FSD v12.x (end to end AI)

[stopcrazypp]

Again, as I said, I think it is fairly clear there is more to this issue than compliance with NHTSA. It’s hard to explain stops of one or two seconds (or whatever it is - needs to be measured) that way! I think all FSD beta users have experienced what I am referring to.

There’s also a perceptible jerk when coming to a complete halt when using the correct method. I’m not sure why this solution was not chosen.

We’ll see what their strategy here is for v12. Seems like one of the elements of everyday driving which is a good test of the autonomy solution.

Maybe you need to specify how many seconds you feel is "insane". 1-2 seconds seems pretty reasonable to me and likely matches what NHTSA would feel is "perceptible".

I should note it's pretty clear what NHTSA is requiring is beyond what is bare minimum legal. From what I can find, the law only requires a "complete stop", which even an almost imperceptible stop (like the 0.2s stop suggested above) qualifies for. Instead, the recall pushes for a "perceptible stop" which is a different criteria. The bit about the display not showing 0mph may be the core issue. I don't have FSD myself, but I noticed also just manual driving myself, even if I come to a stop, the car takes a while to display 0mph (it doesn't if I take off too fast). This may be fairly inconsistent with the actual speed of the wheels (as Elon noted).

Given NHTSA is likely evaluating the fix with human drivers and not with instrumented testing (that's where the "perceived duration" comes in), I suspect criteria is something like the car display showing 0mph and driver counting at least "one-one thousand". The actual wheel speed and amount of seconds the car is actually stationary may vary if the test criteria is this.

 

[stopcrazypp]

Tesla intentionally allowed over speeding too. Why not ban that ?

There are states that implicitly allow speeding as long as you are going with the flow of traffic. Given the highest posted speed limit is 85mph in the US, the current 90mph limit is reasonable.

You are not supposed to travel in left lane unless overtaking (below speed limit) - why not ban that ?

There are many states where you can camp on the left lane, even if not overtaking (California is one of them).

At least allow for a toggle switch that requires driver override to do rolling stops.

The issue with rolling stops is there is no state that allows them, nor does federal law. That is likely why NHTSA clamped down on this issue and not the others you brought up. NHTSA isn't tasked with enforcing state-by-state limits.

BTW, the law is when there is ANY vehicle in a roundabout you are supposed to wait. Tesla clearly recognizes vehicles in the roundabout and still enters. How about banned free right turns that only recently Tesla started following ?

How about not slowing down in school zones !!!!! Or overtaking stopped school buses ?

Those are design limitations as opposed to things Tesla can easily toggle. NHTSA is ok with design limitations (common example is ACC not responding to stationary cars). The rolling stop function however was a deliberate function and it's also fairly trivial to add a delay for stopping in general (how long to stop is basically a deliberate choice).

All these are a lot more important than whether the car comes to a complete stop or a rolling stop.

Nope. Just egotistical.

 

[sleepydoc]

The most useless of all things NHTSA has recommended. I mean FSD tries crashing into vehicles in the roundabout quite regularly and they are worried about coming to a full stop ? Just a lunatic / egotistic complaint.

Is there any data evidence that not coming to a full stop results in accidents - as opposed to say overspeeding that Tesla allows ?

Dan O'Dowd?

 

[AlanSubie4Life]

Maybe you need to specify how many seconds you feel is "insane". 1-2 seconds seems pretty reasonable to me and likely matches what NHTSA would feel is "perceptible".

2 seconds is far too long to be fully stopped when no traffic at all is present - but as I've said (repeatedly now!), I think for this particular aspect of the complaint, we'll need to quantify with video evidence just what it is that is being talked about. It's too easy to talk past one another on this otherwise. I know for sure I've discussed it many times, but I don't have an actual timed number for the delay. I think it can exceed 2 seconds at times. I am certain it frequently exceeds 1 second. I'll grab some videos at some point in the new year when I'm back to my car, and then we can see. The delay is definitely variable - I'm not sure what affects it. I have a couple quick videos from the latest version and it looks to be about 2 seconds at times (but they don't show the screen so I'm not going to post them)

In any case I'm fairly confident this is a definite issue (excessive delay), because there seems to be broad agreement that this is a problem (and I'm not hearing any pushback on it here!).

The NHTSA requirement to stop may seem like a contributor to this problem, but I think that’s a clear distraction here from understanding what is happening at stop signs - it just masks the underlying issue and allows people to conveniently think that somehow it's the NHTSA requirement causing all of this!

It’s very strange to me that people seem to think the two are related - especially since this behavior clearly existed before the recall. There are many discussions about this prior to January 2022. They sound exactly the same. You just have to look them up!

As I've said, the NHTSA requirement simply masks some of the pre-existing hesitancy issues. Even without the actual stop, there were problems. Some of these problems have been solved, and some still exist, these are all from 2021:

not even sure what benefit it provides since FSD Beta does such an awful job at stop signs, anyway. It'll jerk on the brakes early, roll freely for a bit, then jam them on again, possibly stopping prematurely way before the line, or possibly just rolling through the stop sign.

It seems like it wouldn't be asking too much to come to an aggressive & bold, yet predictable, smooth gradual full stop at the line, using only regen, then continue to creep from there as needed (often there is need to creep). How hard can it possibly be to stop reliably at a certain point on the road and minimize jerk in doing so?

At stop sign it seems to decelerate fast a few yards before the stop sign to ~15 mph and then slowly crawl to a stop.

I think it is reasonable to expect that it can come to a stop without using the brakes (except on steep downhills) precisely at the stop line (or the appropriate location if there happens to be no stop line) every time, regardless of slope, with minimized stopping distance/time (with the caveat that only regen is used).

(it consistently slows down way too early, then creeps up to the stop line, requiring the use of the accelerator for nearly every single well-marked stop sign)

End of 2021 complaints.

My recent complaint:

Specifically if: it can come to a sensible brisk complete stop without using the brakes except possibly below 1-2mph. And 2) actually start moving immediately after stopping, rather than WTFing.

Again, coming back to the point - notice there are many complaints here!!!

- Using the brakes excessively above 1-2mph for no reason, stopping with too much jerk. (Old problem)
- Not slowing down fast enough. (Old problem)
- Not using the brakes to come to a complete stop promptly, below 1mph or so. (Not sure why they don't do this for NHTSA)
- Waiting around (I think probably for over 1-2 seconds at times), but this needs to be captured) after the stop for no apparent reason.
- I'd add that sometimes it stops twice (with latest version). (Old hesitancy problem, masked by stopping)

Anyway we can revisit this topic with video. Looking back at the old complaints has confirmed that this is not related to NHTSA (that just changed how it manifests). Now we just have to quantify the issue.

 

[AlanSubie4Life]

but I noticed also just manual driving myself, even if I come to a stop, the car takes a while to display 0mph (it doesn't if I take off too fast).

I thought this too (for quite a while, and I've posted as much), but I've looked at it closely a bit more recently, and I no longer think this is the case. It would be cool to have concurrent video of the display and instrumented wheel speed in any case.

When I've looked at this recently, it's possible it isn't rounding down, but when the car actually stops, it seems to show 0mph immediately. I don't think there is any significant delay.

It's a very hard thing to determine with a simple video, though. Need another data stream, or a time-synchronized camera pointed at the wheel externally.

Instead, the recall pushes for a "perceptible stop" which is a different criteria.

Yeah, I remember reading the text of the recall at the time. It's not clear from the recall text what exactly is being required. You could argue that the recall is actually to fix a stop that is perceived as too long!

Unfortunately we'll never know.

But anyway it's a side issue, not related to the current complaint.

I don't have FSD myself,

You'll have to trust me on this one, then. (And note that there aren't a whole bunch of people coming out and saying that stop signs are somehow ok with FSD Beta!) Haha.

The issue with the stopping time & delay & hesitancy goes far beyond any possible delay on display of 0mph, even if somehow that is an actual contributor.

It's important to recognize that sometimes it properly executes stops at stop signs and proceeds with an appropriate delay.

 

[MP3Mike]

When I've looked at this recently, it's possible it isn't rounding down, but when the car actually stops, it seems to show 0mph immediately. I don't think there is any significant delay.

It can never round down, as I think that the requirement is that the speedometer never displays slower than you are actually moving. (Which is why it reads 1 MPH faster than you are actually going when measured via radar or GPS.)

 

[AlanSubie4Life]

It can never round down, as I think that the requirement is that the speedometer never displays slower than you are actually moving. (Which is why it reads 1 MPH faster than you are actually going when measured via radar or GPS.)

Yep. Seems consistent with what I’ve seen.

 

[DrChaos]

> It'll jerk on the brakes early, roll freely for a bit, then jam them on again, possibly stopping prematurely way before the line

This is probably a safer way to do things given limits of perception without radar. Humans have stereoscopic vision to provide absolute estimation for closer objects but the car does not.

Therefore for the driving system it's safer to approach up to oncoming stopped cars at a slower speed as it needs to integrate multiple frames to estimate distance from monocular vision (it has to estimate the absolute physical size of the car in front of it only from images) and there is some latency. And given that it makes errors in where to stop (as you show), it probably also makes errors in the other direction where the target to stop is behind where it ought to be, risking rear-end collision. So they offset the desired point further back to account for this error, and then creep forward in a safer situation.

This is a case where radar, with low chance of false returns in this instance (knowing there is definitely a car ahead), would work great, giving direct physical distance and velocity measurements of the target.

Remember also that HW3 has significantly lower resolution than human foveas as well and is monocular.

I agree that the behavior is less than desirable, and the sensor set way less than desirable.

 

[AlanSubie4Life]

Therefore for the driving system it's safer to approach up to oncoming stopped cars at a slower speed

The behavior occurs with no lead cars.

I don’t think this particular issue is because they are being conservative with lead cars.

And I don’t think the behavior would be different with a radar (since there would be no lead car involved to bounce radar off of, even if the radar was sophisticated enough to detect and discriminate between stationary objects).

We’ll see what happens with v12!

 

[EVNow]

There are states that implicitly allow speeding as long as you are going with the flow of traffic. Given the highest posted speed limit is 85mph in the US, the current 90mph limit is reasonable.

Going above 25 mph when the limit is 25 mph on a road is illegal. Tesla knows the limit, shows it on the screen and lets you set a higher limit. Infact they have made it easy to set limits above max limit !

There are many states where you can camp on the left lane, even if not overtaking (California is one of them).

But illegal still in a lot of states.

The issue with rolling stops is there is no state that allows them, nor does federal law. That is likely why NHTSA clamped down on this issue and not the others you brought up. NHTSA isn't tasked with enforcing state-by-state limits.

Nope. Most likely explanation is that someone *powerful* complained to NHTSA person in charge and they decided to take it up. Purely arbitrary.

Those are design limitations as opposed to things Tesla can easily toggle. NHTSA is ok with design limitations (common example is ACC not responding to stationary cars). The rolling stop function however was a deliberate function and it's also fairly trivial to add a delay for stopping in general (how long to stop is basically a deliberate choice).

Tesla does a lot of stuff that are potentially illegal - infact deliberately designed. And NHTSA does nothing about them.

 

[stopcrazypp]

Going above 25 mph when the limit is 25 mph on a road is illegal. Tesla knows the limit, shows it on the screen and lets you set a higher limit. Infact they have made it easy to set limits above max limit !

As I noted, there are states that allow flow of traffic exceptions instead of absolute top limits, so that offset is ok.

The recall did include changes in terms of how it responds to posted limits:
"Adjusting vehicle speed while traveling through certain variable speed zones, based on detected speed limit signage and/or the vehicle’s speed offset setting that is adjusted by the driver"

Update Vehicle Firmware to Improve Certain FSD Beta Driving Operations | Tesla Support

Tesla has issued a voluntary recall on certain Model S, Model 3, Model X and Model Y vehicles that have installed or pending installation of software that contains the Full Self-Driving (FSD) Beta feature. This recall affects only US and CA vehicles. Affected vehicles were identified based on...

www.tesla.com

But illegal still in a lot of states.

But not in all states, nor federally, nor even in where there are the most Teslas (California), so it's not an issue NHTSA has a say in (although theoretically, individual states may be able to bring a case about).

Nope. Most likely explanation is that someone *powerful* complained to NHTSA person in charge and they decided to take it up. Purely arbitrary.

I'm not saying NHTSA is fair or consistent (what stands out to me is all the attention nannies they are forcing Tesla to add, while there are other systems that have almost no attention detection, like the Infiniti I linked before in the past), just that I can see why they chose this issue as opposed to others you pointed out.

Tesla does a lot of stuff that are potentially illegal - infact deliberately designed. And NHTSA does nothing about them.

NHTSA has limited investigative resources, so they can't possibly catch every little thing, but the whole rolling stop thing was way too obvious, given it was even in the menu. Basically the issue is rolling stops are not allowed in any state, nor under federal law.

 

[johnm]

BTW, the law is when there is ANY vehicle in a roundabout you are supposed to wait.

Which law is this?

 

[dtdtdt]

BTW, the law is when there is ANY vehicle in a roundabout you are supposed to wait. Tesla clearly recognizes vehicles in the roundabout and still enters. How about banned free right turns that only recently Tesla started following ?

The California Drivers Handbook says:

Driving in a roundabout:
• Slow down as you approach the roundabout.
• Yield to pedestrians and bicyclists crossing the roadway.
• Watch for signs and/or pavement markings that guide or prohibit certain movements.
• Enter the roundabout (heading to the right) when there is a big enough gap in traffic to merge safely.
• Travel in a counter-clockwise direction. Do not stop or pass.
• Signal when you change lanes or exit the roundabout.
• If you miss your exit, continue around until you return to your exit. …

 

[swedge]

Going above 25 mph when the limit is 25 mph on a road is illegal.

In California at least, this may not be exactly true.

It has been a while since I last defended myself and drilled down, so things may have changed, but a ticket for going over the posted speed limit might be successfully challenged if the actual speed was safe for the conditions.

There are several absolute limits in the law (California Vehicle Code) including 100 mph anywhere (CVC sec 22348), 65 on divided highways (22349), unless posted as 70 (22356), 55 on undivided highways, 25 in residential, business, school zones (22352).

But otherwise, it is legal to drive at a speed which is "reasonable or prudent". (22250, the "Basic Speed Law")

I call it fit'n and proper.

Here in Oakland, the freeway is 65 mph, but hardly anyone drives that slow any more. It is pretty clear the the CHP policy is to not enforce the 65 mph limit. I really can't see NHTSA mandating compliance. How many cars would Tesla sell if the cars had a built in RoboCop which emailed a ticket every time you exceeded 65 mph?

My hunch is that the full-stop thing was not so much about controlling how FSD behaves, but rather the feds putting Musk on notice that they could control his behavior. I'm happy to tap the go-pedal when it is clear.

 

[Mardak]

In California at least, this may not be exactly true

Now we're getting into the difference among what a judge will decide in your situation vs what law enforcement can pull you over or even give a ticket vs what's generally acceptable. For example, even the CA driver handbook might not be technically correct in that roundabouts typically have yield signs, so you could theoretically be given a 21803a ticket for failure to "yield the right-of-way to any vehicles which have entered the intersection" especially if you caused an accident entering a roundabout (unless I suppose the handbook referring to "big enough gap" means a gap large enough that there's no other vehicle in the roundabout).

I would guess end-to-end will by default be trained on enough examples to get to generally acceptable human patterns and only when NHTSA or others push back will Tesla specially gather training data to bias the behavior to be legally correct, but that still probably leaves things open for cops to pull you over if they really wanted to.

 

[diplomat33]

 

[Mardak]

Sounds like initial reactions to leaked 12.1 videos is that it's not a significant regression from 11.x that some were expecting many more months to even get to that quality. Driving 11.x in unfamiliar areas for the holidays reminded me of all the unnecessary micro-adjustments that it can make suddenly, so hopefully 12.x improves significantly on that smoothness and comfort, but it's probably hard to tell from videos.

 

[stopcrazypp]

Now we're getting into the difference among what a judge will decide in your situation vs what law enforcement can pull you over or even give a ticket vs what's generally acceptable. For example, even the CA driver handbook might not be technically correct in that roundabouts typically have yield signs, so you could theoretically be given a 21803a ticket for failure to "yield the right-of-way to any vehicles which have entered the intersection" especially if you caused an accident entering a roundabout (unless I suppose the handbook referring to "big enough gap" means a gap large enough that there's no other vehicle in the roundabout).

Yielding the right of way does not mean waiting until there are no vehicles at all in the intersection before entering. It simply means what it says, merging into traffic (without presumably cutting anyone off).

Look at the second part, which should make it clear:
(b) A driver having yielded as prescribed in subdivision (a) may proceed to enter the intersection, and the drivers of all other approaching vehicles shall yield the right-of-way to the vehicle entering or crossing the intersection.

I don't know how you treat yield signs (in general, not talking about roundabouts specifically), but generally if there is a big enough gap in traffic where you can merge in without the car approaching needing to slow down (or if the approaching car slows down actively to allow you to merge in), that is generally regarded as properly yielding the right of way. It doesn't mean waiting until there is no traffic at all.

I would guess end-to-end will by default be trained on enough examples to get to generally acceptable human patterns and only when NHTSA or others push back will Tesla specially gather training data to bias the behavior to be legally correct, but that still probably leaves things open for cops to pull you over if they really wanted to.

 

[swedge]

Dan O'Dowd?

On July 19, 1989, United Airlines flight 232 crash landed in Sioux City Iowa. 112 died, but 184 survived. 52 children were aboard, 4 of whom were "lap children" who sat in an adult's lap. 11 of the children died, including one of the lap children.

While the crew was heralded as heroes for having saved so many lives, a campaign was launched to save children from the unsafe policy of allowing infants to sit in a parent's lap. The proposed solution was to require an individual seat and restrain for each child.

In a remarkably rational review, the NTSB realized that such a policy would increase the cost for families to fly, which would cause many to drive instead, a much less safe alternative.

The FAA estimates that a regulation that all children must have a seat would equate, for every one child's life saved on an aircraft, to 60 people dying in highway accidents.

The proposed policy was scrapped, because rather than saving lives, it would have cost lives.

So, to this day, in the US, children under 2 can fly without their own seat.

An alternative would have been to require airlines to provide free seats and child seats for infants, but, well, that was not going to fly, so to speak.

Running stop signs at full speed is dangerous. Rolling past the limit line at 2 mph, no so much. I expect that Tesla's voluminous data will eventually provide actual evidence on which to base safety decisions, and perhaps regulators will follow the data. One can hope...

United Airlines Flight 232 - Wikipedia

en.wikipedia.org

(moderator edit)

 

[nikolajz1]

Looks like a V12 FSD video leaked on Twitter. Does not seem much better than (V11) as expected. Still makes some weed decisions. Turning seems a bit smoother but that's hard to tell over video

https://twitter.com/x/status/1740809025548333398