What you are thinking of is Swarm Robotics. And you misrepresent the field quite insultingly: it was far from a case of plonking down some simple robots and noting "wow, they do all sorts of things we didn't expect!". The entire
point was to confirm that the emergent behaviour that had previous been simulated with virtual swarm agents, and prior to that theorised as the cause of insect behaviour, was possible to
replicate, and to codify if the physical medium of interaction added any notable additional factors to the swarm behaviour (it does mainly due to computational imitations of the simulation, and computational robotics is now a big and mostly biomimetic field, e.g. using mobile driven whiskers for impact sensing).
What does this have to do with autonomous vehicles? For a start, while autonomous cars are in the minority they will be interacting almost entirely with human drivers. This is the WORST CASE scenario for autonomous vehicles. Once you have a large proportion of vehicles being autonomous, you can begin to have communication between vehicles, and produce behaviour like convoying that is impossible for human-operated or human/autonomous mixes to perform. There are plenty of simulations of large numbers of autonomous vehicles around, mostly to see how to optimise behaviour to produce superior traffic flow; for example removing junction signalling and allowing autonomous vehicles to freely and continuously merge and cross each other is far more efficient than turn-by-turn streamed releases. Papers and reports on this sort of modelling abound.
Here's one that turned up after about 5 seconds of googling, with
a few hundred thousand of its cohorts available.