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

[Az_Rael]

They're releasing robotaxis this year! It's going be to be tough to avoid liability...

You are correct, LOL. I also forgot about the sleeping in my car by 2020 (if I recall the quote correctly) after that feature complete was done end of 2019.

 

[rnortman]

For all of the healthy skepticism of Tesla's optimism around the HW3 hardware achieving any higher level of autonomy, a thought did occur to me this morning.

If you want to buy a car today that has a possibility of being autonomous at some point in the future, I think Tesla is the only option. It may not be a big possibility, but given how inflexible so many other car operating systems are, Tesla is the only one that offers any sort of feasible upgrade path.

And if you want to buy a scrap of paper for a few dollars that has a possibility of making you a millionaire, I think a lottery ticket is the only option.

The current generation of Tesla vehicles will never exceed L3 autonomy without extensive hardware retrofitting, and Tesla is certainly not going to be the one doing the retrofitting. (Nobody will, because it won't make economic sense.) IMO they will never even reach true L3 -- meaning the car gives you at least 10 seconds to take over; none of this take over now stuff. You can say there is some very slight chance of them achieving better than this, but see the lottery analogy above for my take on that.

 

[diplomat33]

But GPS is used to pinpoint what 3D maps to load, if the difference between what you see and what the GPS tells you is too big, then what?

I don't think the difference would be so big that you would load the wrong HD map. As long as you know what city you are in, you can load the right HD map and then you rely on your non-GPS methods for determining your exact location. For example, your cameras see what lane you are in, they see signs that say "Parker Ave" and "Madison Ave", they see a post office to the left, and they know the distance from each, you can calculate your position on the HD map. Lidar can also help in calculating the exact distance and angle to those landmarks, also giving you your exact location on the HD map. You don't need GPS.

This is basically how the leaders in autonomous driving do accurate localization and why they use LIDAR and HD maps.

Now, it may be different for Tesla since Tesla does not use HD maps or lidar. Tesla might be trying to use GPS that is not accurate enough for good localization. It might explain some of the issues NOA has, like missing exits.

 

[Mardak]

I have no doubt that Tesla can add a lot of capabilities with software and will. I just suspect that there might be some cases that require additional hardware, not to do, but to do well. Remember, I am not talking about merely doing a feature but about doing the feature well enough that the driver is not required.

Curious, what is your level of "required?" For example, if Autopilot correctly stops for a car with a door open partially blocking its driving lane resulting in Autopilot waiting for the path to clear whereas a human might decide to cross over into oncoming traffic after waiting a few seconds. Is this "feature complete" or even acceptable for "level 5" as I've definitely seen human drivers decide to wait for the door to close instead of passing.

 

[emmz0r]

I don't think the difference would be so big that you would load the wrong HD map. As long as you know what city you are in, you can load the right HD map and then you rely on your non-GPS methods for determining your exact location. For example, your cameras see what lane you are in, they see signs that say "Parker Ave" and "Madison Ave", they see a post office to the left, and they know the distance from each, you can calculate your position on the HD map. Lidar can also help in calculating the exact distance and angle to those landmarks, also giving you your exact location on the HD map. You don't need GPS.

This is basically how the leaders in autonomous driving do accurate localization and why they use LIDAR and HD maps.

Now, it may be different for Tesla since Tesla does not use HD maps or lidar. Tesla might be trying to use GPS that is not accurate enough for good localization. It might explain some of the issues NOA has, like missing exits.

Yeah but you need a lot of heuristics to map it on and those can change. If you have to do all that advanced stuff then I don't see much more use for it than some planning ahead. Oops road works, you're screwed. Then it has to understand its environment completely anyway.

I just think that you end up with so many correction and coping mechanisms that the HD map becomes somewhat of a crutch.

I was actually thinking of this quote "High-precision GPS maps for self-driving cars are a “really bad idea,” according to Musk, resulting in a “system [that] becomes extremely brittle” by being too dependent and not being able to adapt."

 

[diplomat33]

Curious, what is your level of "required?" For example, if Autopilot correctly stops for a car with a door open partially blocking its driving lane resulting in Autopilot waiting for the path to clear whereas a human might decide to cross over into oncoming traffic after waiting a few seconds. Is this "feature complete" or even acceptable for "level 5" as I've definitely seen human drivers decide to wait for the door to close instead of passing.

As long as it can handle the situation safely and reliably, that's all that really matters. A L5 car would need to know what all the objects around it are doing and have a good driving policy for whether to wait or whether to pass and if it passes, to do it safely when the other lane is clear.

And just to give you an idea of what safety standard is needed to remove the driver: Mobileye is aiming for camera only autonomous car that can do 10,000 hours of driving without a safety critical failure as a starting point. Then, with the added redundancy of radar and lidar, they aim to increase that to 10 million hours of driving with no safety critical error to make it much safer than a human driver.

 

[diplomat33]

Yeah but you need a lot of heuristics to map it on and those can change. If you have to do all that advanced stuff then I don't see much more use for it than some planning ahead. Oops road works, you're screwed. Then it has to understand its environment completely anyway.

I just think that you end up with so many correction and coping mechanisms that the HD map becomes somewhat of a crutch.

I was actually thinking of this quote "High-precision GPS maps for self-driving cars are a “really bad idea,” according to Musk, resulting in a “system [that] becomes extremely brittle” by being too dependent and not being able to adapt."

No, that is not how HD maps work. Like I have said before, HD maps don't include every little object, and the car does not rely on the HD maps as the main source to drive. HD maps only include the stuff that does not move, like roads, signs, traffic lights, etc.... And HD maps are just a back up. So the car still relies on its sensors to actually drive and navigate but it can use the HD maps to double check and make sure it is doing the right thing.

I think Elon dismisses HD maps because he figures the car will use vision anyway and if vision is good enough then HD maps are irrelevant. But pretty much every leader in autonomous driving uses HD maps because they recognize that having a back up that the car can use to double check what it is doing, is a good idea if you want a more reliable system. After all, there will be cases where visibility is poor or maybe you encounter an edge case with a weird looking traffic light that the camera is not sure about. It is a good idea to have a HD map that can help your vision out.

 

[daniel]

That's what Tesla is working on now. Yeah, it's a tall order. It requires excellent localization, mapping, and vision to see exactly where the other objects are and what they plan to do and have excellent driving policy to dictate how our Teslas should respond. It will be the first real test of Tesla's FSD approach. City driving is the hardest part of autonomous driving. It's why other companies have started with city driving and pretty much ignored highway driving. But it is doable with the right hardware and software. Mobileye can already do what you talk about:

Tesla has a version of "city NOA" working now in their internal development but it's probably not good enough yet. I expect it will probably be released to the public later this year if all goes well.

That video is damn impressive! I really hope I'm wrong about Level 4 being a decade away.

Looking at how Tesla has deployed Autopilot features like Automatic Lane Change, I would guess City Street driving will be closer to the overly cautious side initially. And it wouldn't be hard to classify stopping at a stop sign then missing multiple opportunities to enter the intersection before finally entering as "feature complete" as it satisfies both stopping and determining when it's safe to enter the intersection albeit rather poorly.

This is my concern as well. My EAP brakes when a car half a block away makes a left turn in front of me. It should not require HW3 to figure out that that car is too far away to be a concern.

And if you want to buy a scrap of paper for a few dollars that has a possibility of making you a millionaire, I think a lottery ticket is the only option.

Excellent analogy.

Though I've always thought that Waymo was way ahead of Tesla, and the video above, if it is genuine, suggests that Mobileye is also. (Though if Mobileye is part of Microsoft, then I'm suspicious.)

 

[diplomat33]

That video is damn impressive! I really hope I'm wrong about Level 4 being a decade away.

That was just the highlights. If you thought it was impressive, check out the full unedited demo. Yes, it is very impressive.

These demos are why I think we are closer than 10 years from L4. Mobileye is promising it in 5 years.

 

[diplomat33]

Though I've always thought that Waymo was way ahead of Tesla, and the video above, if it is genuine, suggests that Mobileye is also. (Though if Mobileye is part of Microsoft, then I'm suspicious.)

First, it is genuine. I also sent you the unedited video so you can see for yourself. Also, Mobileye is owned by Intel. I don't think they have anything to do with Microsoft.

 

[DanCar]

... Mobileye is promising it in 5 years.

I hope mobileye delivers, but they've made comments before that they have missed badly.
The Evolution of EyeQ - Mobileye
That mobileye webpage still says Level 3 in 2018 and Level 4-5 in 2020. I'd be embarrassed if I where mobileye.
5 years in limited fashion in Jerusalem and 10 years in the U.S.?

 

[willow_hiller]

These demos are why I think we are closer than 10 years from L4. Mobileye is promising it in 5 years.

Is this using some sort of tie-in to the local stoplight system? At 3:27 a red stoplight appears in the upper-right hand corner of the screen, quite a few turns before the stoplight would come into view. At 3:51 the stoplight is visualized at the intersection on their UI, before they turn a corner and could actually see it. They don't come to a stop at the stoplight until 4:05 or so.

 

[DanCar]

First, it is genuine. ...

On reddit was described why it was not so genuine. I don't remember much of the details, but something like:

1. Several iterations before they got the video they wanted
2. No unprotected left turns at busy intersection
3. Closed off part of traffic for part of the video.

 

[diplomat33]

I hope mobileye delivers, but they've made comments before that they have missed badly.
The Evolution of EyeQ - Mobileye
That mobileye webpage still says Level 3 in 2018 and Level 4-5 in 2020. I'd be embarrassed if I where mobileye.
5 years in limited fashion in Jerusalem and 10 years in the U.S.?

Those dates are when the chips will launch, not when they are promising autonomous driving. The CES presentation said that L3/L4/L5 autonomy is coming to consumer cars in 5 years.

 

[diplomat33]

On reddit was described why it was not so genuine. I don't remember much of the details, but something like:

1. Several iterations before they got the video they wanted
2. No unprotected left turns at busy intersection
3. Closed off part of traffic for part of the video.

Uh?! The demo literally shows an unprotected left turn at a busy intersection.

 

[emmz0r]

I think Elon dismisses HD maps because he figures the car will use vision anyway and if vision is good enough then HD maps are irrelevant. But pretty much every leader in autonomous driving uses HD maps because they recognize that having a back up that the car can use to double check what it is doing, is a good idea if you want a more reliable system. After all, there will be cases where visibility is poor or maybe you encounter an edge case with a weird looking traffic light that the camera is not sure about. It is a good idea to have a HD map that can help your vision out.

Yes that was my point. Static mapping other than the basic route kinda seems like waste of time because either you have to rely on it, or have to have facilities that are good enough to create the map in the first place, so what good use is it then?

 

[SOULPEDL]

That was just the highlights. If you thought it was impressive, check out the full unedited demo. Yes, it is very impressive.

These demos are why I think we are closer than 10 years from L4. Mobileye is promising it in 5 years.

It's definitely pulling data of lane/road position. There was a road that went way off straight then right, but in the scene there was no way you could see past the next light.

So new rule: All construction must also notify 99 companies of the change pending? Who knows.

 

[msanghavi]

"Coming soon"!

https://electrek.co/2020/01/13/tesla-full-self-driving-feature-complete-soon-elon-musk/

 

[diplomat33]

Yes that was my point. Static mapping other than the basic route kinda seems like waste of time because either you have to rely on it, or have to have facilities that are good enough to create the map in the first place, so what good use is it then?

I don't think HD maps are useless. HD maps can help your car "see" something that is blocked or help your car when your sensors fail. As a result, the industry actually considers them to be necessary for autonomous driving.

Here's what the "Safety First for Automated Driving" paper has to say about HD maps. This might help you better understand their role:

"An in-vehicle map has never played a safety-related role as it could do in automated driving. For a relatively long period of time, the capabilities of onboard sensors alone will be insufficient to meet the high reliability, availability and safety requirements of the automated vehicle system in certain situations. A HD map is therefore necessary as a reliable off-board sensor containing carefully processed a-priori information to “detect” features that are not easily detectable by on-board sensors or to provide a redundant source of information for on-board sensors, including location-based ODD determination, environment modeling in adverse conditions and precise semantic understandings in complex driving situations. In situations where on-board sensors cannot reliably detect features, the HD map can be utilized as a more reliable redundant source of information."

 

[Daniel in SD]

I'm waiting for the 1 million hour uncut demo video. Then I'll be convinced that L4 is close!

Yes that was my point. Static mapping other than the basic route kinda seems like waste of time because either you have to rely on it, or have to have facilities that are good enough to create the map in the first place, so what good use is it then?

You rely on HD maps to improve the performance of the system. For example Tesla maps overpasses to reduce phantom braking. Sure they could just slam on the brakes every time they approach an overpass but that would make Autopilot useless and unsafe. They could just never brake on stationary radar return (like old school adaptive cruise control) but that would also be less safe.