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

[timewasted]

I understand people having disappointment when something they paid for is not available. Imo when you signed up with Tesla, you signed up for Elon’s timeline. This is not the first time delays are being experienced with a Tesla product. If you purchased a Tesla product, you must have followed Tesla products along with their delays before throwing down your deposit.

Whenever I see people post #Elon letdowns, I just shake my head. You want a letdown? Climb back into your ice and see what a letdown it is. This guy has shaken the automotive world and created vehicles that crush anything I have driven in 30 years. He has done this in a time span of 10 years. I haven’t seen any manufacturers come out with FSD that is ready now. Cadillac? please....a small dot on the map that can be used. EAP has improved alot since I first used it and now it is a joy to use. I could really use less nags for sure, but I understand them. Elon may over promise, but I trust him to always do the right thing. I can’t say that for any other automotive manufacturers. The price drop is going to create some noise for a while, but overall I think Tesla will make us whole.

FSD features will come soon imo and if it initially comes with nags, it’s still better than nothing at all. You still have to pay attention, but it’s less stressful driving having autopilot on with 360 degree vision. Last week driving late night, a truck came into my lane and autopilot reacted quicker than I would have. I’m a believer/fanboy because the tech has improved my life.

Thanks for letting me vent

This post upsets me greatly, because the first paragraph is hardcore victim blaming, and that’s just not OK. Throughout the entire process of buying my $80,000 car I was assured that self driving is right around the corner, especially since we have some favorable self driving rules/laws here in Nevada. I mean, there are laws preventing companies with a $50B market cap from making wildly untruthful statements, right? But no, it’s clearly our fault for falling for another practical joke from that lovable scamp Musk! I seriously don’t understand how Elon/Tesla get away with half the things their fans let them off the hook for.

I really do love my car, but the ONLY reason I put down so much of my money is because of the promise of a car that drives me to work in the morning. Electric is fun and great for the environment and all, but it’s not compelling enough for me on its own. Had anyone throughout the buying process been even remotely honest about the self driving aspect, I’d have instead bought a $30k Prius and banked the extra money until self driving was actually a thing.

 

[SteveD]

If full self driving was paramount for you, you should have waited for it to become a reality. I'm not surprised it is taking far longer for that to happen than anticipated. No offense intended, but anyone plunking money down on hype is foolish. I bought the EAP and could not be happier with the result. Exceeded my expectation. Full self driving will happen and Tesla will be one of the first companies to offer it. But it will take a lot of regulatory review and testing before it happens. Look at the SpaceX program. Big rockets landing perfectly on the landing pad, recent launch of the Dragon capsule that has to match up perfectly with the space station. The technology is here. It's only a matter of time. I think Tesla offers much more than just good gas mileage. BTW, I traded in my Prius to buy a Tesla.

 

[Bladerskb]

I think you are putting your spin on things. You are making it sound like Tesla just took the current autopilot, added traffic light and stop sign recognition and slapped an "FSD" sticker on it. That's not the case.

People don't realize how easy it is for a self driving team to create a "generalized self driving solution" as Elon calls it. So they say and will keep "you can use tesla anywhere but waymo and others are in a tiny geofence city" which is just plainly wrong. Its literally just an evolution of current ADAS. All you need is navigation based lane keeping with your added driving policy you created for your L4 geofenced system.

 

[Byegas]

Time wasted,

That was not my intent at all and sorry if I offended you. I’m honestly asking anyone that before they owned a Tesla product, have they heard of “Elon’s time”. His time is not our time and I have learned to accept that. I don’t look at it as giving him a pass, I look at it as this is who he is and how he thinks. In the end, he delivers a product that “crushes” the competition and I mean crush.

I know what I signed up for and while I’m not happy waiting, I understand the complexity that they have to solve. I enjoy the product I have now and look forward to the product that will come. Been using NOA the past month and it has added to the enjoyment of EAP. Incremental improvements and added features keeps me happy and I do get excited when a update pushes through.

 

[diplomat33]

People don't realize how easy it is for a self driving team to create a "generalized self driving solution" as Elon calls it. So they say and will keep "you can use tesla anywhere but waymo and others are in a tiny geofence city" which is just plainly wrong. All you need is navigation based lane keeping with your added driving policy you created for your L4 geofenced system.

I am not sure I follow. Are you saying that Tesla can create a L4 system by tightly geofencing what they currently call "FSD"?

 

[mongo]

People don't realize how easy it is for a self driving team to create a "generalized self driving solution" as Elon calls it. So they say and will keep "you can use tesla anywhere but waymo and others are in a tiny geofence city" which is just plainly wrong. Its literally just an evolution of current ADAS. All you need is navigation based lane keeping with your added driving policy you created for your L4 geofenced system.

Sure.... iff the larger region is merely copies of the smaller region with no differences... plus high res maps of how the roads are at that exact moment...
If you need maps for anything other than finding a route to your destination, it's not generalized.

 

[Bladerskb]

I am not sure I follow. Are you saying that Tesla can create a L4 system by tightly geofencing what they currently call "FSD"?

No i'm saying that basically any SDC company with geofenced L4 car can create elon's "generalized self driving solution". Because its simply an evolution of ADAS. Take lane keeping on a two way lane for SDC for example. You can have a good ADAS lane keeping with added navigation for where to drive mixed with driving policy you already developed for a L4 geofenced system and voila you have a generalized solution.

I can't speak for every SDC coompany but i can speak for Mobileye. There's nothing in their driving policy that is specific to a specific route.

Example #1; Two way lane, car pulls out and encroaches and blocks part of lane. ADAS today would fail. However you can port A L4 geofenced system driving policy to handle it.

Example #2; Two way lane, a truck is packed and blocking a whole lane. ADAS today would fail. However you can port A L4 geofenced system driving policy to handle it.

Elon Musk: The Recode interview

My point is, most tesla fans says Tesla system is better because its not geofenced. But actually that's not true.

 

[mongo]

My point is, most tesla fans says Tesla system is better because its not geofenced. But actually that's not true.

Don't know how you polled Tesla fans, but my concerns would be overfittng to the geofenced area and lack of data regarding regional variations.
A general (national) solution would inherently cover all regions, but there is no data to support that the limited system does also (and also no data to show it doesn't).

If other regions are just extensions, why do they geofence? Why does the L4 needed ported in your example? Is it general or not?

Examples 1 and 2 would be handled by a ported to street level NoA...

 

[zmarty]

Autopilot now can recognize stoplines : teslamotors - by @verygreen

 

[diplomat33]

No i'm saying that basically any SDC company with geofenced L4 car can create elon's "generalized self driving solution". Because its simply an evolution of ADAS. Take lane keeping on a two way lane for SDC for example. You can have a good ADAS lane keeping with added navigation for where to drive mixed with driving policy you already developed for a L4 geofenced system and voila you have a generalized solution.

I can't speak for every SDC company but i can speak for Mobileye. There's nothing in their driving policy that is specific to a specific route.

Example #1; Two way lane, car pulls out and encroaches and blocks part of lane. ADAS today would fail. However you can port A L4 geofenced system driving policy to handle it.

Example #2; Two way lane, a truck is packed and blocking a whole lane. ADAS today would fail. However you can port A L4 geofenced system driving policy to handle it.

Elon Musk: The Recode interview

My point is, most tesla fans says Tesla system is better because its not geofenced. But actually that's not true.

I appreciate all the thought you put into answering my question. I don't think I really disagree with anything you wrote.

As I understand it, the reason Musk calls Tesla's system a "generalized SD solution" is because it is capable of self-driving without geofencing. it is essentially a "self-driving equation" that does not depend on location. The equation is to take auto steer (controlling direction) and adaptive cruise control (controlling speed) and add navigation (controlling where the car should drive) and add the right driving policy (controlling how the car should drive), and you get self-driving that is not dependent on location, hence why it is a generalized solution. That's essentially Musk's "self-driving equation". And depending on how well you "solve" the equation, will determine if you get to L4 or not. So yes, you are correct that any L4 system can be "downgraded" to duplicate what is presumably Tesla's FSD. But no, I would not say that Tesla's system is better because it is not geofenced. It is not inherently better or worse. It's merely an "equation" for creating self-driving that depending on well you solve it, will determine how good your self-driving is.

And regarding your two examples. Yes, a traditional ADAS would fail. And we have not seen Tesla's FSD in action to know how it would handle those scenarios. Presumably, Tesla's FSD would have the right driving policy not to fail in those scenarios because otherwise, it would not be self-driving.

 

[diplomat33]

Autopilot now can recognize stoplines : teslamotors - by @verygreen

This feels me with confidence. The naysayers can say that Tesla does not have what it takes to do FSD but it's evidence like this that proves that Tesla is incrementally adding the necessary features to eventually make FSD.

 

[electronblue]

This feels me with confidence. The naysayers can say that Tesla does not have what it takes to do FSD but it's evidence like this that proves that Tesla is incrementally adding the necessary features to eventually make FSD.

Both can be right though — it seems like Tesla today does not really have what it takes for FSD, especially at a SAE Level above 2. That does not mean they are not making progress or that they can never get there. Also there is the relative aspect: we were lead to believe Tesla was so much ahead of others and now they are only just implementing things others have had for a long time.

But most importantly history has shown that there is reason to doubt Tesla’s progress for current generations of cars. AP1 was left with many unfulfilled promises and even announced features with timelines not delivered (like Navigation-based exit-taking for AP1 for December 2016). Of course Tesla is fulfilling many of its AP1 promises with AP2.5 but that does not help on AP1...

When will those with AP2 find themselves in a similar situation? Where the limits are for AP2 or AP2.5 (even in an upgraded form) may be different from where Tesla as a company itself will eventually get to with future products and this is a genuine concern for someone who forked out money back in 2016 on the promise of ”Level 5 capable hardware”.

Many of us are just trying to keep the discussion realistic. Don’t let the hype get the best of you and you’ll likely fare better with Tesla.

As said Tesla’s Level 2 features are still interesting from a user perspective. That is why many of us here do drive Teslas.

 

[rnortman]

If full self driving was paramount for you, you should have waited for it to become a reality. I'm not surprised it is taking far longer for that to happen than anticipated. No offense intended, but anyone plunking money down on hype is foolish.

This is a very common deflection of blame from Tesla to the customer. In other words, victim blaming: "You should have known Tesla was lying when you bought the car, it's your own fault!"

Can you be clear about what you mean by this post? Is that a defense of Tesla or not?

 

[rnortman]

This feels me with confidence. The naysayers can say that Tesla does not have what it takes to do FSD but it's evidence like this that proves that Tesla is incrementally adding the necessary features to eventually make FSD.

Do you have any idea how easy it is to recognize most stop lines most of the time? This is like an undergrad intern ML project.

The question is how reliable is it? What is the mAP (mean average precision) when measured over a diverse data set including poorly marked and completely unmarked intersections? (And yes, in principle, you could train a stop line detector to detect unmarked stop lines! It would take a substantial labeling effort.) And then the next question is -- what does this actually buy you in terms of L3+ autonomy? The answer is -- very little. Could be useful for toy L2 gimmicks.

This should instill you with despair that this is a new feature 3 years into FSD development. This is a sign that they are aiming for an L2 system for local roads.

 

[diplomat33]

But most importantly history has shown that there is reason to doubt Tesla’s progress for current generations of cars. AP1 was left with many unfulfilled promises and even announced features with timelines not delivered (like Navigation-based exit-taking for AP1 for December 2016). Of course Tesla is fulfilling many of its AP1 promises with AP2.5 but that does not help on AP1...

When will those with AP2 find themselves in a similar situation? Where the limits are for AP2 or AP2.5 (even in an upgraded form) may be different from where Tesla as a company itself will eventually get to with future products and this is a genuine concern for someone who forked out money back in 2016 on the promise of ”Level 5 capable hardware”.

Those are legit concerns. And the answer is most likely that yes, AP2 owners will be left behind at some point too. Computer tech is moving so fast that it becomes obsolete pretty quickly. We see this to some extent with our own home computers where 2 years later, our computer is already behind the curve. Luckily, Tesla is promising to upgrade FSD buyers to AP3 to do right by them. So AP2 owners will at least get these L2 "FSD like" features. But I have no doubt that eventually AP4 will come out and then AP5 will come out and previous hardware suites will be left behind.

Many of us are just trying to keep the discussion realistic. Don’t let the hype get the best of you and you’ll likely fare better with Tesla.

Keeping the discussion realistic is good.

Do you have any idea how easy it is to recognize most stop lines most of the time? This is like an undergrad intern ML project.

The question is how reliable is it? What is the mAP (mean average precision) when measured over a diverse data set including poorly marked and completely unmarked intersections? (And yes, in principle, you could train a stop line detector to detect unmarked stop lines! It would take a substantial labeling effort.) And then the next question is -- what does this actually buy you in terms of L3+ autonomy? The answer is -- very little. Could be useful for toy L2 gimmicks.

Yes, reliability is the key. It needs to be incredibly reliable, essentially indistinguishable from 100% to be useful for L3 or L4.

This should instill you with despair that this is a new feature 3 years into FSD development. This is a sign that they are aiming for an L2 system for local roads.

Well, I am a half glass full kinda guy. Considering that for years now, we had no idea where Tesla was in their FSD development. All we knew for sure is that they were apparently struggling to duplicate what they accomplished in the 2016 FSD video, I am just glad to see some concrete evidence of their progress.

 

[mongo]

But most importantly history has shown that there is reason to doubt Tesla’s progress for current generations of cars. AP1 was left with many unfulfilled promises and even announced features with timelines not delivered (like Navigation-based exit-taking for AP1 for December 2016). Of course Tesla is fulfilling many of its AP1 promises with AP2.5 but that does not help on AP1...

AP1 was also MobilEye based. Plans pre July 2016 would of course be disrupted.

Do you have any idea how easy it is to recognize most stop lines most of the time? This is like an undergrad intern ML project.

And thus why Tesla starts out with supervised operation.

The question is how reliable is it? What is the mAP (mean average precision) when measured over a diverse data set including poorly marked and completely unmarked intersections? (And yes, in principle, you could train a stop line detector to detect unmarked stop lines! It would take a substantial labeling effort.)

Well, if you have hundreds of thousand if car available along with years of data, you can built up quite a real world data set to test against. If you hired an outside company, you can get lots of labling on the data. If you hired 3D gane engine developers, you can even alter the data you do have to make test cases.
So, if you have the infrastructure already, you can determine reliability and also train quickly.

To validate, you release in supervised more. Or, for minimal data, set up a shadow trigger to get pass/ fail judged against known intersections via GPS maps .

And then the next question is -- what does this actually buy you in terms of L3+ autonomy? The answer is -- very little. Could be useful for toy L2 gimmicks.

Stopping at stop lines is a legal requirement of driving, so it makes your car compliant with laws on local streets. Something you need in L1-L5 (but not L4 if you exclude surface streets).

This should instill you with despair that this is a new feature 3 years into FSD development. This is a sign that they are aiming for an L2 system for local roads.

Why? How's it starting on highway driving and moving to local roads any different than geofencing to a region and then expanding interns of despair level? By your statement, highway only systems should despair...

To follow laws on local roads, it must detect stop lines ...

Further, to judge rate of progress you need to time it against the most recent shift in AP NN training approach, not the original start of development.

 

[Bladerskb]

Autopilot now can recognize stoplines : teslamotors - by @verygreen

This feels me with confidence. The naysayers can say that Tesla does not have what it takes to do FSD but it's evidence like this that proves that Tesla is incrementally adding the necessary features to eventually make FSD.

Honestly that doesn't look good at all. false positives and inaccuracies all over the place.
Mind you that Mobileye was working on a long range stop line detection and a short range stop line detection network way back in EQ3 days.
Now these two networks are both in EyeQ4 which released in 2017. This is literally tesla catching up, this doesn't give confidence to any prediction that elon is making, this actually takes the confidence away.

 

[iTech]

Honestly that doesn't look good at all. false positives and inaccuracies all over the place.
Mind you that Mobileye was working on a long range stop line detection and a short range stop line detection network way back in EQ3 days.
Now these two networks are both in EyeQ4 which released in 2017. This is literally tesla catching up, this doesn't give confidence to any prediction that elon is making, this actually takes the confidence away.

I would agree with that, but I want to see how it would look in the HW3 + new NN. I wonder if there's a huge difference.

 

[Krypto Kat]

the ONLY reason I put down so much of my money is because of the promise of a car that drives me to work in the morning. Electric is fun and great for the environment and all, but it’s not compelling enough for me on its own.

I am with you on this. My thoughts exactly.

 

[diplomat33]

Honestly that doesn't look good at all. false positives and inaccuracies all over the place.
Mind you that Mobileye was working on a long range stop line detection and a short range stop line detection network.
These two networks are both in EyeQ4 which released in 2017. This is literally tesla catching up, this doesn't give confidence to any prediction that elon is making, this actually takes the confidence away.

The false positives and inaccuracies can be fixed. Keep in mind this is just the 2019.4 update and it's using the AP2 chip that Tesla has admitted is not good enough for this. It definitely does not represent what the public will get. So I am not worried about that. With AP3 chip, we should see some pretty big improvements.

Yes, Tesla is playing catch up. But that's not surprising since, as you probably know, Tesla has not been working on FSD for nearly as long as the other SDC like Waymo. And actually, if you look at EAP development, Tesla's catch up has been quite remarkable. Tesla released EAP in Dec 2016 (if memory serves) and it was utterly broken. It could not even do good lane keeping. Then EAP had to catch up to just AP1. Now, just 3 years later, EAP has solid lane keeping, solid adaptive cruise control, auto lane changing, navigate on autopilot on the highway and can recognize stop lines, traffic lights, stop signs etc... In fact, it's progressed to the point where Tesla is preparing to release what they call "FSD" that can fully automate driving on highways and city streets (L2 "self-driving"). Doing all that in 3 years is pretty good if you ask me.

And correct me if I am wrong but is there a mass market car that the public can buy today for $35k that can do city "self-driving" even at L2? If so, I am not aware of any. So, if Tesla does release their "FSD" software even at L2, it will be big because they will be first to market with a system that can fully automate driving on highway and city streets. And with so many cars on the road, Tesla will be able accelerate its validation to L4. Tesla is basically taking the approach of releasing a system at L2 to all their cars, using the massive feedback to validate the system and then with a mere software update and maybe some hardware updates as well, upgrade all the cars to L4.