What I think they should do (what I would do if I were in charge at a AI-car co.): There's a whole ersatz city out in the desert somewhere the whole purpose of which is to allow in situ modelling of new "smart" hardware. (This is a real thing, but I forgot what it's called and I'm too lazy to look it up, I apologize for my barbarism.) The people that "live" there are actually paid employees. That's where you test your self-driving cars.
Further, I would have started by making self-driving golf carts made out of nerf that can't go faster than, say, two miles per hour, and then iterated. It's reckless to immediately attempt to make Knight Industries Two Thousand, in my opinion. The potential for mayhem and death goes up with the kinetic energy, eh? Both speed and mass contribute to the "killer robot" aspect of these machines. Start small and light.
As far as I'm aware, all L4 AV companies have done years of closed course testing. You can (not) see these vehicles being tested in places like GoMentum station in Concord, Altamont Raceway in Livermore, and TRC in Merced, not to mention other courses in Vegas, Seattle, Tahoe, AZ, Florida, Michigan, and China.
Of course, companies also started with small, slow vehicles like the Waymo firefly and the Nuro R vehicles. Voyage (acquired by Cruise) was doing their testing in a low speed access controlled retirement community, with essentially golf carts.
Probably because "training" implies direct and constant supervision, which would almost completely avoid any chance of incidents because the human is in charge at all times.
Deploying an unsupervised robot on the general public for training purposes is even worse than for testing purposes.
