2018 was an eventful year for autonomous vehicles (AVs). As the year began, optimism was high that AVs would soon be available to the public at a meaningful scale. After a number of setbacks, including the fatal crash of a self-driving car being tested by Uber in Arizona, there seems to be an emerging consensus that automation will be deployed more gradually, rather than leaping to scale in the next year or two. Capital markets and public attention have reacted by shifting some resources and attention to domain-constrained AV solutions like low-speed shuttles, autonomous trucks and even to lower-tech micromobility solutions such as (non-autonomous) bikes and scooters.
While technology and business development will remain key short-term goals for the industry, 2018 saw an increasingly prominent focus on the “safety conversation.” As self-driving technology matures and approaches broader commercialization, we are seeing an increase in conversations that seek to better quantify how safe AVs are. This has been manifested in several ways: The federal government continues to develop AV policy, and is specifically prioritizing the question of how you both measure and communicate the safety of AVs (although government progress on this front is slowed by its failure to pass AV legislation).
The private sector has increased its focus on formalizing and communicating its efforts on safety. As of this posting, nine companies have submitted Voluntary Safety Self-Assessments to the National Highway Traffic Safety Administration (NHTSA). One leading developer created its own system for measuring the safety of an AV’s driving behavior, and another sponsored a report by the RAND Corporation on how to best measure AV safety. Companies have also sought to burnish their safety credentials and processes by hiring former senior government safety officials.
There is a clear reason for this: While AV developers advocate that development and refinement of self-driving technology can eliminate the crashes and roadway deaths caused by human driving error, they must grapple—currently without detailed regulatory or legal guidance—with how safe their systems must be before being broadly deployed to consumers. Ultimately, responsible technology developers will need to prove to themselves—and likely at some point to regulators as well—how safe AVs are. This requires better understanding of how to perform “safety assurance” (demonstrate that any means of measuring AV safety is sufficiently robust to ensure performance, as well as communicate this assurance to internal, public, and government stakeholders).
Safety assurance is not a new concept invented for AVs. Society must constantly grapple with questions of balancing safety and innovation: How safe are the medicines we use? How secure are the deposits we make in financial institutions? How reliable is the security of websites we visit and through which we transmit personal information? At their core, these are questions of determining an appropriate balance between safety and other public and consumer benefits and creating mechanisms so that the public is confident that they can reasonably rely on these technologies. To win public acceptance, AVs must match—and will likely need to demonstrably exceed—the safety of vehicle transportation today. The proper approach to safety assurance will communicate to the public that this standard is met and that travel in AVs not only has a strong value proposition, but is reasonably safe as well.
Over the last few years, there have been enormous strides in developing self-driving technology, with AVs reportedly at the cusp of significant consumer deployment. Even so, it is not clear that there has been a lot of progress on the core question of safety assurance: How do we, including the companies developing the technology, know when AVs are safe for full deployment on public roads and how will this be communicated to the public?
An increasing number of seasoned industry observers have noted that a more robust approach to regulation, safety assurance, and communication will be necessary to “industrialize” AVs and achieve scale. Through its research, advocacy and public education work, SAFE has advanced the view that AVs can succeed as a commercial enterprise if, and only if, they deliver material improvements on roadway safety. This is the first of a series of blog posts which will examine a variety of themes and questions on the topic of AV safety and safety assurance.
A Deeply Multifaceted Question
With all of the conversation about safety, data sharing, standards, and certifications, it’s difficult for industry, regulators, and certainly the public to disentangle the complex and overlapping issues necessary to build an AV safety framework. This post outlines a high-level framework to address a simple question: when should AVs be allowed to deploy on public roads? A first step is clearly formulating what “equities” are raised by this question and who is positioned to address them.
Our analysis shows that the issues raised by the question of “how safe should AVs be before deployment” really fit into one of three distinct but highly-related topics, as illustrated by the below figure.
This figure illustrates that there are multiple stakeholders and perspectives even for what seems like a relatively straightforward question. Deciding when AVs can be on the road is not just a value judgment, but a political and scientific one as well. By separating out each sub-question, it will be far easier to make progress in guiding both industry and government on the path forward.
Technological innovation is essential to the American economy, but all change entails some degree of risk. The guiding principles for managing technology-induced change should be based on the risk that society should—and will—tolerate in the deployment of this new technology. This question is especially pointed for AVs, as developing the technology necessitates testing and deploying on public roads, which exposes the public to some level of risk. The decision to deploy AVs requires a value judgement on what degree of risk should be borne by the public in order to advance the deployment of AVs.
The deployment of AVs requires weighing the competing values of the technology’s benefits against the different—even if smaller—risks posed by AVs. What risks should be tolerated, both initially and over time? An answer might provide different answers for different communities and segments of society; there is an important role for advocates and policymakers to articulate the values of the communities they represent. Ultimately, though, an answer to this question might be phrased as a simple statement of values. One illustrative answer, and one which has actually been advanced, is formulated as “AVs should be deployed only when as safe as a human driver.” Some experts have claimed that as AVs might need to be one thousand times as safe as a human driver to gain acceptance. Obviously, such a benchmark would pose significant obstacles to deployment.
Given the multiple perspectives, stakeholders would likely benefit from attempts to articulate and formalize the risks they believe worth taking and socializing them with the public. By doing so, AV developers can dispel any notion that AVs will achieve perfect safety and that deployment even while safety performance continues to improve is in the public interest.
Ultimately, answering a question of values—or deciding which approach to take—requires a decision process. The details of this decision process are a political question.
More specifically, who has the authority to define the safety requirements for AVs, stipulate the conditions for verifying them, and manage the process? This question is already a fault line in the present automotive industry. Today, cars sold in the US are self-certified as safe by manufacturers who measure compliance against a series of motor vehicle safety standards created by the federal government with considerable industry and advocacy input. In Europe, vehicles must undergo “type approval” before going on sale, giving regulators considerable input in every model.
For AVs, this question will be critical: Will the government require a subjective approval process before AVs are allowed on the road? Will they be presumed safe unless evidence emerges to the contrary? Who decides if AVs already deployed are unsafe and should be removed from roads – and how should that determination take place? How will different voices (federal government, states, industry, advocates for the public interest, etc.) be weighted in deciding which benchmark and procedure is employed?
Well-formulated answers to these questions will also help promote acceptance and development of the technology. A robust regulatory system that still promotes innovation will help the public weather inevitable crashes by instilling confidence that the safety of AVs is well-vetted. In turn, this will allow developers to move ahead with activities that have received the regulatory imprimatur; with a relatively light regulatory environment, companies are taking a conservative approach to deployment.
Even if everyone could agree that AVs should be as safe as a human driver, or twice as safe, or any other specified standard, there is a critical and as-yet-unsolved scientific question: How do we measure the safety performance of AVs from a small sample size in a way that is predictive of how they will perform in broader use? We evaluate human drivers with a short driver’s license test and use the assessment of trained observers guided by rubrics to decide whether the driver is “good enough” to be allowed to drive on public roads.
What is the right method for assessing the safety of AVs? For humans, we generalize—passing a driver’s test in New York City allows you to drive in rural California and often even other countries because of reciprocity. Will AVs have one assessment for all operating conditions and locations—or will assessments need to be specific to potential operating locations and environments?
As was discussed earlier, the AV industry itself has not yet clearly formulated or standardized a way to measure the safety of an AV. It is likely that answering this question will require significant cooperative research and development and public funding.
The three categories of questions are starting points for deeper conversations around AV safety. However, it is clear that even superficial answers to each question category will be interdependent on each other.
A choice for the correct safety requirements for AVs will be influenced by how any level of safety will be assessed—one can only implement requirements that can be measured! Similarly, even if the political preference is to apply the current system of self-certification to AVs, the ability to realize self-certification will be constrained by the ability of technology to outline objective procedures that are not dependent on subjective or deployment-specific judgments (as subjective judgments cannot be implemented in self-certification frameworks!).
2019 is shaping up to be an important year for AV deployment and for further development of an AV safety framework. In further posts in this series, we will use this high-level framework to identify and contextualize narrower questions that will be answered in greater depth. Keep an eye on this space.
Amitai Bin-Nun is the Vice President of Autonomous Vehicles and Mobility Innovation at SAFE.
Kristen Hernandez, a Policy Analyst at SAFE, contributed to this post.