Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Mobileye's taxonomy to replace the SAE Levels

This site may earn commission on affiliate links.
Mobileye published a blog that details their new proposed taxonomy for autonomous driving that they introduced at CES.

They propose 4 axes:

(i) Eyes-on/Eyes-off
(ii) Hands-on/Hands-off
(iii) Driver versus No-driver
(iv) MRM requirement

You get 4 categories that can be actual products:

1) Eyes-on/Hands-on: this category covers all the basic driver-assist functions. The driver is still responsible for the entire driving task while the system monitors the human driver.
2) Eyes-on/Hands-off: this is a driver-assistance function where the driver’s hands can be off the steering wheel while the system takes control of the driving and the driver supervises the system within a specified Operational Design Domain (ODD).
3) Eyes-off/Hands-off: the system controls the driving function within a specified ODD without the human driver needing to supervise the driving. Once the ODD comes to an end, and if the driver does not take back control, the system is able to conduct a full MRM and stop safely on the shoulder of the road.
4) No Driver: when there is no human driver present, say in a Robotaxi, the role of the human driver is replaced by a teleoperator who can intervene to resolve non-safety situations.

An Eyes-off system should be governed by three principles: (i) usefulness, (ii) safety, and (iii) scalability.

Usefulness: A good Eyes-off system should operate in an ODD that enables prolonged and continuous periods of use such that going in/out from an ODD does not happen frequently.

According to this requirement, Mobileye sets the minimum useful ODD threshold to be freeways up to 80 mph, including the ability to navigate on-ramps and off-ramps.

Safety: A safe Eyes-off solution should have no systematic errors, i.e., an error that can be reproduced in a certain emergency situation that is within the system’s ODD.

Scalability: We believe the operational domain of a full Eyes-off/Hands-off vehicle can best be thought of as a stack of ODDs – starting from highways, then adding arterial roads, signaled intersections, unprotected turns and so forth, that eventually add up to autonomy everywhere.



Untitled design.png

Admin note: Image added for Blog Feed thumbnail
 
Last edited:
Also Doesn't there need to be a L5 where there is no driver or driving position (like steering wheel/pedals) and there is no teleoperator? I know we are YEARS away but isn't that the ultimate goal?
 
Also Doesn't there need to be a L5 where there is no driver or driving position (like steering wheel/pedals) and there is no teleoperator? I know we are YEARS away but isn't that the ultimate goal?

No, that is a misconception. Removing the steering wheel/pedals does not make the system L5. It is possible to have L4 with no steering wheel or pedals, for example a robotaxi with no controls like the Cruise Origin and Waymo-Geely vehicles. And you can also have L5 that still has manual controls, for example a consumer car that has the option of manual driving or autonomous driving. Both L4 and L5 are the same autonomous driving, just with a different ODD. If the autonomous driving has a limited ODD (ex, geofencing) then it is L4. if the autonomous driving has the full ODD of human drivers, then it becomes L5. So L4 is basically L5 just with artificial limits placed on the ODD. Conversely, you can think of L5 as L4 but with all the ODD restrictions removed. ODD means operational design domain. It is the when and where the autonomous driving is allowed to be used. ODD can be road types, geofence, weather conditions, time of day, etc...

Since the only difference between L4 and L5 is the ODD, I think Mobileye's taxonomy works for both. If you have "eye's off" but only for highway ODD, then it is essentially L4. If you have "eye's off" for a full ODD then you essentially have L5. But I personally think L5 is a fantasy. For one, the levels are left up to the manufacturer. So to be L5, the manufacturer would have to certify that their autonomous driving works everywhere and any time a human can drive. That means the human could use the autonomous driving anywhere and any time a human might drive and the manufacturer would have to accept liability for an accident or malfunction. I doubt any manufacturer will accept liability for such a broad ODD. Furthermore, it might be hard to define if the ODD is truly L5 or not. Imagine having autonomous driving that works on 95% of US roads, that would technically still be L4 since it is missing 5% of US roads but it would still cover a very large ODD. So I think it makes more sense to think of L4 + ODD or "eye's off" + ODD. In other words, tell me if I need to supervise and the when and where. That way, I know if the system will be useful for me or not. For example, if the system works on 95% of US roads, that would be fantastic for me. I would not care that it is technically not L5. So I don't think we will see L5. Instead, I think we will see L4 or "eye's off" just with ever bigger ODD that will get closer to L5 ODD. That is why I call L5 an asymptote, something we will get closer to, without every actually reaching. But once the L4 has an ODD that is close enough to L5 ODD, it won't really matter because it will be good enough.

Removing the steering wheel/pedals restricts the AV to always stay inside the ODD since there is no way to drive the vehicle outside the ODD. If your ODD is a geofence, a vehicle with no steering wheel or pedals can never leave the geofence. That is ok for robotaxis since they only need to do ride-hailing inside their service area. They don't need to ever leave their geofence. So you can have a L4 robotaxi with no steering wheel or pedals. In the case of a L4 consumer car, it needs to have a steering wheel and pedals since the human needs a way to drive the car to go outside the ODD. So in the case of consumer cars, it needs to be L5 (full human ODD) in order to remove the steering wheel and pedals. That is why people have the misconception that L5 means no steering wheel or pedals. But you can also have L4 robotaxis with no steering wheel or pedals. Removing the steering wheel and pedals can work for either L4 or L5, it just depends what type of vehicle you are doing and whether the vehicles will need to leave its ODD or not.

In terms of teleoperators, do you mean someone remote controlling the AV? Obviously, that would not be autonomous driving. So yes, you would need to remove that to have autonomous driving. But you would need to remove that for both L4 and L5. It is not exclusive to L5. And nobody uses remote controls of AVs. Companies like Waymo or Cruise do use remote assistance where someone monitors the car remotely and can provide guidance to the AV. But this is only used for non safety issues, the AV must be able to handle all safety critical issues on its own. And remote assistance never controls the AV. The AV remains in autonomous mode the whole time and simply acts based on the guidance. Since the AV stays in autonomous mode the whole time and the remote assistance never controls the car, it is still considered autonomous driving. Both L4 and L5 can have remote assistance. So again, removing remote assistance does not make it L5.

I don't think anyone is considering removing remote assistance because you will want some back-up to assist the AV just in case the AV encounters an unexpected non safety issue. If your AV has no steering wheel/pedals and no driver position, without remote assistance, there would be no way to help the car if it gets stuck. So you need some remote assistance just in case. But remote assistance should be rare. Perhaps, some day driverless cars will be so intelligent that they never ever get stuck but that is a long ways off.
 
Perhaps it's just due to the newness of their proposal but I do not find it simpler. What we have seems fairly easy for me to understand, which I believe is:
  • L1 - (ME's eyes on/hands on) minimal driver assist, driver in control and responsible for all, hands on wheel at all times - i.e., cruise control
  • L2 - (ME's eyes on, hands off) advanced driver assist, driver in full control responsible for all, hands could be taken off the wheel as all normal vehicle functions (starting, stopping, accelerating, braking, speed limits, collision avoidance, etc.) are performed, monitoring system and nags/driver attenti(on monitoring of some kind is required
  • L3 - (ME's eyes off, hands off) advanced driver assist covering almost all of normal traffic conditions (construction, debris in road, accident, road closed, etc.), driver does not need fulltime attentiona and can be reading a book but must be "quickly" available if needed.
  • L4 - (ME's No Driver), driver does not need to be present, limited scope (i.e., highways only), likely fall back to L3 out of scope.
  • L5 - (ME's No Driver), same as L4 except no scope limitation
The two definitions are close. I think ME's point is that with "Eyes On/Hands On" it is the system monitoring the driver. While in "Eyes On, Hands Off" it is the driver monitoring the system. OK, that makes sense. To me that's what L2 is, while they suggest such a level is missing from the SAE standards.

From a technical standpoint, I'm not sure I like ME's proposal as it appears to be a fine line of difference. From a sales and marketing perspective, I don't like it even more. You want to sell a car to someone with something like Tesla's FSDb can refer to it as a "eyes on, hands off" system as a label? That's just as bad as calling it Full Self Driving.
 
Perhaps it's just due to the newness of their proposal but I do not find it simpler. What we have seems fairly easy for me to understand, which I believe is:
  • L1 - (ME's eyes on/hands on) minimal driver assist, driver in control and responsible for all, hands on wheel at all times - i.e., cruise control
  • L2 - (ME's eyes on, hands off) advanced driver assist, driver in full control responsible for all, hands could be taken off the wheel as all normal vehicle functions (starting, stopping, accelerating, braking, speed limits, collision avoidance, etc.) are performed, monitoring system and nags/driver attenti(on monitoring of some kind is required
  • L3 - (ME's eyes off, hands off) advanced driver assist covering almost all of normal traffic conditions (construction, debris in road, accident, road closed, etc.), driver does not need fulltime attentiona and can be reading a book but must be "quickly" available if needed.
  • L4 - (ME's No Driver), driver does not need to be present, limited scope (i.e., highways only), likely fall back to L3 out of scope.
  • L5 - (ME's No Driver), same as L4 except no scope limitation

This is not quite correct.

L2 can be eyes on/hands on OR eyes on/hands off. Your basic L2 would be eyes' on/hands on. But systems like GM's Super Cruise which are "eyes on, hands off" are also technically L2. ME used to call their Super Vision system "L2+" to try to distinguish it from your basic L2 but that does not really work.

Both L3 and L4 can be eye's off, hands off. It would depend on whether the driver is the fallback or not.

L4 can be eyes off/hands off OR no driver. L4 is not just "no driver". Any system that does all the driving and allows eyes off in a limited ODD would also be L4.

I think this is why ME has a problem with the SAE levels. The SAE levels have too much overlap where your basic lane keeping and cruise control and GM's Super Cruise are both labeled L2 even though the latter might be more advanced. And your consumer car system that allows eyes off could be L3 if the driver is the back up but could also be L4 if the driver is not the back up. And the consumer car that does not require driver supervision and the driverless robotaxi are both L4. This is what creates public confusion.

To me that's what L2 is, while they suggest such a level is missing from the SAE standards.

The SAE L2 definition is a system that handles both lateral and longitudinal control but only part of the OEDR. Yes, that implies it needs driver supervision but the problem is that the SAE L2 definition covers a very broad range of capabilities. You could have a system that does 10% of the OEDR or does 90% of the OEDR and both would be L2. So your basic lane keeping and cruise control is L2. But FSD beta which handles a lot more driving is also L2. There are a lot of different systems that can fall under "driver supervision". This is why they feel the SAE is missing a level. They don't like that basic lane keeping and cruise control and hands-free end to end systems are both labeled L2. They want a "level" that can cover those systems that are almost autonomous but still require driver supervision.

From a technical standpoint, I'm not sure I like ME's proposal as it appears to be a fine line of difference. From a sales and marketing perspective, I don't like it even more. You want to sell a car to someone with something like Tesla's FSDb can refer to it as a "eyes on, hands off" system as a label? That's just as bad as calling it Full Self Driving.

IMO, ME's taxonomy rearranges the SAE levels in a way that is easier to understand to the end-user. Most people don't know the difference between L2 and L3 or between L3 and L4. But I think people will understand the difference between "eyes on, hands on" and "eyes on, hands off" or "eyes off, hands off".

Saying a system is "eyes on, hands off" makes it clear that while you don't have to hold the steering wheel, you still have to keep your eyes on the road. So it makes clear that driver supervision is still required. So I don't think it is as bad as calling it FSD which implies the car is driving itself.
 
  • Like
Reactions: mylapore and Tam
No, that is a misconception. Removing the steering wheel/pedals does not make the system L5. It is possible to have L4 with no steering wheel or pedals, for example a robotaxi with no controls like the Cruise Origin and Waymo-Geely vehicles.....
I'm not referring to SAE levels. I'm referring to the 4 Mobileye category/levels. So let me try again .

4) No Driver: when there is no human driver present, say in a Robotaxi, the role of the human driver is replaced by a teleoperator who can intervene to resolve non-safety situations.

So what if no human driver is present and there is NO teleoperator? The car is on its own to do all driving. EXAMPLE: Your car coming to pick you up without anyone in it OR in Elon's CRAZY example delivering itself new from the factory to your house. 🤣 What is that category called?
 
4) No Driver: when there is no human driver present, say in a Robotaxi, the role of the human driver is replaced by a teleoperator who can intervene to resolve non-safety situations.

So what if no human driver is present and there is NO teleoperator? The car is on its own to do all driving. EXAMPLE: Your car coming to pick you up without anyone in it OR in Elon's CRAZY example delivering itself new from the factory to your house. 🤣 What is that category called?

I think I get what you are asking. You want a level for a car that is completely on its own all the time, no driver, and no remote help even? That does not exist yet. We do have cars that can drive themselves. And we have cars that can pick you up in certain areas with nobody in them. They are SAE L4. But the autonomous driving is not perfect so we can't guarantee that the AV will never ever need help. That is why we have remote assistance, to help the robotaxi just in case they get stuck. We are not at the point where we can just let the car drive itself everywhere with no driver and no remote monitoring or any help whatsoever. That will likely be awhile.

Also, SAE L4/L5 can apply to both driverless with remote assistance and driverless without remote assistance. So we already have a level for driverless with no remote assistance. But I guess you think those should be two separate categories? At this point, I don't think they need to be two separate categories since we don't have driverless with no remote assistance yet.
 
Last edited:
Mobileye published a blog that details their new proposed taxonomy for autonomous driving that they introduced at CES.

They propose 4 axes:

(i) Eyes-on/Eyes-off
(ii) Hands-on/Hands-off
(iii) Driver versus No-driver
(iv) MRM requirement

You get 4 categories that can be actual products:

1) Eyes-on/Hands-on: this category covers all the basic driver-assist functions. The driver is still responsible for the entire driving task while the system monitors the human driver.
2) Eyes-on/Hands-off: this is a driver-assistance function where the driver’s hands can be off the steering wheel while the system takes control of the driving and the driver supervises the system within a specified Operational Design Domain (ODD).
3) Eyes-off/Hands-off: the system controls the driving function within a specified ODD without the human driver needing to supervise the driving. Once the ODD comes to an end, and if the driver does not take back control, the system is able to conduct a full MRM and stop safely on the shoulder of the road.
4) No Driver: when there is no human driver present, say in a Robotaxi, the role of the human driver is replaced by a teleoperator who can intervene to resolve non-safety situations.

An Eyes-off system should be governed by three principles: (i) usefulness, (ii) safety, and (iii) scalability.

Usefulness: A good Eyes-off system should operate in an ODD that enables prolonged and continuous periods of use such that going in/out from an ODD does not happen frequently.

According to this requirement, Mobileye sets the minimum useful ODD threshold to be freeways up to 80 mph, including the ability to navigate on-ramps and off-ramps.

Safety: A safe Eyes-off solution should have no systematic errors, i.e., an error that can be reproduced in a certain emergency situation that is within the system’s ODD.

Scalability: We believe the operational domain of a full Eyes-off/Hands-off vehicle can best be thought of as a stack of ODDs – starting from highways, then adding arterial roads, signaled intersections, unprotected turns and so forth, that eventually add up to autonomy everywhere.

This is much simpler for regular folks to understand.
These type of descriptions should be a legal requirement for all driver assistance systems, just like nutritional information on food.
Just like in "Full Self Driving"tm where "Self" never means the car, rather you drive it yourself :D
 
  • Like
Reactions: diplomat33
Great conversation, simply put, definitions in the engineering world are often debated and changed. In reality Mobile Eye is simply trying to categorize in a simple manner, with words instead of Levels. I think I like it, consumers will have a better understanding of eyes on / hands on then L1 and or eyes off / hands off then L3.

There is a place for both. One with more definition as to required underlying specifications and one as a better consumer classification system.
 
So no difference between NOA and me driving a beat up 1970s Honda ;)

No, that is wrong. These categories are for ADAS systems. "eyes on/hands on" refers to a ADAS system, not manual driving. A 1970's Honda does not have a "eyes on/hands on" ADAS system so it would be outside these definitions.

Their definitions just benefit them. Not customers, not the industry.

I don't deny that the definitions benefit Mobileye but to say that they don't benefit customers or the industry is completely wrong IMO. The definitions benefit consumers because they make it easier for consumers to know what they are buying. For example, if you buy a car and it says that it comes with a system that is "eyes off/hands off" on highways, you will know that from on-ramp to off-ramp, you can take your hands and eyes off the road. Knowing that about what you are buying is a pretty clear benefit. And it benefits carmakers because it gives them an intuitive way to market their systems to customers.
 
These categories are for ADAS systems. "eyes on/hands on" refers to a ADAS system, not manual driving.
Wrong.

The driver is still responsible for the entire driving task while the system monitors the human driver.

In NOA human monitors the system which is driving.

Basically ADAS on freeways has the following ... how do you differentiate them ? MobilEye doesn't care about these differentiations, customers & the industry obviously does.
- Cruise
- TACC
- Lane keeping
- Changing lanes
- NOA / Navigation
 
Wrong.

The driver is still responsible for the entire driving task while the system monitors the human driver.

That is not talking about manual driving. It is talking about L1/L2. With L1/L2, the system performs some of the driving tasks but the driver is still responsible for the entire driving task while the system monitors the driver with a driver monitoring system to make sure the driver remains engaged. And the line that you left out literally says "this category covers all the basic driver-assist functions". So it is talking about ADAS. A 1970's Honda has NO ADAS AT ALL. It is total manual driving. And a 1970's Honda does not have a system that monitors the human driver.

Basically ADAS on freeways has the following ... how do you differentiate them ? MobilEye doesn't care about these differentiations, customers & the industry obviously does.
- Cruise
- TACC
- Lane keeping
- Changing lanes
- NOA / Navigation

Yes, Mobileye cares about these differentiations. If you only have basic Cruise or TACC that is "eyes on/hands on". TACC+Lane Keeping can be either "eyes on/hands on" or "eyes on/hands off". NOA could also be either "eyes on/hands on" or "eyes on/hands off". That is why it is important to specify if the ADAS is "eyes on/hands on" or "eyes on/hands off". It makes a difference.

And I think you might be missing the purpose of the taxonomy. The purpose is not to define individual features but to define the type of overall automated driving system. A carmaker would still specify the features that are part of the ADAS package. Mobileye is not suggesting that you don't need to define individual features. Mobileye is simply saying that you also need to specify if the system is "eyes on/hands on" or "eyes on/hands off or "eyes off/hands off" because that is super important information.
 
Last edited:
system monitors the driver with a driver monitoring system to make sure the driver remains engaged.
That is not part of the original post. Is it in their plan or did you add it ?

If they just say "system monitors the driver" - it could be as simple as monitoring how someone drives and giving a warning when the driver sleeps off or crossing the line.
 
That is not part of the original post. Is it in their plan or did you add it ?

If they just say "system monitors the driver" - it could be as simple as monitoring how someone drives and giving a warning when the driver sleeps off or crossing the line.
Context here is important. Its very obvious that the context is in relation to driver assistance systems.
Also seems obvious what the intent of the text was within that the context.
Not sure how it could be taken that "L2" is somehow more descriptive.

Funny how the context in this case is ignored so as to make it that MobileEye is just trying to take advantage - but Tesla can use "full self driving" to mean nothing of the sort. Especially when the general public sees those three words and assumes it means the car is full self driving and therefore capable of driving itself. Seems that a certain Wozniak agrees.
 
The MobilEye proposal is worded to be more intuitively understandable than the SAE levels, but I don't think it's so clear and unambiguous that it won't also be the subject of endless debate. I think it's a helpful contribution but maybe not the last word. And no, I wouldn't propose Tesla's definitions as a solution either!

I also agree that the No Driver category should not absolutely mandate a tele-operator backup. Obviously it could be helpful, and of course it might be a practical solution if the system fails often enough.

But if the technology evolves to the point that failures are rare, then there's no overriding legal or public safety reason that it has to be burdened with the cost of a ready tele-operator network. Any more than we mandate a roadside assistance or mobile-repair service contract now, for the tens of thousands of cars that break down every day.

Particularly for consumer-owned cars, it should be a business and market decision. If Mobileye wants to develop the service or even includ it as standard for vehicles using their technology platform, more power to them and it might be a market advantage - until someone comes along and advertises a cheaper, better platform that doesn't need a help desk.
 
This is all excellent discussion, and I love Mobile Eye's taxonomy.

But at the end state, at eyes off/hands off or L4/5 capability, why not still have a driver monitor, for those yet unknown situations, forever? Makes the car that much safer.

Because driverless cars don't solve any problems the world is actually facing right now, except for driver convieninece (oh, sad face emoji, I'd rather watch a you tube video than actually pay attention while I drive home). WHAT ever.

So what? Does not help congestion, maybe makes it worse. Does not help climate, not in the critical next 7-10 years. Might help 30-50 years from now when the cars could conceivably be ubiquitous and trusted.

If TODAY, everyone in the planet ONLY bought EVs, used, new whatever (all car purchases, new used private or dealer transactions) it would take 20 years to replace the global fleet of cars.

Autonomous cars are forever going to be decades behind in fleet percentages. AV is no solution to any problem we need to solve right now.

Might as well be talking about AI writing poetry so you don't feel so bad as the world collapses around you, for all the effort we'd need to put out to get AV to work and scaled to meaningful levels.

Sorry, rant off. This is all irrelevant. Our only hope is E bikes (or NEVs or similar) eBusses, and then figure out how to get mass transit to work for everyone, really.

Nice planet you had there, too bad you wasted all your effort debating all the wrong things when you still had a chance to fix it.
 
This is all excellent discussion, and I love Mobile Eye's taxonomy.

But at the end state, at eyes off/hands off or L4/5 capability, why not still have a driver monitor, for those yet unknown situations, forever? Makes the car that much safer.

Because driverless cars don't solve any problems the world is actually facing right now, except for driver convieninece (oh, sad face emoji, I'd rather watch a you tube video than actually pay attention while I drive home). WHAT ever.

So what? Does not help congestion, maybe makes it worse. Does not help climate, not in the critical next 7-10 years. Might help 30-50 years from now when the cars could conceivably be ubiquitous and trusted.

If TODAY, everyone in the planet ONLY bought EVs, used, new whatever (all car purchases, new used private or dealer transactions) it would take 20 years to replace the global fleet of cars.

Autonomous cars are forever going to be decades behind in fleet percentages. AV is no solution to any problem we need to solve right now.

Might as well be talking about AI writing poetry so you don't feel so bad as the world collapses around you, for all the effort we'd need to put out to get AV to work and scaled to meaningful levels.

Sorry, rant off. This is all irrelevant. Our only hope is E bikes (or NEVs or similar) eBusses, and then figure out how to get mass transit to work for everyone, really.

Nice planet you had there, too bad you wasted all your effort debating all the wrong things when you still had a chance to fix it.
Why not equip every car with two sets of controls for two drivers? Maximum safety. :p
It is hard to see how AVs won’t result in more driving so it is hard to see the environmental benefit. I suppose there could eventually be special lanes for AVs so they could draft off each other and increase capacity.
 
  • Funny
Reactions: Twiglett