> because if the driver hadn't been watching Harry Potter

This was proven to not be true: https://jalopnik.com/the-ntsb-to-partially-blame-teslas-auto...

That's the thing, how do you make an UI that unambiguously conveys that the car hasn't seen an object in front of it, so that the driver can react to it?

Maybe with AR the car could highlight the objects detected so that the driver can brake if the car doesn't see an object it knows about, but a head mounted display seems pretty clunky.

The problem is that the "paying attention" part of driving is actually one of the most mentally taxing for the driver and it seems the pay attention 100% of the time is basically the same as saying 5% of the time an event might occurs that requires your immediate action. The problem being that if you have to jump into action to avoid something catastrophic 'sometimes' you basically have to be paying full attention all the time in order to collect enough contextual information about the incident. It's the same as watching someone toss a baseball to you and catching it vs your friend yelling "Headts up!" as a baseball is already flying through the air at your head.

Japanese train drivers are famous for pointing at things to always keep attention. May be similar system can be used for cars: for example car could narrate objects it sees and ask driver to point at them with hand, then check with depth sensor and gesture recognition that user pointed correctly, and then slow down and stop if user pointed wrong.

But that has the problem of "computer recognizes 'car', asks fleshy meat-bag to point at 'car', sees meat-bag point at 'car'".

It doesn't learn anything about the environment from that, so to me it doesn't change anything.

So wait, rather than just pay attention to the road when driving my old beater, I could constantly be pointing at things and struggling with poor voice recognition? Somehow this sounds like more work than just driving the car myself.

Well, how does an airplane do it. A loud robot voice that goes “PULL... UP...”

If a loud robot voice goes "PULL UP", airplane pilots still have much more than a couple seconds to do something about it.

I don't think people realize that "Tesla Autopilot" will never, ever, ever stop for an object in front of it. Tesla has said repeatedly it isn't designed to do so. If you run a Tesla toward a brick wall with Autopilot on, it will crash into it 100 out of 100 times. It will hit ANYTHING in the lane ahead of you. It isn't designed to stop in such cases.

So there's no need for a UI. If there's something ahead of you, you must stop or you will hit it full speed without braking.

They definitely can and do stop. There are plenty of videos of accidents avoided because of stopping, and this is a rather common feature on most mid to high-end cars now, commonly called a collision avoidance system and grouped under the umbrella of pre-crash safety systems.

The avoidance feature is designed to stop the car in the event of an unavoidable obstacle directly ahead. It doesn't guarantee a stop but it will still apply the brakes faster and more fully than a human can react and thus shed more speed from the impact. Pre-crash safety will also tighten the seatbelts, move seat positions, close windows, and inflate auxiliary airbags in preparation, and these are all well-tested and proven features.

What? That's not true. It can and does brake for objects in front of you that it senses. Its ability to sense objects is sadly very incomplete, but it does (quite often, but not often enough) successfully sense objects and slow or stop for them.

See page 93 of this owner's manual, for example: https://www.tesla.com/sites/default/files/model_s_owners_man...

(What it won't do is steer around obstructions in a lane, either full or partial.)

What is the difference between “objects in front of you that it senses” and “obstructions in a lane”?

Well, I guess the second doesn't presuppose that it senses them. But I think you concentrated on the wrong thing:

Autopilot is not designed to steer around obstructions.

Autopilot is designed to brake for obstructions.

It sometimes fails to brake for obstructions. But it sometimes succeeds in braking for obstructions. It never steers around obstructions, even if it senses them perfectly.

Emergency braking is designed to lessen the impact, not prevent crash. See manual. This is mostly for avoiding crashes into a car that suddenly stops in front of you, I.e. in lower speed city traffic and jams. For this, detection within a few meters is all you need - ultrasound sensors might do. If the velocity delta is 120km/h, you’re not going to meaningfully slow down when already this close.

I think the idea is that the UI needs to force you to always pay attention.

The NTSB report said that the driver was not watching any movie.

> Of course, that's where the UI comes into play, because if the driver hadn't been watching Harry Potter, he would have seen the tractor trailer as well.

Reports at the time said he was watching it on a portable DVD player, not the Tesla screen. Not sure about the NTSB report someone just cited though. If that report really said he wasn't watching anything, I wonder how that rumor started?