In This Issue
Spring Bridge on Sustainable Smart Cities
March 15, 2023
The world’s cities face increasing threats from natural disasters, aging infrastructure, traffic, and resource constraints. The articles in this issue examine smart infrastructure, sustainability, net zero carbon options, and autonomous driving, among other approaches to smart and sustainable cities.

Op-Ed A Better Way to Validate Autonomous Vehicle Safety

Wednesday, March 15, 2023

Author: Peter Ciriscioli

The National Highway Traffic Safety Administration estimates that in 2022 nearly 32,000 people died in traffic crashes in the first nine months of the year.[1] Autonomous vehicles (AVs) promise to dramatically reduce that number. But, although AV developers have reportedly spent more than $75 billion to bring improved safety and other AV benefits to the public, the vehicles are still not generally available for public use.

Despite the fact that AVs have driven millions of miles on test tracks and public roads, numerous media reports have documented that AV developers have struggled to demonstrate that the vehicles are safe by test driving them on US roads and highways (CBS News 2023; Noyes 2023; NPR 2022; NTSB 2017). In fact, a recent study found that AVs would have to be test driven hundreds of millions—even hundreds of billions—of miles to demonstrate their reliability in terms of fatalities and injuries (Kalra and Paddock 2016).

It is becoming clear that trial-and-error test driving of AVs on public roads and highways may never demonstrate acceptable safety.[2] AV road test failures are creating unacceptable delays for investors. AV startups have ceased operations. Failure to achieve acceptable safety levels has become a major obstacle to AV deployment.

The public is concerned because public AV test drives have frightened, injured, and killed people (Noyes 2023; NPR 2022; NTSB 2017). AVs are now considered by some to be more dangerous than vehicles operated by an impaired human driver.

It appears certain that AV deployment will not be successful without public consensus that the vehicles are safe enough.” To that end, a better way than test driving on public roads and highways is needed to determine whether AVs are safe enough” to be used without restriction in the public domain.

But how safe is “safe enough” for AVs? Since the public already accepts licensed human drivers as “safe enough,” I think most people would accept that an AV is safe enough if it demonstrates safety performance at least as safe as that of the average human driver. With that in mind, the next step toward developing consensus that an AV is “safe enough” is to determine the specific criteria and the methods used to validate that level of safety performance.

Fortunately, in the United States (and probably other countries) there is an established acceptable precedent for introducing a new human driver into the public domain: a driving test. Such a test is required for an individual to become, and periodically to remain, a licensed driver.

In a typical driving test a person must demonstrate that they

  • can operate a vehicle safely,
  • have safe driving habits,
  • can apply their knowledge of traffic laws in real-life situations, and
  • have the ability to safely compensate, as needed, for any physical condition (e.g., loss of a limb, poor hearing, or the need for vision correction).

An AV driving test could similarly be used to determine whether the vehicle can demonstrate that it

  • follows applicable traffic laws and operates safely under both routine and unforeseen circumstances;
  • exhibits safe driving practices—for example, recognizes the differences between bicycles, motorcycles, pedestrians, cars, and trucks, and knows to stop when approaching an emergency vehicle stopped in the road; and
  • can safely compensate for any loss of capability (e.g., sensor failure, software problems, safety-critical hardware failure).

The use of the driving test standard in the United States has yielded a large amount of data about human driving safety.[3] An analysis of the data will reveal both the safety performance of the average human driver” and the causes and circumstances of vehicle accidents that result in damage, injuries, and fatalities. The results can be used to inform AV driving test “safe operation” requirements and corresponding test methods. This approach will produce an AV safety validation process that mirrors human driver competency validation procedures. An additional benefit is that an AV driving test based on human driving behavior can be readily incorporated in existing regulatory processes.

Im confident that a driving test designed to validate that AVs are as safe as the average human driver will validate AV safety, reduce investor risk, and reduce time to deployment.


CBS News. 2023. 1 dead, 5 injured including firefighters when Tesla slams into fire truck on I-680. CBS San Francisco, Feb 18.

Kalra N, Paddock SM. 2016. Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability? Rand Corporation.

Noyes D. 2023. New video of Bay Bridge 8-car crash shows Tesla abruptly braking in “full self-driving” mode. ABC7 News, Jan 11.

NPR. 2022. Automated tech factored in 392 car crashes in 11 months, regulators report. Jun 15.

NTSB [National Transportation Safety Board]. 2017. Collision Between a Car Operating with Automated Vehicle Control Systems and a Tractor-Semitrailer Truck Near Williston, Florida May 7, 2016. Washington. 


[1]  NHTSA Estimates: Traffic deaths third quarter of 2022, traffic -deaths-2022-third-quarter

[2]  There is no consensus on what constitutes “acceptable safety” for AVs. Each AV developer appears to define “safe” differently. The most widely published metric developers use is miles driven on public roads; the idea seems to be that someday enough miles will be driven to demonstrate a level of safety that the public accepts. This hasn’t happened yet.

[3]  Federal and state departments of transportation track accident reports, and insurance companies quantify driver safety to determine insurance rates. These data, if made available, could be analyzed to develop an appropriate AV safety test that effectively determines whether an AV performs safely when confronted with circumstances associated with accidents involving human drivers.

About the Author:Peter Ciriscioli is a consultant on innovation and new product development.