Can someone speak to Tesla's approach of collecting real-world data, and Google's approach of "simulating" roads and conditions and running self-driving models on that (so technically their vehicles drive millions of miles on simulated roads).
Intuitively Tesla's approach makes more sense, but would love to hear someone with domain knowledge on how much of a difference it can actually make (after all, you need quality training data and Tesla may now have to navigate through significant more noise).
Not an expert, but what you describe is testing different parts of the system. Real world data tests that the sensor system is creating an accurate model of the environment, whereas simulated data ensures that the vehicle makes correct decisions for any given input model.
e.g. real world data tests that the vehicle can detect a stop sign. Simulated data tests that the vehicle stops at a stop sign.
Intuitively Tesla's approach makes more sense, but would love to hear someone with domain knowledge on how much of a difference it can actually make (after all, you need quality training data and Tesla may now have to navigate through significant more noise).