A key selling point for self-driving vehicles is that they should be safer than conventional vehicles. Computers don't get drunk, distracted, or tired, so they should be perfectly alert at all times. And self-driving cars have advanced sensors that give them a 360-degree awareness of the world around them.
In the last couple of years, Google and other companies have started testing their self-driving vehicles on public roads, giving us a chance to measure their real-world performance. And new research from the University of Michigan comes to a surprising — and seemingly paradoxical — conclusion: self-driving cars get in more accidents, per million miles traveled, than conventional vehicles. Yet so far, every single accident involving a self-driving car has been the fault of the person driving the other car.
There are several important caveats to these findings:
- Self-driving cars still haven't logged enough miles to give us a large sample of autonomous vehicle accidents. After logging more than a million miles autonomously, self-driving cars have been in a total of 11 crashes in self-driving mode. With such a small sample, it's hard to draw statistically rigorous conclusions. It's possible, though unlikely, that the high rate of accidents observed so far is just a fluke.
- There's some uncertainty about the number of accidents conventional cars get into, because drivers often choose not to report minor accidents to the authorities — especially in cases where there are no injuries or property damage. The authors of the study used a Department of Transportation estimate on the number of unreported accidents to adjust for this, but that estimate might not be right. In contrast, companies testing self-driving cars are required to report every accident, giving us extremely accurate data on their performance.
- While self-driving cars seem to get in more accidents in general, these accidents tend to be less severe, on average. Most of the accidents happened because another vehicle rear-ended the self-driving car. Self-driving cars have not been involved in any head-on collisions, and crashes involving self-driving cars were less likely to cause injuries than crashes involving conventional vehicles — just two out of 11 crashes led to injuries, compared with 28 percent of conventional vehicles.
- On the other hand, self-driving cars aren't currently tackling all of the challenging driving situations human drivers have to deal with. For example, all of the data examined in this study came from states like California or Texas where winters are extremely mild. So self-driving cars haven't had to deal with the challenges of snow, ice, and sleet. It's also possible that self-driving cars are used in less demanding traffic situations, on average, than conventional cars.
Still, even taking all of these factors into account, it's still surprising that self-driving cars were involved in more accidents despite not causing any. So what's going on?
In a phone interview, lead author Brandon Schoettle stressed that there isn't enough data to draw any firm conclusions. But he suggested one possibility: that while self-driving cars aren't legally at fault, they may be behaving in ways that surprise human drivers who are used to interacting with other human beings.
For example, it's possible that self-driving cars stop more quickly than human drivers in some situations, making it more likely that they'll be rear-ended.
If that hypothesis is confirmed by future data, it could mean there will be a messy transition period, as a mix of human and automated vehicles lead to extra confusion on the road. On the other hand, the problem might be easy to solve if programmers can figure out how to make computers drive a bit more like human beings.
The Michigan team's research also underscores another important fact about self-driving cars: they're no longer a theoretical concept, at least for Google. The search giant has logged more than a million miles of self-driving on the road, providing the lion's share of the data used in the study. We can expect Google and its competitors — which include Tesla, Nissan, Volkswagen, and BMW — to continue ramping up their real-world trials of self-driving technology in the next couple of years, giving us increasingly precise data about the safety record of self-driving cars.