This is something I’ve been noticing for a while, and it’s becoming common enough that I think a whole post is in order to clear some things up. I’m talking about the way autonomous vehicles are classified, as defined by the Society of Automotive Engineers (SAE) in their SAE J3016 Levels of Driving Automation standard. This standard defines a series of six “levels” of automated driving, in hopes of making classification of the capabilities of these vehicles easier. It’s a good idea. Too bad it’s completely misunderstood by so many people.
I really noticed this last week, in comments and tweets relating to an article I wrote about Tesla’s self-driving efforts, like their Level 2 system known as Autopilot and Full Self-Driving (FSD).
In that article, I talked a lot about why I (and others) think Level 2-based semi-automated systems are a dead end, an actively flawed way to approach the problem of automated driving, primarily because they rely on a human driver who must be ready to take control at any moment.
A good number of the reactions I got to the article that disagreed made the point that even if I have issues with a Level 2 system, it’s still a necessary step to get to more advanced Level 4 and 5 systems in the future, and the data gathered and development from the Level 2 step is important.
Here’s some examples of this sort of response:
Every single one of these people entirely misunderstands what the SAE autonomy levels are about. To read these comments, it appears that the people who made them think that the SAE levels are a sort of scale of capability, with Level 2 being less capable than Level 3, and Level 4 being able to drive better than the ones behind it, and Level 5 being able to drive the best of all.
The problem is that’s not what the SAE levels are describing at all.
I think part of this is the SAE’s fault for picking an inherently qualitative-suggesting word like “level” instead of something like “class 2" or a category type that doesn’t imply an improvement or degradation of ability. The Levels are not talking about how good the self-driving system is — the levels describe the parameters of how the system interacts with the humans in the car, and the environment around the car.
Here, let’s look at the chart:
If you actually read what each level is about, you can see that the levels are divided based on two main things: what is required of the driver, and where the system can operate.
Level 0 is like the archaic shitboxes I drive that can’t even shift their own gears, let alone have anything as decadent as cruise control. Level 1 is for cars with just adaptive cruise control or maybe some lane-centering features. There’s no question that the human in the driver’s seat is in full control.
Level 2 — which is where nearly everything on the market currently is — can hypothetically and effectively do all the actions needed to drive in most situations, but requires a human driver to remain constantly alert and ready to take full control with little or no warning.
Level 3 is especially confusing, because its parameters are that the vehicle can, in certain situations, drive itself with no need for human input, but there are times when it may request the human to drive. A crucial criteria of Level 3 systems, I believe at least, is that the system must be able to get out of active traffic lanes when requesting the human to drive, freeing the human from having to remain constantly vigilant and ready to take over with no warnings.
Level 4 is completely self-driving in very specific areas and environments, and Level 5 means the car is able to drive absolutely anywhere, in any condition, like KITT from Knight Rider. Here, let’s take a moment to remember KITT:
That was a nice little break. Okay, back to the SAE levels: you may notice that the one thing not mentioned in any of these levels is, well, how good the system is at driving.
Sure, if you want a Level 4 or 5 car, you’d need a very capable system, because it never has to require any driver intervention. But, fundamentally, the SAE levels are not about the quality of the self-driving system, they’re about what parameters they operate under.
That’s why all of these people saying that Tesla must progress through Level 2 to get to Level 5, or that somehow Tesla (or anyone) needs to deploy L2 systems to get to Level 5, or that Tesla’s FSD is so advanced it’s really closer to L3 or L4 are all dead wrong. None of these statements are true, because the SAE levels simply don’t work like that.
Tesla’s FSD may be extremely advanced and highly capable at driving in many situations, sure. But that simply doesn’t matter to the SAE classification. If it has the potential to disengage and require a human to take over immediately, it’s Level 2. Period. That’s it.
Look, KITT himself — who has proven (well, in fiction) to be a full, magical, Level 5, completely autonomous car — would be considered a Level 2 car if he, say, caught a computer virus that gave him the electromechanical equivalent of narcolepsy, where his consciousness could just shut off without warning.
In that case, poor Michael Knight would have to keep his hands on the wheel and be ready to take over in case KITT nodded off suddenly, making KITT an L2 car, no matter how advanced his self-driving skills remained.
No company has to go through L2 to get to L4 or L5. That’s not how this works.
Sure, L2 cars may be great at gathering data and training AI neural networks how to drive better, but that does not mean they make sense for public commercial deployment. In this context, they’d be best used as special development vehicles.
The better option for drivers and consumers and anyone on public roads would be for carmakers to work on developing systems that do not demand immediate input from drivers, but instead, when they need to disengage, are able to get out of traffic and harm’s way. This capability would put them at Level 3, which really should be the minimal allowed level for commercial sale, if you ask me.
Even if you didn’t ask me, and want to argue with me, then at least take the time to realize that the SAE levels do not refer to the quality of a self-driving system. They never have.
No car maker needs to climb through them linearly to get to their goals, and the actual driving algorithms and logic of a Level 2 and Level 5 system could be identical, but if one has no safe way to deal with disengagement or various failures other than a panicked call to help to the person sitting behind the wheel, then it will always be Level 2.
If you’ve gotten this wrong in the past, don’t be so hard on yourself. It’s confusing! I have some ideas on how to improve things, but, you know, nobody listens to me.