Alphabet’s Waymo subsidiary says its driverless cars have driven more than one million miles on public roads with no human at the wheel.
Waymo marked its achievement with a blog post and an accompanying paper [PDF] analyzing the safety performance of its cars when operated without a driver.
Remarkably, no one was hurt in the making of this milestone and there were only two collisions serious enough to meet the National Highway Safety Transportation Administration’s reporting criteria, which include the need to file a police report and/or towing one and more vehicles away.
The more serious of those two accidents, which occurred in 2020, was due to the teenage driver of another car who was looking at a mobile phone and hit the Waymo robotaxi from behind. The collision was rated to have a 4 percent chance of a MAIS2+ injury on the Maximum Abbreviated Injury Scale (MAIS). Waymo’s researchers characterize MAIS2 level injuries as comparable to a concussion with no more than brief loss of consciousness, a sternum fracture, or the fracture of two or fewer ribs.
The less serious non-consensual car encounter, assessed as having 2 percent chance of a MAIS2+ injury, occurred in 2021. It involved a Waymo vehicle that could not slow down in time to avoid hitting a car that had pulled into its lane and then suddenly braked.
“The Waymo AV exerted maximum brake force to avoid contact, but was unable to stop before making contact with the passenger vehicle, which had entered the intersection and was still braking,” the paper explains. “At the time of contact, the Waymo AV was traveling at approximately 21 mph and the passenger vehicle was traveling at approximately 10 mph.”
The robo car biz reported 18 other “minor contact events” after over a million miles of driving without a driver. Waymo says these would not be serious enough to include in a crash database but are mentioned for the sake of transparency. And human drivers were the cause of most of the incidents, it claims.
“Overall, more than half of all contact events were the result of a human driver hitting a stationary Waymo vehicle,” the company said, adding that none of its cars were involved in the sort of collisions that account for 94 percent of fatalities.
How do we stack up?
As a point of comparison, the fatal crash rate for US drivers, age 25-29, over one million miles during the 2014-2015 period was about 0.02 percent, according to data from the American Automobile Association Foundation for Traffic Safety.
Drivers in that group during that period experienced about 5.26 collisions and 1.5 injury crashes per million miles.
While Waymo automatic vehicles s may be doing well from a safety perspective, they still run into situations their systems can’t handle. According to the company’s 2022 disengagement report, Waymo vehicles reverted to human control 169 times as a result of unexpected maneuvers, conditions, or instrument feedback.
Dr. Alexander Wyglinski, professor of electrical engineering and robotics engineering and director of the Wireless Innovation Laboratory at Worcester Polytechnic Institute, told The Register that he was excited by what Waymo has accomplished, but cautioned that a lot more work needs to be done before self-driving vehicles are reliable enough.
“The key factor in all of this is the fact that these autonomous vehicles have to interoperate with human-operated vehicles,” Wyglinski explained.
People keep their cars for years or decades, he said, so AVs are going to have to deal with existing vehicles. That means emerging vehicle-to-vehicle communication systems can’t be assumed and further work will be required to finesse interaction with human drivers.
Waymo hasn’t been implicated in conflicts with human-driven vehicles as much as has rival GM Cruise, which last year drew the ire of San Francisco Municipal Transportation Board for creating traffic jams with malfunctioning self-driving cars and for its inefficient, inconvenient recovery process for stalled vehicles.
Wyglinski argues that inter-vehicle communication is essential to make AVs viable in the complicated real-world. Dealing with unusual road scenarios and adversarial behavior from human drivers trying to game AI behavior will require much more testing and much more data, he suggests.
‘It’s all about the information,” he explained. “Let’s say this car just experienced a deer for the first time. But other cars have not. How do you communicate its experiences? How do you communicate those databases and not necessarily up to the cloud? Or maybe you can, but that might be too late. It’s about that information exchange. So driving a million miles is one thing, but I would also say, what was the size of the database?”
“The more training the better,” said Wyglinski. “And that’s a frightening truth about artificial intelligence and machine learning decision making.”
Asked to grade Waymo’s accomplishment, Wyglinski demurred.
“This is going to be a long process,” he answered. “Waymo and all these other companies, they’re investing in a lot of the technologies that make sense. I think it’s a matter of time, how this all congeals together into solutions. And I think there’ll be many solutions that will be able to exist on our public roads in the future.”
As a result, Wyglinski said, it’s critical that universities teach about these systems and the technologies involved because today’s students will continue building the world to come. ®