Computer vision (a.k.a. perception) is just one part of the problem. There’s also behaviour prediction and path planning/driving policy, which also can be machine learned and which also therefore improve with a larger and better training dataset.
Anthony Levandowski, formerly a top engineer at Waymo, believes that behaviour prediction is the key problem limiting progress on autonomous driving.
You can’t learn behaviour prediction in simulation because if you don’t have a model that predicts behaviour, you don’t have a model to simulate it. If you do have a model to simulate it, you can use that model to predict it. It’s a chicken and egg problem. You have to learn behaviour prediction from real world behaviour.
Having 500,000 vehicles on the road with Hardware 2/Hardware 3 allows Tesla to do large-scale collection and automatic labelling of training data for learning behaviour prediction.