back
71 comments
I don't doubt that this is a major milestone: gawd knows, I couldn't program a car to do that.

But something that I would like to know is what happens when it fails? Or at least, seems to have failed?

At the moment, if I buy a car and the brakes fail and I crash into someone/something then it's likely there was a defect and the manufacturer was at fault. That seems reasonable when you're making a 2 ton machine capable of hitting over 100 mph, there should be a shit load of safety measures in place.

However, it seems that self-driving software is in its infancy and I've lost faith in tech companies to release a product when it's complete and ready, instead releasing it sooner with a bunch of bugs for the users to test. In addition, if my brakes fail, a sound mechanic would be able to verify that was the case and I'd be off the hook: you would be able to examine the faulty part.

If my self-driving thingy failed on one of these level-3 roads, and I crashed into a bunch of nuns crossing the road would I be on the hook for it? How would I be able to prove in court that it failed and wasn't a result of me being a muppet. Coz you can bet I won't be able to examine the source code after the accident, unlike being able to see the faulty brakes in the earlier example.

We are already there - remember the unintended acceleration problems? Source code was opened to outside review (for the purposes of the trial).
Except that was a scapegoat trial, anyone of the drivers could have pressed the breaks and stopped the car. They were confused and didn’t which resulted in loss of life. So instead of saying what actually happened they fined Toyota a billion
My mom got into an accident with her 2007 Toyota Prius. Now, it could be that she, as an older person, simply didn't push the right pedal. However she claims insistently that she did brake but it accelerated instead.

She didn't kill anyone but did get injured minorly and she got a very nice price for her totaled Prius (above market rates).

I owned a 2002 Celica, the gas pedal got wedged under the floor mat while I was on the motorway. It went from 75 to 95mph before I could work out what was going on. Was I a confused driver? Yes I was. Was it a stupid design fault that could have been solved with a craft knife? Also yes.
One would assume, that like the faulty brake part, the legal process should be able to subpoena all the logs and the video footage taken by your car, and use it to either exonerate or convict you. Alas, I think it will take a few high profile cases of fatalities (in each country) to trigger a change in the law to ensure this.
That's a good level of caution to exercise for Tesla's claims.

Other SDC companies with a lot more promise than Tesla are an awful lot quieter about what they're doing because they know people have the concerns you do and they are downright allergic to over-promising on either capability or safety.

There is no such thing as Level-3 Autonomy, the car can either drive itself safely or not.

A slow moving lane of cars is a dangerous environment where things can turn catastrophic in an instant, with zero time to react for a human driver reading their news. Bikes and motorcycles come out of nowhere and pass you between the lanes. Pedestrians and children will attempt to cross the road, encouraged by the apparent slow speed. You will not see them since your view is blocked by close large vehicles. The same for essential signs and signaling. Slow moving can turn to fast moving intermittently and frustrated drivers will unexpectedly go full throttle to cut in front of you. You might find yourself suddenly stranded in the middle of an intersection with lateral trafic. Trafic is slow because your lane is blocked by construction, which requires carefully observing the human readable cues of the drivers in the other lanes, which might yield to you or not. All manner of bad things can and will happen with a scary probability.

If an automated car can handle all that safely without reverting to human assistance, what exactly can't it handle, and why is it not full self driving?

The 3 and 4 SAE levels were always a dangerous fiction, they imply you can bring to market uncomplete products that take some part of the human effort. This is simply not how the minds of human drivers works: if drivers lose focus to the road they can't regain it instantly in a fraction of a second when prompted by the failing AI algorithm. A half- self driving car is a defective product that is bound to kill someone, regulators should reject any products short of real self driving.

I generally disagree with respect to L4. A car that is fully autonomous under a precisely defined set of circumstances, especially certain types of roads and possibly under certain environmental conditions could still be extremely useful--assuming you can fully hand off control. A lot of driving is fairly predictable highway driving even if it doesn't help people who want a door-to-door robo-taxi.
If you can geofence those specific circumstance and guarantee they won't happen on useful stretches of road, then congratulations, you have a car that reaches level 5 on certain roads.

But that's not what level 4 generally implies, they call it "High Driving Automation" and the expectation is that the driver is on hand to correct the AI in certain limited scenarios.

This is how JD Power describes it:

Referred to as high-driving automation, Level 4 autonomy does not require any human interaction in the vehicle’s operation because it is programmed to stop itself in the event of system failure. Since a human driver is never needed, a Level 4 vehicle may not have a steering wheel and pedals.

And yes, at Level 4, you can take a nap while riding in the vehicle.

Level 4 driving automation technology is for use in driverless taxis and public transportation services. Such vehicles will be programmed to travel between Point A and Point B and restricted to specific geographic boundaries by geofencing technology. Certain conditions may limit or cancel Level 4 autonomous vehicle operation, such as severe weather.

I admit that I thought of L4 as being fully autonomous under specific pre-defined conditions (given that's what would actually be useful and that's probably why I've heard it described that way.) The key would seem to be that the vehicle won't try to hand off control unexpectedly given, in part, that a vehicle can't always stop safely.

Highway driving stops being predictable when a deer steps onto the road.
True. But that's actually the sort of thing I'd expect an autonomous driving system to likely be able to handle as well as a person.
SAE vehicle autonomy levels are mostly a fiction in terms of not being highly contextual. The practical approach for self-driving is to take an actuarial, data-driven approach. For example, on certain streets or neighborhoods at certain times of day, you can be sure with a 99% confidence level that you will not encounter certain types of situations. A Level 3 vehicle in Phoenix is not necessarily a Level 3 vehicle in Manhattan or Nairobi.
Couldn't you imagine a world where infrastructure is adapted for safe self-driving? For instance, what if you had roads (think highways) where self-driving was mandatory? By applying certain constraints, you might be able to create situations where autonomous vehicles are practical.

Maybe it's something like a carpool lane. Autonomous vehicles get the far left lane, and are allowed to drive much faster and closer together than manually controlled vehicles.

This is what I am thinking all the time. Instead of having Cars to adopt to the infinite possibility of real world situation. The infrastructure should be built for it. Like a whole city that only allows self driving cars.

May be building a whole new City is an expensive task, but given the current housing market in US I wonder why no one has put some thoughts into it.

I can imagine it for the sake of argument, but without strong govt action it has a bootstrap problem: no one will create such a lane if there are no such cars, and no one will buy and develop them if there are no roads you can use them on.

Moreover, if the goal is to achieve autonomy with relatively dumb algorithms that fall short of full self driving, you have to do grade separation, fencing, tunnels or overground etc., otherwise you still need full self driving to handle those case where bikes or children roam around in your "dedicated" lane, cars and garbage bins are parked on it, drunk drivers suddenly invade it etc. And if you go the full trouble of dedicated and grade separated infrastructure, why not build a tramway or PRT (like the Booring Co Loop system) that any citizen can benefit from?

All in all, I can't see the scenario you are describing happening in any real city with real politics, maybe some oil prince can decree it.

> If an automated car can handle all that safely without reverting to human assistance, what exactly can't it handle, and why is it not full self driving?

The answer is in the article, and in the definition of Level 3: the automated feature is restricted to a certain operational design domain. For Drive Pilot, this domain comprises specific sections of highway, maximum speed, and certain weather conditions. It refuses to operate outside that domain, even if technically the software might be capable of doing so.

When it detects that it is leaving or has left that domain, it tries to hand back control to the human within a short period, failing which it brings the vehicle to a safe halt and calls for help.

You misunderstood the question. Everything points to the conclusion that solving the self driving problem for that specific operational domain, congested highway, is almost just as hard as solving self-driving in the general case.

The logical conclusion is then that the failure of the AI and calling for help from a human driver will either never happen at all, since you have a safe level 5 algorithm, or will happen abruptly in dangerous trafic situations that the driver cannot be reasonably expected to recover from.

Exactly. Level 5 or bust. It’s fiction to believe anyone can quickly take over — it is not how our brains work. If anything less than Level 5 does a good job much of the time, that only reinforces to the driver that they can completely disengage. It takes many seconds to not only respond to a call that they need to put down their phone and look up, but also to then grok the scene and keep moving. It’s just not cognitively realistic.
I was recently driving a VW ID4 and I loved it.

It only has lane assist and distronic but its very well implemented and especially in heavy traffic its a godsent.

Basically it simply follows the car in front of you but even that is super relaxing, especially in morning taffic. In any major city rush hour traffic is always fun, but this makes it at least somewhat enjoyable.

We have come a long way with assisted driving and as a german I very much enjoy that because it means that I can still "become one with my machine" ;)

This is a clear step up and I believe that the future will lay in smart roads, even if its just a lane maker that your car can rely on.

I agree with the roads. Why adapt computers to a legacy program when you can just add simple markers or computer read information to it instead?
Because smart roads are a massive coordination problem. Counties and states already can't afford to do maintenance on the existing dumb roads. Many of those roads will see no maintenance or upgrades for the next decade. Even if we had the standards, widespread government support, and commoditized tech to do this today, it'd still be decades away at the earliest. Everyone is targeting rollouts in major cities a lot sooner than that.
I recently drove an ID3 and it was impressive, but it had a very different effect on me. I suddenly realised that I'm an able bodied person using an electric wheelchair. It made me seriously question what I was doing. Why am I, a person perfectly capable of locomotion, relying on a monstrously large, unsustainable vehicle for everyday travel?
> Basically it simply follows the car in front of you

What if the car in front of you heads off on a route you didn't want to take?

I think it only uses the car in front for speed corrections. It will stay in the lane it's in.
then it leaves the lane markings and that would cause my car to at least ping me.

we are not talking about self driving here at all, but more like a "i am aware that you exist and adapt accordingly". There is an old person going 30 in a 50 zone? ok I will slow down and drive with the correct distance to him. He is speeding up? nice i will follow his lead to the maximum speed limit that is set.

When going from a 50 to a 30 the car also reduduces the speed to 30 and when you to to a 50 zone again it will speed up

A news item about Mercedes-Benz's Drive Pilot feature was shared on HN recently: [https://news.ycombinator.com/item?id=29517459]

I'm trying to explain to non-technical people why this feature is a major milestone for self-driving cars. On my website, I try to avoid jargon and assume my readers know nothing about the topic.

Thanks for sharing.

I find it very odd how people are so fixated on construction zones in that thread. Construction zones _are_ confusing and don't follow the typical rules of the road. That's why they always come with slower speed restrictions. People acting like their car is trying to kill them clearly aren't assessing if they should even be using self-driving in these circumstances. In fact, they likely don't realize that many legally-licensed drivers also struggle in these conditions. Many people, especially those with poorer reactions or eyesight, have a challenging time navigating certain construction zones.

I have an L2 and it works amazingly well when there are exactly two clearly marked lines. It's not that hard to understand when I can and can't rely on the car to drive itself.

* Less than 2 lines visibly marked, drive it yourself.

* If there are more than 2 things that look like lines, drive it yourself.

Most highway driving is not in construction zones.

>I have an L2 and it works amazingly well when there are exactly two clearly marked lines. It's not that hard to understand when I can and can't rely on the car to drive itself.

If you mean Tesla you should always pay attention, you are responsible even if the road is perfect, there is a non zero chance the car can swing and kill some innocent person because you are alpha testing.

Honestly, this seems like it could be solved by "simply" (it's a lot of tedious but routine organizational work) having the self-driving features get disabled whenever getting close to a construction zone. Like, those are unusual circumstances, handle the wheel yourself for those five minutes.
Yeah, I had a case like this not too long ago. It was I95 near the Maine/NH border.

At first I wondered why the cars in front of me were half into my lane, then I realized that there were an abundance of poorly painted lane lines and it wasn't clear who was actually in the wrong. It was the weirdest mess of lane lines that I've ever seen, tbh.

It was also obvious that I wasn't the only one guessing, as different drivers seemed to guess differently.

You can't do other things; putting on makeup is not a safe activity for a driver with any of these features.

Why is it that I can trust a level 3 system to not hit the car in front of me in stop and go traffic but I can't trust the level 2 system not to hit the car in front of me at 70mph? The stop and go driving certainly requires faster reaction times in my experience. Are the makers of level 3 systems legally liable when the vehicle in level-3 mode hits another car?

> Are the makers of level 3 systems legally liable when the vehicle in level-3 hits another car.

Currently, as how the NHTSA regulates vehicles, the driver is liable for any accident that their vehicle causes. Your car should always be at your control, and you should be the one responsible for it's actions. Not saying this is right or wrong, but that's how the government views driving.

Absolutely, the point of L3 systems is to move the legal liability from the driver to the car maker.

Else there would be no benefit, would it? As long as I have full responsibility and have to be fully aware of the situation I might as well just drive myself.

Germany had to change it's law for Mercedes to be able to role out L3.

> In July 2021 in Germany, the Federal Act Amending the Road Traffic Act and the Compulsory Insurance Act (Autonomous Driving Act) came into effect. The Act allows motor vehicles with autonomous driving capabilities, meaning vehicles that can perform driving tasks independently without a person driving, in specified operating areas on public roads. Provisions about autonomous driving in appropriate operating areas correspond to Level 4. [1]

[1] https://en.wikipedia.org/wiki/Self-driving_car#Legislation_i...

Though the distances are smaller/closer, I think that typical 70mph freeway driving is more likely to need a fast reaction than stop-and-go traffic (and certainly the stakes are much higher). The likely worst outcome from an error in stop-n-go traffic is an inconvenience.
Because the manufacturer is less likely to get sued for a death in the low speed case. Also, there are fewer chances of needing a high speed avoidance maneuver rather than simply hitting the brakes.
It is not just a matter of reaction times. It is mostly of stopping without harming you, which at 70mph is rather difficult if the vehicle in front brakes too hard.
The problem with Level 3 is that is requires much of the same abilities and sensors as Level 5 in order to determine if the conditions are safe, react to unexpected events, and deal with the driver not taking over.
I'd be comfortable with a car that only attempts self-driving after having driven a road 3-5 times. And only below certain speeds such that the likelihood of a fatality/serious accident for myself, a pedestrian, cyclist, or other driver - would be very, very low (<0.00001% etc.). If that meant the car driving back roads and avoiding even busier town streets, and drove 30 miles an hour max, there are certain situations when I'd be fine with that.

As long as a car can handle all scenarios at 25 mph (a kid running out into the street with a ball, and anticipation of more complex events that would require pulling over and passing back control in an easy, non-jarring manner), I would think this still makes sense. What would this do? It would force new street creation (and perhaps extra shoulder space for pulling over) for these slow-ass autonomous vehicles.

But we'd all be safe and able to work or do other things during our commutes or on the way to do errands, etc.

My model 3 essentially does these things, although it does require you to keep your hands on the wheel (so not level 3 by definition), and I can say it makes extended highway driving a lot less painful on nice weather days.
It really varies by person, IMO. I was never able to feel more comfortable in my Model 3 on the highway with AP engaged than without it. It just doesn't drive defensively enough for me to relax, it loves to get into situations I would never allow. But this is really going to vary depending on the expectations of the driver.

Of course, it also started phantom braking at a couple overpasses I regularly traveled under while road tripping, especially when the sun was out, so I stopped using it altogether (my wife really, really didn't like AP after that).

Right. What Mercedes has done here is to recognize and target that value.

They took the automated driving capabilities, added more sensors, and did the engineering and regulatory legwork around it to make it robust when entering, operating inside, and leaving its design domain.

Result: drivers can rely on it and regulators can enforce the rules.

Agreed. I’ve been driven by my M3 many times in one of the specific situations the author refers to (bumper to bumper on Boston’s SE Expressway) - with hands on the wheel to be sure.
Completely agree, it is great for stop and go traffic.
For it to be "milestone" implies we are moving in the right direction. The AI we have today is just putting increasingly red lipstick on a pig: we need to completely overhaul the way we're doing things if we're going to hit level 5.
Full self-driving (not the Tesla FSD fud, but actual Level 5) is probably not possible with all the current roads and the current approach (bigger and bigger ML models with more and more data). If that's what you mean, then I agree with you.

But this Mercedes announcement is something else: it is a car maker providing an engineering automation solution for a useful, limited use case that people might be willing to pay for. It sits squarely within the SAE Level 3 definition. The vendor is standing behind it; the regulators are ready to enforce the rules. It really doesn't matter what the "I want full self-driving" purists think, since they haven't actually produced anything useful yet.