how does the new end to end FSD work, need a block diagram from of data flow from the fleet to DoJO to an individual's car

[Ben W]

All yall anti-FSDs are starting to get really worried. You gotta beef your game up. You got to come up with better arguments.

Vision took me on a 70 mile trek the other night. Without HD maps and on roads rarely travelled by Teslas.

At this point, the more you dig in, the funnier you are going to look. And right now it's getting to be comical.

Vision took me on a 3500-mile roadtrip the other month, under a wide range of driving conditions, with multiple perception-related disengagements and panics. Wake me when it can do at least that with no disengagements, let alone the million miles between disengagements/failures that will be required for L4.

 

[spacecoin]

8/8?

Elon’s robotaxi reveal

 

[OxBrew]

Just to beat a dead horse: L2 (proba ly L3 also) is pretty useless. Nearly impossible to overcome the complacency. If it fails 2X in 200,000 miles it's doing worse than an average human paying attention.

I think ADAS (currently FSD supervised) should always have the human steering, so they never get complacent. Just fill in for the human mistakes, keep us from crashing. So basically a better ADAS suite.

FSD needs to be L4 or nothing.

 

[Ben W]

Just to beat a dead horse: L2 (proba ly L3 also) is pretty useless. Nearly impossible to overcome the complacency. If it fails 2X in 200,000 miles it's doing worse than an average human paying attention.

I think ADAS (currently FSD supervised) should always have the human steering, so they never get complacent. Just fill in for the human mistakes, keep us from crashing. So basically a better ADAS suite.

FSD needs to be L4 or nothing.

It's ironic that the more reliable L2 gets, the more useless (and arguably dangerous) it may get, due to this uncanny valley effect. Agreed that the system will need to achieve extremely robust L4 to be long-term viable, and I don't see that happening with Pure Vision. Perhaps by 2035-2040 there might be an AI powerful enough to drive a HW4 car with L4 levels of reliability. But if Tesla wants it anytime sooner, and they must, they'll need to have a more robust sensor suite to compensate for the AI limitations.

Just this evening, on my 5-mile commute home (HW3 / v12.3.6), I was greeted by another Red Hands Of Death as the car turned slowly on a residential street (~15mph) and sunlight struck the windshield at a glare-y angle. This is the easiest ODD there is. (Or second-easiest, after a cloudy day.) If Tesla can't solve fundamental perception issues like this, seven years after releasing HW3, there is really not much hope. Unless, as seems blindingly [heh] obvious, to me at least, they overbuild their next sensor suite to overcome it.

 

[Ben W]

And again, on a 2-mile drive to grab dinner this evening, the car first attempted to drive straight from a left-turn-only lane (planning error, not perception), but on the way home it took a corner too tightly in a dark residential neighborhood and curbed the right rear wheel, causing another instant Red-Hands-Of-Death panic. (Lidar could certainly have helped it gauge the curb position more accurately.) I saw it coming, but still let the car do the driving in the interest of science and feedback to Tesla, and because the wheel has already been curbed from multiple previous FSD mistakes. (It's an older Model 3, so this is the car I'm most willing to experiment with.)

For all its amazingness, the current FSD is still many many orders of magnitude away from L4. I expect 12.4 / 12.5 / 12.6 / AI5 to close this gap somewhat, but there is still such a LONG way to go.

 

[stopcrazypp]

Just to beat a dead horse: L2 (proba ly L3 also) is pretty useless. Nearly impossible to overcome the complacency. If it fails 2X in 200,000 miles it's doing worse than an average human paying attention.

I think ADAS (currently FSD supervised) should always have the human steering, so they never get complacent. Just fill in for the human mistakes, keep us from crashing. So basically a better ADAS suite.

FSD needs to be L4 or nothing.

Completely disagree. AP is fairly useless with the human steering. It's very easy to stay alert with it, and in fact it reduces fatigue (I went on long couple hundred mile trips with it and felt far less tired). The type of people that drive distracted would drive distracted anyways even using dumb cruise control. The solution to distracted drivers are better attention detection systems, not to make the system dumber.

 

[OxBrew]

Completely disagree. AP is fairly useless with the human steering. It's very easy to stay alert with it, and in fact it reduces fatigue (I went on long couple hundred mile trips with it and felt far less tired). The type of people that drive distracted would drive distracted anyways even using dumb cruise control. The solution to distracted drivers are better attention detection systems, not to make the system dumber.

Interesting, I find the opposite is true. Babysitting the teenage lane keeping driver takes more effort than just steering myself, for me. I drive 700 or 1000 mile round trips on weekends every 4-6 weeks. Since Fall 2023 I've gone back to full manual driving with short bits of TACC on some wildly boring straight sections.

Potholes alone is enough to make me steer manually. The constant bing bong to change lanes seals it for me.

On top of all that, lane position is also a deal breaker. I really really hate how it centers in the lane. Everyone offsets in the lane, to the point where the road is worn offset one side or the other. I feel like I'm crowding everyone being centered, to the point where I'm constantly disengaging when passing/being passed.

 

[Dewg]

Interesting, I find the opposite is true. Babysitting the teenage lane keeping driver takes more effort than just steering myself, for me. I drive 700 or 1000 mile round trips on weekends every 4-6 weeks. Since Fall 2023 I've gone back to full manual driving with short bits of TACC on some wildly boring straight sections.

Potholes alone is enough to make me steer manually. The constant bing bong to change lanes seals it for me.

On top of all that, lane position is also a deal breaker. I really really hate how it centers in the lane. Everyone offsets in the lane, to the point where the road is worn offset one side or the other. I feel like I'm crowding everyone being centered, to the point where I'm constantly disengaging when passing/being passed.

I love pie. The flavors, the textures, the versatility just can't be beat. But no matter how hard I try, I just can't convince my good friend of this simple fact, and he wholeheartedly disagrees with me, arguing for cake as the best dessert. Perhaps he needs to open a service ticket and have his taste buds recalibrated.

The point being: you'll never get people to switch sides by arguing your experiences. Some people very much enjoy AP and FSDS, and some people hate it. And others are indifferent. Some people have amazing experiences with it, enjoying routine multi-mile trips with little to no interventions, and some people can't get out of their residential neighborhoods without disengaging. Some people find it stressful to use, and others reach their destination relaxed.

I, for one, love my experiences with FSDS. It works for me, and perhaps I have a unicorn car. But I always chuckle, when reading passionate people arguing the system sucks, thinking the guy who really likes it is suddenly going to change his mind and say "OMG you're right, I thought it was great, but now I see it sucks. Thanks for opening my eyes". And vice versa.

 

[stopcrazypp]

Interesting, I find the opposite is true. Babysitting the teenage lane keeping driver takes more effort than just steering myself, for me. I drive 700 or 1000 mile round trips on weekends every 4-6 weeks. Since Fall 2023 I've gone back to full manual driving with short bits of TACC on some wildly boring straight sections.

Potholes alone is enough to make me steer manually. The constant bing bong to change lanes seals it for me.

On top of all that, lane position is also a deal breaker. I really really hate how it centers in the lane. Everyone offsets in the lane, to the point where the road is worn offset one side or the other. I feel like I'm crowding everyone being centered, to the point where I'm constantly disengaging when passing/being passed.

Then AP is not for you, so you can choose to not use it. But clearly plenty of people disagree and the system shouldn't eliminate Autosteer just to match your personal preferences. There is a reason why AP is by far the most used ADAS system out there.

As for the offset, you can try camera calibration if it helps, but yes, the car stays generally dead centered even if other lanes are offset. I haven't found this to be an issue, only deal with it in rare cases when the vehicle next is extra wide or if they are straddling the line in the first place, in which case, I just apply steering to offset the car why passing (which may or may not disengage, in which case it's trivial to double tap to reengage).

For lane changes, I don't have EAP, so the way I do it is use the turn signal, steer (which disengages Autosteer) then double tap to turn back on after lane change is complete. The only extra action is the double tap, which I don't find to be a big deal.

I use AP however with both hands on the wheel, so I never feel any stress about it doing anything unexpected, given I am in full control and have enough grip to resist any unwanted steering. YMMV depending on how you use it.

 

[ewoodrick]

Just to beat a dead horse: L2 (proba ly L3 also) is pretty useless. Nearly impossible to overcome the complacency. If it fails 2X in 200,000 miles it's doing worse than an average human paying attention.

I think ADAS (currently FSD supervised) should always have the human steering, so they never get complacent. Just fill in for the human mistakes, keep us from crashing. So basically a better ADAS suite.

FSD needs to be L4 or nothing.

Interesting, since humans have accidents about every 66,000 miles.

 

[ewoodrick]

Interesting, I find the opposite is true. Babysitting the teenage lane keeping driver takes more effort than just steering myself, for me. I drive 700 or 1000 mile round trips on weekends every 4-6 weeks. Since Fall 2023 I've gone back to full manual driving with short bits of TACC on some wildly boring straight sections.

Potholes alone is enough to make me steer manually. The constant bing bong to change lanes seals it for me.

On top of all that, lane position is also a deal breaker. I really really hate how it centers in the lane. Everyone offsets in the lane, to the point where the road is worn offset one side or the other. I feel like I'm crowding everyone being centered, to the point where I'm constantly disengaging when passing/being passed.

I've got a Cybertruck without FSD and we've put about 5,000 miles on it in the last 2 months.

It is SO MUCH more stressful to drive without FSD than it is with it. And my wife says the same thing.

After growing up with FSD, I can pretty well estimate any issues that it may have.

 

[Ben W]

Interesting, since humans have accidents about every 66,000 miles.

Most of those accidents are due to human drivers not paying attention. (Or driving while impaired, or driving recklessly, etc.) FSD needs to be at least as good as good skilled conscientious human drivers. Not just as good as average/mediocre/impaired/careless drivers.

 

[Dewg]

FSD needs to be at least as good as good skilled conscientious human drivers.

I agree. But where do we find the statistics on accident rates for those humans?

 

[Ben W]

I agree. But where do we find the statistics on accident rates for those humans?

To a first approximation, look at the types of mistakes/failures FSD makes and ask whether a skilled human would make that mistake. Or ask an auto insurance company for statistics on their best/lowest-risk drivers.

 

[JB47394]

I agree. But where do we find the statistics on accident rates for those humans?

Why is that significant? As soon as the accident rate for a given demographic is exceeded by FSD, insurance companies should encourage that group to drive with FSD. Or legislators require it.

Interestingly, that would mean that the youngest drivers would be the first in line. Here are death and injury rate charts per 100,000 drivers by age (not simple accident rates).

It may be that cities would require FSD use. Or highways. Wherever FSD becomes better than humans, it would start to be incentivized or mandated.

Supervised FSD is another matter, of course. There would be a split between FSD operated by attentive drivers and FSD operated by inattentive ones. Assuming people can figure out how to remain inattentive even with cabin monitoring.

 

[ewoodrick]

Most of those accidents are due to human drivers not paying attention. (Or driving while impaired, or driving recklessly, etc.) FSD needs to be at least as good as good skilled conscientious human drivers. Not just as good as average/mediocre/impaired/careless drivers.

BS!

It's the impaired or distracted drivers that are killing people! If Tesla could get rid of just those categories, I'm sure that they would get every safety award and medal from the President. That's over 20,000 people per year if I remember correctly.

 

[ewoodrick]

Supervised FSD is another matter, of course. There would be a split between FSD operated by attentive drivers and FSD operated by inattentive ones. Assuming people can figure out how to remain inattentive even with cabin monitoring.

What I suspect that some auto manufacturers are anticipating and I'm sure the NHTSA is thinking about, is the attention mechanisms that the companies are using to confirm driver attention in the FSD type of vehicles.

It's not going to be long before those mechanisms get moved to ALL vehicles, just to improve the numbers.

 

[stopcrazypp]

What I suspect that some auto manufacturers are anticipating and I'm sure the NHTSA is thinking about, is the attention mechanisms that the companies are using to confirm driver attention in the FSD type of vehicles.

It's not going to be long before those mechanisms get moved to ALL vehicles, just to improve the numbers.

The first thing they are exploring is automatic drunk driver detection (via BAC detection, attention detection, or both):

https://www.reuters.com/business/autos-transportation/us-agency-opens-push-mandate-anti-drunk-driving-technology-2023-12-12/

 

[Ben W]

Why is that significant? As soon as the accident rate for a given demographic is exceeded by FSD, insurance companies should encourage that group to drive with FSD. Or legislators require it.

Interestingly, that would mean that the youngest drivers would be the first in line.

There’s a Catch-22 here. If inexperienced drivers are not allowed to drive, how will they ever become experienced drivers?

I suppose that once the autonomous systems are good enough, they would be able to let the driver do 99.9% of the driving (in order to gain experience), but intervene whenever the driver is about to make a serious mistake. This already exists to some extent (AEB, lane drift correction, warning chimes), but it could be made a lot more robust.

 

[Ben W]

BS!

Wut? I think we were saying pretty much the same thing?