@EVNow told you! Shouldn’t have contradicted the head of engineering, finance, logistics and manufacturing of Tesla!
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
zoomer0056
Active Member
Says "won't include new cameras or updated placements." To me this means all the same cameras in the same places. But I can see how this can be interpreted like you say, and this makes more sense.
I thought I read someone's analysis of the new FSD HW4 computer and the formfactor changed, meaning it won't fit in HW2/3 cars. I don't know if that's true, and we'll have to wait until it's actually released to find out.
stopcrazypp
Well-Known Member
Yes, talking about the rearward looking ones on the front fender.
Maybe the Semi uses all 12 camera connections for now?
- F-SVC (bumper front)
- Wide (fisheye forward)
- Main (forward)
- Spare (narrow forward)
- L-SVC (left mirror facing down)
- R-SVC (right mirror facing down)
- L-FF-Side (left mirror facing left)
- R-FF-Side (right mirror facing right)
- L-FF-Rear (left mirror facing rear)
- R-FF-Rear (right mirror facing rear)
- Selfie (interior)
- Backup (reverse / trailer)
Daniel in SD
(supervised)
DanCar
Active Member
When I've looked at prototype Semis, I see two front camera housings. One near the bottom of the windshield and one at the top. The one at the bottom is approximately at the same height as existing Tesla vehicles.
scottf200
Well-Known Member
Yes, the prototypes had 2 sets of forward triple-cameras at the top and bottom of the windshield, and they also had small triangular "fins" above and behind the doors probably before these cameras were consolidated to the sideview mirrors after Tesla evaluated which camera positions were necessary and/or acceptable.
I couldn't find a great picture of how many forward cameras, so I assumed they were like the prototype with 3 and slotted the "unnecessary" narrow/far camera into "Spare." So yes, it could very well be 2 cameras dropping the count to 11 leaving the spare port unused. This is the best view of the delivered Semis showing off the forward cameras, and it does seem like it could be 2:
EVNow
Well-Known Member
I posted this picture in a different thread for X delivered with HW4 where the service manual confirmed unused purple and black ports:
It seems like FSD Beta will need to be supporting multiple vehicle sizes (like HW3 for X vs 3) with slightly different camera positions, multiple camera attributes (HW4 higher resolution and RGGB(?) sensors) as well as above's multiple exterior camera setups (8x with HW3, 7x with current HW4, potentially up to 12x for Semi and Cybertruck).
Potentially to address all this complexity, this might have been what Ashok Elluswamy was hinting at during Tesla AI Day 2022 when he talked about a "weekend project" for Radiance Fields as Vision's Foundational Model. Maybe instead of training the neural networks to predict things from video coming from certain fixed camera positions, an internal representation could be formed from the numerous camera setups, so that Tesla doesn't need to "retrain from scratch" for say the Cybertruck's additional higher resolution cameras in different positions.
This would be quite the architectural change probably worthy of a FSD Beta 12 version bump. Tesla is probably doing something custom for Optimus with the much smaller form factor, but theoretically this unification could bring aligned neural network prediction improvements to vehicle robots, humanoid robots, and probably many other potential vision-based robots too. For example, learning to accurately predict shapes of random objects in a garage is probably useful for Optimus to pick it up or for a vehicle to avoid.
Purple: F-SVC & Spare (2x unused)
Black: L-SVC & R-SVC (2x unused)
White: R-FF-Rear & L-FF-Rear (repeaters)
Blue: Selfie & Wide (dummy in Far housing)
Black: Main & L-FF-Side & Backup & R-FF-Side (pillars)
It seems like FSD Beta will need to be supporting multiple vehicle sizes (like HW3 for X vs 3) with slightly different camera positions, multiple camera attributes (HW4 higher resolution and RGGB(?) sensors) as well as above's multiple exterior camera setups (8x with HW3, 7x with current HW4, potentially up to 12x for Semi and Cybertruck).
Potentially to address all this complexity, this might have been what Ashok Elluswamy was hinting at during Tesla AI Day 2022 when he talked about a "weekend project" for Radiance Fields as Vision's Foundational Model. Maybe instead of training the neural networks to predict things from video coming from certain fixed camera positions, an internal representation could be formed from the numerous camera setups, so that Tesla doesn't need to "retrain from scratch" for say the Cybertruck's additional higher resolution cameras in different positions.
This would be quite the architectural change probably worthy of a FSD Beta 12 version bump. Tesla is probably doing something custom for Optimus with the much smaller form factor, but theoretically this unification could bring aligned neural network prediction improvements to vehicle robots, humanoid robots, and probably many other potential vision-based robots too. For example, learning to accurately predict shapes of random objects in a garage is probably useful for Optimus to pick it up or for a vehicle to avoid.
Similar threads
