This is awesome, I built a toy ML project which ran in a simulated traffic system as a way to explore ML. Like the author, if the individual traffic lights modulate to traffic density "waves" emerge in the simulation of vehicles going through intersections. And, as noted by the author, you only have to do this at the light, it doesn't need any other inputs except an ability to recognize cars coming toward it.

One of the surprising thing is its "disaster" response. Which is to say to totally wipe out like 6 intersections to through traffic. The lights with a way around begin cycling more rapidly, this makes the flows that the lights downstream "see" have more openings in the direction of the exit so they start cycling more vehicles that way, and if the simulation is set up so that vehicles begin to diverge to open roads if enough time has passed with them sitting at an intersection. All of the cars begin to flow around the obstruction.

I couldn't get my "car model" to do an illegal u-turn (as I've seen many cars do in real life when stuck at a traffic light where they could go back the other way) so there were always vehicles that remained stuck trying to get through to impenetrable divide but it was a nice example of emergent behavior from a complex system with automaton type rulesets.