Hello, rear radar :-)

[lunitiks]

Another notable change from about that time is the park system hardware got a new revision (the ultrasonics) with hw2.5.

So a rear-facing radar AND a new park assist ecu?

 

[BigD0g]

Who knows. Evidently their current attempts did not pan out to produce the results they were hoping for, so I guess they went back to the drawing board?

 

[ChrML]

With traditional software engineering, that's true. It becomes less so with machine learning, though. No matter what, each type of car is going to need a different set of networks, due to differences in the positioning of the sensors(it's possible to make a single set of networks that can generalize between hardware configurations, but then you're wasting substantial portions of the capacity of the network on configurations any one particular car will never see). If you're already supporting different networks, then using radar vs a camera just becomes swapping out(or adding) one of the inputs and letting the network learn how to use it. Adding functionality to the existing networks is a matter of just adding additional layers or units in the latter part of the network and letting the training process determine if/how to use the radar vs. rear camera.

The most likely scenario, I think, is still the second one I gave: that some difference(maybe resolution or angle) in the rear camera made it more necessary to have rear radar on the Model 3, but the pessimistic one I gave here isn't as onerous as one might expect.

EDIT: I'll add this: I really don't think it's likely that this radar is something that's being determined to be necessary for FSD. Human drivers do fine the vast majority of the time even with the entire rear view blocked. It's not a terribly useful viewpoint, generally, because there's not usually any needed reaction to anything happening directly behind you. A simple camera should do fine here, and I've successfully used the rear camera display, myself, to stand in for a blocked rear view mirror multiple times. At absolute most, I could see the rear radar being used for some kind of rear-ending-avoidance functionality in the future.

I don't agree. Creating, validating and approving camera-based FSD is extremely hard on a single hardware setup. Doing it on multiple setups is even harder and more work, which would be a waste of resources and increase the company risk.

Machine learning also does not work like that. You can't connect any sensor to an NN blackbox and expect it to learn how to use it. To teach a machine something, you need some way to validate your results and provide feedback to the network. That cannot be done in a single car basis.

Best way of achieving FSD is using the cars to gather a massive amount of data/footage. The footage most likely have to be manually tagged. Train the image recognizing NNs on the server, download the pre-trained networks to the cars and start verifying bits and pieces. Train small modules to do their work perfectly individually, then put them together with conventional software code. Then verify what the computer would have done matches driver input on every car. Now your halfway, now you have to verify and convince yourself and the government that the FSD is 100% reliable in every situation on your particular hardware setup (oh multiple setups? poor you...).

 

[lunitiks]

@verygreen Anything about the cameras???

Speculation in my 2.0 cameras capabilities thread about this

 

[lunitiks]

Not just speculation BTW, a member here posted an all new part number for the side-repeater or b-pillar cam. HW2.5 car

 

[verygreen]

@verygreen Anything about the cameras???

Speculation in my 2.0 cameras capabilities thread about this

The only solid thing I know about the cameras is that there's now code to have different color filters in addition to RCCC we observed on HW2. This seems to corroborate all the reports that HW2.5 cameras are different from HW2.0 - perhaps different color filter (referred to as 'cameras are not color'), but we need to capture a snapshot from hw2.5 to know what it actually reports.
RCCB and RGGB patterns are now supported too

 

[lunitiks]

Here it is - new part number for repeater cam: AP2.0 Cameras: Capabilities and Limitations?

 

[ChrML]

The only solid thing I know about the cameras is that there's now code to have different color filters in addition to RCCC we observed on HW2. This seems to corroborate all the reports that HW2.5 cameras are different from HW2.0 - perhaps different color filter (referred to as 'cameras are not color'), but we need to capture a snapshot from hw2.5 to know what it actually reports.
RCCB and RGGB patterns are now supported too

Might be code to normalize the pictures taken between the different cameras (old 2.0 and new 2.5). In order to be able to use them on the same Neural Networks.

 

[lunitiks]

So what we're potentially talking about here is:

1. New APE
2. New camera sensors
3. New forward radar (Conti)
4. New rear-facing radar
5. New e-fuse
6. New power steering ECU
7. New park assist ECU

?????

That’s one heck of a «wiring redundancy»...

Retrofitting this onto a 2.0 car is beginning to look as likely as retrofitting 2.0 to a 1.0 car!

 

[verygreen]

Might be code to normalize the pictures taken between the different cameras (old 2.0 and new 2.5). In order to be able to use them on the same Neural Networks.

that's the weird part. They do use same NN on both, but while they can autodetect what cameras are installed and use them as is, they also have a map of what color filters are valid for what hw configuration and if you have wrong camera - ape would throw out an error (widely reported).

The only theory I have for this is they DO plan to diverge the NNs and have old-style cams and new-style cams NNs be separate eventually, otherwise what they are doing does not make much sense.

 

[ValueAnalyst]

Who knows. Evidently their current attempts did not pan out to produce the results they were hoping for, so I guess they went back to the drawing board?

oh dear... maybe @BigD0g is right

 

[strangecosmos]

Of course, there is always the third choice: They think a rear facing radar improves FSD, but isn't necessarily required for it. That would mean these newer cars get improved FSD, and the older ones are stuck with the older version.

Bingo. The hardware will continuously improve quarter after quarter, year after year, indefinitely. It won’t ever stop. The floor is what hardware enables a car to drive significantly better than the average human. The ceiling is essentially unlimited and unreachable.

 

[chillaban]

Bingo. The hardware will continuously improve quarter after quarter, year after year, indefinitely. It won’t ever stop. The floor is what hardware enables a car to drive significantly better than the average human. The ceiling is essentially unlimited and unreachable.

Another evil possibility: Cars with rear radar will get FSD or EAP lane changing capabilities first, and eventually they will have a camera-only solution. Not unlike the front side of the car, where you're now starting to see MobileEye and others propose camera-only ACC systems not backed by radar or LIDAR, but all shipping products use one of these additional sensors as the speed/distance measurement yardstick.

Is it possible Tesla is planning for HW2.5 or 2.5+ to be the first hardware to receive additional functionality, with the promise that eventually they'll be confident enough to offer the same to the HW2.0 suite? I think that will tick off a lot of customers (if not open the door to additional litigation)….

 

[BigD0g]

Bingo. The hardware will continuously improve quarter after quarter, year after year, indefinitely. It won’t ever stop. The floor is what hardware enables a car to drive significantly better than the average human. The ceiling is essentially unlimited and unreachable.

Except, they haven’t delivered anything working yet, but still adding more sensors, if they had such high hopes why spend all the time and resources and development hours for a platform you haven’t even delivered. Not to mention, if you keep adding more and different sensors your code base becomes quite the mess.

 

[BigD0g]

Another evil possibility: Cars with rear radar will get FSD or EAP lane changing capabilities first, and eventually they will have a camera-only solution. Not unlike the front side of the car, where you're now starting to see MobileEye and others propose camera-only ACC systems not backed by radar or LIDAR, but all shipping products use one of these additional sensors as the speed/distance measurement yardstick.

Is it possible Tesla is planning for HW2.5 or 2.5+ to be the first hardware to receive additional functionality, with the promise that eventually they'll be confident enough to offer the same to the HW2.0 suite? I think that will tick off a lot of customers (if not open the door to additional litigation)….

I think as soon as the release to just say 2.5 folks and it’s not a fast follower for 2.0 the crap hits then fan, since we were all promised the hardware capable of FSD at a level significantly better then a human driver. Unless ofcourse Tesla does the spinal transplant on our 2.0 cars to bring them to 2.x

 

[chillaban]

I think as soon as the release to just say 2.5 folks and it’s not a fast follower for 2.0 the crap hits then fan, since we were all promised the hardware capable of FSD at a level significantly better then a human driver. Unless ofcourse Tesla does the spinal transplant on our 2.0 cars to bring them to 2.x

As a 2.0 owner, I totally agree. Don't get me wrong, at least with the recent AP2 firmware updates I'm happy with the performance of AP1-class features, but the moment that a newer hardware revision gets more features than HW2 without a retrofit path being offered, that's when I'll take out my pitchfork and join the disgruntled crowd.

But yeah, if it turns out they need any of the new hardware, I don't see any route other than a spinal transplant or a really really shiny trade-up offer for those that pre-purchased FSD.

Heck, even for those who didn't buy the FSD option, it sure sounds like Tesla shot themselves in the foot with the "all cars … have the hardware for full self driving" marketing line.

 

[verygreen]

Of course Tesla position of "if it's not in manual, it was not promised" might also mean they'll keep evolign their sensor suite and compute HW hoping nobody notices (it's not like they were exactly forthcoming about their HW2.5 if you remember), and only much later when things start to pan out they'll need to worry about all the older cars (of which some number would not make it to that date for one reason or another, soem would be traded in back to Tesla and such).

We have seen this pattern play out on AP1 to a degree, so I cannot say I am entirely shocked by what's unfolding.

 

[stopcrazypp]

As a 2.0 owner, I totally agree. Don't get me wrong, at least with the recent AP2 firmware updates I'm happy with the performance of AP1-class features, but the moment that a newer hardware revision gets more features than HW2 without a retrofit path being offered, that's when I'll take out my pitchfork and join the disgruntled crowd.

I think it's almost guaranteed that the 2.5 will have more features than 2.0. I don't think Tesla promised newer revisions of their hardware won't have more features. The question is can the 2.0 hardware (other than processing hardware, which Tesla promised to replace if necessary and is relatively trivial to replace) provide all the functionality/features they promised.

But yeah, if it turns out they need any of the new hardware, I don't see any route other than a spinal transplant or a really really shiny trade-up offer for those that pre-purchased FSD.

Heck, even for those who didn't buy the FSD option, it sure sounds like Tesla shot themselves in the foot with the "all cars … have the hardware for full self driving" marketing line.

The FSD is different issue. They can for example have both hardware configurations give some bare minimum FSD functionality, but the 2.5 hardware having more features or be better at it.

 

[chillaban]

I think it's almost guaranteed that the 2.5 will have more features than 2.0. I don't think Tesla promised newer revisions of their hardware won't have more features. The question is can the 2.0 hardware (other than processing hardware, which Tesla promised to replace if necessary and is relatively trivial to replace) provide all the functionality/features they promised.


The FSD is different issue. They can for example have both hardware configurations give some bare minimum FSD functionality, but the 2.5 hardware having more features or be better at it.

I'm drawing a distinction between more features and first features though. The specific scenario that concerns me is what if AP2.5+ (the variant with rear radar) is what's required for the first generation of EAP "automatic lane changes", and it takes then another year or two to develop a vision-only based approach to detecting fast approaching neighboring cars? What would be considered the fair option to AP2.0 buyers? A retrofit of the rear radar, or waiting an indefinite amount of time while newer build date cars enjoy the feature? This would be a feature that was originally advertised at time of HW2.0 launch.

I strongly suspect, much to our dismay, this is likely going to be Tesla's approach.

 

[strangecosmos]

Unless I’m missing something in this thread, what was uncovered was a non-descript black box under the rear bumper of the Model 3, and some people here are jumping to the conclusion that it’s a radar unit. But couldn’t that black box be anything?