3 ms·
> But this chaos—this unpremeditated waltz of oops, no, you go and nope, buster, me first—is reality. It’s how cities work. Which means that if a car is going t
by dmix 9y ago
> But this chaos—this unpremeditated waltz of oops, no, you go and nope, buster, me first—is reality. It’s how cities work. Which means that if a car is going to drive itself, no humans drivers involved, it must get really good at interpreting and anticipating the behavior of human non-drivers.
I'm not sure why it's always pushed as a near unattainable problem to have cars react in a human-ish way in situations as these.
All it needs is machine learning of thousands of situations of purely human drivers or watching other cars around driverless ones, and adapting the models to be able to apply them in particular situations. Such as a 4-way stop or someone rolling through a stop-sign.
It seems this obsession with defaulting to what the idealistic but unrealistic safety-first situation is (as say, designed by idealistic traffic engineering and law enforcement) vs focusing on learning to drive like humans with the least amount of errors, where rules are more flexible and risk is evaluated without much loss of safety.
We live in a very tightly controlled rules based society on paper, which seems like a major roadblock these systems have to confront, as they seem to pretend real safety is following these existing rules exactly. Yet it sounds like it would be far better if it broke the rules about 10-20% of the time... like most humans do for the sake of efficiency and merely adapting to the situation rationally.
We should be giving the machines as much, or more, leway as humans, much like how we hardly give jay-walking or going 10km/hr over the speed limit a second thought - but applied to a much larger group of compromises.
This may be a bit of my own libertarian bias that I have, a worldview where chaos is okay, and encouraged, as long as no one gets hurt (to a reasonable degree). But the actual problem sounds to me that the traffic rules aren't well adapted to machines yet, which tend to take them far more literally than humans, rather this idea being pushed that the machines are navigating the rules poorly.