10.69.3 is the beginning of the end for LIDAR approaches!
Lidar approaches is having a more powerful compute, high res cameras (8+ mp) rather than a low res (1.2 mp) camera, having more cameras to eliminate blindspot, adding other sensors like high res radar and lidar to improve accuracy...that lidar approach?
1. Try the difficult, but generalizable solution first, even if common knowledge tells you it won't be possible.
2. Failing that, try the second most generalizable solution.
3. Etc.
It does lead to failures and setbacks more often than not, but sometimes it means Tesla accomplishes what others thought was impossible.
What exactly was impossible? Please don't say "vision only" because the very first vision only was Mobileye in 2018 (no radar, lidar or ultrasonics).
Tesla deploys a generalized fsd software to thousands of lay people. Extremely extremely difficult. Karpathy says as much. No one else is doing this. It took them a whole year to refine the occupancy network enough to deploy it.
There's not enough respect for deploying software in the wild, esp something radical like fsd beta.
You can praise Tesla without turning it into a series of statements that is demonstrably false.
As there are others that ARE doing this. Namely
Huawei and
Xpeng and others that will join them soon.
Why do you have to turn everything to "Tesla is the greatest"?
Nothing Tesla is doing is unique.
Bird's Eye View is industry standard. Occupancy Network is industry stand. Even Xpeng who isn't even a player in AV, you can see their transformer based BEV network here called XNet that uses multi-video input from all cameras to do static recognition, dynamic recognition and motion prediction.
Transformers are the most overused architecture in ML today, I saw a stat saying something to the effect that almost 90-95% of projects today are based it.
I don't understand where you see this uniqueness or greatest in Tesla.