Elon Musk Says Tesla Is 'Very Close' To Level 5 Self-Driving Technology

[Soda Popinski]

Not even close. Use FSD every day and love it. But NO WHERE near level 5.

Can’t change lanes on surface street.
Can’t keep up with flow of travel
Can’t make turns.
Can’t handle round abouts.
Can’t handle handle side walks with stoplight beta on.
Constant phantom braking.
No going even +1 over speed limit
Anymore.
Speed limit In nav database often wrong with posted signs.

BARELT level 2. NO WHERE near L4 or L5.

I'm pretty sure Elon Musk is talking about the FSD development build, not what our cars are currently doing. Remember, it's a big re-write not based on HW 2 or 2.5, so we can't really extrapolate from what we currently are using.

That said, we also have no idea or definitive evidence on FSD progress besides the advancements in perception Karpathy has shown in his talks. We do know that accurate perception is necessary, but not sufficient for Level 4 or Level 5 autonomy. Case in point, Waymo does a great job in perception, but still isn't confident enough in the way their cars drive to roll their robo-taxies out to the unrestricted public.

 

[powertoold]

'm pretty sure Elon Musk is talking about the FSD development build, not what our cars are currently doing. Remember, it's a big re-write not based on HW 2 or 2.5, so we can't really extrapolate from what we currently are using.

Elon starts off the WAIC questions by saying AP works "reasonably well" in China. He's not stupid. He knows AP isn't the FSD build. Then he goes on to say he's extremely confident in level 5 soon. Either he's totally crazy or their build is somewhat decent.

Even if Elon is overhyping the FSD build, it has to be "2-3x" better than AP, which to me would be great.

Also when does Elon say he's extremely confident of anything? He is a very calculated risk taker and progress assessor.

 

[diplomat33]

That said, we also have no idea or definitive evidence on FSD progress besides the advancements in perception Karpathy has shown in his talks. We do know that accurate perception is necessary, but not sufficient for Level 4 or Level 5 autonomy. Case in point, Waymo does a great job in perception, but still isn't confident enough in the way their cars drive to roll their robo-taxies out to the unrestricted public.

It is worth nothing that Waymo cars can do a lot more than just perception. They can handle quite a lot of driving scenarios. On the Waymo Support page, it says that Waymo cars can do the following:

• Navigate on city streets to get from point A to B safely.
• Avoid hazards including by slowing down, changing lanes, or braking completely for a stopped car.
• Adjust to unexpected changes in the roads like road work or closed lanes.
• Obey traffic laws like stopping at stop signs, and yielding to pedestrians or cyclists.

Learn how Waymo drives - Waymo Help

This allows Waymo cars to be quite good. And yet, Waymo is not ready to permanently remove the safety driver from all their cars just yet. I think this bolsters your point that getting to driverless autonomy is not a given even if you have great perception. Achieving safe and reliable driverless autonomy is quite challenging, especially in dense urban driving.

 

[TLLMRRJ]

Once it's done (if ever), even in rudimentary form, I'm sure it'll wow me as when SpaceX first landed its rocket on a barge.

It's worth realizing that, while landing rocket boosters on a barge is super cool, it is 1 million times easier to do that than make a self-driving car that operates in the current driving world.

Think about it. The tech to land a rocket booster on a barge was perfected in the 60's moon missions and all of the same navigation tech is now built into inexpensive kids toys in the form of drones. There's so little to worry about compared to pedestrians, road conditions, weather, other drivers, etc. when landing a rocket booster in the ocean on a barge on a good weather day (they don't even try to launch rockets in bad weather).

 

[powertoold]

Think about it. The tech to land a rocket booster on a barge was perfected in the 60's moon missions and all of the same navigation tech is now built into inexpensive kids toys in the form of drones.

Ugh. I feel this sense of lumping all tech into one simple pile. Just because some tech was used prior, doesn't mean it's easy, obvious, or anything. Landing on a barge, pinpoint accuracy, on earth, with enough fuel, is not a million times easier.

I don't mean to deride anyone, but this reminds me of diplomat's perspective that since other fsd companies also use ML, then they have all the same potential as Tesla when it comes to developing fsd.

I agree with you that landing on a barge is easier than FSD. What I was more referring to is the sense of awe. From a tech perspective, there's not much comparable to landing a large rocket booster on an autonomous barge.

 

[diplomat33]

I don't mean to deride anyone, but this reminds me of diplomat's perspective that since other fsd companies also use ML, then they have all the same potential as Tesla when it comes to developing fsd.

No. It's not just that other companies use ML, it's how they use it. Other companies are using ML to do the same things that Tesla is doing. They are using ML to solve the same perception problems as Tesla. For example, they also have 3D labeling, object detection, pseudo-lidar, etc.... Just look at Mobileye. Mobileye has over 13 Neural Networks for vision covering object detection, road signs, driveable space and more.

The big difference is that the other companies might also be doing ML with their lidar data in addition to vision ML since they also use lidar as part of their FSD whereas Tesla has to focus all their ML on just vision since Tesla just has camera vision to solve perception. But they are both using similar ML tools to solve FSD problems.

 

[whttiger25]

Not even close. Use FSD every day and love it. But NO WHERE near level 5.

Can’t change lanes on surface street.
Can’t keep up with flow of travel
Can’t make turns.
Can’t handle round abouts.
Can’t handle handle side walks with stoplight beta on.
Constant phantom braking.
No going even +1 over speed limit
Anymore.
Speed limit In nav database often wrong with posted signs.

BARELT level 2. NO WHERE near L4 or L5.

So despite ALL those drawbacks, you still use fsd every day and love it?

 

[whttiger25]

Elon doesn't think in individual features. He thinks in foundations and potential progress. I think he feels confident that Tesla has the foundation to achieve FSD, but they just need to collect more of the edge cases and do the tedious work.

As CEO of big companies, he can't go too much in the weeds. He just focuses on the foundational developments, and the engineers can refine the technology.

I feel like true FSD is about a merger between sensors and AI that is not only quantifiable safer than average human drivers, but also feels safer to passengers. Seems we are nowhere near that, it isn’t about edge cases. Those are important, but there is such a lo g way to go to what level 5 is supposed to be.

 

[ElectricIAC]

I'll believe it when I see it.

 

[whttiger25]

Uhm... You are right, I can't guarantee what he was feeling, I can only go by what he was saying, like FSD features to begin rolling out by end of 2016 (while not even automatic headlights or wipers worked by end of 2016) and Elon's video on their website showing FSD and a claim that Tesla will demonstrate coast-to-coast FSD by end of 2017. Maybe you are right, maybe he didn't actually feel that way, instead was plain lying to people while showing fake videos to get more people to pay for FSD. Maybe he still is not feeling this way today, just lying?


Really? Did you read the description of FSD Elon was selling in 2016?
View attachment 563680
Did you not see Tesla marketing video showing the car driving itself through city streets, intersections, roundabout, parking itself, etc? That was all part of 2016 marketing.


After he said it would be done in 2016, then end of 2017, then 2018, etc... It's always "3 months maybe, 6 months for sure" (actual Elon quote) or "sooner than anyone thinks" (also an Elon quote about FSD from few years ago).


Ah yes, the Texas sharpshooter method, keep redefining the target to be able to claim hitting it. When Elon said during AP1 reveal that the car will find me anywhere on the property, it meant as long as it's no more than 40 feet away in a straight line from where the car starts, and while I hold a dead man's switch. He just "forgot" to qualify that, right. When Elon boasted how P85D was a 700hp car, he mean Elon's custom definition hp, because after few years Tesla published actual hp (as defined by standard) and it turned out Elon's 691hp = 463 actual hp, like the rest fo the world defines it. Now you're saying "To Elon Level 5 means something completely different". Ok, if by Level 5 Elon means whatever the state of the car will be in few months, of course he's guaranteed to meet it, by definition. Using your logic, why don't we just agree that all Tesla's are at Elon's Level 11 autonomy right now, which we can if we don't use any standardized definitions, sound good to you? However, drop all the level stuff, when is FSD going to do everything that Elon sold in 2016 (read the above screenshot again) - no Levels, just plain text from Tesla FSD description.

Dude I gotta say, reading that screen shot made me literally laugh out loud!

I’m all for the Tesla mission and I have loved my cars, but... I’m amazed this copy was approved but an entire team of people responsible for building and marketing a car like The model S. Most had to know that the 2016 model would not be able to deliver this.

 

[whitex]

Dude I gotta say, reading that screen shot made me literally laugh out loud!

I’m all for the Tesla mission and I have loved my cars, but... I’m amazed this copy was approved but an entire team of people responsible for building and marketing a car like The model S. Most had to know that the 2016 model would not be able to deliver this.

I've heard that engineers knew that it would never happen as described with their hardware back then, but Elon's style is to set aside anyone who says whatever he wants cannot be done in the time that he wants (by moving them out of the team and usually firing them afterwards), so he found engineers to tell him it is possible. When faced with a choice of working on really cool stuff or being fired over saying something is impossible, most just chose to agree that there is a chance (and some might have even justified it to themselves that there probably was, however small, maybe 1e-99999999999999999%, but that is still greater than 0%). After couple of years working on Tesla bleeding edge stuff which gets deployed in the real world, they have no problem getting jobs elsewhere, so they leave.

 

[diplomat33]

Here is what I see as a basic road map for Tesla getting to L5:

Stage 1: Finish Perception where the car can reliably see everything it needs to see. Release "autosteer on city streets" and "reverse summon". This should include turning at intersections and enabling autosteer on roads with no lane lines such as campus and residential streets, and rural roads. Reverse summon should be able to do "auto park" without adjacent cars, just lines.

Stage 2: Combine smart summon, AP on city streets, Highway NOA, reverse summon and auto park into a single NOA so that you can input an address in Park and NOA can take you from your driveway to a parking lot across town. A this stage, there will still be disengagements but the car will at least allow you to enable NOA on all road surfaces (parking lots, roads, streets, highways) and the car can follow a route in the nav from start to finish.

Stage 3: Solve all the remaining cases such as pulling over for first responder vehicles, obeying all traffic signs, avoiding road hazards safely, going around parked cars when you have to cross into the incoming traffic lane, handling an intersection when the traffic lights are out, handling city construction zones, etc...

 

[mmanner]

Someone said that FSD won't really be a reality until all cars have it, and I think that is correct, at least on non access controlled highways. Maybe the technology can get to the point it can deal with the complexities of navigating city streets no matter what other vehicles are doing, but I think that we won't have true FSD until some high percentage of cars are similarly equipped. I hope that I am wrong, since I think people are awful drivers as a group and I'd like to avoid getting run over or squashed when going about in my car (or bike or foot).

 

[diplomat33]

Someone said that FSD won't really be a reality until all cars have it, and I think that is correct, at least on non access controlled highways. Maybe the technology can get to the point it can deal with the complexities of navigating city streets no matter what other vehicles are doing, but I think that we won't have true FSD until some high percentage of cars are similarly equipped. I hope that I am wrong, since I think people are awful drivers as a group and I'd like to avoid getting run over or squashed when going about in my car (or bike or foot).

Having all cars on the road be autonomous would certainly make things a lot easier and safer but I don't think it is a requirement to have autonomous driving. You can design your autonomous car to perceive the environment and other vehicles and react accordingly regardless of whether the other vehicles are autonomous. And we already see companies like Waymo and Cruise that have really good autonomous driving that can handle most city driving. And yes, they can handle the craziness of human drivers pretty well by training the car to be more cautious. Of course, it does make autonomous driving harder because you have to program your car on how to be more conservative in a lot more scenarios. But I agree that having all cars be autonomous would certainly be a lot better.

Here we see a Cruise car drive completely autonomously in the city:

 

[mmanner]

Having all cars on the road be autonomous would certainly make things a lot easier and safer but I don't think it is a requirement to have autonomous driving. You can design your autonomous car to perceive the environment and other vehicles and react accordingly regardless of whether the other vehicles are autonomous. And we already see companies like Waymo and Cruise that have really good autonomous driving that can handle most city driving. And yes, they can handle the craziness of human drivers pretty well by training the car to be more cautious. Of course, it does make autonomous driving harder because you have to program your car on how to be more conservative in a lot more scenarios. But I agree that having all cars be autonomous would certainly be a lot better.

I think your last point is really spot on. I just did a 3,000 mile trip in my new Y, mostly Interstate highways, and it did a very good job in general. When there was construction or lots of traffic, it managed, but it did so in a way that was more tentative than a good human driver under the circumstances, which didn't mesh well with the flow. It made me think how smooth things would have been if the other vehicles had FSD systems and were communicating with each other. I do hope that we can get to FSD or close with individual vehicle sensors/software, because waiting for most cars to get it will be a long wait I suspect.

 

[diplomat33]

I think your last point is really spot on. I just did a 3,000 mile trip in my new Y, mostly Interstate highways, and it did a very good job in general. When there was construction or lots of traffic, it managed, but it did so in a way that was more tentative than a good human driver under the circumstances, which didn't mesh well with the flow. It made me think how smooth things would have been if the other vehicles had FSD systems and were communicating with each other. I do hope that we can get to FSD or close with individual vehicle sensors/software, because waiting for most cars to get it will be a long wait I suspect.

I definitely want to see as many autonomous cars on the road as possible. But yes, it will be a long wait before most consumer cars on the road are autonomous. This is because even when the tech becomes available, it will still take years for automakers to design and produce the cars and for people to trade in their cars and buy a new one. According to the IHS, the average age of a car on US roads last year was 11.8 years. Source: Average age of vehicles on U.S. roads hits 11.8 years
That is an average so some cars might be younger and others might be older. I'm sure you've seen plenty of cars around you that are more than 10 years old. Some people still drive cars that are 20 years old. So even when automakers sell autonomous cars, if we just wait for people to trade in their cars on their own and buy an autonomous car, it could still take 10+ years before a majority of cars on the road are autonomous.

This might be another good argument for ride-hailing robotaxis because deploying robotaxi fleets in cities might be a lot faster than waiting for millions of people to trade in their cars and buy an autonomous car.

 

[Soda Popinski]

I definitely want to see as many autonomous cars on the road as possible.

Not that this is your sentiment, but it reminds me of a survey I heard about a few years ago, that most people surveyed wouldn't want an autonomous car, while most wanted other people's cars to be autonomous.

 

[diplomat33]

Not that this is your sentiment, but it reminds me of a survey I heard about a few years ago, that most people surveyed wouldn't want an autonomous car, while most wanted other people's cars to be autonomous.

Yeah, I know that most surveys seem to suggest that a lot of people are skeptical or even afraid of autonomous cars and don't really want them. I am not one of them. I LOVE the idea of autonomous driving and welcome as many safe and reliable autonomous cars on public roads as soon as possible.

 

[DrtyJrze]

I definitely want to see as many autonomous cars on the road as possible. But yes, it will be a long wait before most consumer cars on the road are autonomous.

All cars don’t necessarily have to be autonomous for this to work. If some sort of transponder could be connected to older cars that can communicate with autonomous cars we could net a large portion of the advantages at a significantly lower cost.

I would imagine insurance companies would have a huge incentive to push and/or subsidize this tech to customers. Similar to that plug in box people use now to get reduced insurance rates. The amount of money they would save in claims would offset the cost. Or, they could make the difference between using the plugin or not so high that most people will conform.

 

[diplomat33]

All cars don’t necessarily have to be autonomous for this to work. If some sort of transponder could be connected to older cars that can communicate with autonomous cars we could net a large portion of the advantages at a significantly lower cost.

I am not sure what problem this transponder would fix though. All autonomous cars need to have excellent perception that can track other cars and what they are doing on the road. So having a transponder to tell the autonomous car where other cars are and what they are doing, is not necessary. The problem is that the humans in the older cars are bad drivers. I am not sure how a transponder would fix that.

The best way to fix the problem of bad drivers is to replace the bad drivers with a capable autonomous car.