Google Self-Driving Car Rear-Ended In First Injury Accident

An anonymous reader writes: Google's autonomous car project, as of June, hadn't been in any accidents that involved an injury. That changed on July 1st, though it wasn't the technology's fault. A Lexus SUV that was self-driving while carrying three Google employees was rear-ended while stopped at a traffic light in Mountain View, California. All three employees had minor cases of whiplash, and were quickly checked out and released from the hospital. The other driver had minor neck and back pain as well. Chris Umson, head of the autonomous car project, said, "Other drivers have hit us 14 times since the start of our project in 2009 (including 11 rear-enders), and not once has the self-driving car been the cause of the collision. Instead, the clear theme is human error and inattention. We'll take all this as a signal that we're starting to compare favorably with human drivers." He also posted a short video of how the self-driving car was tracking other vehicles at the time of the crash — including the one that hit it.

Comments

Finally the problem is clear

[Anonymous Coward]

Here if we watch the video we can clearly see the downside to a self-driving car vs. a human driver.

A human driver could easily make the decision to swerve up onto the sidewalk, or even to brake-check and nudge itself closer to the car in front, thus giving the car behind time to stop.

A human can see the collision about to happen and make changes so it doesn't.

The computer can do this in theory but we are many years from that kind of performance, if not decades.

Crash Mitigation

[jazzy82slave]

If Google's self-driving car was able to track the car that rear-ended it, I wonder if there are ways to mitigate this kind of "predictable" crash. Maybe letting off the brakes a tad to lessen the impact, or (out of left field idea) deploy air bags on the bumpers?

Seems like if the real issue is "everyone else" in driving you would think Google could come up with ways to reduce the impact level of inevitable crashes.

Re:Crash Mitigation

[Anonymous Coward]

Yes, this is obviously the solution. Unfortunately it is non-trivial, and upon this hinges the entire fallibility of Google's "we're-better-than-you-so-stfu" mentality about all of this.

A human will, for the forseeable future, be potentially far greater at this kind of improvisational disaster-avoidance than any computer when dealing with limited data in situations where no course of action is clearly favorable.

Remember when Watson made itself look no better than a chatbot on Jeopardy? Eh, probably not - the media downplayed THOSE answers.

Re:Crash Mitigation

[Solandri]

As for letting off the brakes when getting rear-ended, that may not be a good idea - the guy in front of you may not appreciate turning a 2 car wreck into a 3 car wreck. Especially if said impact pushed them into crossing traffic.

I've actually been in this exact scenario. I was on the freeway when traffic in front suddenly stopped due to an accident. I stopped, but noticed in my rearview mirror that the two teenage girls in the car behind me were busy yammering away with each other. They got closer and closer, before finally noticing that I was stopped and slamming on the brakes.

Here's the crucial part. Based on their distance, how fast they were going, and how quickly the brakes were slowing them down, I could estimate that they were going to stop about a meter past my rear bumper. As it turned out I had stopped with a good 3 meters between me and the car in front. So I just scooted forward a couple meters (this is the reason you're supposed to stop far enough back that you can see the rear tires of the car in front). The girls came to a screeching halt just behind me, heads flung forward against their seat belts, bounced up, and they started nervously laughing at each other. No collision, nobody got hurt, nothing got damaged.

A computer-controlled car which knows exactly the distance to the car in front, distance to the car in the rear is, how fast it's approaching, and how quickly it's decelerating. It can easily make this kind of calculation and decide if its better to let off the brakes and scoot forward, or press down hard on the brake to absorb the rear collision but avoid hitting the car in front. I lucked out because I happened to be watching the entire situation develop in the rearview mirror, and could accurately estimate their speed and rate of deceleration. But a computer could calculate this at any time. And if you watch the video, the Google car had enough situational awareness that it could've easily detected cross traffic - it wasn't just tracking the cars immediately next to it.

Re:Northeast winters

[buk110]

I'd be curious to see how it responds to really weird northeast conditions like a snow squall or black-ice. Or my personal favorite, when it's really snowing and you need to make sure you're stopped in a good spot that you can get traction once you can start moving again

Re:Exchanging insurance information

[xxxJonBoyxxx]

Insurance companies have this information at their fingertips. Here are some public numbers:

1.2-1.5 deaths for every 100M miles travelled
185 crashes for every 100M miles travelled (or 300 if only 45% are reported)

From https://en.wikipedia.org/wiki/... and http://www.caranddriver.com/fe...

However, in my experience, those numbers are pretty low. My wife and I have been involved in 3 accidents around a mid-sized city over the past 10 years (probably about 300,000 miles, or 1 crash per 100K miles), and I remember getting into a minor fender-bender (hardly ever reported) with someone about once a year when I worked near Chicago (1 crash per 20K miles).

I've driven behind one of these cars

[rsborg]

They do very unpredictable driving school-level things like slow/stop where deep shadows fall on the road. Like very suddenly. And then they stay there for a few seconds.

I'm not surprised there's finally a rear-ending. I'm actually surprised it took so long.

Re:11 rear enders

[FlyHelicopters]

A more attainable way to improve safety would be to allow people to continue to drive if they want to, but to add intelligent accident-avoidance software to the automobile so that when the person is driving, if the car notices he is about to cause a crash, it can step in and take the necessary actions to avoid or minimize the crash.

Let me fix this for you...

Your auto insurance deductible is $100 when the car is self-driving, but $1,000 when you're manually driving. If the car is self-driving and it causes a wreck, the auto manufacture is liable, if you're driving, then you are.

You don't have to ban people from driving, you just use carrots and sticks to make people want to stop driving.

No options for normal people, Google did 1/2 bette

[raymorris]

> It isn't clear that there were lots of options from the video - perhaps move ahead a foot but it seems like that would at best delay the crash). There were two cars stopped at the light, the Google car was behind it

The way most people drive, they wouldn't have any option to avoid the crash. According to the video, though, the Google car did better. It did as taught is advanced driving classes and left enough room that it could have pulled to the right, into the turn lane, and even driven away, as it detected the fast-moving car approaching from behind. That's taught as a more safe way to stop - think car jackings, a cement truck coming up fast from the rear, or an ambulance trying to get through. You can get out of the way or leave the area entirely if you leave six to eight feet between you and the car ahead.