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.
All the details is in the thesis (link in the url, or https://highwayflocking.github.io/Flocking_for_Road_Traffic_...).
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.