I don't expect any automated driving system to actually reach full level 5.

[Max Spaghetti]

Level 5 automation is no steering wheel, accelerator, or brake pedal. All humans are passengers.

Actually that's Level 4, in my interpretation. Level 5 is just the same thing without ODD restrictions.

Level 3 basically describes a car equipped with controls that can be used outside the ODD, but doesn't need to be monitored inside the ODD.

Waymo (and any robo-taxi with controls) is probably a special case that is effectively a Level 3 vehicle for a staff member who is authorised to use the controls (or remotely assist), but effectively a Level 4 for a regular passenger who is not.

Level 5 would make more sense if it was defined as "can drive in any situation in which an average human could reasonably be expected to drive". It shouldn't have to be "super-human" to be classed as Level 5. There will always be corner-cases though. A great example is a grass field which has been temporary turned into a car park for a special event. There are marshalers guiding traffic to park in neat lines, but there are no line markings. For a car to be Level 5 (or even Level 4 without controls), it will have to be able to comply with the marshalers verbal or hand signals and park. Problems like that are hard to solve.

This is why I believe that private vehicles will remain defined as Level 3, while Level 4 without controls is the realm of the robo-taxi.

 

[mborkow]

Actually that's Level 4, in my interpretation. Level 5 is just the same thing without ODD restrictions.

If the vehicle’s use was limited to certain service areas level 4 wouldn’t require wheel or pedals. Outside of those areas, Level 4 would require a driver.

A great example is a grass field which has been temporary turned into a car park for a special event

Level 5 would have to be able to deal with that as the vehicle would not be equipped for a human driver: “under all conditions and on all roadways.” A grass field purposed as a parking lot would fall under “all conditions.” As I said, Level 5 won’t be any time soon…but never is a really long time.

https://www.nhtsa.gov/sites/nhtsa.gov/files/2022-05/Level-of-Automation-052522-tag.pdf

 

[diplomat33]

My guess is SAE probably added level 5 for the sake of completeness to their taxonomy. I say this because level 4 is already full autonomy, just in a limited ODD. So you could probably just have level 4 as the highest level and just make sure to specify the ODD. Some level 4 might have a small ODD while others might have a very large ODD. As long as you are clear about the ODD, you know everything you really need to know about the system because you know it can do all the driving without a human driver and you know when and where the system can be used. But level 5 is a like a capstone on top of level 4. The SAE basically said "what if level 4 had no design based ODD limits, let's make that a separate level so that we have a name for that". So if the full autonomy only works in say a geofence, we can call it level 4 but hypothetically, if we had full autonomy that worked everywhere like a human, we have a level to describe that, it's called level 5. But again, level 5 is hypothetical since we are far from achieving it. And for practical or business reasons, we may never need it because level 4 might be enough to describe the full autonomy that we actually use. But level 5 provides a name for that ultimate goal so that if we ever do reach it or if we just want to be able to refer to that ultimate goal, we already have a name for it.

 

[mborkow]

My guess is SAE probably added level 5 for the sake of completeness to their taxonomy. I say this because level 4 is already full autonomy, just in a limited ODD. So you could probably just have level 4 as the highest level and just make sure to specify the ODD. Some level 4 might have a small ODD while others might have a very large ODD. As long as you are clear about the ODD, you know everything you really need to know about the system because you know it can do all the driving without a human driver and you know when and where the system can be used. But level 5 is a like a capstone on top of level 4. The SAE basically said "what if level 4 had no design based ODD limits, let's make that a separate level so that we have a name for that". So if the full autonomy only works in say a geofence, we can call it level 4 but hypothetically, if we had full autonomy that worked everywhere like a human, we have a level to describe that, it's called level 5. But again, level 5 is hypothetical since we are far from achieving it. And for practical or business reasons, we may never need it because level 4 might be enough to describe the full autonomy that we actually use. But level 5 provides a name for that ultimate goal so that if we ever do reach it or if we just want to be able to refer to that ultimate goal, we already have a name for it.

I think it is achievable but not by Tesla (I don’t think vision-only gets you there) and not for our current cars.

 

[enemji]

My guess is SAE probably added level 5 for the sake of completeness to their taxonomy. I say this because level 4 is already full autonomy, just in a limited ODD. So you could probably just have level 4 as the highest level and just make sure to specify the ODD. Some level 4 might have a small ODD while others might have a very large ODD. As long as you are clear about the ODD, you know everything you really need to know about the system because you know it can do all the driving without a human driver and you know when and where the system can be used. But level 5 is a like a capstone on top of level 4. The SAE basically said "what if level 4 had no design based ODD limits, let's make that a separate level so that we have a name for that". So if the full autonomy only works in say a geofence, we can call it level 4 but hypothetically, if we had full autonomy that worked everywhere like a human, we have a level to describe that, it's called level 5. But again, level 5 is hypothetical since we are far from achieving it. And for practical or business reasons, we may never need it because level 4 might be enough to describe the full autonomy that we actually use. But level 5 provides a name for that ultimate goal so that if we ever do reach it or if we just want to be able to refer to that ultimate goal, we already have a name for it.

Level 5 is akin to infinity, immortality. In science world We tend to infinity…..

 

[JulienW]

Level 5 is akin to infinity, immortality. In science world We tend to infinity…..

"To infinity and beyond" Looks like we need a L6 for "beyond" infinity.

Infinity quiz:
Whole numbers are infinite but there are only ½ as many even numbers. So is there a ½ an infinity of even numbers?

 

[bradtem]

I get that the SAE levels might be flawed but I think it is helpful to have some sort of taxonomy to describe autonomous driving. There are many different automated driving systems that range from basic driver assist, to hands-free highway driving to driverless robotaxis. Also, the term "self-driving" is often thrown around to mean a lot of different things. Some people consider a system like AP or Super Cruise to be "self-driving" since the car is doing the steering and braking. Others only consider a driverless car to be "self-driving". You need clear definitions of terms. J3016 does that. People may not understand J3016 or may misinterpret it. But that is not J3106's fault, it does spell everything out clearly. And you need a way to communicate things like the role of the human, when and how are the human is asked to take over, can the AV perform the fallback and if so in what conditions, what is the ODD etc... Again, J3016 does that. So, I do think the SAE levels are useful as a starter. I also like Mobileye's "eyes on/hands-off" taxonomy and defining 6 key ODDs. The bottom line is that I think we need a way to classify different automated driving systems to help frame the discussion.

They aren't a taxonomy of autonomous driving. They are a taxonomy of roles of humans in partially autonomous systems. If that's useful, it's only for a short time. They mostly confuse people, and make them think that a self-driving car and a driver assist car are two "levels" of the same thing. They aren't, not at all.

 

[diplomat33]

They aren't a taxonomy of autonomous driving. They are a taxonomy of roles of humans in partially autonomous systems. If that's useful, it's only for a short time. They mostly confuse people, and make them think that a self-driving car and a driver assist car are two "levels" of the same thing. They aren't, not at all.

I guess I should be more careful with my words. Automated driving and autonomous driving are not the same thing. J3016 is a taxonomy of automated driving. Automated driving includes advanced driver assist and autonomous driving. A self-driving car and a driver assist are 2 levels of an automated driving system.

 

[Daniel in SD]

They aren't a taxonomy of autonomous driving. They are a taxonomy of roles of humans in partially autonomous systems. If that's useful, it's only for a short time. They mostly confuse people, and make them think that a self-driving car and a driver assist car are two "levels" of the same thing. They aren't, not at all.

What is the difference between Tesla FSD and a self-driving car? As far as I can tell the only difference is the error rate.

 

[Max Spaghetti]

What is the difference between Tesla FSD and a self-driving car?

One is a marketing term and the other is an undefined phrase.

Hence some people prefer to talk SAE Levels so that the terms are at least defined somewhere by a standards organisation and not left up to creative and varied interpretations and mis-interpretations.

 

[bradtem]

What is the difference between Tesla FSD and a self-driving car? As far as I can tell the only difference is the error rate.

Well, while it's not a settled question, the prevailing opinion is that Tesla FSD as it stands won't ever be able to self-drive. I would not say it is impossible, but with that hardware setup it seems low probability. There's a silly canard about how because humans can drive with just their eyes, surely a computer can as well, but that computer doesn't remotely have the power of a human brain. And just because birds fly with flapping wings doesn't mean that planes are designed that way.

But if you just want to look at it from a numbers standpoint, Cruise was probably doing about 20,000 trips in a row per incident, and even then they got pulled from the roads. Waymo is about 3 times better probably and has not been pulled. Teslas are lucky to do a handful of trips in a row without incident. That's not just a difference of degree but of kind.

Now the biggest difference perhaps doesn't apply. When you can rely on a human supervisor, you design your product differently, and the constraints are very different and the result is different. But Tesla hopes to drive without the human supervisor so they are doing some design efforts as though they won't have one some day.

 

[Mullermn]

There's a silly canard about how because humans can drive with just their eyes, surely a computer can as well, but that computer doesn't remotely have the power of a human brain.

This is a massive logical disconnect. Explain the connection between a choice of cameras as sensors and the amount of computing power required for self driving? You’re implying that a different choice of sensors would require less computing power?

If that’s not the connection you’re drawing then why mention the choice of cameras at all? Seems like you’re starting with the conclusion that FSD can’t work and working backwards to justify it to me.

And just because birds fly with flapping wings doesn't mean that planes are designed that way.

This is so far beyond dumb I don’t know where to start.

If you genuinely are a robocar consultant it’s for cruise at best.

 

[Daniel in SD]

You’re implying that a different choice of sensors would require less computing power?

Check out the darpa grand challenge. The first self driving cars used lidar with way less compute than is available now. It makes sense that sensors like lidar and radar require less compute since they give direct distance and speed measurements whereas modern computer vision relies on massive neural nets.

Everyone agrees that vision based FSD can work (humans are an existence proof). The question is whether the technology to make it work will be developed any time soon.

The point about birds vs airplanes is just that often the engineering solution to a problem does not mean mimicking nature. Sometimes you need to take a first principles approach instead of looking at what already exists.

And of course all robotaxi designs also use cameras in addition to other sensors.

 

[diplomat33]

This is so far beyond dumb I don’t know where to start.

It is not dumb at all. Brad is using the analogy of birds to make the point that technology does not always mimic nature. Airplanes don't fly the same way birds do. Likewise, autonomous driving does not have to be vision-only just because that is how humans drive.

 

[enemji]

There's a silly canard about how because humans can drive with just their eyes, surely a computer can as well, but that computer doesn't remotely have the power of a human brain. And just because birds fly with flapping wings doesn't mean that planes are designed that way.

I agree that it is idiotic to build robots that look and work like humans (Boston Scientific) when there are better and simpler ways to get the same thing done with robots built specifically for the task at hand.

 

[JulienW]

I agree that it is idiotic to build robots that look and work like humans (Boston Scientific) when there are better and simpler ways to get the same thing done with robots built specifically for the task at hand.

Boston Dynamics , funny there was a 73 year long company named Scientific Atlanta but was bought by Cisco.

EDIT: Just for the hell of it did a search of Boston Scientific and there is a compony by that name that makes medical devices.

Also I need my Pleasure Bot to look human.

 

[Daniel in SD]

I agree that it is idiotic to build robots that look and work like humans (Boston Scientific) when there are better and simpler ways to get the same thing done with robots built specifically for the task at hand.

It depends what the application is. Would be good if you need something with the shape and mobility of a human. They’re also a very interesting research project. Not everything needs to have immediate practical applications.

 

[VT_EE]

The level 5 definition is comical. Must operate in "all" conditions. Neither humans nor the vehicles they drive can operate in "all" conditions. There are plenty of snow storms I've seen where you aren't going out driving without tracks.

 

[Daniel in SD]

The level 5 definition is comical. Must operate in "all" conditions. Neither humans nor the vehicles they drive can operate in "all" conditions. There are plenty of snow storms I've seen where you aren't going out driving without tracks.

That’s not what it says.

“Unconditional/not ODD-specific” means that the ADS can operate the vehicle on-road anywhere within its region of the world and under all road conditions in which a conventional vehicle can be reasonably operated by a typically skilled human driver. This means, for example, that there are no design-based weather, time-of-day, or geographical restrictions on where and when the ADS can operate the vehicle. However, there may be conditions
not manageable by a driver in which the ADS would also be unable to complete a given trip (e.g., white-out snow storm, flooded roads, glare ice, etc.) until or unless the adverse conditions clear. At the onset of such unmanageable conditions the ADS would perform the DDT fallback to achieve a minimal risk condition (e.g., by pulling over to the side of the road and waiting for the conditions to change).

 

[enemji]

It depends what the application is. Would be good if you need something with the shape and mobility of a human. They’re also a very interesting research project. Not everything needs to have immediate practical applications.

build it and they will come