This is a cool idea. It seems like currently only the major players have troves of real data to feed their ML algorithms. Google has clocked how many real hours of driving for their cars? But if an upstart wants to bring a better driving AI/algorithm to the market they won't be able to because they don't have a Google-scale fleet with unlimited cash to burn on driving hours. In those cases simulation becomes a really good substitute. Driving is one example but I can imagine other cases where simulating the real word as data for an ML model is just as good as clocking the real world hours.