Tesla.com - "Transitioning to Tesla Vision"

[JHCCAZ]

Yeah, but at this point you're saying that all Tesla needed to do was get NHTSA to change the FVMSS all so they could optimize an L2 car, which needs a human PAYING ATTENTION at all times. This from an organization that has blocked matrix headlights.

We are nowhere near the car driving itself. The human is still 99% of this system. The human needs high beams. A system which only lets the computer see when AP is on because it has dim lights and a superhuman camera is a completely broken L2 system. And we all know this system will never get to L4, so any optimization for when the human isn't needed is pointless.

Plus, it's clear here that Tesla did this because of supply chain issues. They had zero way to work with any regulator, or re-design headlights, or switch to different cameras in the middle of a massive semiconductor shortage. They had about a month to pull this off. Adding a line of code to interlock AP with the existing high beams is the move fast break things solution.

I don't think I'm arguing with you. I was one of the first to propose that this situation is all about supply chain issues and I completely buy your last paragraph. Also, yes any purpose-designed alternative is going to be realized on a longer timescale, not in two weeks or two months.

The part where I think you are mischaracterizing my proposal is

A system which only lets the computer see when AP is on is a completely broken L2 system.

The pertinent complaint is that new-AP is forcing High-Beams on even though the human driver wouldn't have done so.

I didn't propose lights too dim for human monitoring! I proposed normally-bright dipped beams, that the human drivers would choose, plus a modest long beam that the camera can use but wouldn't have been much good to the human. IMO this is very sensible and takes nothing away from what was previously expected by the human operator. What it takes away is only a schizophrenic high beam that everyone hates and is overkill for the camera.

 

[orion2001]

I am curious, are the 3 "purple" sections of the video labeled as Tesla-vision related speed reductions unique to the new radar-less Tesla systems or does this also happen with existing cars that have radar? I've heard of phantom braking but this is not phantom braking (and the YouTuber says as much in his video description).

I don't recall reading about such things happening with AP in the past, but I also have no first-hand experience beyond my test-drive. If this is another artifact of "Tesla-vision", it only makes this whole rollout look even more half-baked.

 

[stopcrazypp]

I think we are in agreement even if I didn't articulate my viewpoint as well as I would like. My point about the lens was the same as yours. A faster lens (lower f/ stop) would typically be advantageous). Hence the reason so many people use prime fast lenses in low light conditions.

The point still remains, Tesla sourced this stuff years ago. Even then it was to use in conjunction with radar I would assume. So most if there were looking at it holistically, they probably relied on radar for areas where the Tesla Vision had significant enough limitations that radar was necessary. There have been significant improvements in almost every area of technology in the last 5+ years. It looks like Tesla hasn't upgraded the kit over that period of time, based on what I've read so far.

Another point that I think gets lost in the discussion is outside of North America they will keep the radar as apparently some countries required it from a compliance perspective. So did they mandate radar be used or was a pure Tesla Vision not good enough because their testing is more demanding?

North America's AEB/FCW/LCW requirements are "voluntary". That means Tesla doesn't have to have the system tested yet by NHTSA to release it (but as the recent events show, they do lose checkmarks). This may not be true in other jurisdictions. And if such testing is mandatory, it may be better to do it as part of a model year change (as is almost always the case, CR pointed out what Tesla did is rare in the industry).

 

[rjpjnk]

The old hardware is my concern...

I just looked at the sensor specs. WTH, 720p? That is pretty terrible resolution. I am staking my life on a 720p alone?

My guess is that 720p is more than adequate resolution. I think the problem with visual light cameras is not that the resolution is too low, but that they don't perform well in bad weather and low light. Also, it is more difficult to extract range and velocity information compared to radar (in any weather).

I don't know what sort of radar Tesla was using, but if it had any angular resolution at all (not sure it did) it must have been a phased array, and given the size of its aperture and wavelength it would have had a lot lower resolution than even a poor quality 720p camera.

It's just that radars measured speed and distance so effortlessly, and did it even in poor visibility.

 

[Veritas]

I really don’t understand how teslas $65 worth of forward cams can be fsd ever? The more sensors, the more data. Yes that would need greater computational power and better software. Thats where tesla should be focusing. Dont get rid of radar, improve it. They had just filed a patent a few months ago for a new radar system. What an incredible turn of events. Im guessing radar will eventually be back. That’s likely why they are keeping it on the rich folks S and X models. Unfortunately and unbeknownst to me at the time of ordering, i will be getting a radarless m3p next week.

 

[Veritas]

Heres a look at the outgoing radar, if anyone’s interested:

 

[BitJam]

I'm guessing radar will eventually be back.

I agree.

In a recent podcast NO RADAR WTF! Dax and High-Mileage-Rob said that during their years of Tesla driving, having radar see the car in front of the car in front of them do something wacky has saved their bacon multiple times. They said their Teslas would start braking before the car in front of them had reacted.

Dax has a new no-radar Model Y and he says the car in front of the car in front of him is missing from the display unless there is a larger vehicle followed by a smaller one so the cameras can see around the edges of the smaller one. Both commentators seemed to be very Tesla friendly. They scoffed at the idea of people complaining about losing the passenger side lumbar support adjustment.

Assuming Dax and High-Mileage-Rob aren't fooling themselves, it's unclear to me how TV-only will get parity with this feature. I don't see how it can be as safe as a well designed TV + radar system even if you only use radar to get hints about the car in front of the car in front of you.

 

[Knightshade]

Assuming Dax and High-Mileage-Rob aren't fooling themselves, it's unclear to me how TV-only will get parity with this feature. I don't see how it can be as safe as a well designed TV + radar system even if you only use radar to get hints about the car in front of the car in front of you.

So for years folks have complained that the car phantom braking for overpasses and other things caused by radar was dangerous and would cause accidents.

If that's true, it's possible removing that avoids more accidents than "seeing the car ahead of the car ahead" prevented, and net safety improves.

That doesn't mean there'll be NO accidents radar would have prevented- just that it's possible there'd be less of those than the type vision-only prevents.


Not even saying that's the case, but it's an example of "how it can as safe" if radar was causing significant problems and killing the useful signal to noise ratio as suggested.

 

[texas_star_TM3]

So for years folks have complained that the car phantom braking for overpasses and other things caused by radar was dangerous and would cause accidents.

If that's true, it's possible removing that avoids more accidents than "seeing the car ahead of the car ahead" prevented, and net safety improves.

That doesn't mean there'll be NO accidents radar would have prevented- just that it's possible there'd be less of those than the type vision-only prevents.


Not even saying that's the case, but it's an example of "how it can as safe" if radar was causing significant problems and killing the useful signal to noise ratio as suggested.

the solution is sensor fusion .... radar bounces off the bridge and gives a faulty signal (odd though how this doesn't happen with the Q3 / VW Tiguan etc) ... the camera confirms it is "only" a bridge and therefore cancels the radar input when making a decision to keep driving..... just using the camera and no radar could well mean you are ramming a standing still vehicle on the highway at night in the rain because the cameras pick it up way later than radar ever would...

 

[texas_star_TM3]

Heres a look at the outgoing radar, if anyone’s interested:

View attachment 668996

knowing now that this is a continental low/mid tier radar unit ... their decision to ship cars without it was likely *totally* chip shortage related. Continental is suffering massive production output falls due to their inability to source chips.

just funny how thousands and thousands of produced cars being parked on holding lots for weeks for "mysterious" reasons and suddenly Tesla says "nah...radar is not needed anyway" and they get delivered....

 

[Knightshade]

the solution is sensor fusion ....

That's what they did for years.

It causes the exact thing I'm describing.

And it gets WORSE when you are trying to fuse it with 36 fps 360 degree camera input since the radar is only providing low-res 12 fps input.

radar bounces off the bridge and gives a faulty signal (odd though how this doesn't happen with the Q3 / VW Tiguan etc)

It absolutely DOES happen with all cars using radar. It is an inherent limitation of the low-res radars used on cars.

Some worse than Tesla.

Some car makers have been sued or had to issue recalls because of how badly it was happening in fact.

Autobraking tech will be standard in cars by 2022, but drivers complain of "phantom braking"

One woman described how her Nissan Rogue slammed on the brakes on a highway for no clear reason

www.cbsnews.com

Since 2015 there have been seven recalls for auto-braking issues, affecting nearly 180,000 vehicles.

... the camera confirms it is "only" a bridge and therefore cancels the radar input when making a decision to keep driving

The problem with 2 different inputs is how do you know which one is right?

If you think the camera is good enough to trust then why bother with the radar?

If you think the radar is right, well, you get exactly the phantom braking you get today.

..... just using the camera and no radar could well mean you are ramming a standing still vehicle on the highway

Ironically, radar causes this exact problem

And again in ALL brands, not just Tesla

Why emergency braking systems sometimes hit parked cars and lane dividers

Recent Tesla Autopilot crashes hold a lesson for the whole industry.

arstechnica.com

Essentially the fact radar returns a TON of "stationary" object responses on the side of the road that are too low-res to make sense of means the vehicle has to largely ignore them.

And again if your answer is "use the cameras to check" then that means you think the camera is good enough to tell- in which case the radar doesn't help.

knowing now that this is a continental low/mid tier radar unit ... their decision to ship cars without it was likely *totally* chip shortage related. Continental is suffering massive production output falls due to their inability to source chips.

FWIW I agree doing it TODAY is a part shortage thing.

But they've said publicly they're planning to move to a vision-based system for several years now and we've seen development in the code going that way for years.

The shortage probably moved them to it a little earlier than planned, but it's not just "YOLO!" from scratch here.

 

[rjpjnk]

Thanks for the radar pic! So based on the manufacturers datasheet it is in fact a scanning phased array radar with two modes, a wide FOV near scan and a narrow long distance scan. The datasheet shows just how good its distance and speed accuracy are (0.13 m and 0.1 kph respectively). The wavelength is in the 4mm range and its best case angular resolution is 0.2 deg which would yield a resolution cell of about 1 m wide at max range. Better than I expected. I wish they kept this.

 

[texas_star_TM3]

That's what they did for years.

It causes the exact thing I'm describing.

And it gets WORSE when you are trying to fuse it with 36 fps 360 degree camera input since the radar is only providing low-res 12 fps input.

and the decision to suddenly yank radar and go TeslaVision only - before full parity with the old system is achieved - magically coincides with thousands of parked cars unable to be delivered due to radar unit shortages... *rrrriiiiight*

and to make it worse ... if they had planned on vision only all along... why in the world not upgrade the crappy 720p webcams from 6 years ago while you are at it. night vision FLIR or higher resolution / better light sensitivity cameras. fact is... in 2018 they felt the camera system needed radar to do all they wanted and now suddenly the same cameras minus radar is "sufficient" ?

 

[YieldFarmer]

Or mass arbitration requests for current customers. Companies are changing their arbitration rules because so many people are using them against them:

After 75,000 Echo arbitration demands, Amazon now lets you sue it

IT-savvy law firms have swamped retailer with arbitration cases.

arstechnica.com

Yes, Tesla's agreement says they will pay for arbitration. So if a bunch of Tesla owners all decided to file at once.....

Reminder to all new buyers to OPT-OUT of arbitration. You can still arbitrate if you want, but you also retain full rights to sue!

I am curious, are the 3 "purple" sections of the video labeled as Tesla-vision related speed reductions unique to the new radar-less Tesla systems or does this also happen with existing cars that have radar? I've heard of phantom braking but this is not phantom braking (and the YouTuber says as much in his video description).

I don't recall reading about such things happening with AP in the past, but I also have no first-hand experience beyond my test-drive. If this is another artifact of "Tesla-vision", it only makes this whole rollout look even more half-baked.

I noticed a few similar events (slowdowns for no reason) on the new Vision System. Have not driven my local roads in a radar-equipped car though so I cannot compare though.

 

[Knightshade]

and the decision to suddenly yank radar and go TeslaVision only - before full parity with the old system is achieved - magically coincides with thousands of parked cars unable to be delivered due to radar unit shortages... *rrrriiiiight*

My post explictly points out the shortage likely sped up the change to vision only- but it's something they've been moving toward for years.... so it likely sped thing up a few months at most.

Same thing happened with AP1->AP2, they unexpectedly needed to switch sooner than planned due to mobileye partnership falling apart, so it took time before they got new system to parity with old one, and then have far exceeded it.

and to make it worse ... if they had planned on vision only all along... why in the world not upgrade the crappy 720p webcams from 6 years ago while you are at it. night vision FLIR or higher resolution / better light sensitivity cameras. fact is... in 2018 they felt the camera system needed radar to do all they wanted and now suddenly the same cameras minus radar is "sufficient" ?

Because they still feel the current cameras provide sufficient data to "solve" vision.

Again, they planned to do this since at least 2018- vision only is NOT A NEW THING.

Long-term, the car will work purely on vision

 

[gearchruncher]

My post explictly points out the shortage likely sped up the change to vision only- but it's something they've been moving toward for years.... so it likely sped thing up a few months at most.

That's just as much a problem for this "transition" as the unplanned narrative.

If they have been trying to do this since 2018, then they are 3 years in and it wasn't at parity. That means they have no real idea how much longer it will take. It's really unlikely it will be a few weeks if you're any student of history. If you asked Elon how long it would be for vision only in 2019 he would have said 3-6 months, yet it's 24 months past that. He has been telling us v9 is vision only and just a few weeks away for months now, yet it's nowhere to be seen.

You know when they last claimed 2 weeks to reach parity? AP1-AP2. The exact situation you brought up. Like we all know, it took 2 YEARS. Yes, AP2 is now better than AP1, but Tesla customers went half their leases with a much worse product than the one they test drove. All because Tesla was over-using Mobileye's features. Plus, this is a case where they went to indisputably less capable hardware, not more capable.

So, is it really a better story that they have been working on this for years and had to rush to release it, vs the idea that they have only been doing this for a few months? It's really irrelevant- both of them mean Tesla has no idea how long it will take to make this work (if ever), and customers are going to be the ones that pay for it.

...And remember, it's now more than a week into the "weeks ahead" transition they claimed. This clock is already running deep into what anyone would consider a "weeks ahead" timeframe.

 

[texas_star_TM3]

That's just as much a problem for this "transition" as the unplanned narrative.

If they have been trying to do this since 2018, then they are 3 years in and it wasn't at parity. That means they have no real idea how much longer it will take. It's really unlikely it will be a few weeks if you're any student of history. If you asked Elon how long it would be for vision only in 2019 he would have said 3-6 months, yet it's 24 months past that. He has been telling us v9 is vision only and just a few weeks away for months now, yet it's nowhere to be seen.

You know when they last claimed 2 weeks to reach parity? AP1-AP2. The exact situation you brought up. Like we all know, it took 2 YEARS. Yes, AP2 is now better than AP1, but Tesla customers went half their leases with a much worse product than the one they test drove. All because Tesla was over-using Mobileye's features. Plus, this is a case where they went to indisputably less capable hardware, not more capable.

So, is it really a better story that they have been working on this for years and had to rush to release it, vs the idea that they have only been doing this for a few months? It's really irrelevant- both of them mean Tesla has no idea how long it will take to make this work (if ever), and customers are going to be the ones that pay for it.

...And remember, it's now more than a week into the "weeks ahead" transition they claimed. This clock is already running deep into what anyone would consider a "weeks ahead" timeframe.

yup. "they have been working on vision only FOR YEARS" and at the same time AP not working over 75mph and without high-beams at night aren't exactly comforting statements... just tells me that for years they couldn't solve these issues and are now releasing a half baked concept with the promise that a software update in "2 weeks" will fix it all (somehow that wasn't doable in the years prior)

 

[mark95476]

This is probably just FUD, but I know some of the users (and lurkers) on TMC do actual HW/SW/firmware dev work, related or not, and must be laughing and shaking their head at all of the inane things posted here.

I really don’t understand how teslas $65 worth of forward cams can be fsd ever? The more sensors, the more data. Yes that would need greater computational power and better software. Thats where tesla should be focusing. Dont get rid of radar, improve it. They had just filed a patent a few months ago for a new radar system. What an incredible turn of events. Im guessing radar will eventually be back. That’s likely why they are keeping it on the rich folks S and X models. Unfortunately and unbeknownst to me at the time of ordering, i will be getting a radarless m3p next week. View attachment 668995

Heres a look at the outgoing radar, if anyone’s interested:

View attachment 668996

Thanks for the radar pic! So based on the manufacturers datasheet it is in fact a scanning phased array radar with two modes, a wide FOV near scan and a narrow long distance scan. The datasheet shows just how good its distance and speed accuracy are (0.13 m and 0.1 kph respectively). The wavelength is in the 4mm range and its best case angular resolution is 0.2 deg which would yield a resolution cell of about 1 m wide at max range. Better than I expected. I wish they kept this.

View attachment 669045

 

[texas_star_TM3]

This is probably just FUD, but I know some of the users (and lurkers) on TMC do actual HW/SW/firmware dev work, related or not, and must be laughing and shaking their head at all of the inane things posted here.

what's insane? the last OEM who did use vision only was Subaru with EyeSight and they are adding now radar and *never* claimed to be on the path to offer FSD *any moment* now

 

[beowulf573]

I've actually got 30+ years experience with software and hardware including embedded systems and manufacturing. This sudden switch over to vision only boggles my mind and is why I put my order on hold. The only explanation that makes any sense is that they could not get enough radar units, so they reserved those they could source for Model X and S since those are premium cars, and those cars overseas with tighter regulatory requirements, so that quarterly numbers could be met. It also explains why the software is not done and why NHTSA and other orgs like the IIHS were not notified ahead of time.

I may be repeating myself, but I waited years to order a Tesla and was very excited when the stars aligned and I was able to order one in April. Specially given the crappy time all of us have had over the past year plus. To say I'm disappointed is an understatement.