4 ms·

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 t

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.

samcampbell 9y ago
Agreed! This seems like an instance where even if the car knew it needed to stop, it couldn't do so in time.
- Maybestring 9y ago
  Every millisecond of braking reduces the severity of the collision.
danso 9y ago
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?
- Taniwha 9y ago
  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
  - nathanielc 9y ago
    If either of those two cases is true, then the AI is not ready to be driving on public roads.
  - danso 9y ago
    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.
    - taneq 9y ago
      Exactly, it shouldn't matter what the thing is, if it's in front of my car and bigger than a plastic shopping bag I don't want to hit it.
- BurningFrog 9y ago
  The most relevant part is from someone with zero knowledge of what happened?
  - danso 9y ago
    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.
branchan 9y ago
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.
- alistproducer2 9y ago
  If they are just robot versions of humans then I would argue they are far too underdeveloped to be on real roads with real people.
vha32oiwe 9y ago
How can a computer know to handle something when it is relying on us to supply models we don't have answers for? These systems are fancy statistical sorting engines, looking up the most relevant next steps provided BY US. They cannot subvert the laws of physics with their actions. We don't have an answer for this at all because humans just swerve or hit the brakes, regardless of the outcome (often creating an outcome that's worse). What else is there given the constraints of physics? Time and space aren't not mutable, no matter how big the data set we apply.

samcampbell 9y ago

Agreed! This seems like an instance where even if the car knew it needed to stop, it couldn't do so in time.

Maybestring 9y ago
Every millisecond of braking reduces the severity of the collision.

danso 9y ago

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?

Taniwha 9y ago
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
- nathanielc 9y ago
  If either of those two cases is true, then the AI is not ready to be driving on public roads.
- danso 9y ago
  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.
  - taneq 9y ago
    Exactly, it shouldn't matter what the thing is, if it's in front of my car and bigger than a plastic shopping bag I don't want to hit it.
BurningFrog 9y ago
The most relevant part is from someone with zero knowledge of what happened?
- danso 9y ago
  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.

branchan 9y ago

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.

alistproducer2 9y ago
If they are just robot versions of humans then I would argue they are far too underdeveloped to be on real roads with real people.

vha32oiwe 9y ago

How can a computer know to handle something when it is relying on us to supply models we don't have answers for? These systems are fancy statistical sorting engines, looking up the most relevant next steps provided BY US. They cannot subvert the laws of physics with their actions. We don't have an answer for this at all because humans just swerve or hit the brakes, regardless of the outcome (often creating an outcome that's worse). What else is there given the constraints of physics? Time and space aren't not mutable, no matter how big the data set we apply.