Tesla replacing ultrasonic sensors with Tesla Vision

[Knightshade]

something as critical as parking sensors

because, of course, cars were entirely unsafe and inoperable for the roughly 100 years they didn't ever have those at all


I'm not saying your arguments are getting even dumber, but I'm also NOT not saying that.

Doesn’t that make significant assumptions? Like the approach is at a constant rate once out of the field of view? Certainly possible someone may tap the brakes etc when getting closer to a wall/etc.

Are you under the mistaken impression the car doesn't know when the rate of movement changes, when in fact it knows exactly how it changed and by how much?

 

[2101Guy]

because, of course, cars were entirely unsafe and inoperable for the roughly 100 years they didn't ever have those at all


I'm not saying your arguments are getting even dumber, but I'm also NOT not saying that.




Are you under the mistaken impression the car doesn't know when the rate of movement changes, when in fact it knows exactly how it changed and by how much?

Lol. Bless your heart.

just how old ARE you?

By your logic, antilock brakes, airbags, and traction control systems aren’t “necessary” these days either. I get that you didn’t grow up with those unnecessary things, but it’s a new era.

Are you aware there are documented tests that prove the effectiveness (from a safety aspect to children in at least one test) of USS sensor functionality?

Here, let me help

America's cars and trucks are getting bigger, and so are their front blind zones. Children are paying the price.

Every year, many kids get hit and die when their parents can’t see them in their front blind zones. A senator wants to pass a law requiring safety features to prevent these accidents.

www.nbcnews.com

 

[2101Guy]

If I have time later, in order to help you, I will post more documented studies that prove the safety effectiveness of US S sensor functionality.

 

[stopcrazypp]

Doesn’t that make significant assumptions? Like the approach is at a constant rate once out of the field of view? Certainly possible someone may tap the brakes etc when getting closer to a wall/etc.

As others point out, the car knows exactly how much the car has moved via the wheel sensors and IMU, and also if someone is applying the brakes. Even if that weren't the case, there are 8 cameras, so unless you are in a featureless scene, the car can use other objects as a reference point to judge relative vehicle movement.

In fact, that is how humans drive in the first place when parking. You don't have the ability to see through the car, but it's possible to park very precisely by using other visual reference points and your memory of object positions as you see them. Plenty of videos of people doing parallel parking for example with inches to spare. The only issue is many are not skilled at this, but for a computer there is no such variation (once they figure out the program, it'll work largely the same in all cars).

 

[Knightshade]

Question: just how old ARE you? I mean…I’m not saying you ARE a boomer, but if it smells like a boomer..

Congrats, your long uninterrupted streak of being wrong about literally everything remains untarnished!

By your logic, antilock brakes, airbags, and traction control systems aren’t “necessary” these days either.

I mean, they're not, unless you're making up your own definition of "necessary"

They're certainly nice to have, but cars worked just fine for most of a century without them.

Are you aware there are documented tests that prove the effectiveness (from a safety aspect to children in at least one test) of USS sensor functionality?

No, and neither are you.

Here, let me help

America's cars and trucks are getting bigger, and so are their front blind zones. Children are paying the price.

Every year, many kids get hit and die when their parents can’t see them in their front blind zones. A senator wants to pass a law requiring safety features to prevent these accidents.

www.nbcnews.com

Ah, the "THINK OF THE CHILDREN!" logically fallacy.

So- first- your link shows no study at all demonstrating "effectiveness" of USS so that claim was just an outright lie by yourself.

It shows a test where they lined sitting kids up in front of giant trucks and SUVs with terrible front visibility and measured how many they had to line up before you could see one over the hood.

Which had nothing whatsoever to do with USS...or with much smaller vehicles like the 3/Y without those massive hoods and height ruining front visibility.

It DOES mention "sensors" but then goes on to note they didn't actually test them at all-- and that one of the car makers (for the Escalade) specifically told them those sensor

are designed to detect “larger objects” and would not alert the driver to “small children” or “moving people.”

So no... I'm not aware of a study showing what you claimed---and neither are you since your link doesn't claim such a thing and in fact contradicts that claim with a quote from the vehicle maker


Further-

Your link notes 64 children died in a year after being hit by a forward-moving vehicle off of public roads.

For comparison nearly seven times more kids die annually from swimming pools.

Do you support a ban on swimming pools? Is your answer no mainly because Tesla doesn't make them?


There's some other issues here too of course- the big one- your link primarily points out LARGE TRUCKS/SUVS have big front blind spots... this is not true of the 3/Y- which does have a front blind spot of course, but not nearly as large as the one they're concerned about. In fact your own source notes front blind spot is a quite minor issue for things like compact and midsized SUVs (and certainly sedans)-- like the 3/Y.


So as always your actual "source" doesn't really support anything you're claiming, and even contradicts some of it.

Maybe one day you'll starting reading a source past the headline before throwing it out here and looking foolish?

Probably not.

 

[2101Guy]

On a separate note…literally from 20 mins ago…an example of how accurate/well performing TeslaVision is today, relying solely on a camera function to feed software (as if we already didn’t know this by the horrific performance of the auto wipers in the auto headlights).

Can anyone figure out what’s wrong with this picture?

 

[emmz0r]

As others point out, the car knows exactly how much the car has moved via the wheel sensors and IMU, and also if someone is applying the brakes. Even if that weren't the case, there are 8 cameras, so unless you are in a featureless scene, the car can use other objects as a reference point to judge relative vehicle movement.

In fact, that is how humans drive in the first place when parking. You don't have the ability to see through the car, but it's possible to park very precisely by using other visual reference points and your memory of object positions as you see them. Plenty of videos of people doing parallel parking for example with inches to spare. The only issue is many are not skilled at this, but for a computer there is no such variation (once they figure out the program, it'll work largely the same in all cars).

so why isn't it out yet if its so simple

 

[Mr Puzzles]

Honestly: Tesla MUST get a new CEO this year. And one who isn’t a musk puppet. Not offering round wheel option for s/x upfront, was all musk. Removing USS BEFORE replacing same functionality with vision? All musk.

Just two Poor decisions that were avoidable.

Yep. Interestingly he told me he and the people he works with are tired of apologising for him, I was amazed by his honestly.

 

[stopcrazypp]

so why isn't it out yet if its so simple

The Occupancy Network is already out on FSD Beta (I posted up thread some sample output from their demos), and there is evidence they are putting the code as a drop in replacement for USS (simulating the locations and output of it) in 2022.40.4, although they have not added it to the UI yet.

 

[AlanSubie4Life]

so why isn't it out yet if its so simple

Probably not the hard part!

The occupancy network is going to be extremely difficult to impossible to get to work. Easy for a human. Humans have a much harder time with the IMU part.

 

[emmz0r]

The Occupancy Network already out on FSD Beta (I posted up thread some sample output from their demos), and there is evidence they are putting the code as a drop in replacement for USS (simulating the locations and output of it) in 2022.40.4, although they have not added it to the UI yet.

Auto parking etc is not availbale on FSD beta is it?

 

[stopcrazypp]

Auto parking etc is not availbale on FSD beta is it?

Autoparking is the main branch also, nothing to do with FSD Beta. They will probably release everything at once given the current strategy is to just simulate USS output, so they can reuse all the existing UI and functions.

They can support more advanced functionality beyond what USS can do (the occupancy network appears to have much higher vertical and horizontal resolution, although depth resolution is unknown), but that would increase development time a lot more (as it'll require a significant change to a lot of the other code, including the UI).

They also have to consider that they need to continue to support cars with USS and continue giving updates, so this has to be coordinated too.

To bring it back to my original point, basically there's a lot of other software development required to integrate this, but that doesn't mean "a camera which can’t see beyond the hood edge will somehow be able to judge distance to the bumper" is a concern as per my point.

For example, this post shows an example of a cone in the blind spot of the car that is clearly showing up on the Occupancy Network:

Tesla replacing ultrasonic sensors with Tesla Vision

The idea is already here, but the technology is not. Bringing the idea to explain how it can be done is here, but the technology to make that idea successful is not. Ah, I see. Thanks for clarifying your position.

teslamotorsclub.com

 

[emmz0r]

Autoparking is the main branch also, nothing to do with FSD Beta. They will probably release everything at once given the current strategy is to just simulate USS output, so they can reuse all the existing UI and functions.

They can support more advanced functionality beyond what USS can do (the occupancy network appears to have much higher vertical and horizontal resolution, although depth resolution is unknown), but that would increase development time a lot more (as it'll require a significant change to a lot of the other code, including the UI).

They also have to consider that they need to continue to support cars with USS and continue giving updates, so this has to be coordinated too.

To bring it back to my original point, basically there's a lot of other software development required to integrate this, but that doesn't mean "a camera which can’t see beyond the hood edge will somehow be able to judge distance to the bumper" is a concern as per my point.

For example, this post shows an example of a cone in the blind spot of the car that is clearly showing up on the Occupancy Network:

Tesla replacing ultrasonic sensors with Tesla Vision

The idea is already here, but the technology is not. Bringing the idea to explain how it can be done is here, but the technology to make that idea successful is not. Ah, I see. Thanks for clarifying your position.

teslamotorsclub.com

That's just an image for the presentation. What I am concerned with is the accuracy at very short distances. Where I live there are very tight parking spaces and pillars, and it has to be correct down to less than an inch.
Where is the footage of the USS simulation in tight parking spaces?

 

[enemji]

I think it will be a lot of "camera visibility reduced, parking distance estimation disabled".
The backing camera can't be trusted, it's almost always occluded with dirt, especially in winter, so you're left with the repeater cameras.
Also parking front first would be a challenge.

I don’t know where you live and where you drive but it is very rare to have the rear view camera occluded.

 

[enemji]

The sentiments about how Tesla removing features before the replacement functionality is available is crappy customer service, questions about whether it's legal to do this to orders-in-progress and observations about how inconvenient the semi-dodgy wipers/lights etc are to an owner right now, today, are all valid (if MASSIVELY overblown on this whine-fest thread).

Stripping it back to the 'why don't Tesla care?' question, though, I'd say sit down and watch the FSD part of the AI day presentation (Starts 0:58:00. If you understand everything they're talking about you're doing far better than me, but there's a thread of layman-friendly discussion throughout.). Now, imagine you're speaking to that team working on solving the problems that they're working on and imagine telling them they have their priorities all wrong and you need your wipers sorted, ASAP.

Or imagine that you're in charge of Telsa's operations and you have the following options (assuming that the underlying problem is that USS are in short supply across the industry and can't be obtained):
1) Suspend delivery of all vehicles until you can deliver them as specc'd
2) Deliver cars without and prepare for an expensive retrofit program, using your very low capacity service network (vs ICE vehicles that all visit service centres routinely)
3) Recognise that you were dropping this hardware anyway, so drop it a bit early.
(Note that 2 and 3 are the same result to the customer. You can't have USS if they're physically not available, so you'd have had to have had a car without them in a retrofit scenario.)

You'd be insane not to take option 3, with the offer for anyone who doesn't like it to cancel or defer, which is exactly what they did.

PS. Based on what's in that presentation I'd say the TV based parking system does have every chance of being massively superior to USS. I look forward (with moderate confidence) to having a car that doesn't panic at the sight of light foliage and puddles. Whether I'll get it any time soon is another matter.

If you want better crops/yields you burn the existing fields. Elon Musk works in the same manner. The driving force to get it done is higher when you have no backup plan.

 

[KArnold]

I don’t know where you live and where you drive but it is very rare to have the rear view camera occluded.

Except when it rains or snows.

 

[emmz0r]

I don’t know where you live and where you drive but it is very rare to have the rear view camera occluded.

Everyone who lives outside arizona, dallas and california

 

[Boza]

Everyone who lives outside arizona, dallas and california

I don’t know if you follow the news but California just experienced a ton of rain and snow. Stop blaming it for inept software engineering. If Tesla wants to they can test all weather combinations without even leaving the state.

 

[stopcrazypp]

That's just an image for the presentation. What I am concerned with is the accuracy at very short distances. Where I live there are very tight parking spaces and pillars, and it has to be correct down to less than an inch.
Where is the footage of the USS simulation in tight parking spaces?

It might not be correct down to an inch or less (there is no data on that yet, and personally I mentioned previously upthread I was somewhat skeptical if it would be), but that was not the original question. The original question was if the ON can judge distance to the bumper when the camera can't see below the hood, and from their presentations, the answer is clearly yes.

I only wanted to address this point because it has been brought up a billion times on this thread and people may not be aware of how the ON works.

 

[2101Guy]

Tesla visions performance with auto wipers, auto high beams, and auto-direction….let’s me know all that I need to know as it relates to how it can be expected to perform while replacing US S functions…