I think we're still far out. There still hasn't been a solution proposed (that I've seen), of how to avoid multiple disaster scenarios (Ex. hit child chasing ball, or veer off of cliff).

As soon as this question shows up on a human driving test with an agreed-upon correct answer, it'll be fair game as a prerequisite for an L5 autonomous vehicle. For now, though, I can't imagine how it's a significant problem for self-driving cars to be able to solve.

It was on my driving test in Alberta 15 years ago. The correct answer is to remember that the people in the car are protected by a big cage of metal, and pedestrians aren't.

I think the actual question was whether or not you should drive into a wall instead of hitting a pedestrian, not drive off a cliff. But hitting a pedestrian was definitely not the right answer.

I mean, if that's the answer you want to program into a robot car, it doesn't sound hard to do.

Luckily I've never been faced with this choice, and if I did have a fraction of a second to decide I have no idea what would happen. Doesn't stop me from driving though.

That's not a real problem. The correct answer is to stop the car, in almost any realistic scenario. It's also not a big deal, because human drivers are really bad at surprise children jumping into the middle of roads near the edge of cliffs too.

Or the user could choose personality of the AI selfish vs altruistic to match his

maybe this is a joke, but really, no. Car AIs don't encounter trolley problems, because the trolley in the trolley problem is ALREADY out of control and unable to stop (besides the turning lever). If the Car AI is in a situation where it might lose control of the car, it should decide that it isn't safe to drive. Asking a car to be altruistic means the car is making value judgements about what's better to hit, which means there's a mode where the car decides to hit something. The car should not be allowed to decide to hit something, or even swerve wildly out of the way. It should drive at a slow enough speed that it can safely stop for reasonably encountered circumstances. And if someone teleports in front of your car, it should just try (and fail) to stop in time.

It depends. If someone runs a red light and swerving (assuming it knows it's safe) will avoid an accident I don't see why it shouldn't do it. But I agree this is an edge case and is not an important feature. There are more important things to focus on.

The idea is to add an element of randomness to the decision making. I don't mean to pick random decisions, but when there are decisions that weight approximately the same it should pick the decision based on personality/mood. This would enable the algorithm to adapt and be flexible, instead of a deterministic algorithm that could cause chain unoptimal behavior.

By applying deterministic rules to a system an edge case would affect all users.

By introducing variance and enabling the system to adapt, the edge cases would be, in a way, self correcting and more resilient.

The answer is probably that this isn't real. A self driving car with better vision, perfect attention, and instant reactions won't get itself into that kind of situation. And in any other case where an accident is unavoidable, the answer is to just hit the brakes because it won't matter anyway.

Another benefit will be that the self-driving car behind you will also instantly hit the brakes and won‘t cause you neck trauma, making this option a little bit more appealing than it is right now.

There is no need for self-driving cars to have emergency braking assist.

Also, it has about zero understanding of what's going on. And it also has the same laws of physics acting on it. Zero reaction time does not translate to zero braking distance. People understand the world around and can make predictions which gives them time to react.

my personal feeling is L5 won't happen until we have AGI at average human levels.

There's really that much involved with driving a car.