FSD v12.x (end to end AI)

[sleepydoc]

I laughed at this. No it's not made up. It's literally in the UNECE regulation standard for ALKS (AKA L3):

"ALKS’s standard safety concept defines a 10 seconds transition period so that human driver must remain able to respond to a system request so that the human driver assume control of the vehicle when driving system do not do it anymore"

Automated lane keeping systems - Wikipedia

en.m.wikipedia.org

Also what Mercedes explicitly uses as I quoted up thread. That period was not randomly chosen, it was because the regulation that the EU follow a calls for it.

"the buttons in the steering wheel rim turn red, the vehicle requests the driver to retake control within ten seconds."

Will Mercedes jump to level 3 before Tesla? Looks like it.

Sleeping is not desired on 5 minute trips. But your design must still account for it (as well as for intermediate times like 30 minutes) if it explicitly allows sleeping. You can't rely on the "desire" of the owner. To me, a standby driver car is one that legally must leave its ODD while...

teslamotorsclub.com

As I noted multiple times upthread, SAE doesn't use 10 seconds, what they use is "at least several seconds" (J3016 3.12 Note 3; 3.17 Example 2)

SAE J3016 User Guide

In case you don't believe the source posted, I checked and it's right there in page 11 of the PDF document you posted!
"NOTE 3: At Level 3, an ADS is capable of continuing to perform the DDT for at least several seconds after providing the fallback-ready user with a request to intervene. The DDT fallback-ready user is then expected to resume manual vehicle operation, or to achieve a minimal risk condition if s/he determines it to be necessary. "

yes, see my posts above. you're misinterpreting the documents. Go back and re-read carefully.

 

[Daniel in SD]

It's the minimum number before the system starts a "minimum risk maneuver" (i.e. turns on the hazards and pulls over or stops) They do not list a maximum so saying "you need to respond within 10 seconds" is not accurate.

Driving with the hazards on or stopping on the highway for no reason is probably illegal so I would argue that you do need to respond within 10 seconds.
I suppose there is no true maximum though. What if you die while the system is operating? Then you never need to do anything.

 

[AlanSubie4Life]

Source please.

I think in order to implement motion simulators there has been a great deal of research on what a human body can detect and what it is susceptible to. Obviously such devices have been around for decades. This is not a new field of study and it is well established what can and cannot be detected and what is bothersome, etc.

Beyond that, do your own research on control systems for airplanes, trains, elevators, etc. These things have been around for decades and it’s important that certain general limits be adhered to in order to prevent nausea (etc.) in the general population.

Motion simulator - Wikipedia

en.m.wikipedia.org

I’ve mentioned this before and I’m now done cluttering up the discussion.

If FSD can get 5/5 bars on 95% of trips on the insurance apps for stopping score that’s probably an acceptable proxy too (maybe it is not - obviously depends on the app! - my point is that this be quantified, rather than being subjective). I’m sure that to some extent Tesla is using quantifiable measures. Once it gets above (or below, whatever) a certain threshold of g-force& jerk (aka comfort), except for emergencies, no one will really care anymore. At this point Tesla will stop talking about making it smoother. And I hopefully will be long gone here, and I am sure everyone will say good riddance. Basically everything I contribute here is worthless anyway.

 

[Dewg]

Driving with the hazards on or stopping on the highway for no reason is probably illegal so I would argue that you do need to respond within 10 seconds.
I suppose there is no true maximum though. What if you die while the system is operating? Then you never need to do anything.

For some reason your comment made me think that in many states it's illegal to kill yourself, which I find hilarious.

 

[sleepydoc]

They will. No question about it.

From your posting history I have serious doubts.

 

[sleepydoc]

So anyway... has anyone on 12.3.6 with 2024.20.3 having any issues with turn signals either not being used or turning off only after a few seconds before the car even reaches the turn? My Model Y RWD is doing it almost all the time and even if I manually turn them back on, it will even turn them right back off.

I've tried rebooting the car (twice) and though I can't imagine the whole recalibrating the cameras thing would matter, I'm almost to that point.

Yes - turn signals in the last version I had (can’t remember the version) were close to perfect but 12.3.6 is definitely not as good.

 

[sleepydoc]

Mine has become much more tolerant of it with 12.3.6, but she's still not super-happy about it. She especially complains about it missing potholes near certain corners, and the sometimes excessive acceleration taking off from stop signs in low-speed-limit areas. I still get a lot of "I wish you would just drive the car yourself", but she's allowing a lot more miles to happen while she's riding. With previous versions, it had got to the point where I would almost never engage FSDb with her in the car, because it wasn't worth the verbal assault.

My wife isn’t tolerating it; she genuinely doesn’t mind. I’m not the best gauge since I’ve been using FSD for 2-3 years but it really is comfortable to ride in.

 

[FSDtester#1]

My wife isn’t tolerating it; she genuinely doesn’t mind. I’m not the best gauge since I’ve been using FSD for 2-3 years but it really is comfortable to ride in.

My wife isn't a fan of FSD either, however my girlfriend thinks it's pretty cool.

 

[LSDTester#1]

My wife isn't a fan of FSD either, however my girlfriend thinks it's pretty cool.

Your wife doesn’t mind it in my car

 

[rlsd]

City stack can run on highway.
Tesla doesn't need to build 12.5?

https://twitter.com/x/status/1806090094551474519

 

[gsmith123]

Yeah, there are some divided highways in Colorado (typically with 55 speed limits) where V12 will engage. It works just fine, subject to my usual complaints with V12 (inconsistent following distance, lane positioning, and late braking for lead traffic). I've driven around 70 mph for extended periods on V12, on some roads.

 

[SSDTester#1]

I caught @SSDTester#1 voting multiple times for “complaining about FSD!”

Only once

 

[FSDtester#1]

Only once

View attachment 1060158

Let me get this straight, you left the operating room, took off your scrubs, drove home and replied to that post?
No love for me or @AlanSubie4Life ???
We have been posting pages of great material...

 

[stopcrazypp]

yes, see my posts above. you're misinterpreting the documents. Go back and re-read carefully.

No I'm not, read the passage. It says explicitly "The DDT fallback-ready user is then expected to resume manual vehicle operation, or to achieve a minimal risk condition if s/he determines it to be necessary."

This is not talking about the delay for the minimum risk maneuver, it's talking about the time the user must respond. Unless you consider double digit seconds to still qualify as "several seconds" (I don't), I don't see how that doesn't specify that sub-10 seconds is the time you must respond. And if that is the time you must respond, it is not compatible with sleeping (just like how all L3 systems out there currently excludes sleeping as an allowed activity).

And about the "at least" or "minimum" vs "maximum" discussion, all that means is that is the bare minimum standard that manufacturers are expected to meet. Manufacturers are allowed to design systems that give an even longer time for the driver to react.

 

[j0shm1lls]

The real question is...

WWJG12.4.2?

 

[sleepydoc]

No I'm not, read the passage. It says explicitly "The DDT fallback-ready user is then expected to resume manual vehicle operation, or to achieve a minimal risk condition if s/he determines it to be necessary."

This is not talking about the delay for the minimum risk maneuver, it's talking about the time the user must respond. Unless you consider double digit seconds to still qualify as "several seconds" (I don't), I don't see how that doesn't specify that sub-10 seconds is the time you must respond. And if that is the time you must respond, it is not compatible with sleeping (just like how all L3 systems out there currently excludes sleeping as an allowed activity).

And about the "at least" or "minimum" vs "maximum" discussion, all that means is that is the bare minimum standard that manufacturers are expected to meet. Manufacturers are allowed to design systems that give an even longer time for the driver to react.

*sigh* I tire of explaining the English language...

You appear to be citing the page 11 SAE document which says:
"At Level 3, an ADS is capable of continuing to perform the DDT for at least several seconds after providing the fallback-ready user with a request to intervene. The DDT fallback-ready user is then expected to resume manual vehicle operation, or to achieve a minimal risk condition if s/he determines it to be necessary."

"At least several seconds" - it says nothing about a specific time limit and only gives "several seconds" as a minimum for the car to continue autonomous operation. There is no specific definition of 'several' and certainly no legal definition. You your self state that "manufacturers are allowed to design systems that give an even longer time for the driver to react" so you're conceding my point.

 

[stopcrazypp]

It's the minimum number before the system starts a "minimum risk maneuver" (i.e. turns on the hazards and pulls over or stops) They do not list a maximum so saying "you need to respond within 10 seconds" is not accurate.

I looked it up, it seems they changed where they put the 10 seconds in the final spec.

In earlier drafts, the 10 seconds for the transition phase was in other sections:
"2.7.3. During the transition phase, the system shall continue to operate in the automated driving mode (including emergency manoeuvre) for at least [10 sec] unless the driver has resumed lateral control or has brought the vehicle to standstill by braking. The system shall maximize controllability by the driver."

"At the end of the transition phase [of at least 10 seconds], unless the driver has resumed lateral control [and is looking at the road ahead] or has brought the vehicle to standstill by braking, a minimum risk manoeuvre shall be performed."
https://wiki.unece.org/download/attachments/80380926/ACSF-24-08 (EC) draft proposal ALKS regulation based on ACSF-23-02r4.pdf?api=v2

As for what it means for transition demand and transition phase, the definitions:
"2.2. “Transition demand” is a logical and intuitive procedure to transfer the dynamic driving task from automated control by the system to human driver control. This request given from the system to the human driver indicates the transition phase.
2.2. “Transition phase” means the duration of the transition demand."

This is what separates L3 from L4. The driver is expected to take over. In L4 you are not expected to take over. That's why you can't sleep in an L3 car.

And looking at the spec, instead of arguing over the seconds, I see the standard also explicitly mentions and requires a driver availability system (the SAE doesn't). It seems this carried over to the final UN spec.
"6.1.3. Driver availability
The system shall detect if the driver is available and in an appropriate driving position to respond to a transition demand by monitoring the driver. The manufacturer shall demonstrate to the satisfaction of the technical service the vehicle’s capability to detect that the driver is available to take over the driving task."

Note in case you say "that only means it detects the driver is in the driver seat, doesn't mean you can't sleep!"

Nope, there is a separate section for that:
"6.1.2. Driver presence
A transition demand shall be initiated according to paragraph 5.4 if any of the following conditions is met:
(a) When the driver is detected not to be in the seat for a period of more than one second; or
(b) When the driver’s safety belt is unbuckled. The second level warning of the safety-belt reminder according to UN-R16 may be used instead of an acoustic warning of the Transition Demand."

To go back to the original argument, I think that pretty much settles it, the UN spec explicitly requires driver availability monitoring (as well as driver presence monitoring)! It's not even just implicit in the timing of the transition phase. On the flip side, SAE doesn't specify monitoring, but it specifies "several seconds" as the minimal time period for take over, which indirectly forces driver availability monitoring.

It doesn’t even force the need indirectly. What forces it is human stupidity. Since L3 only requires you to take over in a reasonable amount of time you don’t need to be paying attention at all and thus monitoring software should be unnecessary.

See above - they don’t need to detect a sleeping driver, just be able to handle one.

 

[enemji]

Also what Mercedes explicitly uses as I quoted up thread. That period was not randomly chosen, it was because the regulation that the EU follow a calls for it.

and we all are comfortable thinking that Mercedes had nothing to do with developing that regulation?

 

[stopcrazypp]

and we all are comfortable thinking that Mercedes had nothing to do with developing that regulation?

Hard to say, but good point. The German big three probably have considerable influence in the EU, so may have had the ability to influence this regulation (it seems the EU has the strongest influence in such regulations, even though they are UN regulations, as they adopt it most quickly; the US doesn't seem to follow it very often).

 

[JulienW]

My wife isn't a fan of FSD either, however my girlfriend thinks it's pretty cool.

This may be one of the first confirmations of an objective test of FSD 12.3.6 smoothness. So "she" can't tell if you are using FSD or manually driving even deflated and riding in the frunk?