MSc Dissertation: The Future of Mobility Part Three
The Research
Barriers to Adoption
It is often argued that autonomous vehicles are inevitable and will one day saturate the motorways of this planet (Waldrop, 2015) (Cutts, 2014) ushering in a new age in automobile safety (Fleetwood, 2017). Yet to assume success is to invite failure, as is often the case with futurology (Russell-Carroll, 2014). The challenges of implementation combined with the known and unknown negative side-effects of HAVs shall determine the degree and speed of their widespread adoption by the general public. These issues are multifaceted and complex; how they are dealt with in the immediate and near future will determine the extent to which HAVs will be accepted and adopted by the commuters of the future. To better understand and investigate these challenges, this research seeks to group them into three categories of “barriers” which were developed by conducting a literature review on the subject. These barriers are presented as follows.
Legal Framework, Policy, & Regulation
Concern has arisen over the ability of government regulation to keep pace with HAV development (Fleetwood, 2017). Misguided regulation could “delay or curtail” the adoption of HAVs (Thierer & Hagemann, 2014) as government regulation is often reactive (McCormick et al., 2017). Resistance to disruption may arise as the removing of human error from risk calculations will dramatically alter the structuring of insurance policy (Cutts, 2014).
When human drivers are involved in an automobile accident, the degree to which their actions as a driver caused the accident to occur determines their liability for the repercussions. This system of accountability becomes more complicated when a vehicle involved doesn’t have a driver. Clear designations will need to be made for which party is responsible in the event a HAV causes or is involved in an accident, breaks a law, or commits a moving violation (Riender, 2014). Villasenor argues that the “robust” products liability law framework in the US “will be well equipped to address and adapt” to questions of HAV liability (2014).
In a discussion about HAVs and public health, Fleetwood calls for collaboration and discussion between public health advocates and manufacturers, so that HAVs are developed in full transparency within the public awareness (2014). Stilgo argues this will require a certain appreciation for the ways in which autonomous vehicles “will be entangled in their environments” (2017). Society will need to learn much about “technology governance” while simultaneously HAVs will undergo “a test of the powers of machine learning” as they learn about the rules of society (Stilgo, 2017).
Feasibility & Performance of Technology
While R&D efforts around driverless technology are tremendous, a number of technical problems must be solved before HAVs are highly reliable. Technical systems onboard HAVs use a combination of cameras, sensors, GPS (global positioning—used for navigation), and Lidar (light detection and ranging—used for real-time 3D mapping) (Cummings & Ryan, 2013). These systems must enable the HAV to integrate perfectly into the built environments of our cities and highways while also coexisting with human drivers and road users. However, weather conditions such as thick fog or snow tend to prevent sensors from working properly (West, 2016). Precipitation, fog, and dust cause problems for cameras and Lidar (Cummings & Ryan, 2013).
V2V and V2X communications will be achieved via Long-Term Evolution protocols (LTE), placing increased burden on the low-frequency spectrum as facilitated by telecommunications providers (Cutts, 2014). This will also present challenges to ensure high integration of 5G “air interface and spectrum together with LTE and WiFi to provide universal high-rate coverage and a seamless user experience” (Andrews, Buzzi, & Choi, 2014). The available bandwidth of this spectrum must be reliably accessible to HAVs if the benefits of V2V/V2X are to be fully realised. This means the financial costs charged by telecom companies to facilitate and maintain this accessibility must be accounted for.
Many of the sensors and Lidar systems used in SDVs and HAVs are custom or unique, and currently produced at costs which would make them unrealistic for production vehicles (Howard & Dai, 2013). As volume of demand for these expensive components increases, they can be mass produced, lowering these associated costs (Shchetko, 2014). One recent report from Waymo says it has cut production costs of its Lidar units by 90%, down from $75k per unit in 2014 (Naughton & Bergen, 2017).
At L4 & L5 HAVs rely heavily on analysing big data to make decisions and a considerable degree of networked computerisation is required to operate HAVs. This makes them likely targets for hackers and so robust cyber security efforts will be needed (West, 2016). “Cyber-crime tends to follow the trends of market opportunities” and given their expected growth HAVs will be among these targets (IET, 2016). GPS signals can be falsified and jammed, which could be used to misdirect HAVs or prevent their ability to navigate (Cummings & Ryan, 2013). Boundaries for connectivity and levels of security must be embedded in HAV hardware and software (IET, 2016).
Human Factors: Use, Preferences, Control, Trust, & Ethical Expectations with regard to HAVs
As driver assistance systems continue to become more capable in the functions they can reliably accomplish they will also become more pervasive in current production automobiles. This will take the form of incremental additions to standard product packages available in new automobile purchases. It is essential to understand how consumers will respond when faced with increasing complexity in the operation of these systems. Use cases around initiation of control transition (both optional and mandatory) between the driver and the autonomous systems have been introduced to further develop problems and solutions through research design (Happee et al., 2016).
Take the case of L3, which by definition assumes sufficient focus and alertness from the driver. Reimer presents a strong argument that this expectation cannot consistently be met (2013). Focus is a learned skill that comes with regular practice; a good example of this are the stringent requirements placed on airplane pilots. As the complexity of driver assistance programs increase, so must the driver’s motivation to remain focused (Reimer, 2013). As the human driver becomes comfortable with the L3 systems attending to the operational responsibilities of driving their car, they begin to lose focus and neglect their supervisory duty; what begins as an exhilarating experience quickly becomes boring (Lucky, 2014). This is sometimes described as “overtrust” (Ogilvie-Davidson, 2017). Should a situation arise where the driver is suddenly required to resume control, they may not be prepared to handle the transition from autonomous to driver fast enough. One study tested driver response time when resuming control from HAVs and found this process took approximately 15 seconds (Merat et al., 2014). In the 2013 argument, Reimer presented the Yerkes-Dodson Law which illustrates the relationship between human performance and mental alertness.
The human’s ability to supervise an L3 system fails at either end of this spectrum, with serious implications at the zero performance mark (which in driving could result in a crash). For this reason L3 receives criticism ranging from being described as “driver optional” (Cummings & Ryan, 2013) to an outright “myth” (Naughton, 2017). Ford has opted to bypass L3 entirely and work directly towards L4 (Naughton, 2017).
Further down the line, L4 and L5 HAVs will dictate that the human need not be responsible for any control whatsoever. Humans must be willing to fully yield this control and trust the HAV to transport them. Society will have tremendous expectations that this be done safely and people will initially be more critical of HAV safety than they are of other human drivers—indeed, “the idea of a machine killing a human, even accidentally, will likely not resonate with the general public” (Cummings & Ryan, 2013). While an individual may have a positive but brief experience with a HAV enough to inspire their legitimate personal trust (OECD, 2017), they are quick to lose this trust the moment the HAV does something unexpectedly. In the long term, the general public will need to ultimately accept a reality where HAVs outnumber vehicles driven by humans (West, 2016).
User preferences will also impact L4 and L5 designs (Howard & Dai, 2013). Human passengers may want to specify a driving “style” for their HAV, similar to the range of aggressive to defensive styles humans use themselves (Riener, 2014).
Ethical concerns are raised over how HAVs will react when evasive manoeuvring is necessary (Fleetwood, 2017). For example, will the vehicle place its passengers in harm’s way to swerve and avoid pedestrians? The complexity of these decisions, and the speed at which they need to be made, must exceed that of human capability if HAVs are to raise the bar on automotive safety (Lin, 2016). As Fleetwood points out, “The autonomous vehicle, like the human driver, must balance safety, mobility, and legality when those objectives conflict” (2017).
The Research Question
Based on these barriers as identified through the literature review, this research seeks to further investigate the human factors of preferences, use, and trust relative to HAVs by asking the question: “Will people trust highly autonomous vehicles, and how will they prefer to use them?”
Methodology & Research Design
To answer the research question a survey was created titled The Future of Mobility: Trust & Preferences (see Appendix B). Surveys are widely used in social research (Kelley et al., 2003). This survey aimed to collect information about patterns of use and tendencies of trust in current mobility solutions and how these may change when HAVs become an available option in the future. In examining these issues via survey, descriptive research was conducted to facilitate a situational examination of respondents’ behaviours and experiences (Kelley et al., 2003). This offers breadth of coverage and produces a large amount of empirical data in a short time (Kelley et al., 2003); this survey was open for two weeks’ time and was facilitated online through the Qualtrics website.
Anyone over the age of 18 was encouraged to take the survey once, with promotion occurring online through social media (Facebook & Instagram). The researcher’s combined connections on these platforms totalled approximately 1500 people; based on this the target sample size for the survey was between 229 (90% confidence level) and 306 (95% confidence level) at a 5% margin of error. To this end this could be considered convenience sampling, a non-random sampling technique, given these individuals were the most will to respond (Kelley et al., 2003).
Designing the Survey
The content of the research tool was carefully planned in relation to the research question (Kelley et al., 2003). This involved collecting expert opinions through interviews, panel discussions, and keynote addresses, and also observing SDV testing. Data from these conversations, discussions, and observations was collected in raw notation and audio transcription and then organised by which of the barriers from Chapter 2 they fell within. Special attention was made to factors of human trust and use patterns as these directly relate to the research question. Key themes emerged as those which held a common thread throughout this observation process and were ultimately used to inform the content of the survey.
The sequencing of queries was designed such that respondents first focused on their own uses and experiences with driving and other forms of transportation currently available (sections 1 and 2). These sections collected demographic data that was categorically exhaustive and mutually exclusive on a nominal scale (MSG, 2017). This allowed respondents to be organisation by gender, age group, parental status, living environment, experience with driving and owning automobiles, and frequency with which other forms of transportation are used.
Section 3 posed questions of trust, from deterministic trust in the self and other humans (Q3.1-Q3.3) to exploratory trust in driver-assistance systems and HAV systems (Q3.4, Q3.5, Q3.7, & Q3.9). These were accompanied by explanations of the relevant technology in question. A 60-second video was embedded directly into Q3.9 to offer a visual presentation of HAV technology (although there was no mechanism in the survey to requiring the respondent to view it and the video did not automatically play). Queries were made to determine if respondents had any experience using driver assistance systems (Q3.6) or HAV systems (Q3.8 & 3.10). This section objectively measured respondents’ attitudes of trust using a Likert itemised rating scale (MSG, 2017).
Section 4 asked questions about the preferences of the respondent while they were experiencing varying states of control, from driving themselves (Q4.1_1-Q4.1_5), being driven by another human (Q4.2_1-Q4.2_5), to being driven by a HAV (Q4.3_1-Q4.3_5). This section objectively measured respondents’ attitudes of preference using an interval scale to rank importance (MSG, 2017).
Section 5 asked how each respondent generally felt about the future of transportation after having taken the survey (Q5.2) and offered an open field for any comments or questions the respondent may have wanted to share (Q5.3). This was to give respondents an opportunity to share information in whichever way they preferred to write it, which this research later reviewed through qualitative analysis.
Findings
The Future of Mobility: Trust & Preferences (see Appendix B) survey was released for two weeks from 31 July to 13 August 2017. At the end of that period the results were exported from Qualtrics into a spread sheet so the data could be cleaned and prepped for analysis . Of 219 respondents, 204 completed the survey and 15 did not and were removed (see Appendix C). Visualisations were made for demographic and use data (see Appendix D).
Quantitative analysis was performed, including Missing Value analysis using Little’s MCAR test[1] to confirm any data missing was missing completely at random. Missing cases and cases where respondents answered “not sure” and/or “I’d rather not say” were excluded on an analysis-by-analysis basis.
The items in section 3 (Q3.1-Q3.5, Q3.7, and Q3.9) formed a Likert subscale representing degree of trust the respondent had in various control situations. To determine the reliability of this subscale, analysis was performed to calculate Cronbach’s alpha[2], the results of which implied poor inter-relatedness between the item Q3.1 (“How much do you trust your own ability to drive a vehicle?”) and the other 6 items (Q3.2-3.5 & Q3.7, which asked the degree of trust the survey participant had outside themselves, including taxi drivers, other human drivers in general, and trust in driver assistance systems from L1-L3). From this we are able to conclude that trust in the self is not related to trust in others, or in driver assistance systems.
The items in section 4 (Q4.1-Q4.3) formed an interval subscale representing attitude of preference the respondent had with regard to various experiential factors of riding in an automobile. To determine the reliability of this subscale, analysis was performed to calculate Cronbach’s alpha[3] which implied acceptable reliability of the instrument and good inter-relatedness between all items in the set. From this we may conclude that preferences in features about the vehicle itself are interrelated with preferences in the driving behaviour of hired drivers as well as the driving behaviour of HAVs. In other words, participants considered a broad variety of factors in their preferences for vehicles features and operation with equal validity across all of these factors. Automobile manufacturers would be wise to account for this as they proceed with designing vehicle experiences of the future.
A linear regression was then performed to regress trust in L4 systems (Q3.9) on trust in L1, L2 & L3 driver assistance systems (Q3.4, Q3.5 & Q3.7). 46.2% of the variance in dependent variable (trust in L4 HAVs – Q3.9) is explained[4] by the independent variables (trust in L1-L3 driver assistance systems – Q3.4, Q3.5, Q3.7). Two independent variables in particular (Q3.5 p = .005 & Q3.7 p = .000) had a strongly significant impact on the dependent variable. Unstandardized Coefficient β for these same two independent variables (Q3.5 B = .302 & Q3.7 B = .488) showed that for every 1 unit of increase on the scale for Q3.5 or Q3.7 we can expect an increase on the scale for Q3.9 of .302 and .488 respectively: for each whole Likert unit a participant trusted L2, they trusted the concept of L4 HAV .302 units, and for each whole Likert unit a participant trusted L3, they trusted the concept of L4 HAV .488 units. This “measure of the average causal effect” of the independent variables on the dependent variable implies that trust in advanced driver assistance systems (L3) leads to stronger trust in even more advanced driver assistance systems (L4) (Angrist & Pischke, 2009). Bear in mind that a limited number of respondents were certain they had experienced these levels of driver assistance (Q3.6 N = 59 & Q3.8 N = 23).
Qualitative analysis of data gathered from 82 open responses was performed by transposing this information into Excel and undergoing several rounds of coding. The first round identified consistent terms and phrases within the language of each response and noted these as codes. The second round broke these sets of coded terms and phrases down even further, standardising them to the body of responses, sometimes attaching a unique part of the original response to maintain clarity. In many cases one original response was broken down into two or three different coded items. The results intend to reflect the volume of open responses while discerning consistent themes and ideas within the qualities of the data. This process is called “data condensation” and allowed for retrieving the most meaningful material after first assembling “chunks” of qualitative data (Miles, Huberman, & Saldana, 2013).
Data condensation of the open responses resulted in 7 themes: optimism, pessimism, concern, use cases, human-machine interaction (HMI), control, and general responses. Within these were 67 sub-themes, many of which were recurring, for a total of 142 occurrences.
Discussion
The findings of the survey suggest a number of conclusions in returning to the research question, “Will people trust highly autonomous vehicles, and how will they prefer to use them?” Participants were presented with a four lines of questioning and data about their demographics, use of transportation systems, trust in humans and autonomous systems, and their preferences with these systems. This process sought to explore the barrier of human factors that challenge the implementation and widespread adoption of HAVs.
While most respondents trusted themselves “a lot to a great deal” as drivers, a respondent’s trust in themselves was not a good determinant of their trust in other drivers or in HAV systems. Their trust in other drivers was on average “little to moderate”, meanwhile their attitude that L4 HAVs was “somewhat to very good of an idea”. This is a good indication that people are optimistic about HAV technology moving forward. The open responses echoed this—more were optimistic (50 occurrences) than concerned (25) or pessimistic (15) combined. These optimistic responses were also aligned with the major HAV benefits as researched—chief among these was an improvement of safety and the long-term potential of the technology.
Insights were also made into use preferences with transport. With conventional self-driven vehicles, dependability was a top priority followed by safety and then affordability. Shifting to a vehicle driven by another human, the top priority became safety followed the ability to regain control on demand and the ability to control navigation. In imagining a ride in a L4 HAV, the top priority was also safety, followed by ability to regain control and then smoothness of journey. The open responses also included the need to regain control of HAVs at the discretion of the passenger (3 occurrences) while no mentions were made directly to smoothness of journey beyond one occurrence referring to sleeping while being driven in a HAV. The degree of acknowledgement that HAVs could replace human drivers (3 mentions) while no mentions were made directly to smoothness of journey beyond one occurrences referring to sleeping while being driven in a HAV. The degree of acknowledgement that HAVs could replace human drivers (3 mentions) was outweighed by responses expressing humans will always have a certain enjoyment in driving (6 mentions).
Perhaps the biggest limitation of this survey was the homogeneity of the sample population. This is due to the researcher’s limited resources to increase exposure beyond personal, academic, and online social networks. On a positive note, many of the findings were statistically significant, which is a good sign that the survey content itself was balanced as was reflected by the range of data collected (Kelley et al., 2003).
Having done some linear regression analysis to measure the causal effect of trust in driver assistance systems with trust in HAV systems, the case could be made for further research of this type. This could be investigated through in-depth confirmatory research. Participants could be introduced to and experience test rides in vehicles with driver assistance systems. They could then be presented information about L4 and L5 HAV technology and perform a though experiment about riding in such a vehicle. The aim of the research would be to develop a fine-tuned scale in which the units of measure make more sense relative to a more specific query, for example the relationship between human trust in driverless systems and some discrete measure of time or distance could be used to predict how long it takes for a human to trust at HAV in terms of minutes spent or miles driven. Hypothetically, this amount of time is lower than expected, and experimentation done to test such a hypothesis would build nicely on the current research conducted here.
Conclusion
Overcoming the barriers to HAV adoption is a tremendous endeavour. Massive investment of time, energy, and resources in HAV technology continues to pour in. What will likely be the next revolution in mobility is set to improve the environmental sustainability of the transportation industry and expand individual access to personal mobility. Between these first deployments of L1 and L2 driver assistance systems and fleets of L5 HAVs, social scientists will continue to observe human behaviour as individuals interact with autonomous technologies. These observations will inform how innovators and engineers design driver assistance and HAV systems to meet the mobility needs of the future. Early adopters will serve as the first round of test subjects, and their responses to the technology will set new precedents for how humans use automated transportation. Their trust will lead the way for more to follow and the science-fiction concepts of driverless cars will shift into the paradigm of reality.
[2] In performing this analysis for the 7 trust items, N = 173 (31 excluded cases) the result was .553 of the variance is reliable and that score will increase if item Q3.1 is removed. After doing so, the result increased to .704, which improved the reliability of the instrument into a recommended range (Tavakol & Dennick, 2011).
[3] In performing this analysis for the 15 trust items, N = 156 (48 excluded cases) the result was .793 of the variance is reliable, and that score will not increase if any items are removed.
[4] which was statistically significant (p = .000 < .05)
A note about this version:
This version of The Future of Mobility: Trust & Preferences in Highly Autonomous Vehicles by Sean Duggan (submitted to Goldsmiths University London in partial fulfillment for MSc Management of Innovation) has been slightly shortened and edited to produce a more condensed representation of the original document. Certain appendices and the reference list have been omitted but can be found in the primary document upon request, although the reader will notice that the original reference annotations remain within the text of this document. The references included 26 news articles, 5 books, 10 research papers, 17 journal articles, 6 government policy documents, 6 lectures/discussion panels, and 3 interviews with subject matter experts conducted by the author.
Comments
Post a Comment