3 ms·

Are they still all-in on “pure vision” self-driving? I thought they had pivoted back to lidar but can’t seem to find any sources for that.

masklinn 3y ago
The “tesla vision update” faq entry was last updated on October 5th, 2023, and as far as the way back machine surfaces the only changes were removing park assist from the list of disabled features, a few wording changes, and removing some of the more optimistic wording (e.g. “in the near future” around mentions of feature parity). So it seems like they’re still all in on pure vision. I don’t think the xruck has lidar either, only hits I get talk about a lidar-equipped prototype in early 2023.
loueed 3y ago
Pure Vision currently. The last we heard about LiDAR was the issue of combining Lidar Data with Vision in the NN. Which technology should take priority in the NN? For example LiDAR might recognise the shape of a sign but vision will know if its a stop sign or not. I'd be super interested to see if someone has successfully combined RGB data with LIDAR data.
- Mashimo 3y ago
  What does NN stand for?
  - actionfromafar 3y ago
    Neural network
  - hprotagonist 3y ago
    neural network, i would imagine.
  - wkat4242 3y ago
    Neural network i guess
- RugnirViking 3y ago
  > combining Lidar Data with Vision in the NN This is basic university robotics. Sensor fusion ! There are a whole host of techniques for dynamically updating confidence between two or more sensors estimating the same values. Kalman filter being the standard approach (which for reference was used in the apollo missions 50 years ago - developed a lot since then) And yes, it is commonly applied with vision models. There are a host of combined rgb/lidar, structured light, depth camera and more setups in the labs students are working on at my local university, and have been for at least 6 years
  - mrguyorama 3y ago
    For reference, I, a computer science undergrad, learned this kind of sensor fusion theory in an elective class that was just an excuse for a soon to retire professor to play with lego robots You can know more about sensor fusion than Elon does by reading a literal blog post. The NN shouldn't have to "choose" one or the other. It's a classic "Not even wrong" question!
mathgeek 3y ago
Latest update I am aware of is https://www.tesla.com/support/transitioning-tesla-vision https://www.tesla.com/support/transitioning-tesla-vision
cbeach 3y ago
They've never used LIDAR. Tesla have always been all-in on a vision based system, with Elon stating that LIDAR is a "crutch," and a "fools errand" > Andrej Karparthy, Senior Director of AI, took the stage and explained that the world is built for visual recognition. Lidar systems, he said, have a hard time deciphering between a plastic bag and a rubber tire. Large scale neural network training and visual recognition are necessary for Level 4 and Level 5 autonomy, he said. > “In that sense, lidar is really a shortcut,” Karparthy said. “It sidesteps the fundamental problems, the important problem of visual recognition, that is necessary for autonomy. It gives a false sense of progress, and is ultimately a crutch. It does give, like, really fast demos!” https://techcrunch.com/2019/04/22/anyone-relying-on-lidar-is-doomed-elon-musk-says/ https://techcrunch.com/2019/04/22/anyone-relying-on-lidar-is...
- hprotagonist 3y ago
  radar and ultrasonics were removed. https://www.forbes.com/sites/bradtempleton/2022/10/31/former-head-of-tesla-ai-explains-why-theyve-removed-sensors-others-differ/ https://www.forbes.com/sites/bradtempleton/2022/10/31/former...
- infotogivenm 3y ago
  Ah good point, radar/ultrasonics were what I was thinking of, not lidar. Looks like they’re still gone.
laserbeam 3y ago
I'm fairly sure it's: "cameras are cheaper and cars with cameras don't have an ugly thing on top" argument. So yeah, pure vision.
kbos87 3y ago
The most interesting part of their pivot to vision is how none of the disaster scenarios people were so sure of seem to have come true. Meanwhile FSD is better than ever and there are millions of miles per month being logged.
- infotogivenm 3y ago
  I mean, the current requirement for human attentiveness and intervention probably has something to do with avoiding disaster scenarios.
numpad0 3y ago
There was a suspicious news article on a Chinese supplier that specialize in non-rotating LIDAR is investing in a Tesla parts factory in Latin America, which was then suspiciously corrected.

masklinn 3y ago

The “tesla vision update” faq entry was last updated on October 5th, 2023, and as far as the way back machine surfaces the only changes were removing park assist from the list of disabled features, a few wording changes, and removing some of the more optimistic wording (e.g. “in the near future” around mentions of feature parity). So it seems like they’re still all in on pure vision. I don’t think the xruck has lidar either, only hits I get talk about a lidar-equipped prototype in early 2023.

loueed 3y ago

Pure Vision currently. The last we heard about LiDAR was the issue of combining Lidar Data with Vision in the NN. Which technology should take priority in the NN? For example LiDAR might recognise the shape of a sign but vision will know if its a stop sign or not. I'd be super interested to see if someone has successfully combined RGB data with LIDAR data.

Mashimo 3y ago
What does NN stand for?
- actionfromafar 3y ago
  Neural network
- hprotagonist 3y ago
  neural network, i would imagine.
- wkat4242 3y ago
  Neural network i guess
RugnirViking 3y ago
> combining Lidar Data with Vision in the NN This is basic university robotics. Sensor fusion ! There are a whole host of techniques for dynamically updating confidence between two or more sensors estimating the same values. Kalman filter being the standard approach (which for reference was used in the apollo missions 50 years ago - developed a lot since then) And yes, it is commonly applied with vision models. There are a host of combined rgb/lidar, structured light, depth camera and more setups in the labs students are working on at my local university, and have been for at least 6 years
- mrguyorama 3y ago
  For reference, I, a computer science undergrad, learned this kind of sensor fusion theory in an elective class that was just an excuse for a soon to retire professor to play with lego robots You can know more about sensor fusion than Elon does by reading a literal blog post. The NN shouldn't have to "choose" one or the other. It's a classic "Not even wrong" question!

mathgeek 3y ago

Latest update I am aware of is https://www.tesla.com/support/transitioning-tesla-vision https://www.tesla.com/support/transitioning-tesla-vision

cbeach 3y ago

They've never used LIDAR. Tesla have always been all-in on a vision based system, with Elon stating that LIDAR is a "crutch," and a "fools errand" > Andrej Karparthy, Senior Director of AI, took the stage and explained that the world is built for visual recognition. Lidar systems, he said, have a hard time deciphering between a plastic bag and a rubber tire. Large scale neural network training and visual recognition are necessary for Level 4 and Level 5 autonomy, he said. > “In that sense, lidar is really a shortcut,” Karparthy said. “It sidesteps the fundamental problems, the important problem of visual recognition, that is necessary for autonomy. It gives a false sense of progress, and is ultimately a crutch. It does give, like, really fast demos!” https://techcrunch.com/2019/04/22/anyone-relying-on-lidar-is-doomed-elon-musk-says/ https://techcrunch.com/2019/04/22/anyone-relying-on-lidar-is...

hprotagonist 3y ago
radar and ultrasonics were removed. https://www.forbes.com/sites/bradtempleton/2022/10/31/former-head-of-tesla-ai-explains-why-theyve-removed-sensors-others-differ/ https://www.forbes.com/sites/bradtempleton/2022/10/31/former...
infotogivenm 3y ago
Ah good point, radar/ultrasonics were what I was thinking of, not lidar. Looks like they’re still gone.

laserbeam 3y ago

I'm fairly sure it's: "cameras are cheaper and cars with cameras don't have an ugly thing on top" argument. So yeah, pure vision.

kbos87 3y ago

The most interesting part of their pivot to vision is how none of the disaster scenarios people were so sure of seem to have come true. Meanwhile FSD is better than ever and there are millions of miles per month being logged.

infotogivenm 3y ago
I mean, the current requirement for human attentiveness and intervention probably has something to do with avoiding disaster scenarios.

numpad0 3y ago

There was a suspicious news article on a Chinese supplier that specialize in non-rotating LIDAR is investing in a Tesla parts factory in Latin America, which was then suspiciously corrected.