Call me crazy but isn't a big part of the promise of these systems supposed to be they see things humans wouldn't or couldn't? If they're just as surprised as the human behind the wheel, that feels like a problem.
Yes, to me, this is the most relevant part of the posted article:
> “It’s possible that Uber’s automated driving system did not detect the pedestrian, did not classify her as a pedestrian, or did not predict her departure from the median,” Smith said in an email. “I don’t know whether these steps occurred too late to prevent or lessen the collision or whether they never occurred at all, but the lack of braking or swerving whatsoever is alarming and suggests that the system never anticipated the collision.”
We can accept that by the time the Uber AV "saw" the victim, it was too late to brake. But that doesn't mean we can't ask questions about what the AI actually saw and classified. And even if it was a "sudden" event for humans, was it "sudden" relative to computer reaction time? How fast did the victim move into the road that the Uber AV couldn't even brake?
Yeah, but my comment was made back when there were 2 stories about the Uber accident. This one, and the original one. The Bloomberg story has little more information than what the police chief claimed to the Chronicle, and at that point, the Tempe PD was downplaying the importance of her determination.
But everything the police claimed was in context of witness interview and camera footage. Since this is a self-driving car, we know there is more data and more ways to assess performance besides camera vision.
it could be that they never saw someone pushing a bike across a road in tests before, maybe it confused that with someone riding on the road at a different angle
But why is accurate classification needed, when the overriding decision is about how the car should react now that an unidentified, moving object is moving directly into the car's path? Even if someone were riding at an angle, it's still in a direction that may intersect with the car's path.
If the autonomous vehicle performed just as well as the human did (meaning a human would've hit her also), then yes although it can be improved, I don't see how it is a problem.
