Hacker News Reader

Comment by ousta | Hacker News Reader

a quick question. heres the situation I am in a self driving car driving full speed on the highway, a little girl crosses the highway (highly unprobable but still possible) what will my car do? maintain high speed and kill her? or try to avoid her knowing that any attemps to avoid collision at this speed would likely injure me if not kill me? then how the car would react if it is a deer? then how the car will be 100% sure its a deer?

this is where self driving cars are limited and shoudlnt decide for us.

View on HN

We can throw the same question right back at you. What would you do - kill the girl or kill yourself?

These hypotheticals don't really help the conversation. I don't know how the car would react but it would be nice to know if the car is going to try and work with neighboring cars to try and not hurt anyone. Not to mention that the car will already see the girl crossing the highway before she ever does get out there because of how sophisticated the sensors are.

>What would you do - kill the girl or kill yourself?

Unless the AI behind these cars are so advanced that we cannot know ahead of time what it would do and it cannot be put in a simulation to test it, there is a vast difference between a self driving car and a human. Namely, we don't (and with the current level of knowledge can't) know what a human will do til after the situation occurs. And in general, in snap judgment cases like this, we do not punish a person for making a bad choice unless it is outright malicious.

Contrary to this, a self driving car is in a system who we can attempt to simulate (though not at perfect fidelity) and one where we know the underlying algorithms enough to answer the question and even change the algorithms to change the answer before the situation ever occurs.

>Not to mention that the car will already see the girl crossing the highway before she ever does get out there because of how sophisticated the sensors are.

This is attacking the situation instead of answering it. All I would have to do is modify the situation slightly and the question would remain just as valid and thus the response does nothing but postpone.

I'm fine with this. I'll gladly take a trade where my probability of dying, via the general improvements gained from self-driving cars, decreases in most cases, but increases in other very specific, low probability cases. On average, I die less frequently.

Don't let the perfect be the enemy of the good.

I trust a highly trained algorithm to make better and faster decision than a human driver. The algorithm is even aware of all the cars around it and their speed and can decide to take evasive action with a way better chance of success than I ever could.

If the other cars on the road are self driving cars as well they could all communicate and take into account the one car moving to avoid the object in the road, thereby also reducing the chance of damage to the vehicle swerving to avoid the object

It should be a three tier decision. First, preserve the life in the vehicle. Second, preserve life outside of the vehicle. Third, minimize property damage.

Deer vs little girl does not really matter, that is just an emotional ploy.

I'd hope the algorithm was designed to crash the car instead of running over a pedestrian. A pedestrian has little chance to survive getting run over compared to a passenger in a car with a seatbelt, airbags, crumple zones, etc.. to protect him.

How do we take into account emotional harm from killing a little girl causes the driver to be harmed (they weren't driving at the time as they let the car to stake over, but they blame themself for it regardless). A parent would rather they die than hit their own kid.

I guess once these cars begin being sold people can start testing situations using dummies and such.

Trolley problem philosophical discussions notwithstanding, I still suspect that the default mode (as it is with humans under most circumstances) will be to lay on the brakes and hope for the best.

That said, there are certainly circumstances where swerving is the right decision and a vehicle with 360 degree sensors is potentially in a better position to make an optimal decision than a human making a snap judgement based on mostly what they see in front of them. Probably the closest I ever came to being in a really serious accident happened on a highway when someone slammed on their brakes to avoid rear-ending the car in front, ended up drifting into the next lane where another car had to swerve to avoid them. I ended up swerving into the breakdown lane. Amazingly with cars swerving and braking all over the road at 70 mph, no one collided but it was a very near thing.

I strongly suspect that, however many decades away true self-driving is, good assistive driving and collision avoidance systems especially on highways are going to be a pretty big near-term win.

It will better be able to make a decision than a human can. But what decision is optimal isn't well defined because never before have we had the need to define it. So far, only directly malicious behavior is known to be bad. But prioritizing your safety over the girls safety is something we have neither decided to be good or bad.

yes but what if your goal is to get out of this and the goal of the algorithm is to save the deer or the little girl?

Then I trust the safety engineers of that car to have designed a car that will get me out of an accident alive.

I don't know what speed you drive, but I had wildlife run out in front of me at 70mph (the speed limit).

I managed to reduce the speed to about 40 mph before hitting it by which time it didn't do too much damage.

Your peripheral vision can see stuff coming from the side quite well.

Reply on news.ycombinator.com