Hacker News new | ask | show | jobs
by tzs 11 days ago
I was curious how people working on self driving cars handle the large variation in local and region traffic laws and did a bit of searching.

For example in some places a car making a turn that will cross a bike lane is required to merge into the bike lane before the turn (California and Washington for example). On others (Oregon for example) the car must not do so.

School buses are another good example. On a road with lanes in both directions when do you stop for a school bus heading the opposite way that is stopped with its red lights flashing and stop sign extended?

In some place the answer is "always". In others it depends on how many lanes there and whether or not there is a barrier like a median strip between the two directions.

One approach is to not let your self driving system operate in places where you have not explicitly added all the local and regional rules to your system.

Another approach is to try to learn the area with AI. It sees lots of humans making turns from the bike lane, it makes its turns from the bike lane too.

An issue with that approach is that a lot of humans violate traffic laws, so you have the danger that your self driving system learns to violate traffic laws.

1 comments

> An issue with that approach is that a lot of humans violate traffic laws, so you have the danger that your self driving system learns to violate traffic laws.

Phrased another way, a benefit of that approach is that if the local conditions necessitate or advise breaking certain traffic rules, your system learns to do so while retaining plausible deniability that it isn't intentional.

It's not that plausible, unless the idea is that differing traffic laws is some obscure concept.
The point isn't whether it's genuinely plausible or not. If there is some law against something, there is some group of people who believe very strongly that this law should be followed and enforced. If the law is not actually followed or enforced, then the system as a whole does not agree with that small group. The point of this "plausible deniability" is to allow whichever regulator is concerned to give a pass to rule breaking without explictly doing so and hence angering this small group.