This is meant for self-driving cars, right? Because I can not see this happen with human traffic.
I was also waiting in the video to see busier traffic simulations, but didn't see any. I suspect the model of keeping distance then breaks down completely and creates lots of congestion, ifnot collisions (if the road is full and you need to evade an oncoming vehicle...). On the upside, you use the whole road dynamically instead of only a fixed few lanes, that is a nice improvement.
I think you under-estimate the numbers of cars on the road.
I don't remember the exact throughput used for the video, but I think it was around 5500. On a equivalent road the minimum distance between two cars is about 2 seconds. That gives a throughput of 1800 cars per road per lane (3600/2). For two lanes (resembling the one used in the video) this would give a absolute maximum throughput of 3600 roads per hour.
Which means that our system, with about 5000 cars per hour, give a much higher throughput than the maximal on a standard highway.
could you easily simulate human speed control, with gentle variance and effective repulsion between vehicles? or overload the road and see how it breaks down.
Yep, in our thesis we have tried with throughputs from 2000 to about 15000, and with different types of load. We assume that all vehicles are automatic, so no human speed control.
How feasible would it be to simulate non-automatic vehicles with the models you've provided? My hypothesis is that humans, too, could implement the proposed steering behaviors (albeit likely with some degree of imperfection) and thus could feasibly implement some or all of the proposed flocking strategy, but I wouldn't know where to begin with testing things like blind spots and slower reaction times.
At the very least, I reckon the "yield to higher-priority traffic", "maintain distance from other vehicles", "maintain distance from road boundaries", and "drive forward" behaviors to be (relatively) easy for human drivers to implement, seeing as they do these things already (albeit with lanes as visual reference points), and do so even better with modern driving assistance technologies like blind spot indicators. Dynamic adjustment to asymmetric load would be harder, but I suspect even this could be feasible.
I agree that humans would have a hard time with dealing with oncoming traffic. The rest of this, though, sounds hardly different from what human drivers already do; they're expected to maintain a safe distance from other cars, yield to emergency vehicles, drive in the correct direction, and not drive off the side of the road.
Currently all vehicles fit in the space of a lane. Unless your road is really large, the most efficient packing of cars really should be neatly aligning them alongside. Vehicles are also very stable in the direction across the road, so stimulating separation in the perpendicular direction seems unnecessary.
I mean, if we are using AI to drive cars might as well use the most efficient configuration (and not one that only interacts with local neighbors as is required for fish).
Dynamic lanes seem useful though (if a bit scary).
Separation in the perpendicular direction is required to allow room to maneuver left or right in the event of unexpected events (or even expected ones like oncoming traffic or emergency vehicles). This applies for both self-driving vehicles and human controlled ones.
Autonomous vehicles could coordinate to allow manuevering. I mean, if there is enough space, surely you can keep cars at whatever separation, but in heavy traffic the most efficient would be leaving a lane on the right or left exclusive for emergency (or maybe coordinating vehicles for clearance).
One of the original authors here. In our thesis we actually have cited a paper that used an analysis of the traffic from India for something resembeling what we have done.
"Tall, ugly buildings loomed above them to shoulder out the sky. The black streets were wide but very crowded, and the Moties drove like maniacs. Tiny vehicles passed each other in intricate curved paths with centimeters of clearance. The traffic was not quite silent. There was a steady low hum that might have been all the hundreds of motors sounding together, and sometimes a stream of high-pitched gibberish that might have been cursing.
Once the humans were able to stop wincing away from each potential collision…"
Not to be a naysayer, but birds and fish can move in three dimensions, which naturally makes collisions much rarer. There are more than two directions to turn to avoid a collision.
The ideas still applies in two dimensions, we just have one fewer dimension. (Actually, looking at Particle Swarm Optimization, ideas from flocking is used from 1 dimension, to many more than 3).
* The 72% increase in throughput was achieved with diverse asymmetric traffic.*
Does that mean "If you have most of the cars going one way it's more efficient to use more of the road for traffic moving in that direction"? It sounds like it does. And, well, duh.
I was also waiting in the video to see busier traffic simulations, but didn't see any. I suspect the model of keeping distance then breaks down completely and creates lots of congestion, ifnot collisions (if the road is full and you need to evade an oncoming vehicle...). On the upside, you use the whole road dynamically instead of only a fixed few lanes, that is a nice improvement.