Data pooling

Scientist, innovator and entrepreneur Sebastian Thrun, a pioneer in AV development, describes the importance of AV safety data: ‘When one of the self-driving cars makes an error, all of the self-driving cars learn from it. In fact, new self-driving cars are “born” with the complete skill set of their ancestors. So collectively, these cars can learn faster than people’.18 This underlines the unique potential of AVs to learn from one another, where data associated with a safety issue experienced by one AV could be analysed and disseminated so that lessons learnt could be adopted across the entire fleet.19 20

The value of data pooling has been demonstrated in the field of reliability engineering, in which reliability is defined as the ability of a product or system to perform its functions without failure or safety incidents. When data are systematically collected and pooled, lessons learnt from past operational performance can be applied to future designs to address reliability issues. This process, known as reliability growth, follows the general principles of the learning curve and results in continuous improvement.

Reliability growth is not restricted to products from a single manufacturer; if data are shared between manufacturers, the learning rate increases and the entire industry benefits. The aeronautics industry, with its strong emphasis on safety and reliability, has successfully pursued data pooling efforts through two ongoing efforts, the Commercial Airline Safety Team (CAST) consortium and the Government-Industry Data Exchange Program.21 22 These programmes improve the total quality and reliability of operations, systems and components by pooling technical information from multiple manufacturers and vendors.

Data pooling programmes are credited with a number of critical safety enhancements for commercial aviation, including refinements to warning systems that alert pilots that they are flying too close to mountains.23 The refinements corrected a situation where the systems had been excessively conservative and pilots were desensitised to the warnings. Data pooling can be effective, but the programme need to be designed with incentives for participation. Corporations may be reluctant to contribute data that reveal proprietary information. Such issues would need to be addressed before a system could be implemented across the AV industry. However, the aviation example shows that such an achievement is feasible.

Graduated approach to risk exposure

The effectiveness of graduated risk exposure in reducing crashes among novice teenage drivers has been well established. Graduated licensing programmes have been adopted in some form by all US states and these policies have demonstrated fatality reductions of up to 20% among 16-year-old drivers.24 Human drivers learn most effectively and safely when they are initially exposed to relatively simple environmental conditions, such as lower speeds, less dense traffic, good lighting, fewer distractions and better weather. Later, as their skills increase and they demonstrate competence, they are exposed to progressively more complex and challenging conditions.

The same approach could be extended to the testing of AVs. Similar to human drivers, there are several dimensions of risk that could be adjusted during the initial testing phase for AVs. For example, suburban arterials with heavy, fast traffic and frequent vehicle interactions have relatively high crash risk per mile, while limited access divided highways have much lower risk.25 Geospatial crash data could reveal differences in crash risks across road segments within a city and inform route choice for AVs in a graduated testing approach. Visible roadway lane markings and the predictability of road user behaviour also vary considerably among roadway types and locations.26 An inventory of appropriate infrastructure and an analysis of the makeup of road users within a city could provide additional dimensions of risk for gradual exposure to AVs.

Several jurisdictions and companies have already taken this type of phased-risk approach for AV testing, requiring demonstrated competence in low-risk circumstance before permitting use in more complex driving environments.27 Waymo’s decision to deploy their vehicles in Chandler, Arizona, where the streets are flat and wide, there is little rain or fog, and detailed mapping data are available, was a design choice to reduce risk.28 More could be done in managing risk while accumulating operational experience with AV testing, and this caution could help prevent future crashes involving AVs and avoid the subsequent erosion in public trust.

This analogy between driver training for teenagers and AVs could be extended to include the requirement for a supervisor to be present during the learner period. During initial on-road testing, AV companies include a ‘safety driver’ in the vehicle. This is necessary to ensure that a ‘back-up’ safety mechanism remains in place as a computer system learns how to safely handle the wide range of environments and road user situations encountered on public roads. Some companies have extended this concept to allow any employee the authority to ground the self-driving fleet when safety issues are detected in the product development and testing phase.29

Harm reduction

A concept that has been used alongside graduated licensing is harm reduction. When applied to teenage drivers, this approach attempts to compensate for higher crash risk by providing them with the safest possible vehicle (rather than the smaller, older and inherently less-safe vehicles that are too often given to young drivers).30 If applied to AVs, the same approach would have AV companies voluntarily using vehicles with the highest safety ratings (which include pedestrian protections), at least until a predetermined level of demonstrated safety competence had been achieved. Similarly, AVs that have been custom built (rather than conventional vehicles retrofitted with instrumentation) could, for example, be voluntarily crash tested at higher speeds than mandated for passenger cars as a further measure to build public trust. Harm reduction principles would also suggest that initial testing environments for AVs should meet minimum safety standards (eg, adequate sidewalks for pedestrians, well-illuminated roads and the presence of protected bicycle lanes).

Harm reduction principles differ from graduated learning in that rather than attempting to lower crash risk, they acknowledge that some crashes may occur and seek to compensate by providing a safer environment. While AVs that have been carefully developed are likely to present a lower crash risk than human-driven vehicles and therefore may not justify such measures, harm reduction approaches could be worthwhile because they may bolster public confidence by demonstrating an abundance of caution.

Inclusion

Strategic early deployment of AVs in rideshare programmes or other innovative configurations promises to extend mobility by making transportation services cheaper and more convenient. This could mean that mobility services will be economically feasible in underserved areas and that health dipartites could be reduced. Enhanced mobility can impact disparities by providing access to healthy foods, healthcare, economic opportunities, and recreation and exercise, which influence quality of life.31 32 However, since the roadway environment—both physical infrastructure and road user characteristics—can be quite different in these underserved areas than in more affluent neighbourhoods, it is important that AVs develop safety competencies in such areas as quickly as possible. The critical issue here is developing testing strategies that will facilitate AV deployment in underserved neighbourhoods.

The underlying hypothesis is that AV deployment will be quicker and safer in areas that mimic the real-world setting where testing took place. There is a parallel in clinical trials, in which a medical product is tested for efficacy and safety in a targeted population and subsequently approved for the same population. Use of this medical product on a substantially different population (eg, prescribing a drug to children, when the drug was only tested on adults) is unproven and therefore discouraged until clinical tests can be conducted with this population.

We acknowledge that the need to test in underserved areas may sometimes conflict with the need to begin testing in a low-risk environment. However, because of historical mobility deficits, underserved areas should be included in geospatial analyses of crash risk and given priority as testing locations when consistent with safety needs. Moreover, testing should continue until such areas are addressed by the graduated risk approach described previously. This will help ensure that AVs fully deliver their potential public health benefits.

Educating the public

Most people do not think about crash risk in a rational way and education may improve understanding of relative safety risks. Many drivers have an optimistic and inflated sense of their ability to drive safely,33 and generally place a premium on being able to control a situation, rather than on delegating it to another entity (even if the other entity is more capable).34 By implication, many people will demand a much higher standard of safety from AVs than they would expect from other drivers, including themselves. The ease with which one can bring to mind examples of an event (such as highly publicised AV fatalities) can also influence perceptions of risk and the likelihood of an event happening to themselves. This natural human bias, which psychologists termed the ‘availability heuristic’, might explain some of the uneasiness among the public with the thought of riding in an AV or sharing a road with them.35 Public education initiatives, such as the Partners for Automated Vehicle Education, represent steps towards addressing misconceptions about the safety risks posed by AVs.36

In addition to driver education and campaigns, public information comes from media attention concerning the safety of AVs. Statements from trusted local sources, including highway safety officials and law enforcement officers, either through their own information outlets or through media interviews, are key in determining public attitudes. A recent white paper from the Governors Highway Safety Association outlines information needs, operational requirements and other prerequisites that highway safety officials and law enforcement agencies will need in order to effectively perform their functions when AVs are operating in their jurisdiction.37 For example, highway safety officials need information they can use in advising drivers of conventional vehicles concerning how to interact with AVs on the road. Law enforcement officials need a uniform way to identify AVs and procedures to interact with them at crash scenes. Organisations that are testing AVs should respond to the needs of these local officials in order to maintain their trust and confidence.

Smart policy

Regulation is a potentially powerful tool but needs to be used carefully, especially in a dynamic area such as the development of automated driving technology. Regulation of conventional motor vehicles and safety devices has been extremely effective. An evaluation by the National Highway Traffic Safety Administration of 21 federally mandated safety technologies concluded that, between 1960 and 2012, these technologies—from energy-absorbing steering columns implemented in the 1960s to electronic stability control mandated in all cars in 2012—saved a total of 613 501 lives.38

However, the pace of innovation and development of automated driving systems is incredibly rapid compared with these conventional technologies. Serious questions have been raised about the ability of traditional regulatory approaches to provide safety assurance without stifling innovation and safety progress. Many questions remain regarding the most appropriate substitute for traditional motor vehicle safety standards for assuring the safety of highly automated vehicles. Alternatives have been considered based on experience in other fields, including aviation and drug safety, but there is no consensus yet on the best method to assess the safety performance of these technologies or on the most reasonable threshold for acceptability.39

The history of automotive regulation suggests that collaboration may be more efficient than traditional adversarial relationships among regulators, safety advocates and industry. A recent example is the collaboration between National Highway Traffic Safety Administration and the Insurance Institute for Highway Safety’s initiative to work with 20 automakers representing more than 99% of the US auto market to make automatic emergency braking a standard feature on all new cars.40

Another relevant example can be found in the history of airbags, where 40 years passed from the time federal regulators first pursued the goal of equipping cars with air bags to their introduction as standard equipment.41 Interestingly, the history of airbag development also includes a compelling example of cooperative effort between government and industry in advancing vehicle safety. When the problem of passenger-side air bag interaction with children was discovered in the late 1990s, industry, government and safety groups mobilised in a coordinated campaign that resulted in an unprecedented behaviour change requiring children to be in the rear seat, along with technology changes that resulted in the near elimination of this problem.