Self-Driving Cars

Aspirational product naming gets you in legal trouble in most sensible parts of the world. Yes I recognise that in this case the US has decided to not be sensible, but the rest of us don’t have to join you in a race to the bottom.

And no people complaining about a name not being accurate for a product won’t look silly if/when the product is changed such that name becomes accurate. Voltaire wouldn’t have retroactively looked dumb if come 1,800 the HRE became a theoretic state with oversees colonies and a ruling class made up of time displaced personnel from 400AD Italy.

It's like...an aspirational name like that only makes sense if you have no other capacity to like, just tell people what your plan is. Hell, Tesla could easily have trademarked multiple names for the roadmap and moved between them as needed, because that's exciting and it shows actual forward movement.

It's been called "Full Self Driving" for eight years and it's only now in what could be called a beta phase. That's pathetic as hell, when Tesla could have just called it something else. It's generally not a good idea to highlight your company's general issues with keeping to schedules like that.

Edited by Zendervai on Apr 4th 2024 at 3:17:07 PM

Next you're going to tell me that SpaceX's Dragon spacecraft is not an actual dragon, or that Ford's Mustang is not an actual horse. Where will the confusion of brand names end? Tide laundry detergent is not actual ocean tides? Slim Jim isn't a slender gentleman? Call the lawyers!

I'm sorry, but courts apply the standards of a "reasonable person" to these situations, and no reasonable person would see "Full Self-Driving Beta" and interpret that to mean the product is fully developed and ready for widespread use. We can argue about language all day, but "beta" has a commonly understood meaning.

You have the timeline wrong. It has been in development for eight(?) years and is only now at a release candidate state. Tesla's ability to keep its timelines has literally nothing to do with this conversation except as a convenient gotcha.

Edited by Fighteer on Apr 4th 2024 at 3:32:10 PM

You know what the difference is if you aren't being a tedious pedant for no reason?

Calling a rocket "dragon" doesn't say anything about what it can do. Same with "mustang". Calling a program "full self driving" is a claim about what it can do because it literally says "full self driving".

I want you to explain what concrete real world claim that "dragon" or "mustang" is making, because otherwise, you're just being tediously pedantic for no reason. For that matter, what is "slim jim" as a name saying about the product? That it's slim? It is, so there's no issue there. Non-literal names are a thing because you will always get people who will take names literally if given the chance.

Also, it's been available as a feature for five years and they added the beta name after it had been available for a while.

Edited by Zendervai on Apr 4th 2024 at 3:42:05 PM

With all due respect, I don't think your logic checks out.

It's a bit like how Subway once caught a lot of flack for claiming that a Subway Footlong was not in any way guaranteed to be a foot long and "Footlong" was just a brand.

Not every product name is intended to be descriptive, but it's usually pretty obvious when one is. And Full Self Driving absolutely sounds like a descriptive name. And the fact it's not accurate now and likely won't be for the foreseeable future is a genuine problem.

And even with sticking "beta" on full self driving, is it unreasonable for someone to see that and think "oh, it's full self driving, it's just a little glitchy and there'll be places it won't be available"? The software isn't at that level yet (legally, at least), but you're telling me that everyone is supposed to realize that Beta means "it's not fully feature complete or legally allowed to operate this way"? Because that's not really what beta is usually meant to mean anywhere else.

Oh yeah, that's absolutely true too. It'd be pretty easy for someone to miss the beta bit, especially with the legal precedent that no one reads EULAs, making them unenforceable.

Edited by Zendervai on Apr 4th 2024 at 3:51:50 PM

Except Tesla’s marketing (which consists mainly of Musk’s tweets) is hardly consistent in including the “Beta” part of the name, that’s a big part of the problem.

You have the timeline wrong. It has been in development for eight(?) years and is only now at a release candidate state. Tesla's ability to keep its timelines has literally nothing to do with this conversation except as a convenient gotcha.

The only difference between a release candidate and a beta is that you think you've got most of the bugs and UI issues ironed out for a RC. What Tesla's been selling, seeing as they apparently aren't aiming for a perpetual beta, has been an alpha build. And knowing what any of these distinctions actually means relies on a level of technical understanding that, whilst perhaps not uncommon in devoted Tesla users, is not going to fit neatly into a generic 'reasonable person' test.

Yeah, if you ask the average person what a beta is, you'll either get something unrelated to software or you'll get something like "it's basically done, it just needs bug testing." Is that an accurate description of FSD?

All that matters is whether this hypothetical "reasonable person" would understand that the product is intended to be used in a supervised manner. This has been the subject of quite a few lawsuits, with varying outcomes. In the majority of cases it has been determined that the driver is fully responsible for the vehicle regardless of the marketing.

Elon Musk's social media posts are not evidence in favor of your position because at no point has he claimed that the current version of the software is fully autonomous and can be used without supervision, nor has Tesla ever marketed it as such.

But has he ever clearly stated it wasn't autonomous?

Multiple times. Twitter/X doesn't have room to attach pages of legal disclaimers to every post, but regardless, the majority of the claims at issue are clearly aspirational, describing future capabilities rather than present ones. Either that or they highlight specific features and capabilities without making any overall claim of fitness.

Edited by Fighteer on Apr 4th 2024 at 4:10:59 AM

Speaking of X, and with apologies for the double-post, Tesla executive Rohan Patel has been doing a lot of posting lately about Supervised FSD and its potential rollout in Europe. In these posts, Patel notes that the main obstacle is regulations that require autonomous driving software to get user approval before initiating any maneuvers.

Admittedly, I only have his word for this and haven't dived into the details, but it sounds like a nightmare for a self-driving vehicle to have to ask a human before it does anything.

Edited by Fighteer on Apr 5th 2024 at 10:35:32 AM

So I’ve taken a look at the regulations and it seems that they want to put “Supervised FSD” under the UN category of being a “Driver Control Assistance Systems”, which is a sub-catagory of “ Advanced Driver Assistance Steering System” as laid out in the 2017/2018 UN regulations and “means a system, additional to the main steering system, that provides assistance to the driver in steering the vehicle but in which the driver remains at all times in primary control of the vehicle.”

The regulations lay out how the different version of “Automatically commanded steering function (ACSF)" work for an ADASS vehicle and they’re limited to: Category A for self-parking, Category B1/2 for lane keeping, Category C for driver initiated lane changing, Category D for driver approved lane changing and Category E for “a function which is initiated/activated by the driver and which can continuously determine the possibility of a manoeuvre (e.g. lane change) and complete these manoeuvres for extended periods without further driver command/confirmation.”

I’m still going through the regulations and don’t yet get how a DCAS actually fits within the ADASS but honestly I feel like what Tesla has sits well under ADASS, ACSF Category E.

I’ll edit in when I’ve read more regulations.

Edit: Think I’ve found the problem. Going through the UN regulations you get to the following:

With a bunch of stuff within all the sections apart from 5.6.3. So logically after the last section I’ll have 5.6.5 for Category D and 5.6.6 for Category E ACSF, right?

Nope, it’s.

So my take is that the UN regulations have laid out a conceptual pathway (ACSF Category E) but not a regulatory one because nobody have filled in the rules for Catagiry E (or D or B2) with Tesla currently trying to fit the system under a different concept DCAS.

On a completely separate note, I was digging into Wikipedia and found an answer to a question we were going over.

Mercedes apparently have declared (back in December) that they will alert other road users to an active level 3 system by having a vehicle using Drive Pilot display “Turquoise exterior lights”. As of the article I found they are the only ones doing this because they’re the only ones with an approved level 3 system, Tesla are keeping their systems classified as level 2.

Edit again:

I’ve also found that the U.K. Motor Industry’s trade association has actually published guidance on advertising self-driving/driver-assist software, it says that.

Which seems like a good standard to me.

Edited by Silasw on Apr 5th 2024 at 5:39:09 PM

Requiring the driver to initiate an action seems like a reasonable restriction for a level 2 system. That would be something like using your turn signal to tell your car to change lanes while you're using adaptive cruise control, yeah? That seems fair to me. The system still actually does the lane change for you, but you have to tell it you want to change lanes.

If your system is fully driving itself and only needs human intervention when it notices a situation it cannot handle, then that's a level 3 system. A level 2 system, by definition, can't be trusted that much. Making the rule that the driver must actively initiate system actions is a good way of enforcing that. The alternative is letting the system act like a level 3 but making drivers pinky swear that they're definitely monitoring it 100% of the time and can take control instantly and without warning when necessary, and I think we all know how realistic that is.

Technically that's a Level 4 system, depending on exactly how you draw the line. A Level 3 system can drive itself under a limited range of carefully defined conditions and requires driver attention at all other times. For example, Drive Pilot only works on geofenced areas of highway and in adequate weather/lighting. A Level 4 system can drive itself under most/all conditions and only needs to alert the driver to take over if it encounters a situation it can't handle.

Tesla FSD is aiming for Level 4 or equivalent, but is still officially regulated as Level 2. It's tough to create uniform standards when the covered population is one. As far as I know, nobody else is going for an end-to-end AI solution, and certainly not one based on vision alone.

By contrast, Waymo and Cruise are going for Level 5 from the start, but both have had to scale back their operations as they run into the limits of their approach.

Edited by Fighteer on Apr 5th 2024 at 5:47:16 AM

From a regulatory perspective I’m not sure the method matters that much, what’s important is the capability and reliability of the system.

So it’s really not a population of 1. Tesla are one of many commercially available level 2 systems, Drive Pilot is the only commercially available level 3 system (though Honda and BMW have such systems registered I don’t think they’re commercially available) and while Waymo claim to have level 4 we’ve discussed before that what they’re doing may be mroe accurately described as a level 2 system with a remote supervisor, akin to a UAV.

Tesla can aim for level 4 all day long, right now their legal position is that they have a level 2 system so it’s going to be regulated as a level 2 system. If they want to apply to be regulated as a level 4 system I’m all for it.

Edited by Silasw on Apr 5th 2024 at 11:25:56 AM

Because vision alone is honestly.... kind of dumb... like even the cheap maze solving robots I worked with in university used more then just a camrea for navigation...

Its a robot, not a person, it doesn't need to navigate like a person, the goal should be to do it better then a person... and unlike a person you can mount more sensors....

A radar/lidar should be the bare minimum extra for measuring distance without having to make guesses bassed on depth perception.... ultra sonic also helps with collision avoidance....

The eternal point it comes back to is "we can maybe do it with only cameras because it's cheap doesn't mean we should do it with only cameras". -Waves at number of Tesla vehicles produced that could really have pet an impetus on cheaper lidar.-

and consequently how the FSD faults seem to be a 50/50 split between "didn't realise thing was there" and "couldn't read sign". <_>

Tesla seems to believe that vision-only can work, and Supervised FSD seems to be proof of that. Ask Waymo and Cruise how the lidar-based solutions are working out.

Whether it works or not doesn't change that working entirely off of a single human sense probably isn't the most effective way, even if you're using ML to train your system.

The counterargument is that humans drive using only vision and they do fine. Our roadways and signage are all designed for eyes, so that's all an autonomous vehicle should need, too.

I mean, humans also have other senses (although I admit it might be hard to tell what anything except hearing or proprioception are going to do for driving), and they manifestly don't do fine all the time because of hitting things. Which is half why we've even considered self-driving in the first place, the other half being laziness. Point there being that "humans manage okay" doesn't mean having a nearly objective measure of your distance to everything around and not one interpreted through visual recognition alone wouldn't be better.

This isn’t a hypothetical, though. Tesla has already demonstrated that AI vision is capable of building precise 3D models of its environment. Plus, it can see better than humans in variant lighting conditions.

You keep focusing on range and velocity like they're gotchas. Those are not only solved problems now, but are almost trivial compared to the fundamental challenges of self-driving: situational awareness and contextual decision-making.

Cost is an important element as well: for autonomy to be usable it needs to be cheap enough to deploy in nearly every car. LiDAR is coming down in price but it still represents additional complexity and parts that can break. Why add a system you don’t need just to satisfy paranoia?

Edited by Fighteer on Apr 6th 2024 at 2:29:47 PM

This isn't the slam dunk you seem to think it is. Waymo and Cruise have operated fully autonomous vehicles (without a human driver for backup) on public roads. So they've achieved actual level 4 service (fully self-driving with no human intervention, will pull over and safely stop if it encounters a situation it can't deal with) — at least in specific locations and conditions. Tesla remains very insistent that, despite offering something like a level 3 feature set (fully self-driving, but may require a driver to intervene), it's still only a level 2 system.

So which is closer to releasing an actual level 3 product to the public? It's hard to say, and I'm not an industry expert. But it seems to me that that question is more about marketing than technology. Waymo and Cruise have functioning level 3 systems — they operate fleets of test vehicles with safety drivers. But the product they're trying to sell is driverless vehicles, so they're not making their level 3 system (which is really just a testbed for their level 4 system) an actual product that anyone can buy. Technologically, they're already there, but whether they can just slap a public-friendly UI on it and sell it as a level 3 system I couldn't say.

Meanwhile, Tesla has a level 2 system that it keeps adding level 3 features to, without actually calling it a level 3 system. When will it be certified as a level 3 system? Who knows. This is Tesla's main problem: level 2 is "you are driving, the system is assisting". Level 3 is "the system is driving, it may transfer control to you if it gets confused". That's a categorical difference. In the most tldr way possible, a level 3 system can drive itself (within certain limitations) and a level 2 system cannot drive itself under any circumstances. Tesla's approach is to make a level 2 system more and more capable until it's become a level 3 system (and presumably then on to level 4 and 5), but it remains to be seen if this is a viable approach. They may just end up having to build an independent, brand new level 3 system on top of their existing level 2 software, which would make that approach basically a waste of time.

The lidar stuff is a symptom of that. Waymo and Cruise said "what hardware do we need for level 4 capability?" and decided to add lidar sensors. Tesla said "well, we can do level 2 with just visual cameras, then we'll build up from there". And if it turns out that you do need lidar (or something like it) for level 3 or 4 systems? Then Tesla shot itself in the foot by spending a lot of time and effort trying to make it work without them.

Edited by NativeJovian on Apr 6th 2024 at 3:22:28 PM