In the early stages of their development, self-driving cars average 9.1 crashes per million vehicles miles traveled, while human drivers average 1.9 crashes per million vehicle miles traveled. That number may be low because many people don't report minor crashes to police, but even adjusted for unreported crashes, the researchers say humans average 4.1 crashes over a million miles traveled.
Lest you think the results mean autonomous technology is unsafe, they note the self-driving cars have actually been the victims in the collisions examined by researchers. Of the 11 accidents included in the study that occurred in autonomous mode, humans were at fault every time.
Currently, 94 percent of all traffic accidents are caused by human error, according to the National Highway Traffic Safety Administration, and there is high hope from federal officials and transportation experts that self-driving cars could sharply reduce the number of traffic deaths. These early findings may dampen that hope. But even with a higher frequency of crashes, one possibility raised by Michigan's researchers is that self-driving cars might still be safer.
"It could be that you have this higher crash rate, but end up with safer situations than a conventional vehicle." - Brandon Schoettle
Autonomous cars were involved in more serious types of crashes at lower rates than humans. In conventional driving, angle crashes account for 31.4 percent of collisions and head-on ones represent 3.6 percent of the total. In autonomous accidents, angle crashes account for only 9.1 percent of the total and there has not been a known head-on crash involving an autonomous car on California roads, where the research was focused. All injuries in collisions involving self-driving cars were minor.
"Depending on which side of the fence you're on, there's certainly positive messages to take from this," Brandon Schoettle, the report's lead author, tells Autoblog. "The self-driving cars didn't have the most severe type of accidents, head-on, pedestrians. It's a little hard to tell now, but there may be a confusing message to discuss – it could be that you have this higher crash rate, but end up with safer situations than a conventional vehicle."
Schoettle and co-author Michael Sivak culled data for the study from three of the 10 companies authorized to conduct autonomous testing on California public roads: Volkswagen, Google and Delphi. They noted some limitations on the study, namely that testing is still in "early stages" and the 1.2 million miles conducted in autonomous operations are negligible compared to the nearly three trillion miles that human motorists drive each year on American roads.
Currently, 94 percent of all traffic accidents are caused by human error.
Because the sample size is comparatively low, Schoettle said they could not rule out the possibility the actual rates for self-driving vehicles are actually lower than for conventional vehicles. But considering self-driving cars are tested under only limited geographic and weather conditions, there's a possibility it could also be higher.
In addition to avoiding the more serious kind of accidents, most of the accidents involving autonomobiles occurred at low speeds. Nearly three quarters occurred when the vehicles were traveling less than 5 miles per hour or stopped. These rear-end collisions are much more common for self-driving cars. Less than half of conventional accidents, 48.3 percent, are of the rear-end variety. The discrepancy raises questions about why human beings are more likely to rear end a self-driving car as opposed to a human-driven one.
"We don't know," Schoettle said. "One possibility, and we're just kind of speculating, is they're not quite reacting in certain situations the way a human driver would. Hypothetically, maybe they're stopping quicker, or maybe they're stopping in situations that humans might not stop. So they're taking drivers by surprise. It could just be bad luck or drivers not paying attention. We're just not sure."
Google, which has logged more than 1.2 million miles of autonomous operations, believes it's the latter. In the company's monthly newsletters that provide updates on its self-driving car project, it has noted it believes rear-end crashes are underreported and that human inattention is to blame.
"Not once has our self-driving car been the cause of a collision," Google said in a statement following the release of the Michigan study. "We publish the details of all crashes we've been involved in on our website each month, and there's a clear theme of human error and inattention."