In order to make vision-only depth sensing work well, you need ground-truth data. Buying a few thousand lidar to get a ton of ground-truth training data for the NN to learn distance on its own from vision is an obvious move. They're not switching to Lidar.
source: i debugged lidar and built localization and behavioral planning algorithms for self driving cars
Yes, 10 years after the initial announcement, Tesla has just now got around to ordering some LIDARs for capturing ground-truth training data. They then went out of their way to get the one LIDAR made to look palatable to end customers instead of the superior industrial ones everyone puts 8 of on their recording cars.
If you make millions of them per year, lidar won't cost $1k per car. OTOH existing lidar designs are not designed for mass market manufacturing or serviceability.
If you had multiple cameras for stereoscopic vision. Couldn't you figure out the depth data vision only without lidar? I'm sure there was a good reason Tesla didn't go with a stereoscopic camera system (at least front facing) from the get-go.. they already have 3 cameras up there with 3 FOVs. Could add another camera there in a stereo setup for depth data.
My reading is this is only for their own taxi fleet. $2M at $1000 per car = 2000 cars.
Waymo has fully autonomous self driving that actually works for their taxi fleet. Waymo uses lidar. Tesla building a taxi fleet, Tesla now needs to spring the extra dosh for a bit of lidar.
That must clearly be true, but I'd like cars to driver better than humans and make use of all the tech that we can't. Academically vision only is an interesting feat, and as a backup mode for when sensors fail it makes sense to pursue vision only operation, but it seems to me that having more sensors should be superior.
That was always a stupid argument because if you replicate how humans do it, you're going to replicate our failure modes too. Things like being crappy at night driving, subject to optical illusions, etc.
I can assure you that a human with LIDAR and RADAR built in would get around even better than without.
As far as I know Musk and Tesla's position on this hasn't changed. They use Lidar for validation during R&D, they don't ship it on production vehicles.
It's an interesting argument, I'm sure there are lots of reasons why it's a bad argument, but the one I think about is that depth perception isn't actually fully understood, mostly, but not fully, and, an important system that is always overlooked in that argument is that the vestibular system (among others) is also involved in depth perception. You can say binocular vision is just LIDAR, but that's not the whole story.
source: i debugged lidar and built localization and behavioral planning algorithms for self driving cars