I updated my post with screenshots earlier into the video when the curb on the left is still visible and immediately after it becomes not visible. The Camry is much further away. That seems to confirm the curb would be invisible to this camera as you are making a right turn (or left turn from one way).
The Camry isn't that far from the curb! Anyway you need to be able to detect the positions of object this close to the car otherwise you're going to have problems.
I mean the Camry is far away from the Tesla (not that it is far away from the curb). That means the curb becomes invisible to this camera well before the Tesla crosses into the intersection, so if it was making a turn it would also be invisible.
What is the Tesla seeing with the yellow ultrasonic lines? Is that not the curb it's seeing? Or is it the grass?
How many feet in front of the car does the camera suite need to be able to see the curb in your estimation?
If something like a curb is obscured by objects (cars etc) then I'm not sure what any sensor is supposed to do ... they are all going to be blinded. At that point you are relying on the NN being able to extrapolate the curb from the bits it can see (which it can do) and/or handle the object continuity as the curb becomes obscured (which Tesla are working on for cars, and I would assume will also apply to other objects).
groundtruthautonomy.com
Because Chandler is not a major city. They're just playing word games. Kyle Vogt is even worse. In the interview Phil LeBeau says it's "definitely" the first time in the US anyone has done paid rides with no driver and Kyle just rolls with it. Never mind Waymo who has done it for two years. Kyle later says "no major city has had driverless cars running around" even though Waymo is doing driverless in downtown Phoenix (just not paid rides). And Yandex has done driveless in Las Vegas, also a major city. Of course it's not Kyle's job to do Waymo's marketing for them, but that doesn't mean he has to lie.
Ok so I realized that measuring this is actually quite simple since we know the view is masked by the shroud. I used a flashlight and positioned on the ground it as close to possible to the car while still illuminating the fisheye lens. I got a range of about 7 feet directly in front of the car and 8 feet a bit to the side (as a curb would be). It can actually see a foot or two closer than I do sitting in the driver's seat.
I haven't seen this case in any FSD Beta videos where people hit curbs. Does it really follow cars that closely? Seems too close.
I'm not sure I'm understanding correctly, but 8 ft to the side is pretty far way from the curb to be making a tight right/left turn. That would guarantee that while the car is doing the turn, that camera can't see the curb unless you turn really wide.
Does that show up the whole way? I noticed your curbs are angled upward (not 90 degree) and that might help expand the range a bit, because it may end up still reflecting some signal from the sensors back up even though it is low.
No I’m saying it can see the position of the curb better than someone sitting in the driver‘s seat, about 8 feet ahead and 2 feet to the side. Software is already far superior than humans at doing the “extrapolating” you’re talking about. The car knows the exact wheel speed and steering position so it could easily navigate ten feet of travel by dead reckoning, no AI involved. But it also has all the cameras that it can use for localization and an accelerometer.
Yep, he might not even have to get out in all cases. Could just pop the button, poke his head out, give the thumbs up, then close things back up again. Might be able to latch onto that hood somehow to pull it closed from the inside. Probably will be some aftermarket kits.
Frunk Optimus also solves Chuck’s left turn without the need for a camera retrofit.
He/she/it just walks out into the road .. cars stop in amazement .. Chuck turns left.
If you do need a camera upgrade, Frunk Optimus will install them for you.
Except it's not in this case, given Teslas are curbing wheels when in FSD mode. Not saying it's not impossible to fix this in software, but having parking cameras are a trivial way to address it.
It does have an accelerometer, but that is not enough for dead reckoning, it would also need a gyroscope. The accuracy matters too, given Tesla seems to tune to have a relatively small margin of error (not leaving a whole lot of space between wheels and curb, instead of turning wide). I'm skeptical an IMU would be able to beat a simple parking camera in this particular application (especially when your initial guesstimate of curb position may be wrong in the first place, and you get zero updates as you are making the turn given the cameras are completely blind to it).
It certainly helps for automated parking, which a tight right/left turn pretty much is the same thing. That's why parking cameras are implemented this way. If what you say was true, there would be no need for parking cameras.
Or the perception NN can’t tell the position of a curb 10 feet away or a bollard directly in front of the car. Because the state of the art in camera based computer vision isn’t sufficient for self driving and nobody knows when it will be.
