Autonomous Vehicle Reality

I take no pleasure in saying "I told you so" concerning autonomous vehicle risks, but I do hope that last week's fatal crash involving a self-driving car will introduce some reality to discussions of autonomous vehicle benefits and costs, as discussed in my recent column, "The Many Problems with Autonomous Vehicles."

Let's be realistic about the reliability and risks of autonomous vehicles. How frequently does your telephone, computer, websites or GPS navigation system fail to perform as expected? Probably daily. Do those failures cause death? Probably not, thank goodness. However, vehicle operating system failures can have fatal consequences so it is important to evaluate them critically.

The United States currently averages a little more than one traffic death per 100 million vehicle-miles. So far, vehicles have driven an estimated 5-10 million vehicle-miles with autonomous operation (about two million by Uber and four million by Waymo. Tesla claims much more, but its Autopilot only offers limited autonomy), and caused two deaths, giving them a very poor safety record compared with human operation. 

Autonomous vehicle system designers will certainly learn from these accidents and hopefully will reduce future risks. However, operating a motor vehicle in mixed traffic involves many layers of complexity, which will continually introduce new risks, so like a wack-a-mole game, new hazards will pop up.

Optimists claim that, because human error contributes to 90% of crashes, autonomous vehicles will reduce crash rates and insurance costs by 90%, but that claim overlooks additional risks [pdf] these technologies are likely to introduce:

• Hardware and software failures. Complex electronic systems often fail, and even small vehicle operating system failures—a false sensor, distorted signal, or software error—can have catastrophic results. Self-driving vehicles will certainly have failures that contribute to crashes; the question is their frequency compared with human drivers.
• Malicious hacking. Self-driving technologies can be manipulated for amusement or crime.
• Increased risk-taking. When travellers feel safer they tend to take additional risks, called offsetting behavior or risk compensation. For example, if autonomous vehicles are considered very safe, fewer passengers may wear seatbelts and other road users may take greater risks. Pedestrians may become less cautious.
• Platooning risks. Many potential benefits, such as reduced congestion and pollution emissions, require platooning (vehicles operating close together at high speeds on dedicated lanes), which can introduce new risks, such as human drivers joining platoons and increased crash severity.
• Increased total vehicle travel. By improving convenience and comfort autonomous vehicles may increase total vehicle travel and therefore crash exposure.

These new risks will probably cause crashes, so net safety impacts are likely to be smaller than the 90% reductions that advocates claim. Sivak and Schoettle [pdf] conclude that autonomous vehicles may be no safer per mile than an average driver, and may increase total crashes when self- and human-driven vehicles mix. Groves and Kalra argue that autonomous vehicle deployment is justified even if they only reduce crash rates 10%, but their analysis indicates that net safety gains are significantly reduced if this technology increases total vehicle travel. For example, if autonomous vehicles reduce per-mile crash rates 10% but increase vehicle travel 12%, total crashes, including risks to other road users, will increase.

Many factors will affect these safety impacts, including how vehicles are programmed, and whether they increase or reduce total vehicle travel. For example, to maximize mobility autonomous vehicles can be programmed to operate faster, take more risks in unpredictable situations, and platoon; to maximize safety they can be programed to drive slower, be more cautious, for example, stopping for human instructions in any unexpected situation, and high occupant vehicle (HOV) lanes and road pricing can encourage vehicle sharing.

The following table is my summary of the expected benefits and costs of autonomous vehicle technologies. All of these factors should be considered when evaluating autonomous vehicle policies.

Autonomous Vehicle Potential Benefits and Costs [pdf]

Some of these impacts, including increased vehicle costs, reduced driver stress, and increased productivity, directly affect users. Others, such as changes in accidents, congestion, roadway costs, pollution emissions, and mobility options for non-drivers, are external impacts. The total magnitude of these impacts will depend on how autonomous driving technologies affect vehicle travel: improved convenience and productivity cause vehicle travel to increase, traffic problems such as congestion, accidents, pollution emissions, and roadway costs may increase. But, if more affordable shared mobility options allow households to reduce their vehicle ownership, and therefore leverage reductions in total vehicle travel, traffic problems should decline. This suggests that, to maximize net benefits, public policies should focus on resource-efficient and shared urban mobility, which autonomous vehicles can facilitate, rather than autonomous vehicle technology as an end in itself.

As planners, our quest is for a realistic assessment of total benefits and costs in order to identify the policies that will result in the greatest net benefits to society.