My Self-Driving Car Crash – The Tesla was driving perfectly–until it wasn't
theatlantic.com
theatlantic.com
For aircraft it's very rare to receive control back from the autopilot in an upset-state. An exception is a trim-runaway, which is a very serious emergency that gets trained for.
In cars you don't have that luxury, it's much more likely that you have to act immediately with very short notice. That does not work for human drivers.
> which is a very serious emergency that gets trained for.
Experiencing this in pilot training is something that really changed my perspective on driving. Some people might get a bit of that in drivers ed, but most don't, and once you have your license you're not re-training. Yet tire blowouts, black ice, hydro-planing, and many other things can happen. If you drive enough then you get real-world training, but many people don't. Even then, every year in the mountain west, the first snow is a wreck-fest on the highways because people forget they need to drive more cautiously in the snow.
In an age of mostly self-driving cars, we've got a bad combination.
But likely not frequently enough. There’s a reason airline pilots have flight simulator sessions that involve various conditions that each are unlikely to occur to them, ever: they may encounter one of them in their career and if they do, they have to know how to handle them.
This is the same feel good mental trap that the peddlers of vehicle inspections indulge in. Say you wave a magic wand and solve those issues. Congratulations, you've solved a rounding error.
The overwhelming majority of vehicle accidents are the result of people mis judging something and setting up a situation that cannot be harmlessly solved within the laws of physics. Like no amount of practice sliding around on ice or water is going to save you when you only have a few feet of wiggle room to regain control before the ditch. Like at best you might reduce the conversion rate from those situations into accidents from the realm of blackjack house odds to poker house odds. That's a poor ROI. Better to train people to just not get into those situations in the first place.
Aircraft take a very different approach because they generally have a ton of space to work in, but on the flip side they can't just stop what they're doing or slow down a ton on a whim nor do they operate in close proximity the way automotive traffic does. While you can make a comparison between automotive and professional aviation, there's just such a wide gulf between them that porting solutions in either direction is fraught with so many caveats as to be not really meaningful and so the comparison is only really an appeal to authority.
>every year in the mountain west, the first snow is a wreck-fest on the highways because people forget they need to drive more cautiously in the snow.
This framing is questionable at best.
If you don't do a thing for 6-8mo you're gonna get measurably worse at it. Just how it is. No amount of smug remembering is gonna replace muscle memory and feel.
It's not the muscle memory and feel that matter so much. The second paragraph of your comment seems to agree ("Better to train people to just not get into those situations in the first place."). What matters is the training, such as remembering to slow down in poor conditions (smugly or not) and stay out of those situations.
Can't wait to see how this plays out with "your job is now code review instead of writing code"
What a meager existence ... and, in my experience, a downgrade from actually just being a driver.
Lane assist and cruise control were always enough (for me) and have been around for decades (certainly for the latter).
So why pay extra to take risks?
> The car was making a turn. Something felt off—the steering wheel jerked one way, then the other, and the car decelerated in a way I didn’t expect.
I use the latest FSD in an M3 and I have noticed it behave indecisively when changing lanes, not so much when turning, but I believe the author's account.
> I turned the wheel to take over. I don’t know exactly what the system was doing, or why. I only know that somewhere in those seconds, we ended up colliding with a wall.
The author disengaged FSD (reasonable when concerned) and ran into a wall.
I almost never let go of the steering wheel when FSD is driving. I want to be able to take over with the minimum delay. I don't know that I'll ever trust it to drive unsupervised.
It's an unbelievable driver-assistance system. But you need to treat it as such. Tesla may market it, and name it, otherwise, but anybody using FSD should quickly realize it has limits.
I wouldn't be so hasty to exonerate the software with suboptimal vision only sensors.
That said, I only notice this behavior in the "Standard" FSD profile. When I bump it up to "Hurry" or "Mad Max" the confidence is 100x. Not sure why, exactly, but this may help you. The only downside is that "Hurry" loves to speed and "Mad Max" even more so.
>I don’t know enough about what actually happened during my accident to say that Tesla’s technology crashed the car.
The cause may have been the combination of FSD and human takeover. When the car fucks up, the driver can take over poorly. Sort of how 'overcorrection' on a highway could spin you out of control.
There is a gray area where a car's guidance drives stupidly, yet would not actually result in an accident. The hot take is that a driver with his face buried in his phone the whole time may have had a better outcome.
In a modern car with modern tires in non-inclement weather taking a left turn at the speeds autopilot will drive (i.e. not 99th percentile race car behavior) there should be enough time to avoid this outcome if you're even only casually observing it.
FSD cannot take the blame if the driver took over controls and simply cannot remember anything that happened after doing so.