Elon: "Feature complete for full self driving this year"

[Daniel in SD]

I think this is exactly why @verygreen felt that this particular behavior will ONLY be released to early access.

I think (hope?) even Tesla realizes that stopping for green lights isn’t something that the average customer can react to and overrule reliably, so they’ll keep this rollout limited until they can be more confident about green light behavior.

I know if I got this early update, I’d have a few routes I’d like to use it on for fun, but imagine this would get mentally exhausting if you used it daily. Add to the fact that it won’t “navigate” (take turns), it’s really only useful in limited circumstances, so I suspect this is mostly a “data gathering” release.

I suspect that the wide release will also stop at green lights (doesn't it stop at all intersections?).
I think all the data for this feature could be collected in shadow mode. How is this different than comparing the driver's reaction to intersections to the light detection neural net running in the background?
I don't think people getting rear ended will be an issue, there will be plenty of honking to train the driver's neural net.

 

[Mardak]

I think this is exactly why @verygreen felt that this particular behavior will ONLY be released to early access.

The dangerous behavior is more of Autopilot incorrectly identifying a stop line with a traffic light or stop sign because 1) the driver wouldn't expect the car to slow down and 2) Autopilot probably will panic and brake hard. At least with the FSD visualization preview, drivers might get a sense of where these false positives might occur, e.g., random overhead lights, poorly angled signs, confusing road markings.

The common case behavior of approaching a traffic light has a few seconds for the driver to confirm the green light with a slight tap of the accelerator before the car even reduces speed. And if the driver didn't notice the intersection ahead of time, there's then a few seconds of the car starting to slow down as a physical/tactile reminder. Then comes even more aggressive braking to stop, so for a person to not realize Autopilot has stopped at a green light, perhaps Tesla is doing the right thing because clearly the driver hasn't paid attention to the road for ~15 seconds at this point.

I've been using Autopilot extensively on city streets and know it can get confused with no notice especially at intersections where it might drive into oncoming traffic or failing to take sharper curves. So my main concern is actually people learning about stopping at traffic lights and using Autopilot for the first time on city streets thinking it's as confident as highway driving.

 

[alstoralset]

Not sure it will need to pass Turing test to reach nuanced decision making. A machine doesn’t necessarily need to think like a human to achieve end result.

Passing the Turing test is (probably, hopefully) unnecessary. But I believe the test has less to do with how the machine thinks than how the machine is perceived.

 

[alstoralset]

I don't think the car will need to pass the Turing test. It will need to perceive the difference between pedestrians, cyclists etc which perception can already do in autonomous cars and have the right driving policy for how to act around each one.

I think I agree, but have become more pessimistic about the whole thing after watching the M3 misperceive or respond inappropriately to the environment, over and over again. It seems far, far away from being able to, say, detect a rolling ball and realize a child will likely come running after it.

But, the recent Mobileeye demo (I think you were the first to post it?) was encouraging.

 

[Battpower]

It seems far, far away from being able to, say, detect a rolling ball and realize a child will likely come running after it.

For me, that's certainly a concern, but more so how determined the car is to accelerate aggressively just as you are approaching a hazard like a sharp bend, roundabout or intersection, purely because the speed limit says its OK. And it's often is working to an incorrect speed limit too!

 

[diplomat33]

I think I agree, but have become more pessimistic about the whole thing after watching the M3 misperceive or respond inappropriately to the environment, over and over again. It seems far, far away from being able to, say, detect a rolling ball and realize a child will likely come running after it.

But, the recent Mobileeye demo (I think you were the first to post it?) was encouraging.

That's because Tesla's FSD is actually not as advanced as we are led to believe. And Mobileye has better FSD.

 

[emmz0r]

Turing test doesn't mean anything here either. I think a dumb chatbot made it, because the test panel thought it was close enough to human responses

 

[DanCar]

That's because Tesla's FSD is actually not as advanced as we are led to believe. And Mobileye has better FSD.

I don't agree, but understand why you believe that. The proof is in what is released. Mobileye will suffer a similar fate that Waymo has, unwilling to have fatal accidents so they won't release. Tesla will release, and Tesla has demonstrated willingness to accept fatal accidents. In other words: having good prototypes, doesn't do much good.
I can agree Mobileye has better technology, but I'm dubious it will released in the next few years.

 

[mongo]

I think I agree, but have become more pessimistic about the whole thing after watching the M3 misperceive or respond inappropriately to the environment, over and over again. It seems far, far away from being able to, say, detect a rolling ball and realize a child will likely come running after it.

But, the recent Mobileeye demo (I think you were the first to post it?) was encouraging.

If you need a rolling ball to precede a child to avoid hitting them, your system is not very safe. Most children do not have such things taking point for them.

 

[boonedocks]

We got some updates from Elon on FSD. Elon is saying "reverse summon" (where the car drops you off and parks itself) is coming later this year:


As always, take the timeline with a big grain of salt. After all, Tesla promised FSD was coming "later this year" last year, 2019. But I like at least knowing what features Tesla is working on.

On April 13, Musk also hinted at using the fleet to map parking lots:


This one, I think particularly interesting because it seems like using HD mapping for the very thing that fans always said was not necessary. But I do hope Tesla implements this because it should improve the reliability of Smart Summon.

I know you are not old enough (in Tesla years) but that promise you quoted is off by over 2 years. Try 2017....sadly enough

 

[diplomat33]

I know you are not old enough (in Tesla years) but that promise you quoted is off by over 2 years. Try 2017....sadly enough

Oh I know about the 2017 FSD promises.

I only said "last year" because that is when Tesla actually attached "coming later this year" to the traffic light feature that is now being promised for this year.

 

[InternetDude]

Elon also said the Trans-Canada highway should be complete (superchargers) by end of 2017 in spring 2017. They did get it done a few days before Xmas 2019. I'm done with Elon's BS personally. I regret buying FSD on my P100D in June 2018. Hell I regret buying my P100D. Won't make that mistake again.

Buy what's available right now, not what's "coming soon" when it comes to Tesla. Lesson learned the hard way.

 

[Dutchie]

Elon also said the Trans-Canada highway should be complete (superchargers) by end of 2017 in spring 2017. They did get it done a few days before Xmas 2019. I'm done with Elon's BS personally. I regret buying FSD on my P100D in June 2018. Hell I regret buying my P100D. Won't make that mistake again.

Buy what's available right now, not what's "coming soon" when it comes to Tesla. Lesson learned the hard way.

What's wrong with your P100D?

 

[Dutchie]

What baffles me is that I read stories about how relentless Elon is as a manager to work for. How he sets strict timelines and puts enormous pressure on teams to achieve and show results.

In the case of FSD, however, time lines slip and slip, and progress is super slow. Yet, Elon appears to have utmost confidence in the whole project as seen by organizing an investor's day on the project, Twitter feed and the regular increase of price for FSD.
What is it? Will there be suddenly a whole new software upload for just FSD cars and everything magically appears? is there already a whole "feature complete" running in shadow mode?
Something is off...

 

[diplomat33]

What baffles me is that I read stories about how relentless Elon is as a manager to work for. How he sets strict timelines and puts enormous pressure on teams to achieve and show results.

In the case of FSD, however, time lines slip and slip, and progress is super slow. Yet, Elon appears to have utmost confidence in the whole project as seen by organizing an investor's day on the project, Twitter feed and the regular increase of price for FSD.

I don't see a contradiction. By all accounts, Elon is a tough manager who sets tough deadlines. I've seen accounts of the AP team consistently working day and night to get features out. And recently, he's set the deadline for the first robotaxi at the end of this year. So, he is still setting what many would say are impossible deadlines.

But, if they miss deadlines and progress seems so slow, it is because autonomous driving is a lot harder than Elon seems to think. Here's one example:

So we see that in Dec 2018, Elon said that they were testing traffic lights, stop signs and roundabouts and concluded that our cars would be able to self-drive from our home to work with no driver input at all "soon". Now, I don't doubt that Tesla was testing these things in their development software. I don't think Elon was making that part up. But it has taken 1.5 years for Tesla to go from testing traffic lights and stop signs internally to actually sending it to Early Access. Also, even with this traffic light feature, we are not exactly ready for "go from your garage at home to parking lot at work with no driver input at all" just yet. So Elon does seem to underestimate what it takes to do autonomous driving. I think he underestimates edge cases and underestimates how long machine learning takes.

Will there be suddenly a whole new software upload for just FSD cars and everything magically appears? is there already a whole "feature complete" running in shadow mode?

Based on the FSD demo on autonomy day, Tesla does seem to have "feature complete" software in development that has a lot of the different pieces together which I am sure they use to test how FSD is working as a whole. But I doubt it is reliable enough for the public. Tesla seems to roll out each individual piece in order to validate each feature one by one. So I doubt that Tesla will just one day roll out a big "robotaxi" update all at once. More likely, we will get each piece separately.

 

[Battpower]

everything magically appears? is there already a whole "feature complete" running in shadow mode?
Something is off...

If we knew, they'd have to kill us!

But it's hard to feel that what we have seen over the past 4 years is suddenly about to turn into viable FSD.

 

[InternetDude]

What's wrong with your P100D?

My problems are with the company, not the car. The car is amazing although there is a bit of a suspension rattle over bumps that they basically refuse to fix.

 

[CRASHER]

Oh I know about the 2017 FSD promises.

I only said "last year" because that is when Tesla actually attached "coming later this year" to the traffic light feature that is now being promised for this year.

Elon is man of his words;
In 2017, he promised FSD by end of the year
Come 2018, He held his promise, "FSD by end of the year"
In 2019 again, he did not break his promise, "FSD by end of the year"
Now 2020 again he is sticking with his promise, "FSD by end of the year"
I can assure that he will not break his promise of "FSD by end of the year" for many more years to come.

 

[Battpower]

Elon is man of his words;
In 2017, he promised FSD by end of the year
Come 2018, He held his promise, "FSD by end of the year"
In 2019 again, he did not break his promise, "FSD by end of the year"
Now 2020 again he is sticking with his promise, "FSD by end of the year"
I can assure that he will not break his promise of "FSD by end of the year" for many more years to come.

Then I need to know for sure if he said '....... the year.' or '........ this year.'

Tesla's future and reputation encapsulated in one simple question!

 

[murc585]

Longest review of 2020.12.5.6 I've seen to date. Much better than the 30 second clips I've seen.