It was a "tough day" for Chris Urmson, Google's self-driving car director, when his prototype vehicle collided with a bus last month while cruising along at 2 mph in Mountain View, California.
The slow-motion accident, after all, involved a bus, which is a lot more common on roads than some of the other things that Google's robo-car has been trained to deal with.
In a presentation at the South by Southwest conference on Friday, Urmson recapped everything the car has learned to dodge:
- A woman riding in an electronic wheelchair and chasing a duck with a broom
- A streaker running in front of the car
- A group of people playing "Frogger" and leap-frogging through a crowded intersection
- A pedestrian jumping and rolling over its hood
- A cyclist coming in the wrong direction over a curb
- Three cars in Austin driving the wrong way
His team even taught the car to avoid "snailing," an internet phenomenon where people lie face down and try to crawl like a snail across the street. This one it hasn't seen in the wild.
Given its prep with dealing with unknown scenarios, it was still a bad day when the car caused its first accident in self-driving mode on Valentine's Day after more than 4 million miles on the road.
The car had moved onto the right side of the street and out of the lane as it prepared to make a right-hand turn. The cars originally were programmed to stay in their lanes to make turns, but they were soon taught that most people will pull out of the lane to move right if there is room available, Urmson said.
In this case, the Google car pulled to the right for the turn, but then encountered a pile of sandbags. After the light turned green, it waited for a few cars to pass before it saw a bus. To the bus driver, there was plenty of room to pass, Urmson said. But to Google's car, the software told it that it should move back into the lane. That caused a collision.
The Google car was going only 2 mph, but its sensors got lodged in the bus door as it was dragged across the intersection.
"That was a tough day for us," Urmson conceded.
While it may have been trained to avoid something as crazy as snailing, Google's self-driving cars still needed a software update so it knew that larger objects are much less likely to yield.