In this report, we use the RAND Model of Automated Vehicle Safety to compare road fatalities over time under (1) a policy that allows HAVs to be deployed for consumer use when their safety performance is just 10 percent better than that of the average human driver and (2) a policy that waits to deploy HAVs only once their safety performance is 75 or 90 percent better than that of average human drivers—what some might consider nearly perfect.

We find that, in the long term, under none of the conditions we explored does waiting for significant safety gains result in fewer fatalities. At best, fatalities are comparable, but, at worst, waiting has high human costs—in some cases, more than half a million lives. Moreover, the conditions that might lead to comparable fatalities—rapid improvement in HAV safety performance that can occur without widespread deployment—seem implausible. This suggests that the opportunity cost, in terms of lives saved, for waiting for better HAV performance may indeed be large. This evidence can help decisionmakers better understand the human cost of different policy choices governing HAV safety and set policies that save more lives.

—Kalra and Groves (2017)

Given the many uncertainties about the future of autonomous vehicle performance and use, the calculations were made by estimating road fatalities over time under hundreds of different plausible futures and different safety requirements for autonomous vehicle introduction.

Our work suggests that it is sensible to allow autonomous vehicles on America’s roads when they are judged to be just moderately safer than having a person behind the wheel. If we wait until these vehicles are nearly perfect, our research suggests the cost will be many thousands of needless vehicle crash deaths caused by human mistakes. It’s the very definition of perfect being the enemy of good.

—Nidhi Kalra, co-author of the study and director of RAND’s San Francisco office

Developers of autonomous vehicles are testing the cars in cities such as San Francisco and Pittsburgh, while federal lawmakers are considering a variety of new regulations and updates to existing regulations to govern their deployment and encourage their use. But what remains unknown is how good the vehicles have to be before they are made available for use to all consumers.

The allure of driverless cars is based partly on convenience and partly on the potential to eliminate costly human errors, such as driving when drunk, tired or distracted. More than 90% of crashes involve such driver-related errors, according to the National Highway Traffic Safety Administration (NHTSA).

Researchers acknowledge that even if autonomous vehicles are proven safer than the average human driver, the vehicles would still cause crashes. They remain vulnerable to other hazards, such as inclement weather, complex traffic situations, and even cyber-attacks.

This may not be acceptable because society may be less tolerant of mistakes made by machines than of mistakes made by people. But if we can accept that early self-driving cars will make some mistakes—but fewer than human drivers—developers can use early deployment to more rapidly improve self-driving technology, even as their vehicles save lives.

—David Groves, study co-author and co-director of RAND’s Water and Climate Resilience Center

Kalra hopes the study will enable policymakers and the public to better weigh potential risks and benefits of autonomous vehicles. Key considerations include how to measure the safety of the vehicles and what should constitute a passing grade.

The report builds upon past research that found road testing under real traffic conditions is impractical for proving autonomous vehicle safety prior to deployment because it would take decades or longer to drive the requisite miles.

Resources

• by Nidhi Kalra, David G. Groves (2017) “The Enemy of the Good: Estimating the Cost of Waiting for Nearly Perfect Autonomous Vehicles”