Tesla.com - "Transitioning to Tesla Vision"

[stopcrazypp]

I'm speaking as an owner, but also as someone technical enough in the field to know progress from BS. I do want to believe they have progress, but its a higher bar to actually SHOW it.

I think enough, here, are technical enough to appreciate a dual-stack drive thru demo series.

tesla DOES have to prove this; if we are to trust our lives, YES, they DO have to prove it and prove it and keep proving it for a few years before that seemingly asymptotic level 5 can be reached.

But you are talking about comparing radar to vision only. They don't really need to prove anything on that, just like none of the L2 ADAS systems out there really have any rigorous comparison tests being done (some of them use vision only also, while others use radar or a combination). People still happily buy them and use them.

When Tesla wants to move to L4/L5, sure, they will have to prove safety to the public (and perhaps regulators), but it'll mainly be based on the stats and incident numbers. It won't be based on an A/B test or a "dual-stack" drive. Vision-only being better or worse than radar in such a test doesn't really tell you if the system is safe enough for L4/L5.

 

[Rattletrap]

With the latest firmware, are "Vision-only" 3s and Ys still gimped to 80 MPH during basic AP? I'm still on 2021.4.18.10 and don't see a reason to update until they allow 90 MPH AP. With the way the traffic moves out here, the feature is useless on freeways until they remove the nerf.

 

[emmz0r]

people of science trust numbers, not word of mouth.

and again, shadow mode is BS; what I am describing is true parallel mode. one codebase in one cpu and one in another. dual array of sensors so that you truly have 2 cars on 1 chassis (computer wise). you run them both and you log data to both.

THAT is what I'm talking about. until we see that, there is zero proof to the user community. they should be proud to show us proof if this is actually real.

What's the point of that? If you run 2 sensors in the same environment, you A/B their outputs. Simple as.
It's not necessary to make things harder than it needs to be on purpose.

 

[linux-works]

you run both VERSIONS concurrently. if you can y-cable things (so to speak; not always easy for many reasons) then fine; but usually, you just dupe everything on a mule and run that and log the results that they both get. the only challenge is: if you were to give control, which one would it be. still, having a consistent course and test setup (which is what we're talking about) it would let you switch to system-A as the driver and B as the passive, then switch back and redo.

that is how you KNOW that you are making progress.

I've done this kind of thing before and its how I feel confident that the new code or new hw is better. test-driven approaches are a must, here. and repeatable test sequences and the results of the A and the B will show when false braking is finally fixed.

at least show us that. people are conjecturing that vision will help fake-braking. well, show us with actual proof.

for those that will buy, no matter what, I'm not talking to you. go have a juice-box; the adults are talking, here.

 

[emmz0r]

you run both VERSIONS concurrently. if you can y-cable things (so to speak; not always easy for many reasons) then fine; but usually, you just dupe everything on a mule and run that and log the results that they both get. the only challenge is: if you were to give control, which one would it be. still, having a consistent course and test setup (which is what we're talking about) it would let you switch to system-A as the driver and B as the passive, then switch back and redo.

that is how you KNOW that you are making progress.

I've done this kind of thing before and its how I feel confident that the new code or new hw is better. test-driven approaches are a must, here. and repeatable test sequences and the results of the A and the B will show when false braking is finally fixed.

at least show us that. people are conjecturing that vision will help fake-braking. well, show us with actual proof.

for those that will buy, no matter what, I'm not talking to you. go have a juice-box; the adults are talking, here.

I agree that they don't have proof that it's actually better. Because pure vision can react on totally different things.
And I start to see you point in running them separately 100%. But then they would need to notify the users. Or at least split the software releases.

 

[linux-works]

let me be more clear; this isn't instrumenting the user cars. these are mules; test chassis that are equipped with dual systems for testing.

its how I would do a before/after test on each nightly. each new version would run the challenge tests that used to fool the radar systems. repeat until all the false-triggers no longer trigger. or at least show that you are just making parity.

but 'trust us, lack of one sensor system is actually better' - that's an extraordinary claim and it has a higher burden of proof before anyone should believe it.

for me, its a real actual issue; I'm stuck on pre-xmas-2020 version that I know works, I know its foibles and until I'm convinced the new system is better, I'm holding the updates back on my car. it will take real proof before I voluntarily press THAT button.

people talk about The Button(tm). for me, any upgrade is a dangerous thing right now, since I'm not at all convinced I'm getting BETTER software. not at all convinced in that, and in fact, I'm quite convinced that all versions since mine have gone downhill. for the things I care about, like, uhm, the level2 features.

disney can blow goats; I care nothing about such things. give me more safety when I drive. keep your farts to yourself, elon.

 

[S4WRXTTCS]

With the latest firmware, are "Vision-only" 3s and Ys still gimped to 80 MPH during basic AP? I'm still on 2021.4.18.10 and don't see a reason to update until they allow 90 MPH AP. With the way the traffic moves out here, the feature is useless on freeways until they remove the nerf.

To explain to people not aware of this Oregon place.

It's a place where the speed limit doesn't really exist, but they love putting speed limit signs up. Like if you drive on I5 in WA state people are mostly going 65-70mph in a 60mph zone, but the second you hit Oregon all bets are off on what speed people will go. Maybe 60mph in a 50mph, or maybe 70.

They'll change the speed limit for no reason like from 55mph to 60mph, and then back to 55mph. To the uninitiated you might think you're supposed to slow down to when going from 60mph to 55mph, but nope just leave it.

Now you'd think its just people driving fast to get to place faster, but that's totally completely wrong as there is always a traffic jam ahead so its really just speeding between traffic jams for no particularly good reason.

We can't actually self driving cars because of Oregon.

The AI would simply turn on us, and run us over for not being logical.

 

[emmz0r]

you run both VERSIONS concurrently. if you can y-cable things (so to speak; not always easy for many reasons) then fine; but usually, you just dupe everything on a mule and run that and log the results that they both get. the only challenge is: if you were to give control, which one would it be. still, having a consistent course and test setup (which is what we're talking about) it would let you switch to system-A as the driver and B as the passive, then switch back and redo.

that is how you KNOW that you are making progress.

I've done this kind of thing before and its how I feel confident that the new code or new hw is better. test-driven approaches are a must, here. and repeatable test sequences and the results of the A and the B will show when false braking is finally fixed.

at least show us that. people are conjecturing that vision will help fake-braking. well, show us with actual proof.

for those that will buy, no matter what, I'm not talking to you. go have a juice-box; the adults are talking, here.

That is what they are doing for one specific thing: The radar. It provides the same input Tesla Vision is providing, so they are compared on the spot. No separate setup needed.
In fact this is way better since it runs within the same code loop so it can be compared more directly and see where they disagree.
The radar was the "ground truth" for distance measurement.

Tesla cars have been seen using a full LIDAR setup, probably to compare there also.

 

[thewishmaster]

Have over 100 miles of NoA use on my new vision-only 3 and so far I’m a believer… Controls acceleration/speed smoothly, slows down and accelerates predictably based on what’s visible in front. No weird lag when a gap opens up in front or when changing lanes and AP needs to accelerate to the set speed. Haven’t encountered any tricky situations yet but outside of typical NoA logic failures it seems quite solid and an improvement in the light-traffic situations I’ve been in so far.

 

[Rattletrap]

To explain to people not aware of this Oregon place.

It's a place where the speed limit doesn't really exist, but they love putting speed limit signs up. Like if you drive on I5 in WA state people are mostly going 65-70mph in a 60mph zone, but the second you hit Oregon all bets are off on what speed people will go. Maybe 60mph in a 50mph, or maybe 70.

They'll change the speed limit for no reason like from 55mph to 60mph, and then back to 55mph. To the uninitiated you might think you're supposed to slow down to when going from 60mph to 55mph, but nope just leave it.

Now you'd think its just people driving fast to get to place faster, but that's totally completely wrong as there is always a traffic jam ahead so its really just speeding between traffic jams for no particularly good reason.

We can't actually self driving cars because of Oregon.

The AI would simply turn on us, and run us over for not being logical.

Honestly not sure if you're being facetious because we do have wild swings in posted speed limits even across rural straightaways. They used to use them for speed traps until Portland police were de-funded, now they have their hands full keeping the left-wing and right-wing protestors from killing each other and razing homeless camps. They stopped pulling vehicles over entirely 18 months ago so you can drive as fast as you want and only have photo radars to worry about (and our local drivers have realized this by now). It may be a lawless dystopian wasteland but it's perfect for driving a Tesla if you know where the potholes are and can avoid the drugged-out zombies walking in the street trying to kick moving cars. I just really need the 90 MPH autosteer that I was promised "in the weeks ahead".

 

[thewishmaster]

Grain of salt etc but supposedly the S/X are being built vision-only now:

https://twitter.com/x/status/1432390677476876288

 

[BogStandard]

So I took a long 400 mile journey for work this week. It was an extremely frustrating drive. So many braking events.
In the 7300 miles I now have on my May 2021 M3, I think I've found the patterns of unintended/phantom braking using TACC. (no autopilot, I rarely use it)

1. Cresting hills on 2-way roads. Happens occasionally with no other traffic, but happens ~80% of the time if another vehicle is oncoming on the other side of the hill either at the top or the bottom. As soon as the camera sees it, it brakes. This happens on very shallow rolling hills and also steep hills.
2. Oncoming traffic on 2-way roads. Related to above, but even on perfectly flat sections, unintended braking occurs roughly 10% of the time regardless of conditions.
3. Night time traffic on 2-way roads. I have only driven a significant distance a few times at night now. But in every case, I'd say 50% or more of oncoming vehicles triggered braking, didn't matter if hills, curves, or flat and straight. The only pattern I could find was older cars with dim headlights didn't trigger an event.
4. Sunny conditions. I think I've figured why I get phantom braking when no one is within close range. Mirages. Yesterday I was following a car 1/2 mile ahead of me and kept getting a lot of phantom braking events when it was only me and him/her on the road. Occasionally I could see a reflection of the car in the mirage that looked much closer than it actually was. I wonder if this was fooling the cameras to see the car much closer that it was.
   I haven't noticed a difference between when it's high noon or sun closer to the horizon. Seems to be the same result.
5. The times I rarely get braking events are on one way roads, especially if TACC is actively following a car. (Although hills sometimes still trigger it.) Or when it's fully overcast/raining. Which leads me to believe mirages are tricking the system as described above, or perhaps the cameras are getting temporarily washed out from the sun, I'm not sure. Would be interesting to see a live view of the forward facing cameras.

I can confirm that completely blocking the front cameras locks out both autopilot and TACC. I thought maybe I could trick the car into defaulting to a dumb cruise system, but no dice. Just doesn't have that capability now.

I also get pillar "camera is blocked or temporarily blinded" alarms fairly often when they're not dirty or sun shining on that side. I wonder if I'm having camera failures... I've been thinking about calling the service center to have them take a look. And maybe they can also clear the 200 bug reports I recorded that I thought was being actively sent to Tesla's big brain center.

 

[gearchruncher]

I thought maybe I could trick the car into defaulting to a dumb cruise system, but no dice. Just doesn't have that capability now.

FYI, this has never worked, it's not new. Tesla's cruise control does not ever degrade to "dumb" cruise control. It's either functional, or no cruise. This was true even back in 2016, where some snow on the bumper caused "obscured radar, cruise control unavailable" messages.

 

[BogStandard]

FYI, this has never worked, it's not new. Tesla's cruise control does not ever degrade to "dumb" cruise control. It's either functional, or no cruise. This was true even back in 2016, where some snow on the bumper caused "obscured radar, cruise control unavailable" messages.

I figured it wouldn't but I had a roll of tape on my workbench sitting right there and just decided to try it on a whim to see what would happen without researching anything to satiate my own curiosity. The car did not like it. The glass heating elements turned on and the wipers went crazy trying to get it off, lol.

 

[JHCCAZ]

So I took a long 400 mile journey for work this week. It was an extremely frustrating drive. So many braking events.
In the 7300 miles I now have on my May 2021 M3, I think I've found the patterns of unintended/phantom braking using TACC. (no autopilot, I rarely use it)

1. Cresting hills on 2-way roads. Happens occasionally with no other traffic, but happens ~80% of the time if another vehicle is oncoming on the other side of the hill either at the top or the bottom. As soon as the camera sees it, it brakes. This happens on very shallow rolling hills and also steep hills.
2. Oncoming traffic on 2-way roads. Related to above, but even on perfectly flat sections, unintended braking occurs roughly 10% of the time regardless of conditions.
3. Night time traffic on 2-way roads. I have only driven a significant distance a few times at night now. But in every case, I'd say 50% or more of oncoming vehicles triggered braking, didn't matter if hills, curves, or flat and straight. The only pattern I could find was older cars with dim headlights didn't trigger an event.
4. Sunny conditions. I think I've figured why I get phantom braking when no one is within close range. Mirages. Yesterday I was following a car 1/2 mile ahead of me and kept getting a lot of phantom braking events when it was only me and him/her on the road. Occasionally I could see a reflection of the car in the mirage that looked much closer than it actually was. I wonder if this was fooling the cameras to see the car much closer that it was.
   I haven't noticed a difference between when it's high noon or sun closer to the horizon. Seems to be the same result.
5. The times I rarely get braking events are on one way roads, especially if TACC is actively following a car. (Although hills sometimes still trigger it.) Or when it's fully overcast/raining. Which leads me to believe mirages are tricking the system as described above, or perhaps the cameras are getting temporarily washed out from the sun, I'm not sure. Would be interesting to see a live view of the forward facing cameras.

I can confirm that completely blocking the front cameras locks out both autopilot and TACC. I thought maybe I could trick the car into defaulting to a dumb cruise system, but no dice. Just doesn't have that capability now.

I also get pillar "camera is blocked or temporarily blinded" alarms fairly often when they're not dirty or sun shining on that side. I wonder if I'm having camera failures... I've been thinking about calling the service center to have them take a look. And maybe they can also clear the 200 bug reports I recorded that I thought was being actively sent to Tesla's big brain center.

Can you comment on weather you think these breaking events occur the same with or without cars behind you? We've had some discussion about this, as many reports of phantom braking mention "good thing no one was behind me at the time". So one of the questions is whether the Tesla takes this into account and maybe errs on the side of cautious braking when there's low chance of a rear end collision, but errs on the side of not-braking if someone is following closely.

 

[gearchruncher]

So one of the questions is whether the Tesla takes this into account and maybe errs on the side of cautious braking when there's low chance of a rear end collision, but errs on the side of not-braking if someone is following closely.

Tape over the rear camera (the only source they would have of a car behind you) and see if you get any AP errors.

 

[BogStandard]

Can you comment on weather you think these breaking events occur the same with or without cars behind you? We've had some discussion about this, as many reports of phantom braking mention "good thing no one was behind me at the time". So one of the questions is whether the Tesla takes this into account and maybe errs on the side of cautious braking when there's low chance of a rear end collision, but errs on the side of not-braking if someone is following closely.

It's definitely happened with other vehicles directly behind me. I'll just hover closely or keep the accelerator slightly pressed when I anticipate it to happen until they pass me. I'm not going to speed excessively just because the doofus behind me does.
The first time or 2 it happened it probably was a good thing no one was behind me as I didn't know why or what to do to stop it from slowing down. I mean, if they were paying attention then it only would have been a major annoyance to them. But one of the first times it was more than just a full regen slowing, it was mechanically assisted hard braking. My lunch bag flew from the front seat and smacked the glove box. Scared the bejeezus out of me.

 

[powertoold]

I'm curious if there are any videos or dashcams showing the phantom braking events.

 

[rt29781]

Can you comment on weather you think these breaking events occur the same with or without cars behind you? We've had some discussion about this, as many reports of phantom braking mention "good thing no one was behind me at the time". So one of the questions is whether the Tesla takes this into account and maybe errs on the side of cautious braking when there's low chance of a rear end collision, but errs on the side of not-braking if someone is following closely.

I have had severe PB while using TACC and the lorry behind was almost sideways trying to stop so I don’t think the software takes into account what is behind. There is a well publicised vehicle accident in Norway where a M3 brakes for no valid reason and is hit by the following van which was in turn is hit by the following lorry.

 

[BogStandard]

I'm curious if there are any videos or dashcams showing the phantom braking events.

I haven't looked. It would be easy, I could just drive 10 miles out from my house and come back with at least a few events. I don't have any recording equipment though, other than my phone, but I don't have any mounts or anything. I'd have to film my feet somehow as well, otherwise I just know people would say I'm doing it with the pedals.
I'm just not motivated enough to buy a couple phone mounts and edit a bunch of video footage together.