I am a bit skeptical of using software to test 99.99999% of possible scenarios. Software tests are great for testing software, because you’re testing software in the environment that it will be deployed. Even videogames, however, are nowhere close to real life. The physics engines are good, but you can’t program an AI to drive like a human, only like an AI. So you aren’t necessarily getting a good idea of how AI cars would respond to human unpredictability.
I think we need a lot of self driving cars at a level 3.9 — fully autonomous, not in need of human takeover, but with the capability of human takeover. We need several years, at least, of humans teaching cars these corner cases. This could be a great, albeit temporary, paid position at companies like Uber.