I've commuted a bit on the German highways. I only run Autosteer, which is cruise-control + lane-control, and it disengages almost every time there's roadwork.
It is very bad at handling ambiguity when the road has both white and yellow navigation lines. There's a number of scenarios, some of which involve criss-crossing lines, some where they run mostly in parallel.
The car sometimes wants to follow the white lines when the yellow are actually overrides. When the yellow lines are wrong (and should rightfully have been scraped off), this always leads to disengagement. For stretches where the yellow lines are correct, and the older white lines are off by half a meter so that there's effectively not enough space between the white lines and the concrete side, autosteer will sometimes try to switch to the white lines, which would lead to the car smashing into the concrete side.
And sure, autosteer != FSD. But it's a much simpler problem that the Tesla still basically fails at.
Sometimes, in fog, cruise-control will abrubtly break. Very nerve-wrecking.
Relying fully on vision has its downside. Listening to Kaparthy on Lex Fridman's podcast, it sounds like this was cost-cutting. I can't believe that "nope, that big white thing isn't a wall" is not objectively better navigation.
Bit by bit, it gets harder to figure out how buying a Tesla is ever the right answer. Autosteer (beta) is worse than the competition, traffic aware cruise control is timid to the point of braking for things which don't even exist, you don't get basic things like ultrasonic sensors, radar, rain-sensing wipers, or even carplay. And even when they do include something that sounds halfway cool, it turns out to be underwhelming -- the matrix headlights are all of 100 pixels, which would have been cool about 10 years ago. And they aren't even enabled. Given how poor automatic high beams are, do we really want to see how Tesla matrix headlights would work in real life?
Autosteer on our Tesla Model 3 is actually pretty great. I use it all the time because it makes driving a lot less tiring, especially over long distances. I don't know where you get your information.
As far as competition, our other car Kia Forte 2022 now comes with Autosteer like feature as well. With that, unless, the lines are pristinely painted, you are literally taking your life into your own hands.
"All cars" is absolutely not true. If I drive up to a Tesla Supercharger with my Mach E in the US, I'd have no way to charge my car. The Tesla charger has a proprietary plug that needs a proprietary app with a proprietary payment network to start the charge.
Tesla chargers are not open, in the slightest. Them publishing specs online on the physical plug design doesn't make it open to everyone. There's still lots of Tesla IP that covers those specs and designs. Tesla won't license it to other automakers unless other automakers essentially never enforce any of their IP.
Thanks for pointing that out. I genuinely was under the impression that it was open to all cars, so I learned something new today. It made me realize I fell prey to Tesla's PR.
People will pay more for cars that project images they want to project. This explains pretty much all the car preference patterns for people who have enough money sloshing around that they don't need to purely focus on maximizing utility for cost but who don't have "I don't care what you think" money.
Autosteer: I haven't tried the competition; Tesla's is pretty great on the highway, but not when there is roadwork or in cities where the roads will have unexpected swerves (e.g. with cars parked in the road, or blockades).
Cruise-control: In heavy traffic, I don't use it. I like to contribute towards reducing congestion, and the cruise-control accellerates too much at once, causing lumps. But for non-heavy traffic it's pretty great. The "braking for things that don't even exist" is a separate concern, IMO.
Ultrasonic, radar: Yeah, I'm disappointed here.
Rain-sensing wipers: It actually has that, and they do sort of work. What's annoying is, apparently in Model Y, you can't disable automatic wipers if you're on autosteer. So when you're on a very regular stretch of road, and the rain-sensing wipers think it's raining and it's not, you can either choose between autosteer + wiping a dry windshield at the lowest frequency, or driving manually.
Automatic high beams: They work pretty good in my experience.
Carplay: I get that Apple wants to be my car's operating system. It may seem great, but I'd prefer if my devices got along without total platform dominance from one party.
My experience is that Autosteernis significantly better than the competition. And the Tesla absolutely has rain-sensing wipers.
I find it to be a fun, well performing and comfortable car. I am shopping for a second car currently and have yet to see anything close. There are no other electric cars on the market that tick the price, performance and range boxes, before I even start looking for the creature comforts I'm used to.
Talking about "lane assist" on a German Autobahn - I had a bad, if not worse anecdote with a French (the model) rental car one or two years ago. I did not know that assistant was on, but had I known I would have thought nothing of it.
Because what happened really surprised me: Entering a construction zone the assistant took the wheel away from me(!!!). I was steering manually, normally, but the assistant insisted I follow the previous line in the road, that lead right into the construction zone. Which was behind a concrete divider.
I actually had to fight the "assistant", with not insignificant use of force on the wheel, to not crash sideways into that divider wall.
I swear my car was (gently) fighting my steering here due to the lane keep feature before I turned it off. I frequently drive around potholes and dead animals and don't need the car trying to fix that for me.
I tried VW(+Skoda), BMW, and Stelantis(Citroen/DS/Peugeot models) "Autosteer" in past year(changing comanies, waiting for new company car to be build while having Hertz rentals), plus I had autosteer with adaptive CC in Skoda from 2018.
It got a lot better over the 2018 thing, that tried to crash itself on several roadworks, I actually almost scratch the car in first 100km I had it because of it.
My latest BMW works perfectly fine in German roadworks with bright yellow lanes,
Also all the stelantis cars I've driven (all 2020+ models) worked fine in that conditions.
Camera system has problems when there's fog, or direct sun in precise angles.
Radar has problems where there's lot of rain, and rainsplash from trucks, or snowy conditions, and the front radar gets frozen.
Skoda/VW lane assist feels like it would bounce you back and forth between the lane markings if you just let it do it's thing; more useful a backup system to stop you crossing the line when you have a lapse in concentration (which is how they describe it I believe).
Yep, its function is basically: if your child distracts you and you don't pay attention for a few seconds then it'll try preventing you sliding into the other lane. Which is fine.
They have several tiers of it I believe, the best one also offers automatic lane change, like BMW, it was actually working perfectly fine for me, with no back and forth bouncing - it had the annoying <steer a bit or I don't think you are holding the steering wheel> feature (Karoq 2018) - but newer models I think also have capacitive steering wheel, so you can only touch it.
This is an example of the risks of vertical integration. Some of what Tesla needed/chose to do itself is paying off. More integration into a single electronics system puts Tesla ahead of an industry trend with many unique features.
But rolling your own ADAS and FSD might result in having spent a lot for a result that is inferior to what ADAS suppliers can sell to other car OEMs, with no substantial benefit to the FSD effort, which is much riskier, with the added, and perhaps gratuitous risk of trying to do FSD with only camera sensors. The risks in that are not just it is a risky approach to real time sensor data, but that the data all your cars are collecting is worth a lot less for not having a LIDAR point cloud providing 3D data.
I’m most curious how Teslas autosteer compares to CommaAI (or other offerings). I’ve been seriously considering letting CommaAIs set of supported cars to try it out.
It’d be cool if there was a yearly competition to rank the basic lane keeping/speed matching/turning. Eventually when it’s actually full-self-driving they can rank FSD.
My brother uses CommaAI's openpilot on his Honda HR-V, and it's really quite interesting. It does pretty well where we've tested in Virginia and Maryland, but we haven't compared it head-to-head against Tesla's offering.
This comment made me wonder how they train the car for different sets of traffic rules. Some of this goes beyond formal rules. On most US highways you are technically supposed to overtake on the right, but it's totally common. In contrast in much of Europe this is actually enforced and the convention is that you only leave the rightmost lane while overtaking.
Do you do 100% of training on data from the same region? Are there low-level skills that all translate and don't bring the risk of training contradictory expectations? Can you just annotate training data with the rule set/location that applies?
when it comes to decision-making there is a lot of hardcoded rules in these systems. They don't learn the speed limits or to stop at red lights. The learning is all in the perception and object recognition but a lot of the behavior is just explicit rules.
> And sure, autosteer != FSD. But it's a much simpler problem that the Tesla still basically fails at.
They've publicly admitted that the entire autosteer codebase hasn't been touched in over 4 years, since they fully pivoted to getting FSD in a usable level 2 state.
It is very bad at handling ambiguity when the road has both white and yellow navigation lines. There's a number of scenarios, some of which involve criss-crossing lines, some where they run mostly in parallel.
The car sometimes wants to follow the white lines when the yellow are actually overrides. When the yellow lines are wrong (and should rightfully have been scraped off), this always leads to disengagement. For stretches where the yellow lines are correct, and the older white lines are off by half a meter so that there's effectively not enough space between the white lines and the concrete side, autosteer will sometimes try to switch to the white lines, which would lead to the car smashing into the concrete side.
And sure, autosteer != FSD. But it's a much simpler problem that the Tesla still basically fails at.
Sometimes, in fog, cruise-control will abrubtly break. Very nerve-wrecking.
Relying fully on vision has its downside. Listening to Kaparthy on Lex Fridman's podcast, it sounds like this was cost-cutting. I can't believe that "nope, that big white thing isn't a wall" is not objectively better navigation.