Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Red light detection, stop sign detection, stopping!

This site may earn commission on affiliate links.
I wonder how much of the new FSD features they will get running on 2.5? Should be interesting for sure. Will be nice to get to try at least some of the new features without waiting for a hardware upgrade.

I just love the fact that they're sorta working, but not totally.

So it gives room for HW3 to come in, and crush it with the higher resolution images along with the much better neural network.
 
I just love the fact that they're sorta working, but not totally.

So it gives room for HW3 to come in, and crush it with the higher resolution images along with the much better neural network.


HW3 will run on higher resolution images? How do we know this? Or think this?

I thought it was just going to be larger/more complex NNs, higher fps, and redundancy. I don't recall anything about higher resolution.
 
Last edited:
currently ap2.5 downsamples the images to quarter resolution before feeding into (some) NNs. full resolution is people speculating they won't do it for hw3 (but they still do nowadays I think)

Thanks for all the great posts and responses that you provide to the community. Greatly appreciated.

Very exciting to see the new functionality starting to take shape and that it is running on HW2/2,5 and hence with the current NNs.

I am not an NN expert so was wondering how the developments of the existing NNs and the new larger NNs that run on HW3 are managed in parallel? Can anyone help me out? Is the new, larger NN purely on the vision (perception) side with a different NN handling the 'driving'?
 
I am not an NN expert so was wondering how the developments of the existing NNs and the new larger NNs that run on HW3 are managed in parallel? Can anyone help me out? Is the new, larger NN purely on the vision (perception) side with a different NN handling the 'driving'?
new larger NN is mostly a speculation at this point, nobody have seen it in the wild.

Currently HW3 cars just get the same NN that's used in HW2.x car compiled to the other architecture.
 
These videos do really encourage me. I know the features are not ready for the public yet but seeing actual video evidence of the car detecting traffic lights and stopping at red lights and stop signs, shows that Tesla is making real progress. It also shows me that these features are not out of reach. Also, the fact that Tesla can do this on AP2.5 makes me feel like AP3 will be really good.
 
new larger NN is mostly a speculation at this point, nobody have seen it in the wild.

Currently HW3 cars just get the same NN that's used in HW2.x car compiled to the other architecture.

You said "there's like a state when it's greyed out. so the car knows it's there (from maps?) but cannot see it."

You should cover your trifocal camera and drive and see if the greyed traffic icon shows up. Then we can know its 100% getting info that there is one in the intersection from the map.
 
You said "there's like a state when it's greyed out. so the car knows it's there (from maps?) but cannot see it."

You should cover your trifocal camera and drive and see if the greyed traffic icon shows up. Then we can know its 100% getting info that there is one in the intersection from the map.
we know it's from the map because it also detect traffic lights that are not yet visible. so don't need to cover anything (that would produce a bunch of warnings too - they now have good blocked camera detection)
 
Hasn't worked for me. 19.8.3 downloaded a few days ago. I have a 2018 M3RWD LR, and bought the "new" AP 2 weeks ago during the $2,000 sale price discount period. My car has AutoSteer, TACC, but isn't technically EAP like the original AP option that was offered last year.

I drove it in clear weather with a full view of traffic lights, and it didn't recognize any lights. If I hadn't taken manual control it would have gone right through them. Also doesn't recognize Stop Signs. Curiously, I also lost the ability to plan music from any of my rear speakers with this Version install. I went to the local TESLA service center and the tech told me that the problem with the audio system "is a known issue", and to expect future Version updates to correct it. He wasn't aware of the lack of Red Light and Stop Sign reading with the AP engaged. "Maybe it's not included with the new AP feature set" was all he could offer.
 
I drove it in clear weather with a full view of traffic lights, and it didn't recognize any lights. If I hadn't taken manual control it would have gone right through them.
if you read the release notes that opened after update you'd learn that it's only supposedly MAYBE will warn you if you are about tu run a red light while on autosteer.

Anything beyond that is out of scope and you need advanced access to your car to enable.
 
Hasn't worked for me. 19.8.3 downloaded a few days ago. I have a 2018 M3RWD LR, and bought the "new" AP 2 weeks ago during the $2,000 sale price discount period. My car has AutoSteer, TACC, but isn't technically EAP like the original AP option that was offered last year.

I drove it in clear weather with a full view of traffic lights, and it didn't recognize any lights. If I hadn't taken manual control it would have gone right through them. Also doesn't recognize Stop Signs. Curiously, I also lost the ability to plan music from any of my rear speakers with this Version install. I went to the local TESLA service center and the tech told me that the problem with the audio system "is a known issue", and to expect future Version updates to correct it. He wasn't aware of the lack of Red Light and Stop Sign reading with the AP engaged. "Maybe it's not included with the new AP feature set" was all he could offer.

Dear lord. The feature set the OP demoed is not available to the public. At the present time, I promise your car will never stop at any red lights or stop signs. You’ll need to take that action.
 
that was just a "Spherical NN description in vacuum". meaning just the description, no weights, no code to interact with it. We don't even know if it actually compiles into anything.

Okay, forgive me I have been out of the loop past few months... but I am wondering what happened to this network described in this post:
#509
This V9 network is a monster, and that’s not the half of it. When you increase the number of parameters (weights) in an NN by a factor of 5 you don’t just get 5 times the capacity and need 5 times as much training data. In terms of expressive capacity increase it’s more akin to a number with 5 times as many digits. So if V8’s expressive capacity was 10, V9’s capacity is more like 100,000. It’s a mind boggling expansion of raw capacity. And likewise the amount of training data doesn’t go up by a mere 5x. It probably takes at least thousands and perhaps millions of times more data to fully utilize a network that has 5x as many parameters.

Or is this now the current/regular network?
 
  • Like
Reactions: APotatoGod
Okay, forgive me I have been out of the loop past few months... but I am wondering what happened to this network described in this post:
#509
This V9 network is a monster, and that’s not the half of it. When you increase the number of parameters (weights) in an NN by a factor of 5 you don’t just get 5 times the capacity and need 5 times as much training data. In terms of expressive capacity increase it’s more akin to a number with 5 times as many digits. So if V8’s expressive capacity was 10, V9’s capacity is more like 100,000. It’s a mind boggling expansion of raw capacity. And likewise the amount of training data doesn’t go up by a mere 5x. It probably takes at least thousands and perhaps millions of times more data to fully utilize a network that has 5x as many parameters.

Or is this now the current/regular network?
that network was never deployed anywhere where you can see it. It was just a descriptor file dropped into (forgotten in?) firmware. The actual deployed NN was different though.
 
  • Informative
Reactions: APotatoGod