Hacker News Reader

I thought maybe this Oakland video was an outlier in how bad it was, but I've watched several videos from the related videos, and the Tesla subreddit, and the system really seems like it's in bad shape. I'm really surprised that regulators are allowing this to be tested on public roads, and that Tesla hasn't pulled it. It is bad enough that when I see Teslas on the road now, I feel unsafe, because the system is so unreliable right now.

I would encourage anyone to watch the videos and make up their own mind, bc I do feel that people over-correct in their Tesla skepticism sometimes. But in this case, it seems warranted.

> It is bad enough that when I see Teslas on the road now, I feel unsafe, because the system is so unreliable right now.

At the very least you should be required to put a visual sign that your vehicle is in "FSD" mode. Sort of like how we require "oversize load" signs so that others can take extra precautions around those vehicles.

In Japan, trucks are required to have three horizontal green lights above the cab, facing forward[0].

These indicate speed. I would guess FSD should have something similar.

[0] https://www.quora.com/What-is-the-usage-of-three-green-light...

My vote is for a big old Cylon eye in the windshield.

They all just look like knight rider, oh shit was knight rider a Cylon?

This sounds like great crossover TV.

Or Student Driver is a bit more apt.

> It is bad enough that when I see Teslas on the road now, I feel unsafe, because the system is so unreliable right now.

The vast majority of Teslas aren't using this FSD beta but the older and more limited Autopilot software.

You'll only have to start worrying about this in the next few months, or last year, or the year before, etc (depending on which Musk tweet you go by).

> I'm really surprised that regulators are allowing this to be tested on public roads

One should add that in this case it's tested by consumers, not certified professionals.

There is no "certification" to be a babysitter for a self-driving car.

https://www.law.cornell.edu/regulations/california/13-CCR-Se...

Just ride as a passenger through a road with cars parked on the side. When I did that about 1.5 year ago, it missed about a third of the cars and pictured a couple as moving on the driving lane.

It's almost like it's in beta, right ?

So where can I opt out (as a pedestrian or cyclist)?

That’s an excellent point. Our government has failed to handle this situation, and these systems should not be allowed on the road. I also avoid Tesla cars like the plague and do not trust what’s going on inside.

Everything gets to be crap by adding the label beta. Texas electricity system, beta!!

> It is bad enough that when I see Teslas on the road now, I feel unsafe, because the system is so unreliable right now.

Conspiracy theory: Tesla wants to show how bad the system is early on to raise public awareness and make people drive more cautiously around teslas.

And when that catches on, Tesla will start producing cars that look very similar to the ones from other manufacturers, so that people would have to drive carefully around every car as well.

They'd lose a lot of customers if their cars didn't look Musky.

> The video is worth a watch: https://www.youtube.com/watch?v=antLneVlxcs

For contrast, here's Tesla's promotional video[1] for this feature and its caption:

> The person in the driver’s seat is only there for legal reasons. He is not doing anything. The car is driving itself.

It's been up on Tesla's site for almost 5 years now.

[1] https://www.tesla.com/videos/autopilot-self-driving-hardware...

> The person in the driver’s seat is only there for legal reasons.

He's the legal crumple zone.

Did lol

This is probably my favorite post on this site, good one.

The other video that this guy posted is also worth a watch:

https://news.ycombinator.com/item?id=26519495

It's failing to safely do a very basic turn across traffic on a clear flat road with wide open view to oncoming traffic, on a clear sunny day. There's not even any weird concrete barriers, shrubbery, signage, traffic light control boxes or such blocking the view of the oncoming cars.

This is utterly horrifying.

Even if Tesla owners opting into this beta test are willing to put their lives on the, uh, line, I'm not sure everyone else on the road is as keen as they are to be volunteered into this test.

I'm not sure I understand how other posters are watching this video, and are concluding that the FSD system is being 'overly cautious'. To me, it looks like it's doing its best to get T-boned.

It seems to be either overly cautious and just frozen in place, or "I'm gonna get myself t-boned", no middle ground in between.

This is precisely my problem, especially with the new Model S that guesses whether you want to go forward or reverse. I'm sure at least a few people will die for the sake of this "feature", and as a pedestrian I can't simply opt out of that risk by not buying a Tesla.

Which makes me wonder why these features are still allowed in cars in general. Having an aerospace background, I have no idea how certification of these things work for cars, but I would like to think some governing body has to sign of on this stuff. At least in Germany we have the TÜV for cars already on the road, and sometimes changing basic stuff like suspension isn't allowed. For safety reasons, which I get. On the other hand, you have half assed self driving features in production cars (not just Tesla).

It seems to me there are probably laws on the books here in the US that could regulators could use to stop this. For some reason, they all seem to be asleep at the wheel (for lack of a better phrase). I guess they’re prioritizing the company over the public.

It's doing both. It's stopping when it should be moving, and moving when it should be stopping.

That really isn’t good. I don’t like how it sits there with the wheels turned into traffic. If you were to get rear ended you would be pushed into traffic. It is much safer to sit with car and wheels straight and only start your turn when safe to proceed.

This is terrifying. I'm now convinced that the engineers at Tesla aren't just bad programmers, but bad drivers as well.

Agreeing with others here, this definitely isn't good. I'm curious if the car "boxes" in the bottom right view represent all of the cars it can see, or if it can see further out. Because in this case, every time it "went for it", there were no car boxes visible, which makes me believe whatever vision system it is using isn't able to see far enough to safely drive at those speeds.

While watching the video, my thoughts as a software engineer: wow, this is impressive.

My thoughts as a driver of a car: wow, this is trash and should not be on the road.

That's a great way to put it, agreed. Purely from a software perspective, that's quite something. But as a driver on public streets? Astoundingly terrible and very dangerous.

> https://i.imgur.com/BG0gN2R.jpeg

My god, and this is on the roads?

Yea, as a development build

Move fast and crash things.

Move fast and crash into others’ things.

FTFY

Doing ML with stereo vision is very hard in controlled settings, and I don't even wanna think how hard it has to be in moving vehicles with lots reflections and other IRL stuff.

I'm pretty sure Tesla only uses monoscopic vision for its self-driving features, or maybe it's just for Autopilot.

Who cares what the the difficulty involved is when these are going to be on the streets today. You sound like the video's passenger

I wonder if these cars distinguish between "I don't know what's in this space" and "I'm sure this space is empty". I hope they do... If the car doesn't recognize the other car right away it is no problem, as long as it is 100% sure it can stop in time if something appears.

I believe safe autonomous cars should behave a bit strange to human sensibilites: They would drive a LOT more defensive in many situations since they cannot be 100% sure that a child is not going to pop out the bushes and onto the street. But on a freeway, maybe if they negotiated with all surrounding cars to drive as a "fleet", it would be able to drive a LOT faster than the safe speed limit for a human driver. The comfort of the passengers would become the limiting factor in many situations.

If I had a friend who jerked the steering wheel around like that, I would not ride with them anymore. That is practically an SNL skit about a bad driver.

It's Toonces, the driving computer.

Mr. T-Bone?

There is nothing "beta" about this, it's at a "proof of concept" level at best.

And it's questionable if it even proofs the concept.

I don't understand why anyone wants anything less than full ('level five') self-driving - that video looks more fatiguing than just driving yourself. Watching someone drive and being ready to take over is a skilled job, why would I want to learn to do it, and do it every time I want to drive anywhere, and not be paid?

Maybe Tesla should follow the “mock it til you make it” pattern with a lot of these AI companies: just hire a bunch of gig workers to sit in simulators driving your car remotely using the cameras and sensors themselves. Have the model learn from them. Genuinely wonder if this would have better results

LTE bandwidth too low, latency too high, and reliability too low.

Compared to the video of the beta, the results may be the same

I had a car with level 2 self-driving, and it was amazing for the commute, going down from SF to the valley and back. It removed a lot of cognitive load from the boring highway driving, because that part is so predictive that the car could handle it. So that's a big partial win for systems like that.

But if you have to watch and be ready to take over at a moment's notice... it doesn't really remove that cognitive load? Unless you don't watch and aren't ready of course, but that's not really the point.

Except in highway driving, the amount of weird stuff that can happen is very, very limited, compared to city driving. It's predictable, and you learn what your car can handle, and what it can't handle.

I assume from your comment that you haven't actually spent a lot of time in a car with these features, but the best way to describe is that it's like driving in grooves. You're driving, but you don't have to actively work to keep your distance to the car in front of you, stay below the speed limit, or to keep your car centered in your lane. Just like with an automatic gearbox you don't have to actively work to keep your car in the right gear, the car does it for you. Same principle here.

The most disconcerting part of the video is watching real world objects blink in and out of the computer's "awareness".

Yea, I understand there is noise in signals so some jiggly lines are OK, but the detected objects disappearing and jumping around seems horrible.

I think this is ok. The perception system has a threshold for the classifications and all the logic comes after that. It cant make reasonable decisions that humans can undersand on something that is 80% likely to be something. Its where where this tech is at right now. What is not ok is to call this FSD.

Humans gain object permanence at around 3 months of age. This software does not yet seem to have mastered it.

Do we know what that view is showing though? If you tapped in to a 3-month-old's visual system you would see things disappear too, but that's not the full story. Maybe it's the same here?

Possibly, but it seems odd to me that they would hook it up to the raw visual system rather than showing the "current state" of the car's perception.

Honestly surprised people are surprised.

> people walking in the street, etc, I would expect to be problematic

As a baseline, this should not pose a problem.

Driving Instructor Simulator 2021

Thanks for the video link, very enlightening. It is disturbing that these are allowed on public streets, its driving skill is far below even the most nervous beginner.

Seems predictably terrible.

According to what I've seen people saying online, in most of these cases Tesla does identify the vehicle but it identifies it as going in the wrong direction so doesn't think it is a problem for pathfinding. Vehicles going in the opposite direction are marked in yellow on the visualization, so they are hard to see sometimes.

The frame grabs of this incident clearly show no vehicle, yellow or not. There are a couple of other yellow objects visible, so this is not a case where it would be hard to see.

In the video I actually do see the yellow outline if I really turn up the resolution, it's super faint though, hard to tell if it is there the whole time. Easier to see in https://www.youtube.com/watch?v=uClWlVCwHsI , at 4:00 the cars are all yellow when they shouldn't be.

But it usually marks on coming traffic, it seems to be flat out missing the cars in some cases

True! While it's difficult to see, the yellow car is present and turns green when it is passed, so the car appears to be detected by the system.

Very hard to see because the brightness of color yellow is pretty close to brightness of the background, almost no contrast. They should have picked some other color instead.

I don't think optimizing for Youtube is important.

Im both suprised and underwhelmed by the video. Its like this video make me really doubt on the FSD system because he looks very hesitant, but at the same time the 4 "near misses" looks so much as the driver overreacting to something that was perfectly fine (especially the curb and the """head on""") that i had to look if the youtuber didnt have some ties to another car maker.