Hacker News Reader

Comment by Animats | Hacker News Reader

Yet another Tesla crash involving failure to detect a stationary obstacle on a road. Tesla has been killing people this way for years now.

View on HN

After the firetruck incident[0], semi across the road incident[1], and road separator incident[2] this isn't unreasonable to claim.

What is odd is that modern Tesla (3, Y) vehicles have a better sensor suite than most other vehicles, and yet its AEB (Automatic Emergency Breaking) in isolation is seemingly worse than some other modern vehicles.

I can tell you for a fact that Subaru's relatively primitive Eye Sight system based on two optical cameras in parallax (like the human eye, thus the name) will "see" the depth ahead and brake accordingly, it doesn't get confused by a bridges/road signs and reflective stationary vehicles. I've seen it do so frequently.

A Tesla has significantly more sensors and more processing power but these problems continue. Everyone always gets distracted talking about Autopilot this and Autopilot that, but let's not, let's talk only about AEB and why Tesla is somehow both the best at advanced automatic driving and also somehow has uncompetitive AEB?

Maybe it is time to take AEB out of Autopilot and make it its own system that is simpler but can also react to potential Autopilot errors? Or buy AEB off the shelf and put it side-by-side with Autopilot?

[0] https://www.wired.com/story/tesla-autopilot-why-crash-radar/

[1] https://www.wptv.com/news/region-s-palm-beach-county/serious...

[2] https://en.wikipedia.org/wiki/Tesla_Autopilot#Mountain_View,...

I think the problem is that Tesla went all-in on optical sensors and relied on software instead of something "simple" like sticking a lidar on the front of the car.

I am assuming a simple modern lidar and software package with a field of view limited to "What can the car hit if it keeps going straight" and connected to the AEB would be a pretty fool-proof way of stopping this but that might be too much cost for too limited an application.

A pretty critical application

From what I’ve seen a relatively simple radar is what many automakers are using as the main sensor for AEB. You can see them on the front grille, usually a small rectangular patch in the plastic about the size of paperback book.

Though those have had their problems too. Obviously easier for them to detect other vehicles but harder for pedestrians.

Maybe AEB sucks because it's already a separate system from Autopilot, and therefore doesn't get much attention or engineering headcount.

For example, this would explain why Autopark sucks and hasn't improved in years.

I have to assume that tesla's cars aren't confused by those things either, for the most part. The difference is that most car's emergency brake systems act when a driver has failed to respond; the driver and the computer are acting in concert. Tesla has designed and marketed their system as if it worked separately from a focused driver. In the rare instances that their system fails, it is much less likely that the human driver has been paying attention and will brake manually.

I just looked up the subaru system. In the emergency braking case, they clearly claim that their system acts only if the driver has failed to brake. That means, for example, that it will brake much more abruptly than a human would, since it is waiting for the last moment before braking. Their lane following technology behaves similarly. Tellingly, it is called "eyesight", suggesting it is merely an extra pair of eyes, rather than "autopilot", which suggests that it can drive on its own.

I think the points that have been raised in this thread are all quite valid.

Just to add another: I wonder if Tesla's driver assistance systems has been designed to minimise false-positives in comparison with those of other manufacturers. When driving a friend's Subaru on a road trip a while ago we all commented on the lane-keeping alarm being overzealous, and ended up disabling it altogether to shut it up (this is on somewhat narrow b-roads). I've read of similar complaints from automotive journalists regarding various other manufacturers. In comparison, I've barely ever experienced false positives from the model 3's lane-keeping assist, even on similar roads -- maybe once or twice ever.

Minimising false-positives is important for system efficacy, as otherwise people will simply disable it. But I do wonder if this comes at the expense of increasing the chance for false-negatives, which are far more dangerous. Is trading 100 false alarms for a single missed obstacle worth it? I'm not sure.

I drive a Kia e-Soul myself, and while I keep the lane keeping assist/alarm on most of the time, I definitely disable it on narrow, twisty roads. Otherwise, the alarm goes too often, especially when I move out closer to the shoulder to make room for meeting traffic.

After the firetruck incident[0], semi across the road incident[1], and road separator incident[2]

Also the street sweeper incident [1], the overturned truck incident [2] ...

There's a pattern here.

[1] https://youtu.be/fc0yYJ8-Dyo

[2] https://youtu.be/X3hrKnv0dPQ

My guess is that Tesla is running an object classifier on the two front-facing cameras in parallel, and then measuring the parallax of the objects, not of the features of the image. As such, if it can’t recognize something in front of the car as a known object, then the car just plows into it. This is why a lot of the accidents involve things like side-on semis, overturned vehicles, and fire trucks - things that don’t look car-like.

On top of that, they have to discard all of the radar data from objects moving at zero speed, because otherwise they’d be swamped out by the return from the road itself. So, the car has a literal blind spot for stationary objects in its path on the highway, unless it can recognize the object as a car with its cameras.

>Tesla has been killing people this way for years now.

Fo years? Article says 4 deaths. If they "have been killing people for years" one would assume you'd see higher numbers?

Their "autopilot ready" cars still ship with monocular center mounted vision system. I dont know what they are smoking there :/

Here are the top 10:

https://www.autoinsurance.org/which-auto-makes-get-in-the-mo...

This is just a list of crashes by make, with no attempt to normalize by sales or registered vehicles or miles driven or anything.

It's no surprise that the makes with the most sales and most vehicles registered and almost certainly most miles driven also have the most fatal collisions.

I bet DeLorean has zero fatal collissions for most of the last ten years. Doesn't make it a good or safe car, just means the founder got caught (set up?) on drug charges and the company collapsed and there's not too many on the road.

as expected this tracks the number of cars sold.

2020 fatal crashes:

    #1 – Ford        21,262 

    #2 – Chevrolet   20,435   

    #3 – Toyota      12,918 

    #4 – Honda       10,969  

    #5 – Dodge        7,983

    #6 – Nissan       7,956 

    #7 – GMC          4,533

    #8 – Jeep         4,138

    #9 – Hyundai      3,766   

    #10 – Kia         2,834

    ...

    #41 – Tesla          62

2019 sales [1]:

    #1 – General Motors                2,886,904

    #2 – Ford Motor Co                 2,406,188

    #3 – Toyota Motors                 2,383,304

    #4 – FCA US (Fiat, Chrysler, etc)  2,203,663

    #5 – American Honda                1,608,170

    #6 – Nissan-Mitsubishi             1,466,727

    #7 – Hyundai-Kia                   1,325,342

    ...

    #13 - Tesla                          192,250

also, the crashes is all model years so it favors less US market time.

[1] https://carsalesbase.com/us-car-sales-analysis-2019-brands/

So % of sold cars that end up in fatal crashes:

Tesla: 0.0322% (or 1 in about 3100 cars sold)

Ford: 0.8836% (or 1 in about 114 cars sold)

Its late for me, so check my math, but I think it’s right.

That looks about like you took the 2020 fatal collisions and divided by the 2019 sales, but that doesn't make a whole lot of sense. Probably most of the Ford crashes were not from the 2019 model year, they've been selling cars for over a century, and many of their cars are still on the road.

My understanding is fatal crashes are biased towards older models as well. Tesla has barely any cars over 10 years old, let alone 20 or 40 or 50.

More fair would be to look up registered vehicles by make and compare to fatal collisions (which I can't find a free source for, unfortunately).

Isn’t it still likely overall safer to use, especially within its guidance (to keep hands on the wheel and eyes on the road)? I would imagine their self driving features have prevented many more accidents than they have caused.

Most cars have AI technology to prevent crashes. What differentiates tesla is that they have technology that enables the driver to stop paying attention. Automatic braking has likely saved lives, but in this dimension it isn't clear that tesla has gone above and beyond other auto-makers. Enabling drivers to stop paying attention endangers them, and it is only in this dimension that tesla is clearly unique.

The data don’t bear that out. If you compare accident rates for highway miles, Tesla’s are not any safer than other cars, even though nearly all of them have autopilot.

Reply on news.ycombinator.com