Autonomous transportation for a Swedish production facility
Mapping the technological and regulatory hurdles
Karl Fredriksson
Industrial Design Engineering, master's level 2021
Luleå University of Technology
Department of Social Sciences, Technology and Arts
MSc in INDUSTRIAL DESIGN ENGINEERING Department of Social Sciences, Technology and Arts
Autonomous transportation for a Swedish production facility
-‐ Mapping the technological and regulatory hurdles
Karl Fredriksson 2021
SUPERVISOR: Magnus Stenberg REVIEWER: Maili Schönning EXAMINER: Lena Abrahamsson
CIVILINGENJÖR I TEKNISK DESIGN
Master of Science Thesis in Industrial Design Engineering
Autonomous transportation for a Swedish production facility Mapping the technological and regulatory hurdles
© Karl Fredriksson
As per request, the identity of the client has been withheld
Published and distributed by Luleå University of Technology SE-‐971 87 Luleå, Sweden Telephone: + 46 (0) 920 49 00 00
Printed in Luleå Sweden by
Luleå University of Technology Reproservice Luleå, 2016
Abstract
The technology of autonomous vehicles has the potential to provide a significant number of safety, efficiency and environmental benefits to those who are able to harness it. As such, it is only natural that the company which is the subject of this project should want to explore this field, since the company prides itself on being at the cutting edge of both environmental sustainability and technological advancement. This inquiry was therefore launched in order to amass a sufficient knowledge base to enable management to make informed decisions about the possible future implementation of autonomous trucks, specifically to handle the logistics flow between their production facility in Skellefteå, Sweden and the nearby harbour. The first step to achieving this objective consisted of an exploration of the state of autonomous vehicle technology as well as the regulatory framework in Sweden for operating such systems on public roadways. Information was gathered from a vast array of sources, including academic literature, official reports from various authorities, journalistic publications as well as interviews with individuals with competence or experience within this field. While the regulatory situation in Sweden at the moment offers no legal way to operate autonomous vehicles on public roads, it is possible to be granted permission to perform trials of this technology under certain conditions. An investigation was undertaken to determine whether this might be a viable option for the company’s case. As such, hazard analysis was performed on the proposed route in Skellefteå. The method for this was based off of methodology gained from sources who had previously executed safety cases for trials of autonomous technology. A list of potential hazards relevant to the operation of autonomous vehicles was composed, together with variables with which to measure their severity. The relevance and appropriate scope of these hazards and variables was then verified by discussing it with sources with competence in this field. The route was then travelled in order to observe the prevalence of the aforementioned variables. The information was completed and verified through various reports gathered from the Swedish Transport Administration and the Swedish Meteorological and Hydrological Institute. The result of the inquiry was that the autonomous technology on the market today is not sufficiently advanced to handle the specified application with an adequate level of safety. The route is also of limited use in establishing trials for testing of autonomous vehicles. While there are uses for autonomous transportation technology, great breakthroughs are needed before the technology reaches the level needed to handle such complex challenges as would be encountered on the proposed application.
KEYWORDS: AUTONOMOUS VEHICLES, AUTONOMOUS TRUCKS,
ARTIFICIAL INTELLIGENCE, INQUIRY, TRAFFIC SAFETY
Sammanfattning
Självkörande fordon är en teknologi som visar potential för betydande fördelar inom säkerhet, effektivitet och miljömässigt för dem som kan tygla den. Det framstår därför som naturligt att uppdragsgivaren till detta projekt skulle vara intresserad av denna teknik, då företaget är känt för att vara vid både miljö- och teknikfrågornas framkant. Därför lanserades detta utredningsarbete för att sammanställa tillräcklig kunskap för att kunna ta informerade beslut om en potentiell implementering av ett autonomt transportsystem från deras fabrik i Skellefteå till Skelleftehamn. Denna utredning började med att kartlägga hur autonoma fordonstekniken ser ut idag, samt de regulatoriska möjligheterna att driva autonoma system på allmän väg I Sverige. Informationen samlades från en mängd olika källor, inklusive akademisk litteratur, rapporter från officiella källor, journalistiska källor samt från intervjuer med personer som besitter kompentens och erfarenhet av ämnet. Emedan den regulatoriska situationen i Sverige för stunden inte medger något lagligt sätt att operera självkörande fordon på allmän väg så finns det möjlighet att få tillstånd att utföra försöksverksamhet med sådana fordon så länge vissa villkor uppfylls. En utredning genomfördes för att fastslå om sådan verksamhet skulle kunna vara relevant i företagets fall. I och med detta så utfördes en riskanalys på den föreslagna rutten i Skellefteå. Metoden för dess utförande baserades på metodologi som hämtades från källor som tidigare hade utfört säkerhetsbevisningar för försöksverksamhet på autonoma fordon. En lista av möjliga risker framtogs, tillsammans med mätpunkter vilka skulle kunna användas för att fastslå deras betydelse för autonom fordonteknologi. Dessa riskers relevans och lämpligheten av dess omfattning diskuterades därefter med källor med kompetens inom området. Sedan besöktes rutten för att observationer om mätpunkternas förekomst kunde utföras. Informationen kompletterades och verifierades därefter med information från ett antal rapporter från Trafikverket och Sveriges Meteorologiska och Hydrologiska Institut. Det man kommit fram till är att det idag inte finns något autonomt fordonssystem som är tillräckligt avancerat att klara rutten mellan fabriken och hamnen med god nog säkerhet. Rutten är dessutom av begränsat värde vad det gäller att testa sådana system. Även om det finns autonoma system i operation i dagsläget så ligger dock tekniken långt under den nivå som skulle behövas för att ta sig an de utmaningar som skulle uppstå I det föreslagna användningsområdet.
NYCKELORD: SJÄLVKÖRANDE FORDON, SJÄLVKÖRANDE LASTBILAR,
ARTIFICIELL INTELLIGENS, UTREDNINGSARBETE, TRAFIKSÄKERHET
Content
1 Introduction 1
1.1 Background 1
1.2 Stakeholders 1
1.3 Objective and aims 2
1.4 Project scope 2
1.5 Thesis outline 3
2 Context 5
2.1 Benfits of autonomous transportation 5 2.2 The client: Mission and philosophy 7
3 Method 9
3.1 Process 9
3.2 Project planning 10
3.3 context immersion 10
3.4 Literature review 11
3.5 Hazard analysis 12
3.6 Interviews 14
4 Current state of autonomous vehicles 2 4.1 Classification of autonomous vehicles 2 4.2 Decision making hierarchy of self-‐driving cars 3 4.3 Artificial Intelligence and machine learning 4 4.4 Sensors used in autonomous vehicles 6
4.5 Technology readiness levels 7
4.6 Testing 8
4.7 Junction classification 9
4.8 Autonomous trucking systems 10
5 Regulations: current state and looking to the future 15 5.1 Permission to perform AV tests 15
5.2 Application for permission 16
5.3 state regulation: looking to the future 17 5.4 Licensing as a model for certification 19 6 Hazard analysis of the proposed route 20 6.1 The proposed route: scope and limitations 20 6.2 Lane changes and obstacle avoidance 22
6.3 Junctions 24
6.4 Non-‐protected road users 26
6.5 Weather and sun 29
6.6 Communication 33
6.7 Dangerous goods 34
7 Discussion 35
7.1 Av trials at the production facility 35 7.2 The future of autonomous transportation 39
8 Conclusion and recommendations 42
9 References 43
List of appendices
Interview template Swedish Transport Agency 1 page Interview Dr. Missy Cummings 1 page Interview template Einride 1 page Interview template Scania 1 page Autonomous vehicle hazard assessment 3 pages Rules and infrastructure of the junctions of road 372 1page
1
1 Introduction
The production facility of the client is set to become the first large-scale lithium- ion battery production facility in Europe. In order to transport incoming materials and outgoing products as safely, efficiently and economically as possible, the client has requested an inquiry into the possibilities afforded through the use of autonomous trucks. This is a technology which is poised to become central to all logistics operations in the coming decades, with significant benefits to the safety, efficiency, environmental impact and costs. However, as of today, the technology is still at an early stage of its developmental lifespan and is sure to be subject to momentous advancements and innovation in the coming years.
The central tenets by which the client base all their operations are boldness, passion and excellence. By boldly exploring uncharted technologies, passionately embracing the changes they present and, with excellence, efficiently extracting the benefits thereof, this investigation is a reflection of the many factors which make the client one of the most exciting and cutting edge innovators in the world.
1.1 BACKGROUND
In 2016, the client was founded with the goal to create sustainable lithium ion battery production within Europe in order to facilitate the transition towards renewable energy. This mission started with an research and development centre in Västerås, and will continue with the creation of the production facility in Skellefteå. The initial construction on the production facility began in June 2018.
The facility is projected to have a capacity of 16GWh/year.
From the production facility to the industrial harbour of Skellefteå there is a distance of approximately 10 km by existing roads. Once production commences at the facility, an estimated 12000 shipping containers will have to make a round-trip of these destinations every year, and thereafter increasing to a projected 40000. As a company at the forefront of both technology and environmental sustainability, the client is eager to explore the possibilities of using autonomous trucks to automate these transports. The information gathered in this report will be used to help inform their future implementation.
What follows is a master thesis project for the Luleå University of Technology (course code A7009A).
1.2 STAKEHOLDERS Primary Stakeholders
The client Management
- need to factor in the results of the project in their strategy Logistics team
2
- the results of project will have a direct effect on their decision making IT team
- will eventually be responsible to amass sufficient competence to run any of the recommended systems
Operators
- a final human point of contact before the autonomous system takes over: this transition is a point of interest
Maintenance
- will be responsible for maintaining the vehicles Drivers
- will likely initially be necessary to supervise the systems operation Skellefteå harbour
Management
- apart from having some responsibility for the systems success: may gain experience from the systems implementation to use in other parts of their operation
Operators
- a final human point of contact before the system runs autonomously Secondary Stakeholders
Other road users
- have an interest in the system operating safely and efficiently Transportstyrelsen
- will need to make an informed decision on the extent the system is legally allowed to operate
1.3 OBJECTIVE AND AIMS
The project objective is to amass a sufficient knowledge base to enable the client to make informed decisions about the future implementation of autonomous trucks at their production facility.
The aims are:
● To surmise existing research of autonomous trucking technology so as to be easily understood by management.
● To perform a thorough investigation in the current regulatory status for autonomous trucking, as well as a prognosis for how it may advance in the future, as well as recommendations for steps the client may take in order to bring along the necessary changes sooner.
● To present report on the feasibility of utilizing autonomous trucking to handle the logistics flow of the production facility, and the possible timescale of when such technology might be implemented.
1.4 PROJECT SCOPE
Due to time and resource constraints, this project will not delve into any primary research and development of any particular autonomous trucking system: only
3
secondary research and development of more general concepts for the systems implementation will be undertaken.
The type of transportation used will be battery driven heavy lorries. The client has already entered into negotiations with a supplier. As such, only the autonomous system will be investigated; specific trucks will not be considered, neither will alternate forms of transport.
The logistics flow which will be investigated are the products and materials being transported from the facility in Skellefteå to Skellefteå harbour, and vice versa.
Other flows will not be considered.
1.5 THESIS OUTLINE
What follows is a brief description of the contents of each chapter in order to provide a simple way to find the desired information.
1. Introduction
A look at the circumstances which gave rise to the project, as well as establishing the objectives, aims and stakeholders.
2. Context
A review of the reasons that autonomous transportation might be beneficial to the logistics of the production facility. This subject is explored from the perspective of autonomous technology as well as the perspective of the client as a company.
3. Method
An explanation of what was done, why it was done and how the information was verified.
4. Current state of autonomous vehicles
An exploration of the technology of autonomous vehicles, including an overview of how the technology works, highlighting some of the challenges which must still be faced, as well as an examination of the main actors on the market.
5. Regulations: current state and looking to the future
A summary of the possibilities of operating autonomous vehicles on public roads currently llowed through Swedish regulation, as well as an exploration of how the future might look
6. Hazard analysis of the proposed route
An examination of the most significant hazards encountered on the route between the production facility and the harbour, from the perspective of an autonomous transportation system.
7. Discussion
An analysis of the possibilities and realities of operating autonomous vehicles of the proposed route, followed by an exploration of what the future may hold and in regards to autonomous vehicles.
4 8. Conclusions and recommendations
Presenting the most important findings from the discussion chapter in a clear and concise way.
5
2 Context
The context in which this project finds its relevancy can be examined by considering one central question: why should the client want to explore autonomous transportation technology? The answer to this question may be framed in two distinct perspectives: first, what are the benefits of autonomous transportation, and second, how do these fit in with the client as a company? In this section both these questions will be answered.
2.1 BENFITS OF AUTONOMOUS TRANSPORTATION
Although there have been experiments and attempts at developing autonomous vehicles (AV) for the road since at least the late 1930s, work on implementing the concept specifically on trucking is a newer development (Wetmore, 2003).
Some of the earliest attempts at finding real world applications of the technology took place within the mining industry. One of the very first was the Komatsu’s Autonomous Haulage System, where testing of a fleet of 5 Ultra Class trucks was commenced in 2005, in the Codelco Mine Radomiro Tomic in Chile (Moore, 2018).This application was well chosen as the mining industry is one of those with the most widespread use of AV to date. There are several characteristics of AV which make them suitable for this work, as well as other potential advantages which the technology could produce, many of which are beneficial to the potential applications envisioned at the clients facilities.
One of the benefits which are most attractive when talking about logistical transportation applications of AV is the increase in efficiency that the technology may bring. One of the more obvious areas where AV would be more efficient would be the decrease in labour costs. When estimating the generalized transport cost for road freight transports in the EU, the Joint Research Centre arrived at the conclusion that the drivers wages accounted for the largest single component of the GTC: on average 42.1% of the costs (Persyn et al., 2020). If those costs could be mitigated through autonomous technology the impact would understandably be significant; even more so in Sweden, which currently holds some of the highest labour costs in EU and the entire world (Eurostat, 2020).
Another gain in efficiency would be seen in the reliability and scheduling components of logistical operations. By eliminating human errors and needs, such as those of taking breaks to eat, sleep, and go to the bathroom, the logistical system may see great improvements in both time and fuel efficiency (Fagnant &
Kockelman, 2015). Such improvements have already been observed in the cases where AV technology has been implemented in closed off facilities (Van Meldert
& De Boeck, 2016).
More than the benefits to the individual users of the technology, advancements in
6
efficiency bring further systematic benefits to the greater world. One self-evident such benefit is the ecological impact of operating at greater fuel efficiency.
However, there are other, less obvious advantages such a system may bring. Due to the high degree of control possible over an AV motion planning algorithm, simulations suggest that the amount of road wear caused by trucking could be significantly lowered if appropriate algorithms were used (Chen et al., 2019).
Analysis has also revealed that AV technology would likely help prevent congestion and increase the throughput of the roadways on which they operate.
Due to more predictable and controlled stopping and acceleration patterns, autonomous vehicles could increase the stability of the traffic stream as well as helping to alleviate the effects of shockwave congestion (Talebpour &
Mahmassani, 2016).
Another area where AV could make theoretical advancements is road safety.
Advancements in automation, with technologies such as collision warnings, automatic emergency braking and lane departure warnings have already had an impact on road safety, to the extent that insurance companies in Europe already offer premiums on vehicles with such new features (Bartz, 2017).Any autonomous system would combine advanced forms of these technologies with other advantages, such as faster-than-human reaction times and probabilistically calculated behaviour algorithms for consistent and optimal decision making (Paden et al., 2016). In addition to this, AV lack some of the failings human drivers are plagued by. In the US, for example, 94% of accidents are due to driver error. 31%
of accidents involve a driver who at the time was legally intoxicated, 10% of accidents stem from the driver being distracted and 7% are due to drivers falling asleep (Singh, 2018). Naturally, AV systems do not suffer from these deficiencies, and so it is projected that they could have a significant impact on lowering accident rates (Paden et al., 2016).
Almost all of the benefits previously discussed have one thing in common: they scale with the relative proportion of AV on the roadways. This is due in part to the increased predictability of AV in comparison to human drives, but also to the possibilities of Vehicle to Vehicle (V2V) or Vehicle to Infrastructure (V2I) communication, allowing the autonomous systems to make more informed route and motion planning decisions. The benefits to both road wear and congestion, for example, increase substantially with the proportion of AV in circulation (Chen et al., 2019; Talebpour & Mahmassani, 2016). The increased predictability, as well as connectivity, would also make the road safety of the vehicles easier to secure. The reduced congestion would have knock-on effects for fuel efficiency of all travellers (Fagnant & Kockelman, 2015). Furthermore, there are reports and projections of driver shortages in both the European and US trucking industries which self driving trucks would undeniably help alleviate (IRU Report Forecasts Alarming Jump in Driver Shortage in Europe, 2020; Huff, 2020).
7
2.2 THE CLIENT: MISSION AND PHILOSOPHY
As we as a species continue to press on into the first tentative decades of the third millennium, it is becoming abundantly clear that if the future is to be one to look forward to, significant changes need to be made in the way we use the resources of our planet. Lithium ion batteries are a key part of this cleaner future, helping to reduce the amount of fossil fuels being consumed, as well as helping to ease the transition to sustainable energy sources (Sasmal et al., 2016). However, efficient use of this technology is not without its obstacles. First, the materials needed to create lithium ion batteries are obtained through mining, oftentimes with a devastating environmental and humanitarian impact (Katwala, 2018). Second, the manufacturing capacity for the batteries is almost entirely based in eastern Asia.
This results in less sustainable end products due to the inefficiency of transporting them halfway around the world, as well as causing the supply to be unacceptably insecure. The client proposes to solve these issues through a simple central aim: to create local European manufacturing capacity for lithium ion batteries through the cleanest means possible. The company will accomplish these goals through implementing their revolutionary recycling process for batteries in a production facility in Skellefteå, Sweden.1
The production facility is currently being built on a 25 hectare plot of land south east of the city of Skellefteå, Sweden. With a projected capacity of 40 GWh, enough to power 600 000 electric vehicles, the plant will be at the cutting edge of modern manufacturing, with advanced automation technologies, a full recycling plant and utilizing 100% clean energy. The facility has a target of using 50%
recycled materials in new cells by 2030, seen as a step on the road towards the ideal vision of making batteries a closed material cycle. The industrial harbour of Skellefteå is located 11 km from the facility. The most direct route runs on the 372, a major roadway with a maximum speed of 80km/h. Projections show that at full capacity the facility will need to perform tens of thousands of return journeys through this route each year, carrying standard 40ft shipping containers each way. 1 (see fig. 1)
8
Fig, 1: The projected volume for the logistics flow between the production facility and Skellefteå harbour until 20261
On the road to these goals, it is not surprising that the client may wish to look towards autonomous transport as a solution to their logistics flow. As a company at the forefront of technological advancement, it is logical to look towards the technology being touted as the future of the transportation industry. Since a significant portion of their manufacturing process is to be highly automated, there is a certain logic to extending this methodology to their logistics flow in order to maximize the time and cost effectiveness. Further, the environmental benefits promised by AV are certainly in line with their overall mission of creating a cleaner future. In short, autonomous transportation is predicted to be a central part of the same generation of manufacturing and industry which the client is spearheading, and so it is only natural that the client should want to integrate it into their system as early as possible1
1A.BRUNDIN, THE CLIENT, PERSONAL COMMUNICATION,30-10-2020
9
3 Method
As an inquiry, this project in essence consisted of the gathering and structuring of information in a way which will be useful for management to make decisions going forward. This section will describe the different stages of this process.
3.1 PROCESS
The first phase of the project consisted of planning, followed closely by information gathering: first through studying existing literature and research, and then through interviewing persons of interest. When an overview of the available technologies had been acquired, the investigation moved to acquiring and analysing knowledge specifically relevant to the project aims. During such a process it is essential to use the information acquired during the previous stages of the project, as well as to perform further, deeper investigation into areas of study particular to the requirements of the production facility. As such, the project employed the iterative process, with a project cycle which can be seen in fig. 2.
The cycle was iterated throughout the project: the work began with an exploration of autonomous vehicle technology through academic literature. After this the focus was changed to exploring the regulatory situation of autonomous vehicles in Swecen. To accomplish this, a return to the literature research phase was necessary in order to gather information specific to this subject, and so the cycle was recommenced. This cyclical process was repeated in a similar fashion for all of the chapters included in this report.
Fig. 2: The project circle
This cycle was originally based on the structure of Bellgran’s “ramverk”
model for the development of production systems, as seen in table 1. However, the process was repurposed, rearranged and simplified to more closely fit the project, which in its essence is more of an inquiry than a development project. As such, the category of “development” was substituted by “analysis” and “realisation”
by evaluation. “Context” and “performance” were given the much broader remit of “research” in general, and the planning phase was mainly concerned with strategizing what empirical and other data needed to be collected to inform the analysis stage.
Research and literature review
Planning and data collection
Analysis Evaluation
and refocus
10
Table 1 “Ramverk” for the development of production systems (Bellgran & Säfsten, 2005
Context and
performance Planning Development Realisation
Contextual impact Leadership and
direction Preparatory Physical
System performance Structured process Specialised Operational 3.2 PROJECT PLANNING
The project planning was performed on two distinct levels: the project as a whole and on a weekly basis.
For the project as a whole, a Gantt chart was created. (see fig. 3) As a tool, it is regarded by most as an effective method of visualizing the work which must be done for a project to reach its completion. It was useful as a simple means to get an overview of how the project was progressing in relation to external and internal, hard and soft deadlines.
Fig. 3: The Gantt chart for the project
On a weekly basis, the planning took the form of a system of brief reports. These consisted of a list of the work completed since the last session, the deviations from previous planning, and the planned activity due to be done before the next week.
These reports served as important documentation of the progress made, as well as an opportunity to analyse, critique and adjust the direction of the project in light of new information. They were also presented to the internal thesis supervisor before each biweekly review session, and served a dual function as a form of mini agenda for each meeting.
3.3 CONTEXT IMMERSION
This project had two areas where it was important to build up some form of contextual immersion. The first of these was the philosophy and mission of the client. This was an important part of the project as it helped in defining the perspective from which the client motivated their desire to explore the possibilities afforded by autonomous transport. The process started off with an introduction to the company at their offices in Stockholm, which was followed up with onboarding sessions. Contact was also held with the external supervisor through biweekly video conference meetings, where both parties were able to update each other on developments, as well as to comment on the path the project was taking. In order to stay informed and in touch with the general goings on and culture of the company, the weekly logistics department meetings were also attended through video conference software.
11
The next step of context immersion was to gain an insight into the industry and technology of autonomous trucking. Ideally, the preferred way to gain exposure to this field would have been through visiting facilities where this technology was either being tried or had already been implemented. However, this proved impossible due to the circumstances brought on by the COVID-19 pandemic.
Instead, the required knowledge was gained through researching the topic on the internet. The first step was to read through the Wikipedia entries on the subject, starting with the entry titled “Autonomous Vehicles”, then moving on towards
“Autonomous Trucking”, and so forth, progressing to more specialized subjects.
The point of this was to get an overview of the subject, gain familiarity with all the terms and a decent grasp on the history of the technology. While Wikipedia was never used as a direct source for any information presented in this report, it proved immensely useful in order to build up a general understanding of the field of research. Whenever a subject was deemed of specific interest for the project it served as a stepping stone to direct the research, after which verifiable sources could be found to confirm or deny the information, such as academic papers and interviews with authorities on the subject. Once this source of information had been covered, news articles and the websites and press releases of the companies involved in the field served as sources of information. To stay abreast of new developments, as well as staying immersed in the field, Google news was searched on a weekly or biweekly basis throughout the project. Certain key words were used, such as “autonomous vehicles”, “autonomous trucking” and the names of select companies, such as TuSimple and Waymo.
3.4 LITERATURE REVIEW
The literature review was performed with two main purposes. The first was to amass a base of academically verifiable information on autonomous trucking technology, in order to present the client with a summary in line with the project aims. This was done through querying databases such as SCOPUS. The key words used ranged from the more generalised, such as “autonomous vehicles”, on to more specific topics such as “LIDAR”. There was an effort to choose the literature with the most citations, as well as the newest literature. There was also a focus on searching for survey articles, as these often proved useful to gain an overview of a specific subject, as well as a way to discover other useful articles through the references. Often, when such an article could not be found in the databases, it could easily be acquired by contacting the author directly. In most cases, the response was positive and the article was sent over email in a matter of days.
The second purpose was to research specific topics which were deemed relevant to the subsequent portions of the project, including the regulatory survey and especially the hazard analysis. The subjects which required further analysis were in part those directly related to the capabilities of AV technology, such as clarifying exactly under which circumstances the sensors were able to operate. Such cases required the return to the academic databases in a way which may be regarded as a classic example of the iterative process at work. Othertimes, the knowledge needed was on subjects peripherally related, such as the different classifications of dangerous goods or Swedish standards in road design. This was often sequestered from reports acquired from the websites of various government agencies, such as
12
the Swedish Transport Administration and the Swedish Civil Contingencies Agency. Another source of the reports were through the interviews conducted with various individuals. Oftentimes, they could supply reports of particular interest, or otherwise give helpful leads as to where one might look for the necessary information.
3.5 HAZARD ANALYSIS
A central part of the project was the hazard analysis performed on the route between the production facility and Skellefteå harbour by county road 372. The initial purpose of this work was to gather information on the route in order to assess the viability of implementing an AV transportation system or testing such a system on the route. The assessment would also serve as a useful starting point from which future hazard and risk analyses may be launched, as well as a framework upon which to discuss the strengths and weaknesses of the autonomous systems currently on the market. The information may also prove useful when developing those infrastructural elements of the route which the client still have some control over, as well as when making decisions when implementing any non- autonomous transportation system.
The initial step in producing the hazard assessment was to create a plan. This consisted of listing all of the relevant potential hazards which lay along the route, the reasons why they were of particular interest in regards to autonomous vehicles and the data which should be collected in order to assess the hazards. This was then used to plan a trip to visit the actual route in Skellefteå. The purposes of this visit was to collect any data which must be amassed in person on the ground, to get a feeling for the route itself in order to make qualitative statements about its features, as well as to observe the route in an attempt to discover any hazards which may have been missed when producing the initial plan.
The main focus of the trip to Skellefteå was to gather information on the junctions.
This was done by driving along the route from start to finish, and stopping at each junction to collect the relevant data. This method was based off of the reports of safety cases performed in preparation for previous trials of autonomous technology.1 Some of the aspects of the route which were investigated were the traffic rules and laws decided the right of way at each junction as well as what infrastructure and signage existed in order to signal or enforce these. This information was recorded on a spread sheet. Pictures were also taken of each junction, both to refer to later as well as to document any peculiarities and hazards present at specific junctions. The points where the environment broke the line of sight for the vehicles of the 372 was observed and assessed. Unless it was completely clear that these would not be an issue for the visibility on the 372, the coordinates of these would be noted on google maps, the exact locations of them determined through scrutiny of the aerial photographs provided on the software (see fig. 4). This was done in order for further analysis to be performed out of the field.
13
Fig. 4 The route and the coordinates collected during the hazard assessment While the junctions of the route were being investigated, observations were made concerning other aspects of the route. The locations of bus stops, pedestrian crossings, train crossings, bicycle paths and points of concentration of pedestrian traffic were recorded on google maps. Where deemed potentially useful, these features were photographed for future reference. Finally, separate drives were performed to gather information about the lane widths, the amount of lanes, speed limits and the road quality. However, difficulties were encountered when attempting to stop and record these safely. Instead, their approximate location was memorised and verified by other means at a later date.
The data collected on the trip was augmented and completed with data from other sources. One resource which was relied upon heavily was the Swedish Transport Administration. Since the 372 is a county road, its upkeep and infrastructure lies directly under their responsibility. As such, their database Lastkajen contained vast amounts of information in the form of maps, detailing different infrastructure elements and regulations, such as the number of lanes, road width and speed limits.
These were then analysed using mapping software such as ArcGIS, Google Earth and the Swedish Transport Administration’s own map-generating software, Nationella VägDataBasen (NVDB). In addition, the Swedish Transport Administration has recently produced their own hazard assessment report on road 372 (Swedish Traffic Agency & Ramström, 2019).The work is still on-going, but the report was developed in order to take decisions about the infrastructural changes needed to raise the safety of the road. This was deemed necessary due to the numerous accidents on the road in the past decade; the road had at times been
14
known as one of Sweden’s 100 most dangerous (Swedish Traffic Agency &
Ramström, 2019). This report proved extremely useful both in order to validate and verify the information already gathered, as well as to contribute information and statistics on hazards and accidents which could not realistically be gathered through the previously described hazard assessment methods.
Data concerning weather was sourced from the Swedish Meteorological and Hydrological Institute (SMHI). The results from their measuring stations are open to the public, and as such can be accessed freely. The data for some the categories relevant to this report stretched as far back in time as the late 1960s. However, in order to arrive at comparable datasets, only data from 1989 and onwards was used as this was the oldest data available in one of the categories. The measuring stations chosen were those closest to the route which actively recorded the required data: Kusmark D for snow depth and precipitation, and Skellefteå Flygplats for visibility and temperature.
3.6 INTERVIEWS
A large part of the project was accomplished through the information gathered by interviewing people of interest within the AV industry. These interviews were at the most semi-structured, with templates developed with possible lines of questioning, but where the great majority of the information was gathered through improvised follow up questions. Often, these interviews were improvised from start to finish, with a brief description of the client’s case as a starting point to inspire lines of inquiry. The interviews were also a way to verify the information amassed through academic literature and journalistic sources, as well as the reasoning behind the methodology of the project. An example would be the hazard analysis. The basic plan for how to proceed was at first developed by analysing what was known about how AV systems function, and thereafter envisioning those hazards which would likely be most significant for such a system. This list of hazards was then discussed independently with a consultant who had worked with developing risk assesments for AV systems previously1, as well as with a project leader from a company which develops AV systems.2 Both of these sources verified that the reasoning behind the hazard analysis was correct, advised to drop certain variables whose importance had been overestimated, as well as offering other hazards which had been missed in and should be investigated.
2C.WALLBÄCKS,EINRIDE, PERSONAL COMMUNICATION,30-10-2020
1B.ENQVIST,COMBITECH, PERSONAL COMMUNICATION,01-12-2020
2
4 Current state of autonomous vehicles
The first step to making informed decisions in any venture is to amass the required knowledge on the subject. What follows is an exploration of the technology of autonomous vehicles, including an overview of how the technology works, highlighting some of the challenges which must still be faced, as well as an examination of the main actors on the market.
4.1 CLASSIFICATION OF AUTONOMOUS VEHICLES
The degree to which a vehicle is autonomous varies from vehicle to vehicle, and is more than just a function of the sum of all of the autonomizing features included in the machine. The factor which truly lies at the centre of the level of autonomy of any vehicle is the degree to which is can operate unsupervised, as it is through this such a system would create value. SAE International have devised J3016 as a generally accepted industry standard which attempts to categorize the level of autonomy present in an autonomous vehicle on a scale from 0 to 5. Each grade of the scale represents a different level of autonomous functionality and, conversely, driver responsibility (Paden et al., 2016).
SAE J3016 Level 0
Vehicle requires driver input for all functions. The system has only the capability of warning the driver of any dangers its sensors picks up (Paden et al., 2016).
Level 1
Certain basic functions necessary for vehicle operation have been automated.
Examples include adaptive cruise control, anti lock breaking and lane centring. The system can control either acceleration/braking or steering at the same time: not both (Paden et al., 2016).
Level 2
The functions of level 1 have been integrated so as to allow the vehicle to control steering and acceleration/braking at the same time (Paden et al., 2016).
Level 3
The vehicle is able to drive autonomously, but the driver’s full attention is needed as the system can at any point in time require manual input (Paden et al., 2016).
3 Level 4
The system is able to drive completely autonomously as long as certain conditions are met, as well as being able to abort the journey safely once a situation arises which is beyond its limits. In the context of trucking, a driver may still be necessary, but can have his/her attention entirely elsewhere: sleeping or working another job remotely, for instance (Paden et al., 2016).
Level 5
The vehicle is entirely autonomous under all conditions; a driver is no longer necessary (Paden et al., 2016).
.
It is from the perspective of removal of driver engagement where the scale distinguishes the difference in usefulness of the autonomous systems at its different levels. This is also the main source of the usefulness of the scale itself is. For example, the largest breakthrough for autonomous trucking would be reaching a level 4 autonomous vehicle for public roads, as this would open up a new degree of efficiency and productivity (Daimler, 2020).
4.2 DECISION MAKING HIERARCHY OF SELF-‐DRIVING CARS
Autonomous vehicles are in essence self governing decision making machines. In order to function they need real time information upon which to base their actions.
This is acquired through sensors like LIDAR, cameras and odometry, as well as through operator input such as the final destination. Like any other decision making entity they then need a framework of rules upon which to structure and evaluate the input, as well as to determine the actual decisions made. While the details vary from system to system, in general this takes the form of a hierarchy consisting of four components. These are listed below, in the order from highest to lowest level, where the preceding level informs what the decisions the next level will make (Paden et al., 2016).
Route planning
The system attempts to plan the most efficient route through combining information from roadmaps, GPS localisation and traffic reports. The road network is first reconstructed as a directed graph. Each road segment is assigned a cost, which is required to travel along it based on factors such as time-use, distance and elevation. An algorithm is then used to calculate the minimum-cost route to arrive at the final destination (Paden et al., 2016).
Behavioural decision-making
Once the route has been established, it is the responsibility of the system to travel that route whilst maintaining correct and safe behaviour for the situation at hand.
Decisions include engaging in such behaviours as performing stops at stop signs, waiting to observe pedestrian actions and continuing the route one the way ahead is all clear. To determine what behaviour is appropriate, decisions are based upon the position and behaviour of other road users, laws and rules of the road, the layout
4
and functions of the infrastructure as well as other more general factors, such as visibility and road conditions. These are usually analysed through modelling each available decision as a state in a finite state machine, and selecting the decision which brings the optimal outcome (Paden et al., 2016).
Motion planning
The motion planning system is tasked with creating the specific path trajectory which the vehicle will follow. Examples include which line to take in a crossing or how to switch lanes in a highway. This trajectory is determined through creating a virtual model of the world surrounding the vehicle. By using various techniques, the optimal path within the model is identified, whereby inputs may be generated for the vehicle controls in order to recreate this path in the real world. When a path is chosen priority is given to working within the physical limits of the vehicle, avoiding collisions with all obstacles and creating as comfortable a ride as possible for any passengers (Paden et al., 2016).
Vehicle control
The actual operation of the vehicle is performed through feedback controllers. Due to small errors and inaccuracies in the motion planning system and vehicle model, small adjustments and error corrections must be performed continuously throughout operation in order to secure the path and trajectory stabilisation (Paden et al., 2016).
4.3 ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING
While the hierarchy described above presents a procedure for the way decisions have to be made on different timeframes and as a function of different kinds of information, at some level there needs to be a system which determines what decisions are actually going to be taken. Throughout the hierarchy, but especially in the behavioural decision making and the vehicle control levels, this is done through some manner of artificial intelligence (AI) system. In some cases the system has no problem making a decision: the correct decision can be easily determined ahead of time and programmed into the system, or the system receives such good information that a decision is self evident. An example could be what speed to hold on a stretch of empty road with ideal weather and road conditions. In this case the AI could use its GPS localisation system, combined with a connection to some database of the speed limits of all roads in the area, determine what the officially mandated speed limit is. If all other things are equal, the vehicle would then proceed at that speed (Cummings, in press).
In the world domain, however, things are rarely this simple: there is usually some degree of uncertainty which the system has got to factor into the decision making process. The system has to first perceive environmental factors through the information it gathers with its sensors. It then has to act appropriately based on these factors. An example would be if the sensors detected a long thin object on the roadside. The system would need to establish whether the object is a tree, a road sign or a person, and adjust its speed accordingly. The source of uncertainty in this case arises in perceiving and recognizing what object the system is sensing
5 (Cummings, in press).
Although all the systems on the market are likely slightly different, for the most part it can be said that the ability to perceive the world around it is gained through machine learning. Oftentimes a more advanced variant of machine learning called deep learning is used (Cummings, in press). While there are differences, both methods at the most base level use similar methodology to acquire their decision making capabilities. The system is given algorithms with which to reach a desired state. It is then presented with data which to parse and upon which it applies said algorithm. The system is able to change these algorithms based on the results of past runs in order to bring itself closer to the desired state; in essence it teaches itself to become better by trying different decisions in different situations (Grossfeld, 2020).
In the case of AV, the data used to train the machine learning systems to perceive the world around them would principally be images of known objects, such as road signs and traffic lights. When confronted with such an object, the system would likely use edge detection software in order to create an image which is easier to process (Cummings, in press). It would then use an object detection algorithm, such as the viola-jones object detection framework, to identify the features of an object in order to be able to recognize similar objects in the future (Lee, 2020).
Although in theory, with enough computing power and training, such a system would be entirely adequate for the task of controlling an AV. At this point in time though, with the technology currently available, there are serious problems with this methodology. Since the control of an AV is such a safety-critical system, an extremely high standard must be placed on its control system. However, with current existing technologies, the previously discussed perception system when applied to the world domain is brittle to the point of inadequacy. In this context, the term brittleness applies to the inability of an algorithm to “generalize or adapt to conditions outside a narrow set of assumptions”, as Dr. Cummings puts it (in press). As an example, the system may learn to recognize a stop sign. It collects a set of assumptions of characteristics, such as colours and shape, with which it can successfully identify the signs. However, were the sign to be covered with snow, suddenly the systems assumptions do not hold true, and so it may no longer recognize the stop sign. However, even if the system learned that this was also a stop sign, the next one could have a layer of snow and frost in a different configuration to the preceding sign, leading again to the system not being able to identify the object. It would have failed to create a generalized algorithm to identify the object, as the model by which it attempts to do so has been generated through parsing assumptions made by analysing specific stop signs, which end up not holding true due to slight changes in the conditions.
These issues are symptomatic of the “bottom-up” approach by which machine learning systems develop their algorithms. While such a system is working in chartered territory, it works absolutely fine. However, as soon as it faces uncertainty, its algorithms cannot cope. Uncertainty, instead, can be more effectively tackled with “top-down” reasoning, with causal inference to be able to
6
fill in the gaps. A human, by comparison, would be much more likely to inductively identify the sign (Cummings, in press).
4.4 SENSORS USED IN AUTONOMOUS VEHICLES
Although there is variation from system to system, the information an AV requires to operate can generally be divided in two categories: information from external sources, such as GPS positioning, traffic reports and roadmaps, and information from sensors. This information is that upon which the car must model the world around it in order to implement its behavioral decision making and base its motion planning. The information is acquired through a selection of visual sensors, of which the most commonly used are described here, along with a comparison of their most frequently encountered uses in Table 2 (Hecht, 2018).
Lidar
Lidar, which is a portmanteau for “light detection and ranging”, is a relatively new and expensive technology used to detect the distance between the vehicle and surrounding obstacles, as well as in some cases the speed at which these are moving relative to the sensor. It can be used to produce high resolution images. In the most basic form, the sensor works by emitting a laser pulse and recording the time taken for the pulse to be reflected back towards the sensor. The maximum range of the lidars currently on the market is around 200 to 300 metres, although this number varies depending on the reflectivity of the target which the sensor is attempting to detect (Hecht, 2018).
Cameras
Cameras produce a high resolution image of the world surrounding the vehicle.
They cannot in and of themselves provide information about the distance or relative speed of surrounding obstacles, although this can in some cases be accomplished through data analysis, or by using two cameras to produce a binocular image. (Kehtarnavaz et al., 1991) Colour can provide important information, although the data analysis requires a great deal of processing time/power (Hecht, 2018). The sensor has no inherent source of light, and so is reliant on the external conditions, such as time of day and weather. In good conditions, the range can be as far as 1 kilometer (TuSimple, 2020).
Radar
A highly mature technology, which results in relatively cheap sensors, amongst other benefits. The sensor emits a pulse of electromagnetic radiation at a certain frequency, usually in the radio or microwave bands, and measures the time taken for the pulse to return to the sensor. Different frequencies are used at different ranges. In all cases, the sensors can detect both distance and speed. However, the resolution is low, and oftentimes results in a 2D image of the surrounding objects, as seen from a top-down perspective (Hecht, 2018).
Ultrasound
The sensors emit a high frequency sound wave pulse and record the time taken to
7
return to the sensor. These re only useful at extremely close range, making them ideal for use in slow speed, close quarters maneuvering (Hecht, 2018).
Table 2 Comparison of the uses of the different sensors Lidar Cameras
Long range radar
Short/medium
range radar Ultrasound
Adaptive cruise control x
Collision avoidance x x
Emergency braking x x
Pederstian detection x x x
Environment mapping x x x x
Traffic sign recognition x
Lane departure warning x
Cross traffic alert x
Digital side mirror x
Blind spot detection x x
Surround View x x
Rear collision warning x x
Park assistance x x x x
(Hecht, 2018)
4.5 TECHNOLOGY READINESS LEVELS When a new technology is being developed, such as that of autonomous vehicles, it is important to gauge the maturity of the technology in order to find appropriate applications in which to implement it in an effective and safe manner. In the 1970s, NASA proposed the Technology Readiness Levels framework (TRL) in order to accurately assess the capabilities of the new technologies being developed in the space program (Cummings, in press).
The framework, consisting of 9 levels (see fig. 5), may be applied to autonomous vehicle technology to assess the suitability of running such systems on public roads. Dr. Cummings argues that the AV systems currently on the market are at level 6 (in press). This would imply that the technologies are adequate to operate in a “relevant environment”. In the case of AV, this specifically refers to a controlled environment where the levels of uncertainty for both the perception and sensor systems are at a minimum. An example would be the trials performed with autonomous trucking in Southern California, where hub to hub transfers are performed along stretches of relatively empty highway, with negligible pedestrian and cyclist traffic, as well as predictable weather conditions conducive to the
8
systems abilities (Kirsher, 2019). Some of the technologies most in need of scrutiny in the regard are their readiness is the software used for perception of the world, as well LIDAR sensors, which constitute relatively new technologies in relation to others technologies in use, such as radar RADAR (Cummings, in press).
Fig. 5 The Technology Readiness Levels (NASA, 2019) 4.6 TESTING
Autonomous vehicles are a new technology aimed to operate in the public space with serious safety implications and potential economic damage for both users and third parties. Therefore, one of the most important problems to solve before their widespread implementation is to create a functional, secure and proven methodology to test autonomous systems (Cummings & Britton, 2019). While the most basic safety functionalities of the systems have already been developed to the point where level three autonomous systems are able to operate satisfactorily safely on the roads, the standard needed to be attained in order for SAE level 4 technologies to be allowed in public spaces is equivalent or better levels of safety (EBLS) than those currently achieved by human drivers (Cummings, 2018).
9
One might assume that a simplistic approach to achieving EBLS might be to allow the autonomous systems out on the roadways in order to gather enough data to statistically prove this standard is met. However, in order to produce enough data for this conclusion to be statistically significant, an unreasonable amount of testing must be performed. Due in part to the rarity of traffic accidents per total miles travelled by traditional cars, a fleet of 100 cars would have to be driven for over 12 years non-stop in order to statistically prove that their fatality rate is lower.
Naturally, alternative approaches will have to be considered (Kalra & Paddock, 2016).
4.7 JUNCTION CLASSIFICATION
One of the great challenges in operating an AV system, or any vehicle for that matter, is necessarily the handling of junctions; this is one of the main places where the vehicle must cross paths with another vehicle on a regular basis. The Swedish Transport Administration has classified junctions between two public roads into 6 types: A through F. These are ranked depending on the amount of risk-managing features they incorporate. These have been divided into two distinct groups depending on the conditions provided for vehicles traveling on the joining roads:
types A-C are the “minor junctions” and types D-F are the “major junctions”. The main difference between the groups is that the minor junctions do not employ any major features to facilitate travel from the joining roads, whereas the major junctions all employ features which put vehicles on the joining roads on equal footing with those on the main road. The minor junctions may or may not include a shoulder to facilitate right hand turns off of the main road (Swedish Transport Administration & Frid, 2015).
Type A
A simple junction with no features to separate traffic.
Type B
There is a separation of the traffic on the joining road.
Type C
Traffic is separated both on the joining road and the main road.
Space is reserved for vehicles on the main road to turn left without disturbing traffic.
10
Fig. 6 The types of junction seen on Swedish roads (Swedish Transport Administration & Frid, 2015)
4.8 AUTONOMOUS TRUCKING SYSTEMS
On the market today, there are a number of companies currently at work developing autonomous trucking systems. What follows is a description of some of the biggest, as well as those who are of greatest interest for the client.
Scania/TuSimple
Scania, a Swedish company, is an international leading producer of transport solutions, such as busses and trucks for heavy transport applications (Facts and Figures, 2020). In 2019, the company presented two autonomous concepts: the Scania AXL and the Scania NXT. The Scania AXL is an autonomous truck developed for mining applications. The truck is designed to operate with no driver on board, and therefore has no cab (A New Cabless Concept – Revealing Scania
Type D
The roundabout, where priority is always given to those vehicles already in the roundabout.
Type E
Traffic is controlled through stoplights.
Type F
Interchanges, typically seen on larger roads like highways.
