BobbyKings
S85D > M3P > S Plaid
Welcome.
Tesla is focusing on learning how to learn.
You're focusing on individual cases.
All the best for you.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla is focusing on learning how to learn.
You're focusing on individual cases.
All the best for you.
It just needs more cameras. I think they can put two more on the front side cameras facing out and then maybe it needs two more in the back facing sideways. Not to hard to retrofit into existing cars.
jabloomf1230
Minister of Silly Walks
You'll probably get the same response that this suggestion has generated before:
"If you put two cameras on the nose of the car (one facing left and one facing right) they will immediately get blocked with mud, ice and snow."
Somehow, I have faith that Tesla could design such a unit which would stay mostly free of debris. The cameras would just need to be embedded behind a shroud that faces forward. The biggest issue would be that such a unit would increase the car's drag coefficient. Besides, such a unit would not be a primary source of sensor data and would only be useful for unprotected left turns. ULTs are FSD's Achilles heel. I just don't see Tesla ever being able to safely make 100% of ULTs using the present sensor mix.
if they need to retrofit cameras in then they’re screwed with current cars. It would absolutely not be easy to retrofit that stuff in unless they happened to put in wires already in anticipation of that, which is doubtful. Even if the wire harnesses are already installed and waiting, cutting Holes in the bumpers or replacing them entirely with new ones that have holes for the cameras in would not be trivial.
linux-works
Active Member
and what data lines do these new video sources take?
(I dont know, actually; does tesla use gig-e automotive or just regular broad-r-reach eth? if gig-e, then there could be enough bw on existing lines, but that does not mean other systems can use/consume it).
unless they planned for it, you cant just slam in more high bw sources. this isn't a pc. but I'm also pretty sure you were not serious, either
jabloomf1230
Minister of Silly Walks
Chazman92
Member
#FSDBeta Unprotected Left Turn with Distance Measurements for Range Detection
FSDBeta 9.1 - 2021.4.18.13
FSDBeta 9.1 - 2021.4.18.13
Nothing like that is easy. But I do think that there is some evidence that existing cameras struggle. Vision only needs good, dependable vision! Existing cameras already compromised by condensation and blinding in some situations and environments so no point adding dependence on more potentially compromised feeds, but also removal of radar demonstrates that additional types of sensor input are not without issues either.
You said in the video you weren't sure why it aborts with a right turn sometimes, and my guess is that the left path is too low confidence because the neutral network keeps thinking there's no road to the left due to the grassy median. If you do want to make more FSD Beta 9.x unprotected left turn videos, it might be more interesting to find a variety of left turns onto Roosevelt to get a better sense of what aspect is actually failing.
powertoold
Active Member
At this point, I think Tesla is fully aware of the unprotected left issues. I think it's been improving with every update, but it's obviously not there yet, and more risky testing of difficult unprotected lefts is probably not worth it. I guess if people are comfortable with testing them, then that's fine, but personally, I tend to avoid those types of unprotected lefts when I drive.
powertoold
Active Member
V9 does very well in some areas more complex than Chandler, AZ, except for the disengagement at the end :
:
Thanks, great video.
Seems car identifies other cars around 50 meters out, that is 2 seconds at 60 mph/26 m/s or way to little to safely know when to drive. Chuck though has 150 meter of vision, meaning 5-6 seconds to act.
Sad, Tesla has really painted themselves up in the corner with all those promises the last 5 years.
Going to retrofit a million or more cars with front bumper radars or higher resolution cameras will be very expensive.
Does it matter if it is 2021 or 2022 or 2023? When it turns on they can all do it (2017+). It won’t change how many cars are sold between now and whenever it gets turned on or how many cars are Robotaxi capable. All that matters is FSD progress and maximum production. Both are happening.
Not in the bumpers, where the side backwards facing cameras are. You could even design a module that can shift between cameras based on scenerio.
I think they certainly could have designed for it, should have ect, big even if they didn’t there are ways around the issue. Smart switching of the camera sensors, for example.and what data lines do these new video sources take?
(I dont know, actually; does tesla use gig-e automotive or just regular broad-r-reach eth? if gig-e, then there could be enough bw on existing lines, but that does not mean other systems can use/consume it).
unless they planned for it, you cant just slam in more high bw sources. this isn't a pc. but I'm also pretty sure you were not serious, either
powertoold
Active Member
No offense to any of the current fsd beta testers, but my thought is Tesla likely doesn’t need their input for improving the system.
I think the main purpose of having the fsd beta testers was to test the regulatory waters and experiment to see if non-affiliated people can use fsd beta without serious consequences. Most of what I’ve seen reported to Tesla by the current testers are obvious issues. As Elon said, Tesla has spent an insane amount of time on bike lanes alone, so it’s not like Tesla is surprised by the car not signaling when merging or not seeing cross traffic at 50mph+.
The whole fsd beta tester “experiment” is a great and risky idea IMO. The testers generate hype, we all get to be “in on it,” Tesla gets to see how the media / regulatory bodies react, and we all get to see the progress they make (aka witness history being made, kinda like with Starship).
I think the main purpose of having the fsd beta testers was to test the regulatory waters and experiment to see if non-affiliated people can use fsd beta without serious consequences. Most of what I’ve seen reported to Tesla by the current testers are obvious issues. As Elon said, Tesla has spent an insane amount of time on bike lanes alone, so it’s not like Tesla is surprised by the car not signaling when merging or not seeing cross traffic at 50mph+.
The whole fsd beta tester “experiment” is a great and risky idea IMO. The testers generate hype, we all get to be “in on it,” Tesla gets to see how the media / regulatory bodies react, and we all get to see the progress they make (aka witness history being made, kinda like with Starship).
linux-works
Active Member
car companies are the cheapest. well, one of the cheapest. they save 5 cents if they can.
this has a ripple effect. if you dont plan for more camera video sources, there clearly wont be extra headroom left in the design since that costs.
I've had cpus put into designs where we argued for more headroom and they ended up cutting us. they saved a few dollars.
this is typical.
so, likely, whatever the system is balanced for, that's pretty much it until next pcb and cpu rev time.
not to mention the harness.
linux-works
Active Member
jury is still out on the reliability of wifi (if that's what you mean) for high bandwidth mission critical sensors.
personally, I would not ask for this, on my design. I would not want to trust the unreliability of wireless for this.
in fact, I keep arguing for fiber, not copper. but I always get shot down. (but I keep trying, one vendor will take the idea, I'm sure..)
Similar threads
