malm, and A. Ronkainen, “An overview of current safety re-
quirements for autonomous machines–review of standards,” Open Engineering,
vol. 10, no. 1, pp. 665–673, 2020.
[59] M. A. Javed, F. U. Muram, H. Hansson, S. Punnekkat, and H. Thane, “Towards
dynamic safety assurance for industry 4.0,” Journal of Systems Architecture,
vol. 114, p. 101914, 2021.
[60] T. Menzel, G. Bagschik, and M. Maurer, “Scenarios for development, test and
validation of automated vehicles,” in 2018 IEEE Intelligent Vehicles Symposium
(IV). IEEE, 2018, pp. 1821–1827.
[61] S. Riedmaier, T. Ponn, D. Ludwig, B. Schick, and F. Diermeyer, “Survey on
scenario-based safety assessment of automated vehicles,” IEEE access, vol. 8,
pp. 87 456–87 477, 2020.
[62] T. Stolte, G. Bagschik, and M. Maurer, “Safety goals and functional safety
requirements for actuation systems of automated vehicles,” in 2016 IEEE 19th
International Conference on Intelligent Transportation Systems (ITSC). IEEE,
2016, pp. 2191–2198.
[63] L. Duboc, E. Letier, and D. S. Rosenblum, “Systematic elaboration of scal-
ability requirements through goal-obstacle analysis,” IEEE Transactions on
Software Engineering, vol. 39, no. 1, pp. 119–140, 2012.
[64] M. Chowdhury, M. Islam, and Z. Khan, “Security of connected and automated
vehicles,” arXiv preprint arXiv:2012.13464, 2020.

79
Bibliography

[65] M. A. Javed, F. U. Muram, S. Punnekkat, and H. Hansson, “Safe and secure

platooning of automated guided vehicles in industry 4.0,” Journal of systems
architecture, vol. 121, p. 102309, 2021.
[66] N. Schömig, K. Wiedemann, S. Hergeth, Y. Forster, J. Muttart, A. Eriksson,
D. Mitropoulos-Rundus, K. Grove, J. Krems, A. Keinath et al., “Checklist for
expert evaluation of hmis of automated vehiclesdiscussions on its value and
adaptions of the method within an expert workshop,” Information, vol. 11,
no. 4, p. 233, 2020.
[67] A. Pauzie and O. Orfila, “Methodologies to assess usability and safety of adas
and automated vehicle,” IFAC-PapersOnLine, vol. 49, no. 32, pp. 72–77, 2016.
[68] H. Younessi and S. Lauesen, “Six styles for usability requirements.” Proceed-
ings of REFSQ98, Presses Universitaires de Namur, 1998.
[69] N. Bevan, “Usability is quality of use,” in Advances in Human Factors/Er-
gonomics. Elsevier, 1995, vol. 20, pp. 349–354.
[70] J. W. Creswell, “Mixed-method research: Introduction and application,” in
Handbook of educational policy. Elsevier, 1999, pp. 455–472.
[71] F. Shull and R. L. Feldmann, “Building theories from multiple evidence
sources,” Guide to Advanced Empirical Software Engineering, pp. 337–364,
2008.
[72] P. Runeson and M. Höst, “Guidelines for conducting and reporting case study
research in software engineering,” Empirical software engineering, vol. 14, pp.
131–164, 2009.
[73] C. Seaman, “Qualitative methods in empirical studies of software engineering,”
IEEE Transactions on Software Engineering, vol. 25, no. 4, pp. 557–572, 1999.
[74] H. Snyder, “Literature review as a research methodology: An overview and
guidelines,” Journal of business research, vol. 104, pp. 333–339, 2019.
[75] C. Robson and K. McCartan, Real World Research, 4th ed. John Wiley &
Sons Ltd, 2016.
[76] R. Huelin, I. Iheanacho, K. Payne, and K. Sandman, “Whats in a name? sys-
tematic and non-systematic literature reviews, and why the distinction mat-
ters,” The evidence, pp. 34–37, 2015.
[77] R. Popenoe, A. Langius-Eklöf, E. Stenwall, and A. Jervaeus, “A practical guide
to data analysis in general literature reviews,” Nordic Journal of Nursing Re-
search, vol. 41, no. 4, pp. 175–186, 2021.
[78] N. Zhao, X. Zhao, N. Xu, and L. Zhang, “Resilient event-triggered control
of connected automated vehicles under cyber attacks,” IEEE/CAA Journal of
Automatica Sinica, 2023.
[79] L. Xiao, F. Gao, and J. Wang, “On scalability of platoon of automated vehi-
cles for leader-predecessor information framework,” in 2009 IEEE Intelligent
Vehicles Symposium. IEEE, 2009, pp. 1103–1108.
[80] M. Host, A. Rainer, P. Runeson, and B. Regnell, Case study research in software
engineering: Guidelines and examples. John Wiley & Sons, 2012.
[81] A. Galletta, Mastering the semi-structured interview and beyond: From research
design to analysis and publication. NYU press, 2013, vol. 18.
[82] R. K. Yin, Case study research: Design and methods. sage, 2009, vol. 5.

80
 Bibliography

[83] A. N. Ghazi, K. Petersen, S. S. V. R. Reddy, and H. Nekkanti, “Survey research

in software engineering: Problems and mitigation strategies,” IEEE Access,
vol. 7, pp. 24 703–24 718, 2018.
[84] I. Etikan, S. A. Musa, R. S. Alkassim et al., “Comparison of convenience sam-
pling and purposive sampling,” American journal of theoretical and applied
statistics, vol. 5, no. 1, pp. 1–4, 2016.

81
Bibliography

82
 A
Interview Guide

Interview questions
Individual experience
• Can you give us a short introduction regarding your experience in the context of
AGV systems.
Safety
• What do you believe impacts the safety of an AGV system?
• Which if any safety regulations does your AGV solutions comply to?
o How does these regulations impact the requirements of said solutions?
Robustness
• What do you believe impacts the robustness of an AGV system?
• When is an AGV system considered robust?
Scalability
• What impacts the scalability of an AGV system?
• When is a system deemed scalable?
Efficiency
• What impacts the computational efficiency of a centralized AGV system (in the
context of path-finding)?
• What sets the upper limit for computational time?
Reliability/Availability - Stipulates how long the system can run before experiencing a failure.
• What do you believe impacts the reliability of an AGV system?
• What would you see as an appropiate threshold for the reliability of a centralized
AGV system?
Ending:
• Would you consider other non-functional requirements, such as Usability or
Security to be more important than the types that we have discussed?
• We have discussed a number different non-functional requirements. Which ones
do you identify as the most essential?
• Is there anything you believe we have missed that you would like to add?