Profound progress towards FSD

[diplomat33]

To sum up this thread:

1) "turning at intersections and autosteer on narrow streets" is in alpha development.
2) Elon estimates the rewrite with does 4D is "2-4 months" away
3) Elon also said that the latest alpha software can almost do his commute, which includes construction zones, without any disengagements.
4) Elon thinks that Tesla is "very close" to L5.

Some see it as proof that Tesla is making mind blowing progress on FSD. Others are skeptical on the timeline of when we will get these features and think Tesla is still very far from actual L5.

 

[mikes_fsd]

To sum up this thread:

1) "turning at intersections and autosteer on narrow streets" is in alpha development.
2) Elon estimates the rewrite with does 4D is "2-4 months" away
3) Elon also said that the latest alpha software can almost do his commute, which includes construction zones, without any disengagements.
4) Elon thinks that Tesla is "very close" to L5.

Some see it as proof that Tesla is making mind blowing progress on FSD. Others are skeptical on the timeline of when we will get these features and think Tesla is still very far from actual L5.

And some see "feature complete" for Tesla as giving FSD features with supervision by the driver until "extensive software validation" is completed (the march on 9's)
i.e. "feature complete" FSD !== L5 until software proves itself with real life data.

 

[powertoold]

high confidence of the actual 3D scene from just a few time snapshots, let alone one

For driving, having 2-4 non-sequential frames from each camera would probably be enough to make a prediction for object trajectory. As humans (bad comparison I get it), we can do rough trajectory estimation from a few frames (and most of the time, just 1 frame).

 

[ZeApelido]

For driving, having 2-4 non-sequential frames from each camera would probably be enough to make a prediction for object trajectory. As humans (bad comparison I get it), we can do rough trajectory estimation from a few frames (and most of the time, just 1 frame).

I think camera resolution / noise are too high to work on just a few. The images are still too "noisy" and need some smoothing. Just my hunch based on experience with handling lots of time series data.

 

[allelectrified]

Everyone on this forum, or at least those who paid for FSD, must be a bunch of eternal optimists.

Hasn't Mr Musk himself been saying for years that getting to 90% safety is easy, as is 99%? In the big picture, it is only then that it starts getting hard. If it takes a year with "4D" to get to 99% (my numbers here are illustrative; you can map them to intervention rates), then does it take 2 more years or 20 more years to get to 99.9999%?

The correct answer must be that no one at Tesla has much idea about this. Those 9s represent unknown unknowns. Can they be done with the same approach as the first couple of 9s? The next one, yes, the one after that likely no. Which 9 will require a 15-year-old child's common sense? In humans, almost all the intelligence needed for driving has nothing to do with driving. But Tesla's systems are only trained on driving data!!

I listen to Lex Fridman's A.I. podcast, and it's been a long time since any guest has believed that Tesla could achieve FSD this decade.

I bought FSD and love it to bits. It's !@#!# amazing. But I like to be beguiled only a finite number of times.

 

[Ficheh]

I don't see how there could be level 5 automation. How could the car react to a new issue? Let's say you're being driven by the car, so you're in the backseat and there's no one in the front. The car is driving along and all of a sudden the lane lines disappear (maybe they're going to repave), what does the car do? Or if the traffic lights go out and the car decides just to drive through @ 35 mph. The cross traffic expects you to stop (since all intersectuons are now 4 way stops with a power outage) and they go. Or the offramp needed for your destination is closed for some reason. There's no way I'm getting into a car without a driver. I still remember Elon saying in 2016 that FSD would be here in 2018. I sure as hell am not believing 2020. I didn't pay for FSD in my Model 3. I think it's a great car, but I don't think FSD is coming in my lifetime.

 

[dwbeam]

In early June I drove from North Carolina to Grand Canyon. Obviously I drove back also. From the time I got on interstate in South Carolina till stopping in Oxford, Alabama Tessie did all navigation,lane changes with no intervention. The next major area was Dallas Fort Worth and again Tessie was flawless. Pretty much this way for 4 thousand miles.

Just to be clear I could have been asleep.

A couple of times I came up on orange barrels closing a lane and I did not have the guts to see what would happen and disengaged. This was in New Mexico.

Since then have got updates doing stop lights etc. Car now navigates some near 90 degree turns that it used to not be able to do.

Not FSD but much closer.

I have also driven from North of Charlotte to Isle of Palms(240 miles) twice with no disengagements.

The last couple of updates are incrementally better than what I drove to Grand Canyon.

I am guessing this is level 3.

Model S long range.

 

[powertoold]

I don't see how there could be level 5 automation. How could the car react to a new issue? Let's say you're being driven by the car, so you're in the backseat and there's no one in the front. The car is driving along and all of a sudden the lane lines disappear (maybe they're going to repave), what does the car do? Or if the traffic lights go out and the car decides just to drive through @ 35 mph. The cross traffic expects you to stop (since all intersectuons are now 4 way stops with a power outage) and they go. Or the offramp needed for your destination is closed for some reason. There's no way I'm getting into a car without a driver. I still remember Elon saying in 2016 that FSD would be here in 2018. I sure as hell am not believing 2020. I didn't pay for FSD in my Model 3. I think it's a great car, but I don't think FSD is coming in my lifetime.

These potential edge cases are easily navigated by the current Tesla software.

 

[diplomat33]

In early June I drove from North Carolina to Grand Canyon. Obviously I drove back also. From the time I got on interstate in South Carolina till stopping in Oxford, Alabama Tessie did all navigation,lane changes with no intervention. The next major area was Dallas Fort Worth and again Tessie was flawless. Pretty much this way for 4 thousand miles.

Just to be clear I could have been asleep.

A couple of times I came up on orange barrels closing a lane and I did not have the guts to see what would happen and disengaged. This was in New Mexico.

Since then have got updates doing stop lights etc. Car now navigates some near 90 degree turns that it used to not be able to do.

Not FSD but much closer.

I have also driven from North of Charlotte to Isle of Palms(240 miles) twice with no disengagements.

The last couple of updates are incrementally better than what I drove to Grand Canyon.

I am guessing this is level 3.

Model S long range.

I would caution you. It is only Level 2. It is not Level 3 yet. You cannot fall asleep. You just got lucky that the 4,000 miles were stuff that the current AP can handle. But if there had been a case that AP can't handle and you had been asleep, you would have crashed.

Also, the car cannot drive 4000 miles without recharging. So you clearly had to disengage to pull over at a supercharger. So no, it was not autonomous. You still had to disengage. It just happened to handle the boring highway stuff.

 

[powertoold]

I would caution you. It is only Level 2. It is not Level 3 yet. You cannot fall asleep. You just got lucky that the 4,000 miles were stuff that the current AP can handle. But if there had been a case that AP can't handle and you had been asleep, you would have crashed.

A week ago, my AP was driving in early morning direct sunlight on a hilly highway in a construction zone with lane lines swiggling all sorts of directions. It was the most impressive environment I've seen AP drive through. Even other cars were getting confused about the lines. Then again, I get some dangerous phantom braking from time to time from overhead boards. To me, phantom braking is probably the last "issue" with NoA in highways, aside from slowing down for traffic on curvy bridge exchanges.

 

[Battpower]

Just to be clear I could have been asleep.

I think more likely you could have been dreaming!

 

[dwbeam]

The car did drive itself down each exit ramp getting to chargers.
Someone questioned what about stop lights being out. The car would stop and not continue.

It is definitely not ready for me to be asleep.

I think what I have is level 3 on Freeways. Weather detection probably in not full environmental detection. Current car does passes on freeways.

Level 3 vehicles have “environmental detection” capabilities and can make informed decisions for themselves, such as accelerating past a slow-moving vehicle. ... The driver must remain alert and ready to take control if the system is unable to execute the task.

 

[diplomat33]

A week ago, my AP was driving in early morning direct sunlight on a hilly highway in a construction zone with lane lines swiggling all sorts of directions. It was the most impressive environment I've seen AP drive through. Even other cars were getting confused about the lines. Then again, I get some dangerous phantom braking from time to time from overhead boards. To me, phantom braking is probably the last "issue" with NoA in highways, aside from slowing down for traffic on curvy bridge exchanges.

Personally, I tested NOA about a month ago near my house and it could not handle the two exits.

But you raise an interesting point. The fact is that there are going to be cases that AP will handle beautifully. And there will be cases that the NN will really impress in how well it handles. Those are the sort of cases that will make someone marvel at Tesla's camera approach and become very optimistic for the future of Tesla's FSD. So it's not all bad. Far from it. But there will also be cases that the NN can't handle well. That's kind of the tricky thing with L5. There are going to be a ton of cases to solve in the US. And depending on where you drive and what time of day, you might see cases that the NN can handle wonderfully but you might also happen to encounter cases that the NN can't handle. I believe that why some folks can honestly talk about how great AP is, how it handled a long drive flawlessly while somebody can say that AP tried to kill them.

 

[dwbeam]

I think more likely you could have been dreaming!

I was referring to I made no interventions driving through major metro areas (Atlanta and Dallas Forth Worth) .

The first leg was 298.9 miles. More than 270 miles was on Freeway. I am saying had I been sleeping the car would have gone ramp to ramp and not hit anybody or anything. That is no small feat.

 

[Buckminster]

FSD is merely a collection of scenarios that the most sophisticated widely available neural net has to handle one at a time. Once the great apes have labelled things to make sense of its world we are golden. The process is proven for half the scenarios - sit back and watch with positive energy for the miracle that is unfolding. You can believe what your eyes are seeing - it will happen.

 

[pilotSteve]

I don't see how there could be level 5 automation. How could the car react to a new issue? Let's say you're being driven by the car, so you're in the backseat and there's no one in the front. The car is driving along and all of a sudden the lane lines disappear (maybe they're going to repave), what does the car do? Or if the traffic lights go out and the car decides just to drive through @ 35 mph. The cross traffic expects you to stop (since all intersectuons are now 4 way stops with a power outage) and they go. Or the offramp needed for your destination is closed for some reason. There's no way I'm getting into a car without a driver. I still remember Elon saying in 2016 that FSD would be here in 2018. I sure as hell am not believing 2020. I didn't pay for FSD in my Model 3. I think it's a great car, but I don't think FSD is coming in my lifetime.

Well... when I BOUGHT FSD in 2016 I didn't expect to have to wait over 3 years before seeing any significant benefit. BUT.... I also am amazed that my same vehicle (well, with the MCU2/HW3 upgrade) is now recognizing stop signs, traffic lights, etc. and actually becoming competent at lane keeping (mostly).

 

[ram1901]

i applaud your optimism. but ive heard 3-6 month predictions since i jumped in during 2015 and ive seen everyone one of them take 12-18 months. it took ap2 almost 2 years for them to activate surround camera's and be on par with AP1. what kind of bullcrap was that?

now my lease is ending in 2 weeks

nice to see my 9k FSD investment go down the tubes and not get what we were promised.

If you had a lease, fortunately you did NOT invest in FSD. A lease only pays for a portion of the car's use. The balance is carried by the lease company. Those buying their car outright are paying for the development of FSD. If it is completed before they sell their car, they get all the step by step feature additions and the final product at no additional charge. Just to clarify.

 

[insaneoctane]

Most of the impatient customers awaiting FSD improvements are merely listening to Elon/Tesla and their own timelines. I'm sure you can find company estimates that we'd have more long ago (wasn't Tesla supposed to demo a coast to coast drive across the US in 2018?). I certainly didn't make up my expectations on my own. I'm very patiently waiting for some semblance of the FSD I've purchased and hoping it's soon.

 

[emmz0r]

Why isn't there any cut-in handling yet in production?

 

[PhaseWhite]

Why isn't there any cut-in handling yet in production?

It is in production.
From the April 2019 Autonomy presentation: