I was in a Waymo in SF last weekend riding from the Richmond district to SOMA, and the car actually surprised me by accelerating through two yellow lights. It was exactly what I would have done. So it seems the cars are able to dial up the assertiveness when appropriate.
It doesn't seem impossible technically to up the assertiveness. The issue is the tradeoffs: you up the assertiveness, and increase the number of accidents by X%. Inevitably, that will contribute to some fatal crash. Does the decision maker want to be the one trying to justify to the jury knowingly causing an expected one more fatal incident in order to improve average fleet time to destination by 25%?
Reinforcement learning is a helluva drug. I'm sure by now Waymos can time yellows in SF to within a nanosecond, whereas humans will only ever drive through so many yellows will never get that much training data.
A human can know the yellows on a few routes. A Waymo can pull over, observe a given intersection for an hour, and tell every other Waymo that exists precisely how long that light lasts.
It's not just collecting the information; it's the ability to spread it.
I'm sure "timing of yellow" is only a few parameters in its network at this point. And it's continuously training, it can probably one-shot the timing changes ( one taxi ride through maybe 3 lights ).
When red-light cameras are installed at an intersection, the number of rear-end accidents typically increases as drivers unexpectedly slow down instead of speeding up at yellow lights.
The cost of these accidents is borne by just about everyone, except the authority profitably operating the red lights. (To be fair, some statistics also show a decrease in right-angle collisions, which is kinda the point of the red-light rules to begin with.)
That seems only like a temporary problem until people get used to actually stopping at red lights, as they are supposed to. After the initial acceptance phase, it should minimise accidents over the longer term.
Unless there is a warning of how long is left on the yellow light, it’s an unsolvable problem because there is an asymmetric risk of stopping vs accelerating
The lights should be designed so that if you don't have enough space to stop with a mild deceleration you should just go through. If a mild deceleration get you rear ended then of course that's an unsolvable problem
No one wants to risk a ticket with a guess at how long the yellow is going to be, or whether they’ll make it thru or not. That is the unsolvable part. Yellows are inconsistent , and you aren’t accounting for slow-moving traffic ahead of you that might cause you to block the intersection, etc.
There was actually a scandal in Chicago were a study found that the city systematically reduced the length of yellows only on lights that had red light cameras in order to harvest tickets.
I feel like the subtext of all these concerns is that you'd need to drive very carefully to reliably avoid camera tickets... and nobody wants to drive that carefully. I get it, I don't either, and I do get occasional camera tickets. But like: I should also be driving more carefully.
My memory may be outdated or only local to my jurisdiction but my understanding is that yellow means “do not enter the intersection” where “intersection” begins before the box, usually with some alternate street indicator, like broken white lines turning to solid, at a braking distance that accounts for posted speed limit and yellow light duration.
