HW4 sensor suite and hi-res camera updates?

[AlanSubie4Life]

People don't realise that when Elon says "safer than a human" he assumes with a fallback driver, and points at the "Safety Reports" that Tesla uses for marketing. This is L2 after all. HW3 will likely never perform the OEDR instead of the driver (as in autonomous driving, eyes off, L3+) in a meaningful ODD. And Tesla will never assume liability for driving on HW3 cars.

It all makes sense when you think of it this way. (and disregard all the robotaxi and L5 promises he also made).

So true. “Some people” laugh at this post and insist that FSD beta is a Beta of FSD, but such is not the case. Elon definitely lies about this all the time (effectively; even though they are not explicit lies in many cases, they are statements that are not what anyone who is trying to be honest would try to make). But obviously this is all L2 and that appears to be the objective for the next five years. (HW4 is about the same and will not be capable of anything additional, of course - according to Elon, anyway.)

 

[pilotSteve]

I have to say - drove 350 miles today (mostly highway/interstate) and had NO unplanned disconnects, no phantom braking, and very good comfort with how steering, braking and accelerating was quite smooth, never bad and only disconnected when I get into traffic circles, which I know it will screw up. So all in all – helpful and net less stress than I expected.

 

[Mike7564]

Is the HD Radar discussion separate or possibly linked to the HW4 transition? If so, weren’t we expecting some announcement my “mid Jan” to get ahead of the FCC release in early Feb?

 

[WyoDude]

Is the HD Radar discussion separate or possibly linked to the HW4 transition? If so, weren’t we expecting some announcement my “mid Jan” to get ahead of the FCC release in early Feb?

Nobody knows. Speculation that HD radar will be announced before Feb 7th based on the filing but could be something different or they could ask for another extension. Some think it will come with the HW4 update which currently seems pretty far away since it seems to be tied to CT launch. But also could possibly come with Highland update that is being estimated for Q3 or later

 

[Mike7564]

Nobody knows. Speculation that HD radar will be announced before Feb 7th based on the filing but could be something different or they could ask for another extension. Some think it will come with the HW4 update which currently seems pretty far away since it seems to be tied to CT launch. But also could possibly come with Highland update that is being estimated for Q3 or later

Well guess we’re in for an interesting (or boring) week ahead.

 

[spacecoin]

So true. “Some people” laugh at this post and insist that FSD beta is a Beta of FSD, but such is not the case. Elon definitely lies about this all the time (effectively; even though they are not explicit lies in many cases, they are statements that are not what anyone who is trying to be honest would try to make). But obviously this is all L2 and that appears to be the objective for the next five years. (HW4 is about the same and will not be capable of anything additional, of course - according to Elon, anyway.)

I mean people who think that HW3 (with blindspots, zero redundancy, no overlapping sensor modalities et.c) will be L3+ clearly do not understand what L3 is. All the "experts" in the Tesla bubble has been saying it's basically the same as Level 2 or "humans can drive with only two eyes", so that's understandable.

If anyone here thinks that HW3 will ever be driving while the human watches Netflix or reads a book - while Tesla takes responsibility/liability for accidents you need a reality check.

 

[argon2018]

I mean people who think that HW3 (with blindspots, zero redundancy, no overlapping sensor modalities et.c) will be L3+ clearly do not understand what L3 is. All the "experts" in the Tesla bubble has been saying it's basically the same as Level 2 or "humans can drive with only two eyes", so that's understandable.

If anyone here thinks that HW3 will ever be driving while the human watches Netflix or reads a book - while Tesla takes responsibility/liability for accidents you need a reality check.

Agree. And Elon has repeatedly said that. But who knows all of of his companies in particular spacex finally delivered on the last try. But I think on this one he f ed up big time

 

[Mardak]

people who think that HW3 (with blindspots, zero redundancy, no overlapping sensor modalities et.c) will be L3+ clearly do not understand what L3 is

How do those issues (blindspots, redundancy, modalities) prevent something from being a robotaxi? Or are you suggesting regulators and potential passengers wouldn't approve of a deployment with those issues even if it was safer overall?

 

[Mardak]

Elon didn’t say it couldn’t be done, he said it would be expensive

Right, he said "the cost and difficulty of retrofitting HW3 with HW4 is quite significant," and that sounds like in aggregate as do we expect it to be significantly more difficult than existing HW2 (computer and RCCC cameras) upgrade process to HW3 which can be done in about an hour by Mobile Service than say a 2023 vehicle wanting to upgrade to HW4?

 

[spacecoin]

How do those issues (blindspots, redundancy, modalities) prevent something from being a robotaxi? Or are you suggesting regulators and potential passengers wouldn't approve of a deployment with those issues even if it was safer overall?

No it's purely a technical issue.

1) You need a redundant computer, redundant cabling, redundant sensing in all directions if you're going to hand over the complete OEDR (object event detection and response - "identifying and reacting to things") to the car. There is no "take over immediately" in autonomy, including L3, since the whole point of autonomy is to free up time for the human, so she can work, read a book, watch a movie et.c.

2) At present we do not even trust computer vision in radiology without human oversight. The research are many breakthroughs away for vision only systems in safety critical applications. The current state is that it's, by itself, not safe enough. This is why people add physical measurement systems to compliment the ML range estimations (like Lidar, Radar), mapping (for ride quality and occlusions), expensive sensors with high resolution (to work at higher speeds and get better safety), expensive computers (to handle all that sensing input).

3) Tesla has been going at this now for over three years, and it can barely drive. The MTBF has been constantly under 10 miles. You need to 2000x that at the very least (probably a lot more), and if you do there is still the redundancy issue. Every ML system I know of has an S-curve, so adding more reliability gets exponentially harder as a factor of time.

Simply put - this system will never be autonomous to such a grade that Tesla will assume liability. On top of that, there is zero business incentive for Tesla to do so. They've already gotten the money from us upfront, and they decide when it's "done".

 

[spacecoin]

Right, he said "the cost and difficulty of retrofitting HW3 with HW4 is quite significant," and that sounds like in aggregate as do we expect it to be significantly more difficult than existing HW2 (computer and RCCC cameras) upgrade process to HW3 which can be done in about an hour by Mobile Service than say a 2023 vehicle wanting to upgrade to HW4?

My understanding is that the newer MCU3 cars use automotive ethernet instead of CAN-bus, and there are a lot of other differences between the "legacy" S/X and the new platform, such as the placement and form factor of the FSD computer, etc.

HW4 may have a different camera setup (location wise), which makes the upgrade even harder.

As with most hard problems, they can be solved with unlimited cash and unlimited time. But Tesla doesn't have either, nor a real interest to spend money on 4+ years old cars. So based on Teslas most recent communication, I do not think there is any chance of a retrofit.

 

[argon2018]

No it's purely a technical issue.

1) You need a redundant computer, redundant cabling, redundant sensing in all directions if you're going to hand over the complete OEDR (object event detection and response - "identifying and reacting to things") to the car. There is no "take over immediately" in autonomy, including L3, since the whole point of autonomy is to free up time for the human, so she can work, read a book, watch a movie et.c.

2) At present we do not even trust computer vision in radiology without human oversight. The research are many breakthroughs away for vision only systems in safety critical applications. The current state is that it's, by itself, not safe enough. This is why people add physical measurement systems to compliment the ML range estimations (like Lidar, Radar), mapping (for ride quality and occlusions), expensive sensors with high resolution (to work at higher speeds and get better safety), expensive computers (to handle all that sensing input).

3) Tesla has been going at this now for over three years, and it can barely drive. The MTBF has been constantly under 10 miles. You need to 2000x that at the very least (probably a lot more), and if you do there is still the redundancy issue. Every ML system I know of has an S-curve, so adding more reliability gets exponentially harder as a factor of time.

Simply put - this system will never be autonomous to such a grade that Tesla will assume liability. On top of that, there is zero business incentive for Tesla to do so. They've already gotten the money from us upfront, and they decide when it's "done".

So what happens with the promise that the car can be an income generator through Tesla ridesharing network making the vehicle an appreciating asset? It’s obvious to some of us that it will never happen with the current hardware. Even in Southern California with some rain the FSD beta warns that it can’t drive.
I thought these vehicles were future proof from 2018.
What does that do to promises made by Elon?
What does that do to Tesla company? They can’t just brush off all these people or people that are currently buying paying 16k

 

[JulienW]

So what happens with the promise that the car can be an income generator through Tesla ridesharing network making the vehicle an appreciating asset? It’s obvious to some of us that it will never happen with the current hardware. Even in Southern California with some rain the FSD beta warns that it can’t drive.
I thought these vehicles were future proof from 2018.
What does that do to promises made by Elon?
What does that do to Tesla company? They can’t just brush off all these people or people that are currently buying paying 16k

Robotaxi income is not in your contract and Tesla will just say it was an aspirational goal and not a promise. Elon never said" I promise" or "guarantee" you will make money.

Also Elon has said (not promised) literally 1000s of things across lots of industries that haven't happened yet. 2011: Musk: "I'll put a man on Mars in 10 years". Well.............looks like that didn't happen either.

Or Tesla could just keep saying it is still being worked on and will be delivered at some point when it is ready. The old by "the end of the year" every year on and on.

 

[EVRider-FL]

Lots of discussion about hardware requirements for autonomy, but from my experience with FSD Beta and NoA, many of the mistakes the car makes are not caused by hardware limitations. For example, the car often fails to move into the correct lane even though it sees the lane and in some cases knows it’s supposed to be there. The AP software has a long way to go before HW3 becomes the limiting factor.

Maybe that’s why Elon thinks true FSD can be achieved with the current hardware. I’m not saying I agree with him, but FSD could be improved a lot without hardware changes.

 

[spacecoin]

The AP software has a long way to go before HW3 becomes the limiting factor.

I partly agree. It would likely be improved by a lot if they incorporated HD-maps, and the planner is extremely bad right now.

But the fact of the matter is that they ran out of compute 18 months or so again, and are fighting the hardware limitations ever since, juggling NN:s in and out of memory, and using the extra SoC that was supposed to be for redundancy as an active node.

I don't think 1.2MP cameras are safe enough for higher speeds or for details such as pedestrian-intent. Blind spots are a real issue and there's blinding issues etc, but one might limit the ODD to get around some of those limitation.

I'd say "better" is definitely possible. There are a lot of room for improvements, such as handling of the speed signs. Right now it's not smooth at all. More reliable to a few orders of magnitude? Doubt it.

 

[Mardak]

this system will never be autonomous to such a grade that Tesla will assume liability. On top of that, there is zero business incentive for Tesla to do so.

It seems like all(?) Tesla leases now no longer allow buyouts/end-of-lease-purchases potentially originally with the intent that Tesla could build up a robotaxi fleet after first getting an upfront chunk of revenue from new vehicle sales then ongoing revenue from these "used" vehicles. Potentially robotaxis could provide more revenue than just reselling the vehicle as used even when accounting for insurance and liability, so that seems like non-zero business incentive to get it working on existing vehicles.

 

[Mardak]

It’s obvious to some of us that it will never happen with the current hardware. Even in Southern California with some rain the FSD beta warns that it can’t drive.

Do you think HW4 will solve the rain issue and that software with HW3 cannot?

 

[spacecoin]

It seems like all(?) Tesla leases now no longer allow buyouts/end-of-lease-purchases potentially originally with the intent that Tesla could build up a robotaxi fleet after first getting an upfront chunk of revenue from new vehicle sales then ongoing revenue from these "used" vehicles. Potentially robotaxis could provide more revenue than just reselling the vehicle as used even when accounting for insurance and liability, so that seems like non-zero business incentive to get it working on existing vehicles.

Wishful thinking. Look at the data. There is zero chance of HW3 becoming driverless. It's 100% BS. The science isn't here yet wrt computer vision. You can't get it safe enough.

If they get HW4 driverless autonomous in the LVCC loop with ingress/egress without human assistance within three years I'll be shocked, given where they're at right now and the rate of progress. Just take a moment to think about what's needed to get that right, in that small, low speed ODD.

 

[Mardak]

Look at the data. There is zero chance of HW3 becoming driverless.

Which data? People's reports of FSD Beta? Tesla's reported safety metrics? Reports to regulators for automated driving systems?

Failure or projected failure to achieve driveless from observations so far doesn't mean there haven't been learnings to be successful in the future potentially even with the same hardware. Notably, it seems like Autopilot/FSD Beta has gone through many foundational architecture changes that at least from a resourcing perspective, doesn't really make sense to focus too much on temporary comfort workarounds if the foundation is shaky and unsafe potentially requiring a different approach that obsoletes the comfort effort.

Maybe Elon Musk's innovation approach of failing fast and learning to try again wasn't appropriate for getting to driverless? But then again, maybe it would have taken even longer without the current approaches so far with HW3?

We don't really know what HW4 will change, but are you thinking the change won't be significant enough to get to driverless either?

 

[Daniel in SD]

Maybe Elon Musk's innovation approach of failing fast and learning to try again wasn't appropriate for getting to driverless?

That doesn’t sound like the approach Tesla is taking. Other AV companies change their hardware way more often.