shrineofchance
she/her, they/them
There's lots of stuff that's in development long, long, long before it ever hits the public. I'm not sure if HydraNets have even still seen a fleet-wide release?
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Waymo
- Thread starter Daniel in SD
- Start date
There's lots of stuff that's in development long, long, long before it ever hits the public. I'm not sure if HydraNets have even still seen a fleet-wide release?
diplomat33
Average guy who loves autonomous vehicles
"Autonomous testing" is probably just a reference to the current FSD Beta in development, not to some super secret other L4 software. The most logical explanation is that the current FSD Beta is the only software that Tesla is testing. That is the latest software that Tesla has released to Early Access and talked about in their official company statements. And Tesla has told regulators that FSD Beta is L2. If Tesla considers FSD Beta to be L4 when they test it outside of CA but tell the CA DMV that it is L2 to avoid regulations, then they are being dishonest.
True. The datasets are just a sample. They are not the entire data that Waymo has.
I have no idea how many so-called L4 cars Tesla has, or consider they have in hallway discussions (perhaps still zero which doesn't mean that Tesla L4 is many years away). But from what I can gather it's kind of the wrong question to ask at this time.
Until they have a marketable product to claim as L4 FSD, Tesla will not represent it as such to customers, nor to CA DMV who would simply use said claim as an opening to demand data and oversight. I don't see why Tesla would open up to that. It's the old Forgiveness vs. Permission issue. (I'm not at all claiming that I figured this out on my own; it's been well-discussed as you all know.)
Then, the question boils down to whether Tesla derives any engineering development advantage from an internally-considered distinction of L2 vs. L4. I think not.
Doing essentially all testing, alpha / beta / whatever, as claimed-L2 with an engaged human monitor driver, doesn't prevent them from getting the same data. A lot of L2 driven miles, without interventions or disengagements, provide the same confirmation data and/or near-miss data as if the drives had been claimed L4 attempts. As far as I can see, the main exceptions to thar statement come in the famous grey area of 'I intervened because I wasn't sure the car would do the right thing, but probably it would have".
Added to this is the whole idea of Tesla's ability to compare the AI "shadow mode" decisions to the human driver's actual maneuvers. This data-set is available even when no L2 modes are engaged.
So, in my view Tesla is constantly evaluating L4 capabilities whether or not they advertise it as such. One can complain that it's in secret, but I think it's a pretty open secret, and a very sensible one.
shrineofchance
she/her, they/them
L4 vs. L2 is largely a matter of legal liability with regard to consumers, rather than a matter of inherent technical capability.
diplomat33
Average guy who loves autonomous vehicles
Wrong. It is all about inherent technical capability. L4 can do the entire OEDR. L2 cannot. So there is a big technical difference. Read the SAE document again.
Even if they have super secret software, why do we assume that's L4? L3 is still autonomous (and not allowed in CA without reporting). These testers could be testing L3 on highways only. The idea that line in the CA DMV document proves there are L4 vehicles / software being tested is a huge leap.
I'd even argue this is the most logical thing given what we know right now- Tesla has zero reason to give L2 city streets to 1900 employees and 71 customers if they have an L4 capable codebase somewhere. What is the L2 testing teaching you if you already have L4? However, they could totally be doing L3 highways while everyone else tests L2 city. And Texas is a good place for highway work.
And this would also explain how they believe they can suddenly drop no radar highway stuff on everyone us with no public beta test, even in the L2 city group, and why they have different city and highway codebases.
Heck, that actually excites me! Does Tesla maybe, actually have L3 on the horizon somewhere, and somehow Elon has been completely silent about it while tweeting endlessly about how great city streets L2 is?
shrineofchance
she/her, they/them
If L4 can do the entire OEDR, does that mean vehicles with test drivers aren't L4? If they are L4, why do they need test drivers?
Because they can't do the entire OEDR.
diplomat33
Average guy who loves autonomous vehicles
Sure. We know during Autonomy Day that they did the demo with no hands on wheel and reported it as L3. But I was merely responding to Shrine's claim that Tesla has as many L4 cars as Waymo which I dispute. And when I asked him to clarify, he said that Tesla's L4 fleet is outside of CA.
Reliability is an important metric as well.If L4 can do the entire OEDR, does that mean vehicles with test drivers aren't L4? If they are L4, why do they need test drivers?
Because they can't do the entire OEDR.
You can have a system that can do the whole OEDR nominally, but fails at a rate that is unacceptable to release as L4.
In fact, that's what I assume testing would be. You'd believe the vehicle is capable, but you're just out proving it.
Hence, if you say Tesla has a fleet of L4 cars, they have a fleet of vehicle systems they believe are capable of the whole OEDR, they just haven't proven it is reliable enough yet.
If they already know they aren't capable of the whole OEDR, then they aren't L4 vehicles, and it's not correct to say Tesla has a fleet of them. It's just a fleet of L2 or L3 cars that Tesla someday hopes will be L4 after they do a ton of work. Which by that definition and Tesla's current optimism means about 1M cars. Just not in CA
Last edited:
diplomat33
Average guy who loves autonomous vehicles
If L4 can do the entire OEDR, does that mean vehicles with test drivers aren't L4? If they are L4, why do they need test drivers?
Because they can't do the entire OEDR.
No. You do not understand the levels. L4 cars with safety drivers are still L4. L4 cars can do the entire OEDR and still have safety drivers. The safety drivers in a L4 car don't need to do the OEDR. Safety drivers are optional in L4 cars. So if there is a safety driver it is only because the company is still testing the safety of the L4. With L2, a safety driver is required since the safety driver has to do some of the OEDR in L2.
Basically,
L4 does the entire OEDR. Safety drivers are optional and don't need to do any OEDR. But safety drivers may be present for safety reasons.
L2 cannot do the entire OEDR. Safety drivers are required. Safety drivers do some of the OEDR.
shrineofchance
she/her, they/them
I said roughly as many. Tesla might have 100 test vehicles and Waymo might have 600. To me, that's roughly the same. Not 10, not 1,000, somewhere in the middle.
My pronouns are right under my username. She/her or they/them.
Unless it runs across a few cones in the road. Then it is no longer L4 because it can't handle the cones? Or are cones somehow not part of the OEDR?
shrineofchance
she/her, they/them
So, a car that crashes 100% of the time can be L4.
And a car that drives millions of miles in all road environments and weather conditions without intervention can be L2.
Hm.
FSD beta has only 150k miles of city driving experience with city driving enabled. It is, of course, built atop a system that has 3 billion miles, mostly on highways, but with probably at least half a billion miles of city driving from folks who turn on autopilot anyway.
powertoold
Active Member
yup, the levels have nothing to do with safety or performance.
shrineofchance
she/her, they/them
diplomat33
Average guy who loves autonomous vehicles
IMO, 100 is not roughly the same as 600. But even 100 cars would still be wrong. Tesla has zero L4 cars. Tesla has said that FSD Beta cannot do the entire OEDR, hence why it is L2. So unless they have secret software that can do the entire OEDR, they don't have L4.
I am so so sorry. My apology. I was not paying attention.
OEDR means Object, Event Detection & Response. The Waymo handled the OEDR just fine. It detected the cones and avoided hitting the cones just fine. It was a planning problem. The Waymo was not sure what lane to be in. So the problem was not a OEDR problem. And the problem was caused by the remote assistance giving the Waymo bad info. So again, not a problem with OEDR.
shrineofchance
she/her, they/them
diplomat33
Average guy who loves autonomous vehicles
Theoretically yes. The levels have little to do with performance. They are about the role of the driver and the system. So theoretically you could have a L4 car where the car is fully driving and is terrible and crashes all the time. And you could have a L2 system similar to AP, where the L2 system is really good at some tasks and the driver does the remaining OEDR and the combo of the two, the car goes millions of miles without intervention.
Although those are pretty extreme examples that would probably never happen in real life. For one, a L4 car that crashed 100% would never be released. Heck, even a bad L4 would probably avoid some crashes so it would probably do better than 100% crashes. And it is highly unlikely that a L2 system could go millions miles without any intervention.
Then why are they out testing L2 and letting people make videos of this failing when they internally have better stuff,
Heck, is it ethical to sell/deliver customers L2 when you have L4 code? You can call L4 code L2 all day long.
Similar threads
- Article
