[RcouF1uZ4gsC]

I think the big issue is that if you really want autonomous driving over any road, you will likely either need sensors beyond vision or some AGI level model of the world.

For example, on the highway, truck infront of you drops some debris.

What do you do? It can be dangerous to suddenly swerve or brake, but also dangerous to hit something.

What you do In real life is quickly evaluate. You are looking at the object that fell and estimating how deformable it is. For example, if it looks like a wad of paper, you would continue on. Or you estimate if you have clearance based o how big it looks and your mental model of your car’s clearance.

You also look at the traffic beside you and behind you to evaluate risk of braking or swerving.

You are also doing object recognition. If it is an empty box, you may be ok hitting it, but if it is an infant car seat that fell, you will brake hard/ swerve even if it is dangerous.

Also, for people who bring up the safety of human drivers, one thing to consider is the safety of a hybrid approach where the human is still in charge but you have safety features like lane keeping, blind spot monitoring, automatic emergency braking, etc. That hybrid approach may be actually safer than either only human, or only computer driving.

[danans]

> You are also doing object recognition.

Object recognition has been a part of autonomous driving systems for a very long time. The difference between a stroller and a piece of paper is pretty well understood by these systems.

> For example, on the highway, truck infront of you drops some debris. > What do you do?

The answer is to always keep a large safety margin between you and the vehicle in front of you when traveling at high speeds, so you have time to react and take the safest course of action. Seconds make a huge difference in this scenario and it's easy to get them. No autonomous system is going to be hugging a truck bumper, but rather is going to always maintain a significantly greater safety distance than an aggressive human driver would.

> If it is an empty box, you may be ok hitting it,

You can never be sure that the box is empty, so I'd recommend always swerving to avoid.

[klmadfejno]

Events like this are rare, and often not handled well by humans. You have 3 basic options, and in reality some mix of all of them.

1) Hit the object 2) Swerve 3) Brake

Hitting the object is almost always a bad idea. On a highway the best option is almost always 2. On an empty road, the best option is almost always 3.

Saying things like "You also look at the traffic beside you and behind you to evaluate risk of braking or swerving" implies that humans are good at this and computers are bad at this. I'd wager humans are bad at this and computers are good at this. And in either case, situations like these represent a small fraction of accidents on the road. If the computer arbitrarily chose to recklessly swerve to the nearest lane every time it would still likely outperform human drivers in terms of safety if it solves the low hanging fruit elsewhere.

[kingsuper20]

> I think the big issue is that if you really want autonomous driving over any road, you will likely either need sensors beyond vision or some AGI level model of the world.

Somehow people manage without this.

Without any real knowledge of the problems the engineers are hitting, my current take is that to produce a reasonably good autonomous driving system requires killing some people. Slowly adding features will send more information to the mothership and they'll have to push the envelope to learn enough.

[DreamScatter]

Killing people to achieve a technology is unacceptable and unethical ... maybe in nazi germany you could pull that off, but I think you'd eventually be charged with crimes against humanity.