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

[dckiwi]

He's at it again, this time on the ARK podcast: On the Road to Full Autonomy With Elon Musk — FYI Podcast

"I think we will be feature complete FSD this year, meaning the car will be able to find you in a parking lot, pick you up and take you all the way to your destination this year. I would say I'm certain of that. That is not a question mark."

He did then caveat that it would not 'work perfectly' and require observation. Even then - that would be an incredible accomplishment given where Tesla AP is today. They're either sitting on some incredible progress with HW3 or he is way off the mark - again.

 

[dckiwi]

He also mentioned that intersections are their next feature priority. In dev mode it is recognizing stop signs and stop lights, but they're still dealing with some corner cases. He suggested this functionality would be released first in the US.

No big surprises there, but any level of detail is worth mentioning.

He's also qualifying *a lot* of his claims on what regulators will allow. He even says that regulators may be slow to allow this functionality while the driver is supervising. Who knows if that is just a delay tactic.

I am cautiously optimistic that when they roll out HW3 they will add traffic light and stop sign recognition, at least in the US, this year. Just adding that feature would make AP a lot more useful. I drive long distances on either highways or large roads with traffic lights. With traffic light detection, most of my daily commute could be on autopilot.

 

[strangecosmos]

At 14:25: “And we’re really starting to get quite good at not even requiring human labelling. Basically the person, say, drives the intersection and is thereby training Autopilot what to do.”

Sounds like Tesla is using supervised imitation learning. (I wrote about this topic in depth here.)

 

[rnortman]

He did then caveat that it would not 'work perfectly' and require observation.

This is clearly the key. This is the direction Tesla has been moving in for quite some time now -- reposition "Full Self Driving" as an L2 or maybe highway-L3 system requiring supervision by a human driver. That is quite a let-down from the original promises of L5 and your car being able to generate revenue for you as an autonomous taxi in the Tesla Network.

What is the point of my car coming to find me in a parking lot if I must supervise it the entire time? I would sort of have to be standing right there to properly supervise it. Ditto for self-parking when I arrive at my destination.

 

[heltok]

At 14:25: “And we’re really starting to get quite good at not even requiring human labelling. Basically the person, say, drives the intersection and is thereby training Autopilot what to do.”

Sounds like Tesla is using supervised imitation learning. (I wrote about this topic in depth here.)

Could mean many things. Could be that they use imitation learning like the waymo paper about path planning from imitation learning, it could also mean that they train obstacle classification such as plastic bag->drive through, duck->stop automatically and then use their normal heuristic path planning for control. Or in the interesection use the driver action to train if green or red light and then use a neural net to evaluate data and see which one it disagree with the action of the driver and send those frame to manual labelling.

 

[humbaba]

There's also the issue of handling non-perfect conditions. Granted, my Tesla doesn't have HW3, but it requires clear lane markings in order to detect lanes. Once the snow and ice melted there was enough salt on the road that I couldn't see the lane markings. Because it is on my commute I know how many lanes there are and where they are, but my Tesla doesn't.

The Tesla mapping also doesn't exactly know a section of road that was added recently (its been open for a year or so). Its on the map, but route planning avoids it and, if you take it, tries to direct you onto side streets. A major problem with that is the side streets don't actually connect.

A problem for any FSD solution is the ever changing network of roads. They have to be able to adapt intelligently when what their sensors see disagrees with what their internal map data says. And currently it isn't clear how an FSD Tesla would behave when confronted with my morning drive.

 

[mongo]

This is clearly the key. This is the direction Tesla has been moving in for quite some time now -- reposition "Full Self Driving" as an L2 or maybe highway-L3 system requiring supervision by a human driver. That is quite a let-down from the original promises of L5 and your car being able to generate revenue for you as an autonomous taxi in the Tesla Network.

What is the point of my car coming to find me in a parking lot if I must supervise it the entire time? I would sort of have to be standing right there to properly supervise it. Ditto for self-parking when I arrive at my destination.

Feature complete, but monitored (2019 goal) is level 2-3
Driverless maneuvering in a parking lot is Level 4
Sleeping in the car all the way to your destination (2020 goal) is 4 or 5. 4 if, it only works under certain conditions, 5 if if it works under all conditions.

 

[rnortman]

Feature complete, but monitored (2019 goal) is level 2-3
Driverless maneuvering in a parking lot is Level 4
Sleeping in the car all the way to your destination (2020 goal) is 4 or 5. 4 if, it only works under certain conditions, 5 if if it works under all conditions.

Where does the "2020 goal" come from? When did he state that? Sorry, I haven't listened to the podcast that this thread is about, is it in there?

 

[happyzod]

There's also the issue of handling non-perfect conditions. Granted, my Tesla doesn't have HW3, but it requires clear lane markings in order to detect lanes. Once the snow and ice melted there was enough salt on the road that I couldn't see the lane markings. Because it is on my commute I know how many lanes there are and where they are, but my Tesla doesn't.

The Tesla mapping also doesn't exactly know a section of road that was added recently (its been open for a year or so). Its on the map, but route planning avoids it and, if you take it, tries to direct you onto side streets. A major problem with that is the side streets don't actually connect.

A problem for any FSD solution is the ever changing network of roads. They have to be able to adapt intelligently when what their sensors see disagrees with what their internal map data says. And currently it isn't clear how an FSD Tesla would behave when confronted with my morning drive.

Changing roads can probably be crowdsourced.

 

[croman]

Where does the "2020 goal" come from? When did he state that? Sorry, I haven't listened to the podcast that this thread is about, is it in there?

Yes, he said sleep in car "probably" and "maybe" by 2020.

 

[rnortman]

Yes, he said sleep in car "probably" and "maybe" by 2020.

What a dipstick.

 

[mongo]

What a dipstick.

Tesla's don't have dipsticks.

 

[rnortman]

“However, people sometimes will extrapolate that to mean now it works with 100 percent certainty, requires no observation, perfectly. This is not the case."​

Well, you don't say, you call it "full self driving" and people "extrapolate" to think that it will be, you know, full self driving.

How many people here spent the cash on the FSD option thinking that one day, with any luck, it would be able to drive itself while you carefully supervise, ready to take over at a moment's notice?

 

[humbaba]

Changing roads can probably be crowdsourced.

Doesn't matter how you do it, it has to be done. There is currently no evidence that anything is being done for it. Maybe it is fully realized in their internal dev build, but maybe it isn't.

When it comes to these things I tend to be conservative because it is wrong less often. Even reasonable optimism tends to "Elon Time." He is confident, but I'll wait for signs of progress -- just like I do with Waymo and Cruise.

 

[mongo]

“However, people sometimes will extrapolate that to mean now it works with 100 percent certainty, requires no observation, perfectly. This is not the case."​

Well, you don't say, you call it "full self driving" and people "extrapolate" to think that it will be, you know, full self driving.

How many people here spent the cash on the FSD option thinking that one day, with any luck, it would be able to drive itself while you carefully supervise, ready to take over at a moment's notice?

I would help if you included at least a sentence from each side of that quote... Was he referring to the 2019 release or the 2020 sleep in you car one?

 

[humbaba]

“However, people sometimes will extrapolate that to mean now it works with 100 percent certainty, requires no observation, perfectly. This is not the case."​

Well, you don't say, you call it "full self driving" and people "extrapolate" to think that it will be, you know, full self driving.

How many people here spent the cash on the FSD option thinking that one day, with any luck, it would be able to drive itself while you carefully supervise, ready to take over at a moment's notice?

Anyone buying the FSD option is essentially foolish. It has no predicted delivery date and only vague statements about its capabilities. The only rational reason to buy it is on the assumption that 1) it will be released someday and 2) the future cost of activation will be significantly higher than the present cost. Without a significant pricing differential you are losing the utility of having the money now. The longer it takes for FSD to be rolled out the greater the pricing differential needs to be.

Of course, most people are foolish at least some of the time. They may have misjudged both points above, or not even considered it rationally. Such is life.

Now, although I didn't buy FSD, I'd be much happier with a driver assist system than a fully autonomous system. Not least because no one developing autonomy seems to really believe how difficult it is. GM Cruise comes closest, but it is obvious from their own statements that it is superficial -- they don't really believe how difficult it is. This is classic "the last 20% is 80% of the effort" except that it is much more likely to be "the last 1% is 99% of the effort."

And, finally, there is the "small" matter of regulatory approval. In the US expect a lot of feet dragging on Tesla in favor of Waymo and Cruise. There's a lot of money riding on who gets there first and it is so easy to delay approval, or even approve Waymo or Cruise while denying Tesla by cherry picking differences and insisting those are "deal breakers."

 

[Camera-Cruiser]

I can't wait for it to be a reality, but seriously? I drive the same freeways and roads he does in Southern California and I have to disengage AP on my M3 multiple times from Sunset blvd to Hawthorne Blvd. That involves traveling on the 405 and 105 FWYs. He should drive the same firmware/hardware configs that his customers have. He may be surprised that it attempts to run you into the "jersey berm/k-rail" when in the carpool lane in underpasses, or randomly hit the breaks when driving south on the 405 in the early morning with the cast harsh shadows on the FWY.

I can't even imagine it getting you safely out of any of the Hollywood canyons, on to Sunset Blvd, and then the FWY at over 5 mph. Then again, that is often the speed on Sunset.

But, I'm hopeful.

 

[mongo]

Anyone buying the FSD option is essentially foolish. It has no predicted delivery date and only vague statements about its capabilities. The only rational reason to buy it is on the assumption that 1) it will be released someday and 2) the future cost of activation will be significantly higher than the present cost.

Or, they want to give Tesla a little more working capital...

Now, although I didn't buy FSD, I'd be much happier with a driver assist system than a fully autonomous system. Not least because no one developing autonomy seems to really believe how difficult it is. GM Cruise comes closest, but it is obvious from their own statements that it is superficial -- they don't really believe how difficult it is. This is classic "the last 20% is 80% of the effort" except that it is much more likely to be "the last 1% is 99% of the effort."

Oh, I think Tesla knows how hard it is. However, they also have the largest (and growing) instrumented fleet which is feeding back data to make the system better continuously.

I'm a terrible passenger, give me driver assist, or FSD while I sleep....

 

[dckiwi]

Anyone buying the FSD option is essentially foolish. It has no predicted delivery date and only vague statements about its capabilities. The only rational reason to buy it is on the assumption that 1) it will be released someday and 2) the future cost of activation will be significantly higher than the present cost. Without a significant pricing differential you are losing the utility of having the money now. The longer it takes for FSD to be rolled out the greater the pricing differential needs to be.

Sort of but not really.

You're assuming that all features of FSD will be released at once many years from now. If that were the case I would be inclined to agree with you. However the reason I paid for FSD on my Model 3 is that it's almost certain Tesla will start releasing FSD features incrementally in the relatively near future (this year).

For example, what if they roll out stop sign and traffic light detection later this year? Some people would (perfectly reasonably) say that such a feature isn't worth the $3,000 cost of FSD. Others, like myself, might consider that feature alone (along with all the other features of EAP) worth the cost, especially knowing that more is coming.

I can only speak for myself. I am fascinated by self-driving cars and automation in general. I bought FSD for no other reason than to 'go along for the ride', as bumpy and slow as that ride might be. I genuinely delight at even the slightest improvements to AP - I love testing them, seeing where they succeed, seeing where they fail. For that reason my EAP/FSD money was well spent, at least for me.

You really shouldn't label an entire group of people 'foolish' just because they have different priorities for their money or value experiences differently than you do.

 

[mongo]

You're assuming that all features of FSD will be released at once many years from now. If that were the case I would be inclined to agree with you. However the reason I paid for FSD on my Model 3 is that it's almost certain Tesla will start releasing FSD features incrementally in the relatively near future (this year).

My $0.02
Under the Software 2.0 design system, there is only one NN, all test cases and training data get processed together to produce that network. They may add feature enable/ disable switches, but I'm guessing it will all be released as one big update.
The incremental feature additions we are used to to have been due to the old development style as well as the hoops they need to jump through to get the NN to run on the AP2.x's limited HW.