Hacker News Reader

Comment by noncoml | Hacker News Reader

IMHO Tesla's biggest mistake was to move away from Radar/Lidar/etc.. sensors. Their software feels years behind what most cars have these days.

Some examples:

- When you drive in the night on a dark highway the car keeps complaining that the door pillar cameras are blind.

- Parking assist is a joke in calculating distances from other cars/obstacles. Also completely stops in poorly lit place. "Parking assist degraded"

- The emergency line crossed alarm sometimes triggers while I am dead center in my lane

- Blind spot monitor is basically unusable. Why would I want to look at the screen on my right when i try to change to the left lane.

- Windscreen wipers going on randomly when there is absolutely no rain, but don't start when there is light rain.

If they continue like this I can see BMW/Merc eating their lunch in a couple of years

View on HN

Here is a video here showing the Tesla Vision implementation of parking sensors [1].

After all these years if Tesla's ML models are incapable of basic static object detection then I don't see how they belong on cars at high-speed. And it says a lot about Tesla management that this made it into the car at all let alone as a replacement for USS.

[1] https://www.youtube.com/watch?v=kfsRFTpaTwA&t=1s

I drove Teslas for 4 years and recently went back to a humble Toyota. I really don't think the pure-camera FSD will ever work properly. Between a Model 3 and Model X I've seen them do so many goofy things on the road, and get stumped by very common environmental factors such as fog, and strong sun with snow on the ground (glare). I've seen FSD get confused by poorly marked roads and swerve into exit lanes, or drive on the complete wrong side of the road. I stopped trusting it and I think cameras alone are just not fault tolerant enough.

RGB cameras aren't, but there are other kinds of vision they could be using, like IR and polarization cameras or event cameras for motion.

It's certainly not good they've picked systems strictly worse than human vision.

I suspect the greater shortcoming isn't in the sensor suite department, but the intelligence department. Either way, what Tesla's putting on public roads currently seems criminally negligent and I simply don't understand why no TLA has intervened yet.

I think it’s because we see everyone texting while driving. No one wants to enforce safe driving.

How is it "criminally negligent" when the driver has the ability and the obligation to take over if the car starts to do something weird?

You'd have to be actively ignoring reality to not know of Tesla's problems in the autonomy department.

Just search for phantom braking:

https://www.youtube.com/watch?v=6HDbDXeRSPw

That's not even getting into the insanity that is having random members of the public LARP as FSD QA on public roads, shared with other random unwitting members of the public, who given the choice wouldn't want any of this happening around them while driving their kids to school.

That decision was downstream of another mistake: selling a feature before the feature actually worked. They wanted to sell their cars as having all the hardware for FSD before FSD was actually implemented, and they wanted to make a tidy profit while doing it so they were strongly incentivized to conclude that the hardware needed for FSD would be cheap (excluding lidar.)

And blatantly lying about it for years:

https://www.tesla.com/blog/all-tesla-cars-being-produced-now...

In 2016 Tesla claimed "as of today, all Tesla vehicles produced in our factory – including Model 3 – will have the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver".

It was false at the time and it's still false seven years later.

It's crazy to me that people paid for this feature like 5 years ago and never got it, and their cars are now depreciated >50% on top of that.

It's also crazy to me reading this thread where people are talking about how they've bought OTHER teslas with FSD since their first. Suckers are apparently born every minute and some number of them NEVER LEARN.

It’s the people saying things like “Full Self Driving obviously doesn’t mean that it can drive unsupervised!” that get me.

Was it a mistake when it boosted the share price to make Musk the richest person alive with 0 consequences?

You’re completely right, misleading people to make money is not a mistake from musk, it’s his job, he’s very good at it, and he indeed is shielded from any consequences since any crash while using "Full" Self Driving is still fully the driver’s fault.

I think in Elon's book by Issac, it says that if it weren't for their FSD announcement, they would have gone under.

It was only a mistake if their goal is to be honest to their customers and provide the best possible product, but clearly other motives were/are involved ...

Are these sensors really that expensive? It seems incredibly dumb to not use sensors that give you more data than pure vision provides.

As of 5 or so years ago, the lidar used for this on the research side of things started at US$10k per unit (and for "surveying" you can do ok with one-on-top but for navigating you need something smaller and not sticking out of the roof, so then the car itself is a blind spot so you want 2x or 4x units.)

As a robot developer, I was really hoping that high-volume self-driving would lead to lidar mass production (and thus price drops), and there was one promising non-mechanical lidar that was aiming for $1000/unit for an self-driving-grade sensor, but I've lost track of it...

I think at least the ultrasonic parking sensors that are on pretty much every other car are pretty cheap.

They are cheap, but range and resolution are terrible. More comparable are the radar units for the "car in your blindspot" lights in mirrors (not really comparable to lidar at all, just vastly better than sonar.) Tesla dropped those too, with a claim that they were a lot flakier than their optical system (which I don't dispute - the material the published about it was pretty convincing - though mostly it gave the impression that Tesla's radar implementation was terrible, and that optical modeling beating it wasn't as impressive as it might sound.)

Even if we take this at face value, and assume that the camera is better, the implementation is horrible. Don’t force the driver to look at the screen in the middle, put red lights in the mirrors instead.

But as usual Tesla likes to be different for the shake of being different even if this makes the car worse

Yes, and unreliable if low cost mechanical versions are used, or if there are rain drops on the sensor glass. Also, they are blinded by signs.

Reply on news.ycombinator.com