3 ms·
This genre of posts is so tiresome. As the second video in the thread demonstrates, the truck is literally hauling traffic lights. The AI recognition is corre
by ajross 5y ago
This genre of posts is so tiresome.
As the second video in the thread demonstrates, the truck is literally hauling traffic lights. The AI recognition is correct, the only thing worth complaining about is that they're displayed as static objects for the user after recognition, just to be re-recognized a few seconds later in a different place. Note that the car is correctly not detecting they are lit, so not inferring direction (though AP isn't engaged, so I guess we'll never know what it would have done).
No doubt you could play the same game by putting a traffic cone on your bike. The car wants to see important traffic objects, it's literally what it's trained for.
- theamk 5y agoAnd that illustrates the important problem with self-driving car: if you want L5 autonomy, you need to be handle all the weird cases.
- ajross 5y agoThe car handled this just fine, though. (Though obviously AP wasn't engaged so we don't know exactly what it would have done). The bug report here is that the correctly-detected signs, in the correctly-detected state (i.e. off) were presented to the user in a confusing way in the UI visualization. UI bugs aren't blockers for mythical SAE Level 5 Autonomy. You just... fix them.
- theamk 5y agoI am not sure why you think it was UI blocker and not actual problem affecting the driving? In many places, the road rules say that if you see a broken traffic light, you should treat it as a stop sign. So a theoretical L5 car should have stopped there, in the middle of the road -- and again, and again. Seems pretty bad to me. And yes, I agree that you "just.. fix them" -- but of the difficulties of the real road driving is that the number of unique situations like that is very large, and many of them would not happen during test drives / development.
- ajross 5y ago> I am not sure why you think it was UI blocker and not actual problem affecting the driving? I think it's not an actual problem affecting the driving because the driving was unaffected. Autopilot wasn't engaged, we don't have evidence. I'm just pointing out that the autopilot doesn't drive by reference to the dash UI (which is showing a pretty obvious visualization -- the UI understands these to be static objects so once it gets one it "animates" it as if the car was passing it), so bugs there aren't very informative as to its behavior.