Autonomous Car Progress

[stopcrazypp]

What could go wrong? If they do that, or it looks like they're going to, I'm definitely going to sell my stock, or at least the shares that have cost basis over $40!

Well you can start selling it now, because they said as much to the CA DMV already. There is no goal for "Autosteer on City Streets" to be good enough that the driver doesn't have to pay attention (which would make it L3+), so any comparison to being better than a human driver goes out the window. It'll be like AP right now and a L2 feature that merely approximates how a human might drive, but have plenty of situations that it can't handle and will fail miserably if the driver was not there to intervene.

 

[diplomat33]

So you're saying the OEDR capability must be unlimited?

The L5 software must be able to recognize and respond (blah blah) to every single road object?

You're not very clear. You keep bringing up pedestrians, but we all know there are unlimited road objects / events to recognize.

I think I am clear. Pedestrians is just one example of OEDR. The OEDR must cover all relevant objects and events in the designated ODD. OEDR includes roads, signs, traffic lights, road debris, animals, pedestrians, cyclists, motorcycles, cars, trucks, construction zones, school zones, bad weather, etc...

Obviously, for L4, since the ODD can be limited, the OEDR may not need to cover as much, depending on the ODD. But for L5, since the ODD is unrestricted, the OEDR must cover all objects and events, yes.

 

[powertoold]

But for L5, since the ODD is unrestricted, the OEDR must cover all objects and events, yes.

Cool, so if L5 isn't able to recognize / respond to a single road object / volcanic lava, then it's not level 5?

 

[AlanSubie4Life]

Well you can start selling it now, because they said as much to the CA DMV already.

No, it’s likely fine as long as it is aspirational L2. The market mostly understands that it’s an L2 product (I think, but doesn’t really matter - it’s priced in whatever it is). The release is where it gets a bit sticky and consequences are less properly priced in. Anyway, don’t want to divert into that discussion too much - what stock prices do are what they do.

My point was that it seems really dangerous, and I’m not sure Tesla is going to be wild to do wide release - even when it meets your described threshold of quality. They certainly can release it at that point but whether they will is unclear.

I'm not sure it makes any sense to take that risk, if competitors actually end up with true L4/L5 vehicles, broadly deployed (an open question), because then it would make sense for Tesla to (quickly!) reproduce that success, rather than fiddling around with "end-to-end L2" with all the issues that come with it.

But if others don't actually end up with true success in that arena in the near future, Tesla can drag this out for a bit.

 

[diplomat33]

Cool, so if L5 isn't able to recognize / respond to a single road object / volcanic lava, then it's not level 5?

The road object needs to be relevant to driving of course. But if it can't recognize a road object that is relevant to driving no, it would not be L5 (IMO).

 

[powertoold]

The road object needs to be relevant to driving of course. But if it can't recognize a road object that is relevant to driving no, it would not be L5 (IMO).

Ok, this is wrong, and it's not my opinion, lol.

 

[daniel]

Out of beta here means just out of the invite-only limited beta that only the Early Access Program members have access too. This is separate from the marketing "beta" label applied to features already in public release.

Thanks for the clarification.

One of the amazing thing that Mobileye is trying to do by 2025 is have a consumer car that’s L4 and works on the majority of all US/EU roads but may still requires a licensed driver for situations where the system comes to a safe stop after achieving a minimal risk condition. Think of the human driver/passenger doing the job of a teleoperator. That means the driver/passenger can sleep, watch a movie, do homework, prep for presentation and be notified if they need to become the driver to finish the drive or the system runs into a situation that it needs input on.

I hope they succeed. The above is exactly what I want to own before I get too old to drive safely. I'd just note that "trying to do" is not the same as "will do." But the above does agree with what I've been saying all along: That companies like Mobileye that are not car makers still want to make L4 driving available to consumers, probably by licensing technology to, or partnering with, car makers.

Perhaps. End of this year would actually be a year late!
Just to be clear, I am joking!
My prediction is that HW3 will never be achieve robotaxi safety 200% greater than the average human or whatever the current goal is. In other words, FSD on HW3 will never be out of beta.

Thanks. And I agree that HW3 will not be FSD-capable. Maybe HW7 or HW12 will achieve it. And meanwhile, Tesla will continue to build the best cars on the road. I really hope that Tesla is among the first with true autonomy, because I'd hate to have to trade in my Tesla for a Ford or a Chevy just to get real FSD.

 

[idriveacar]

Thanks for the clarification.



I hope they succeed. The above is exactly what I want to own before I get too old to drive safely. I'd just note that "trying to do" is not the same as "will do." But the above does agree with what I've been saying all along: That companies like Mobileye that are not car makers still want to make L4 driving available to consumers, probably by licensing technology to, or partnering with, car makers.



Thanks. And I agree that HW3 will not be FSD-capable. Maybe HW7 or HW12 will achieve it. And meanwhile, Tesla will continue to build the best cars on the road. I really hope that Tesla is among the first with true autonomy, because I'd hate to have to trade in my Tesla for a Ford or a Chevy just to get real FSD.

Out of curiosity, what do you think it is about the hardware that will prevent FSD from operating at 2x human safety on HW3, or, at the very least, HW4? There are known issues with camera positioning, but it seems like those should be corrected by the time HW4 comes out, now that Tesla has much more experience with what it is that they're building.

The compute resources are also at their limits, but I'm sure there is room for optimization still. Some of that optimization will come at the cost of reaction time and planning fidelity of course.

HW3/FSD computer 1 was a real shot in the dark. They've gotten so far with HW3, my personal belief is that even if greater than human safety isn't achieved on HW3, it will be on HW4. They've learned a lot since HW3 was designed.

 

[StealthLSU]

The problem is the levels are not great for sequential progression. I mean technically if you made a car that can only do a single 1 mile route and nothing else, and it is a straight line on 1 road. But It can do it well without a driver or intervention. It is a level 4 AV. It is useful or good? Nope.

Tesla is not competing with waymo, waymo is made to make money as a service that consumers do not own. Tesla is making money selling to consumers so they need it to work everywhere.

Honestly, level 4 is very attainable by many companies and if you are using it as your car there isn't much difference from 2 to 4.

But for every company the jump to level 5 is enormous. Waymo can't do highways or weather or 99% of the country. Tesla can go everywhere but not at a high level. I think all companies are a decade away or so from true level 5(all roads, all conditions)

 

[powertoold]

There is just mass misinterpretation and misunderstanding of the levels.

It's to the point where the levels are worthless in practice because everyone seems to think they have anything to do with performance or AV progress.

The levels only describe a software feature from an autonomy perspective. They don't create threshold criteria for weather / road conditions / safety / passenger comfort / etc.

Before I even fully understood the levels, I already had a negative opinion of them. Now, seeing everyone misinterpret them, it's clear they're worthless and a joke. I've never seen a definition be so confusing to almost everyone, even to people who've read the whole thing.

 

[daniel]

Out of curiosity, what do you think it is about the hardware that will prevent FSD from operating at 2x human safety on HW3, or, at the very least, HW4? There are known issues with camera positioning, but it seems like those should be corrected by the time HW4 comes out, now that Tesla has much more experience with what it is that they're building.

The compute resources are also at their limits, but I'm sure there is room for optimization still. Some of that optimization will come at the cost of reaction time and planning fidelity of course.

HW3/FSD computer 1 was a real shot in the dark. They've gotten so far with HW3, my personal belief is that even if greater than human safety isn't achieved on HW3, it will be on HW4. They've learned a lot since HW3 was designed.

I am of the opinion that the task is so extremely complicated that it will take far more sophisticated software than has yet been developed and far more computing power than present-day microprocessors have. I am also of the opinion that cameras alone will not be sufficient and that more types and numbers of sensors will be needed.

I dearly and desperately hope to be proven wrong.

I would happily wager $5,000 (but not with an unknown person on the internet from whom I could not hope to collect) that at no time in the next ten years will I be able to buy a car for less than $100,000 that can drive me on all my usual trips while I take a nap in the back. If I lose the best, I will win, because I'll have my dream car. If I win the bet, I'll lose, but I'll have $5,000.

 

[stopcrazypp]

No, it’s likely fine as long as it is aspirational L2. The market mostly understands that it’s an L2 product (I think, but doesn’t really matter - it’s priced in whatever it is). The release is where it gets a bit sticky and consequences are less properly priced in. Anyway, don’t want to divert into that discussion too much - what stock prices do are what they do.

My point was that it seems really dangerous, and I’m not sure Tesla is going to be wild to do wide release - even when it meets your described threshold of quality. They certainly can release it at that point but whether they will is unclear.

I'm not sure it makes any sense to take that risk, if competitors actually end up with true L4/L5 vehicles, broadly deployed (an open question), because then it would make sense for Tesla to (quickly!) reproduce that success, rather than fiddling around with "end-to-end L2" with all the issues that come with it.

But if others don't actually end up with true success in that arena in the near future, Tesla can drag this out for a bit.

There are zero competitors that are even approaching end-to-end L2 in the US (you might see some in China, but those are irrelevant to people here), nor are there any production vehicles planned for L4/L5 in the near future (there are only fleets, which is not what Tesla is competing with). Tesla's logic is doing "end-to-end L2" is a necessary step to reach L4/L5, so they must get that out first. They aren't going by the logic of others of skipping directly to L4 or L5.

As for Tesla dragging this out, people are running out of patience, especially with so many public videos available of FSD Beta. If they don't release something in the next year or two, I think the pitchforks will be out. The chances of them reaching L4/L5 in the next year or two approaches zero given their latest statement to CA DMV, so the only logical route forward is to release end-to-end L2 first. And on the subject of hardware, I think Tesla will very much not want to do a HW4 retrofit for HW3 cars, and they clearly don't need it if they stick with end-to-end L2 (they might if they insist on waiting until L4).

As relevant to this thread, Mobileye wants to release end-to-end L2 also, so Tesla is not alone in this.

I think only L3 is something that automakers may see no point in releasing, as either it's going to be extremely limited in scope (traffic jam only in a limited subset of vehicles) or it's going to be too big a liability if expanded in scope (lawyers will probably want L4 to eliminate the handover ambiguity).

 

[Daniel in SD]

Nope, FSD Beta is only for the City Streets function, which is a L2 system. When it is out of the limited beta, it'll be to deliver the "Autosteer on City Streets" bullet point on the current order page. It will not be the features that were promised in earlier FSD order pages that suggested L4. I actually discussed this before with you, with reference to the DMV filing that explained it.
FSD Beta Videos (and questions for FSD Beta drivers)

I also gave my analysis of the 3 different stages that Elon talked about in the past. City Streets (end-to-end L2) can basically be released when it's "feature complete". This does not require 200% better than average human, only basic competence in day-to-day local driving (such that it's a feature rather than an annoyance, for example, it currently failing more than or approaching half the time on fairly basic tasks like unprotected lefts is not competent enough yet).

Stage 2 is L3+:
"feature complete to the degree that … where we think that the person in the car does not need to pay attention"
Tesla was clear in the DMV document that "Autosteer on City Streets" will be purely L2 on public release and not L3+, so it's not going to match this criteria.

Elon thinks that after reaching "feature complete" that it'll be only a matter of gathering data to improve reliability to reach L4/L5, which is likely not realistic (at least in near term time scales), so I don't see Tesla holding out releasing City Streets while waiting for that goal. Their filing with DMV suggests they would release City Streets first before they move to the next step of L4/L5 development.

I'll believe that when Elon says that FSD Beta is not a beta of FSD.

 

[stopcrazypp]

I'll believe that when Elon says that FSD Beta is not a beta of FSD.

So you'll just turn a complete blind eye to what Tesla is saying to CA DMV? I encourage you to at least read it and then see if your mind still says the same.
PlainSite :: Documents :: California DMV Tesla Robo-Taxi / FSD E-Mails
PlainSite :: Documents :: California DMV Tesla Robo-Taxi / FSD Notes

And do you really seriously believe Tesla will not do any further public release of FSD option features (as they have done already, like for example the Traffic Lights features) until they get to L4?

 

[AlanSubie4Life]

And do you really seriously believe Tesla will not do any further public release of FSD option features (as they have already done already, like for example the Traffic Lights features) until they get to L4?

I’m pretty sure he’s read those documents. Who hasn’t here?

It’s interesting to think about what other L2 features they might be able to release - my opinion (hope?) is that they will attempt to not compromise per-mile safety, but I am not sure how accurately they can measure that, with their current highly selected user base with specific demographic characteristics, and they must compare to what they will be replacing (rather than to generic NHTSA data).

I think they might end up getting mostly stuck releasing new active background safety features at some point, or maybe a few piecemeal bits (not sure what) of point-to-point L2 which make it not point-to-point L2, but I’m probably wrong. All depends how many safe miles these testers can rack up, I suppose. And they of course have to carefully analyze interventions to see whether any will be problematic when released to a wider audience.

 

[powertoold]

The whole point of the emails is to persuade the DMV to not require any additional permits for expanding the fsd beta / City Streets feature. And to persuade the DMV that fsd beta isn't autonomous, lol, again, to avoid any permit requirements. I wouldn't place any weight into what Tesla says wrt their future plans.

Fsd beta can recognize and respond to all of the things Tesla says it can't in the emails (except maybe emergency vehicles):

 

[daniel]

I’m pretty sure he’s read those documents. Who hasn’t here?

Me. I haven't. That first document is 300 pages! That's a frigging novel. I read (and enjoyed!) War and Peace. This "novel" I'd never make it through.

 

[AlanSubie4Life]

Me. I haven't. That first document is 300 pages! That's a frigging novel. I read (and enjoyed!) War and Peace. This "novel" I'd never make it through.

I meant the relevant section, of course.

 

[Daniel in SD]

So you'll just turn a complete blind eye to what Tesla is saying to CA DMV? I encourage you to at least read it and then see if your mind still says the same.
PlainSite :: Documents :: California DMV Tesla Robo-Taxi / FSD E-Mails
PlainSite :: Documents :: California DMV Tesla Robo-Taxi / FSD Notes

And do you really seriously believe Tesla will not do any further public release of FSD option features (as they have done already, like for example the Traffic Lights features) until they get to L4?

The description provided to the DMV is entirely consistent. It's certainly not feature complete yet but it is a beta of FSD.
I'm skeptical that they'll release FSD Beta because I think they'll find that it does not improve road safety. I really have no idea and I don't think they do either which is why they're now talking about a slow rollout. They may also release a subset of FSD Beta, perhaps "Autosteer on City Streets".

 

[stopcrazypp]

The description provided to the DMV is entirely consistent. It's certainly not feature complete yet but it is a beta of FSD.
I'm skeptical that they'll release FSD Beta because I think they'll find that it does not improve road safety. I really have no idea and I don't think they do either which is why they're now talking about a slow rollout.

FSD Beta IS "City Street Pilot". Both documents refers to the two interchangeably. I don't think Tesla will care that much if it necessarily improves road safety significantly, especially to a level of 200% or more (vs not using it at all), as that's very hard to prove. Even with the AP safety stats right now, there are lots of naysayers saying there's too many caveats to the data. The fact is they have that feature as a bullet point on the order page that they have to deliver and I don't see them waiting for L4 to release it.

As long as they make it abundantly clear that the feature requires the driver's attention (there's talk of instructional videos in DMV docs, perhaps Tesla may make it mandatory to activate the feature, something some makes in China is already doing for their more advanced L2 features), keep the driver on their toes, and utilize all driver attention detection systems available (like mandatory camera monitoring as they are doing for FSD Beta), this can probably reduce the chance of a severe accident (as seem to have done so far with FSD Beta), such that at least it's not seen a feature that is hugely dangerous (as some people are arguing AP is).

They may also release a subset of FSD Beta, perhaps "Autosteer on City Streets".

The current FSD Beta is a pilot program for "Autosteer on City Streets" where they said explicitly that final full release of the feature will remain SAE L2. So they already said that is what they planning to do. There is zero timeline given for a L3+ feature. Perhaps they will recycle the "FSD Beta" name for the beta for that, but it's clearly not what the current testing is testing for, at least according to their filings.

Please describe and provide any relevant documentation reflecting Tesla’s intended functionality for the final release of FSD City Streets to the general public. Specifically, which of the limitations associated with the OEDR described in the letter dated November 20, 2020 will continue to be part of the final release of FSD City Streets to the general public?
While the current pilot version of City Streets is still in a validation and review stage, we expect the functionality to remain largely unchanged in a future, full release to the customer fleet. We are analyzing the data obtained in the pilot and using it to refine the feature’s operation and customer experience. We will continue to make refinements as necessary, and only after we are fully satisfied with performance, integrity, and safety will we release the feature to the customer fleet. That said, we do not expect significant enhancements in OEDR or other changes to the feature that would shift the responsibility for the entire DDT to the system. As such, a final release of City Streets will continue to be an SAE Level 2, advanced driver-assistance feature.
Please note that Tesla’s development of true autonomous features (SAE Levels 3+) will follow our iterative process (development, validation, early release, etc.) and any such features will not be released to the general public until we have fully validated them and received any required regulatory permits or approvals.