from Hacker News

Amazon's self-driving taxi Zoox under investigation by US after rear-end crashes

by batmaniam on 5/13/24, 6:36 PM with 16 comments

• by tgsovlerkhgsel on 5/13/24, 8:48 PM

  "after two of its vehicles braked suddenly and were rear-ended by motorcyclists."

  While sudden braking definitely contributes to it and is dangerous, this is mostly the fault of the drivers crashing into the self-driving vehicles even if they stopped without a good reason. And those vehicles are rather obvious, i.e. the driver following could have known that tailgating them at an unsafe distance is an even worse idea than with human drivers.

• by ryandrake on 5/13/24, 8:50 PM

  I always wondered how different our auto/road system would be if the NHTSA investigated each and every human driver crash as thoroughly and visibly as they seem to be interested in autonomous driving crashes--and instituted corrective regulatory action on human driving as heavy as we're likely going to see on autonomous driving. Apparently ~2M human crashes/yr in the USA is just a fact of life, but a handful of autonomous crashes/yr warrants investigation?

• by neilv on 5/13/24, 8:57 PM

  Using this as an example of a huge alignment and priorities issue in "AI"...

  Consider a kind of automobile crash for which driver of vehicle A is generally considered not-at-fault. For the sake of this discussion, let's say, a driver who gets rear-ended is generally considered not-at-fault (though I'm not sure that's the case -- just a simple example).

  However, if you have an interest in the safety of occupants of vehicle A, who's at-fault isn't your top priority.

  And there's never any occasion on which you do the math to say how you'd prefer your loved ones be possibly crippled or killed based on who would be considered at-fault.

  OK, now suppose that you're an autonomous driving product developer. The decisions of driving behavior are different, and maybe your top priorities are to avoid bad PR and big payouts. So you'd have big incentive to take the at-fault PR and payout priorities into account when you're developing your autonomous driving sytem.

  For example, your vehicle moving forward to strike something is probably very bad, for PR and payouts. But stopping abruptly, likely causing a vehicle behind to strike you, doesn't get as much of your attention, and/or doesn't weigh as strongly in your balancing.

  Especially if the vehicle behind you is much smaller, like a motorcycle, so maybe your occupants don't even get whiplash, and the crushed motorcyclist would be considered at-fault (and generally motorcyclists don't get much public sympathy).

  (The first examples I heard for autonomous driving were things like "There's an oncoming vehicle or obstruction, so do you choose to crash into it, probably killing/crippling the occupant of the vehicle you're driving, or to swerve into that pedestrian on the sidewalk, probably killing them. The PR calculus is different there, since consumers might secretly favor a system that prioritizes the safety of the buyer while increasing threats to others, without saying it out loud (see SUVs). But at the end of the day, the answer of how the autonmous vehicle drives, and who loses in driver decisions, might come down to the developer's very typical corporate priorities.)