diplomat33
Average guy who loves autonomous vehicles
No they don't. Tesla has released their best software called FSD Beta. And it does not do the entire OEDR. Tesla said so in the emails to the CA DMV.
-
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
No they don't. Tesla has released their best software called FSD Beta. And it does not do the entire OEDR. Tesla said so in the emails to the CA DMV.
shrineofchance
she/her, they/them
diplomat33
Average guy who loves autonomous vehicles
See above. L4 ≠ "better".
By "role of the driver", I don't mean legally. The levels define the driving tasks that the driver and the system are responsible for. That is why the levels are about the role of the driver and the system. They define what the driver and the system do. But they don't define how good the system should do the driving tasks. They merely define what driving tasks the system should do.
shrineofchance
she/her, they/them
As you said: “The levels have little to do with performance. They are about the role of the driver and the system.”
The same software that is L2 in consumer's hands can be L4 internally simply by being put in dev mode.
So therefore if Tesla turns off the driver monitoring system in their software, shifting the responsibility for the driving tasks from the driver to the software, they have changed the intent from L2 to L4.
shrineofchance
she/her, they/them
diplomat33
Average guy who loves autonomous vehicles
The software that is L2 in consumer's hands cannot do the entire OEDR, per Tesla's emails to the CA DMV. So putting it in dev mode would not make it L4. It would still be L2 since it can't do the entire OEDR.
Put differently, the role of the driver and the system is about what driving tasks the driver and system do. Tesla switching to dev mode does not make it suddenly be L4. Unless the dev mode can do more driving tasks than the L2 version that consumer's have, then it is not L4.
diplomat33
Average guy who loves autonomous vehicles
It would still be L2, just L2 with no driver monitoring.
Sure, Tesla could turn off driver monitoring but the current software cannot do the entire OEDR so that would be extremely dangerous to do.
shrineofchance
she/her, they/them
Elon first publicly disclosed the existence of FSD v9 in March. Who knows how long before that the underlying code was in development and testing. Elon is saying a closed beta release of v9 might not happen til June or July. July would be 4 months from public disclosure to closed beta release.
And per Elon, v9 is very different from v8.2:
The lag between Tesla's internal testing and fleet-wide release will probably in this case be at least 6 months.
I want to banish the idea that what Tesla owners are getting in their cars is the same software version that engineers are testing.
And per Elon, v9 is very different from v8.2:
The lag between Tesla's internal testing and fleet-wide release will probably in this case be at least 6 months.
I want to banish the idea that what Tesla owners are getting in their cars is the same software version that engineers are testing.
diplomat33
Average guy who loves autonomous vehicles
You are changing the topic. We are talking about whether Tesla has L4, not whether the newer software will be better. Of course V9 will be better.
And are you saying you think V9 will do the entire OEDR and therefore be L4? That is wishful thinking. We know it will be better but we don't know it will do entire OEDR. We don't know V9 will be L4.
To be pedantic, we know it will not be L4. What we don't know is whether the reason that it will not be L4 is because there are truly required tasks that it is known to be unable to do or merely because they don't yet trust its ability to do them sufficiently to take on the additional liability.
shrineofchance
she/her, they/them
We (and Tesla) are prevaricating between an objective, binary can/can't distinction and an understanding of "can" vs. "can't" as a subjective discernment about what is wise to allow vs. disallow based on objective capability which exists along a continuum.
Take the example of "construction zones" (mentioned in the letter). We see from Tesla owner vlogs that Teslas can autonomously handle some construction zones sometimes. The notion that Tesla's code can simply never handle any construction zones ever is simply false.
What I gather the rep from Tesla is trying to say is: we do not consider it wise to allow the code to driverlessly handle construction zones upon initial public release.
Conversely, we see that Waymo's vehicles sometimes struggle to deal with construction zones, to the point that Waymo carefully constrains routes in order to avoid routes. This is an objective failure, but as we established, even a vehicle with a 100% objective failure rate can be Level 4, since it's a classification not based on real technical capability as we understand it in the everyday sense.
Tesla's code can handle more than 1% of construction zones and Waymo's code can handle 100%. We don't have the data to say what exactly their success/failure rates are.
Bottom line, an account of autonomy software that says "L4 = better" is reductionist and simplistic to the point of being false.
Hypothetically, a company with only L2-classified software could be doing 10x better than a company with only L4-classified software.
So, we can't decide the question of whether Karpathy's thesis that the ceiling on NN performance is simply data based on what software is classified as L2 or L4.
shrineofchance
she/her, they/them
Because L2 and L4 classifications are not based on objective performance, the argument here is not cogent.
shrineofchance
she/her, they/them
It is a separate (but related) topic. I simply want to dispel the common notion that the software Tesla engineers test and that Tesla owners drive is essentially the same. On the contrary, engineers can spend 6 months testing a radical rewrite that Tesla customers haven’t touched yet.
Just so everyone knows @shrineofchance is @strangecosmos and you can see his entire post history here.
Its good that you know who you are really interacting with. You are interacting with someone who believes that what Waymo is doing in phoenix by ferrying the paying general public in a robot-taxi with no human driver is a PR stunt and is trying to pay people to come up with a paper so that he can falsely use in his next seeking alpha article to say that what Waymo's doing in phoenix is a PR stunt and that Tesla is doing exactly the same privately.
This is similar to the articles he has written over the years.
For example:
Tesla Has An Immense Lead In Self-Driving (2017)
Tesla Leapfrogs Self-Driving Competitors With Radar That's Better Than Lidar (2017)
- While competitors remain reliant on lidar, Tesla has dispensed with any need for lidar by developing a radar system with superior performance to lidar.
- Replicating Tesla’s advancements in radar requires fleet learning. No car company or software company besides Tesla has any experience implementing fleet learning in a production environment.
- Since Tesla developed its radar system two years before affordable lidar will be available, and since it has other advantages, it seems likely that Tesla will achieve full self-driving in its production cars one to two years before any competitor.
- The market has yet to fully grasp how far ahead Tesla (TSLA) is in self driving
Is Tesla Building A Moat With HD Maps? (2017)
- Tesla is using its fleet of Hardware 2 cars to create high-definition maps of roadways. No competitor can do this.
- This creates a network effect wherein Tesla's advantage in HD maps allows it to develop better self-driving software and roll it out to more geographical areas sooner
We also know that the "PR spectacle" team that Tesla setup is only 1 version behind what's on Elon's car. Elon uses the same version of software that is on the internal employee testers cars. They simply just skipped 8.3 for their "PR spectacle" team.
There is no major lag between Tesla's internal testing and the "PR spectacle" team release and its definitely not 6 months.
The difference between versions have been week/weeks not months.
8.3 would had released in march before they shelved it.
This put the "PR spectacle" FSD beta team at most 1 months behind the FSD Beta development branch.
Elon and internal employees got access to 9.0 in mid/late April.
There is no major lag between Tesla's internal testing and the "PR spectacle" team release and its definitely not 6 months.
The difference between versions have been week/weeks not months.
8.3 would had released in march before they shelved it.
This put the "PR spectacle" FSD beta team at most 1 months behind the FSD Beta development branch.
Elon and internal employees got access to 9.0 in mid/late April.
diplomat33
Average guy who loves autonomous vehicles
To be pedantic, we know it will not be L4. What we don't know is whether the reason that it will not be L4 is because there are truly required tasks that it is known to be unable to do or merely because they don't yet trust its ability to do them sufficiently to take on the additional liability.
I am just going by what Tesla told the CA DMV:
"For context, and as we’ve previously discussed, City Streets continues to firmly root the vehicle in SAE Level 2 capability and does not make it autonomous under the DMV’s definition. City Streets’ capabilities with respect to the object and event detection and response (OEDR) sub-task are limited, as there are circumstances and events to which the system is not capable of recognizing or responding. These include static objects and road debris, emergency vehicles, construction zones, large uncontrolled intersections with multiple incoming ways, occlusions, adverse weather, complicated or adversarial vehicles in the driving path, unmapped roads. As a result, the driver maintains responsibility for this part of the dynamic driving task (DDT)"
Note the bold parts. Tesla did not say "FSD Beta is L2 because we don't trust it to handle some stuff". They said "FSD Beta is L2 because it is not capable of doing some OEDR".
willow_hiller
Well-Known Member
So far, @shrineofchance has been perfectly forthcoming with her motivations for posting.
And as an aside, using incorrect gender pronouns after she's already politely corrected someone within the last page feels intentional and mean.
But it's rather hypocritical for you to write this entire mini-exposé when I don't believe you've ever shared anything similar about yourself; Other than vague notions that you're an expert, and we should trust your opinions sight-unseen.
Since you've taken all that time to inform us about someone else's background, could you kindly share yours?
mikes_fsd
Banned
@diplomat33 now they do not say "on autopilot" or "with autopilot engaged or disengaged"
They use the word "autonomous" ...
And they make sure to indicate that it's not being tested in California
"various other locations around the world" when replying to California DMV...Again, to me, this shows that they're trying to bypass the California DMV bureaucracy!
diplomat33
Average guy who loves autonomous vehicles
@shrineofchance Since you say you are interested in measuring AV safety and are even commissioning papers to try to measure Waymo's safety, you might be interested in this.
SAFE just published yesterday a report called "A Regulatory Framework for Autonomous Vehicle Deployment and Safety". It discusses how we can determine when an autonomous vehicle can be deployed safely:
Here is the full report:
SAFE just published yesterday a report called "A Regulatory Framework for Autonomous Vehicle Deployment and Safety". It discusses how we can determine when an autonomous vehicle can be deployed safely:
Here is the full report:
Similar threads
- Article
