There are a couple significant areas where you and the industry, including Tesla, get this wrong. Real-world viability and proper simulation.
Yes real-world data is crucial. However it is simply impossible to gather a small fraction of what is needed. Tesla’s 10B miles is almost insignificant. RAND estimates 500B miles and Toyota one trillion miles. It is simply impossible to spend the time and money to get most of what is needed DRIVING in the real-world. Example-LiDAR and cameras have issues with object degradation, patterns and finishes. That involves clothing. You are far better off scraping apparel catalogs and creating combinations, including lighting and environmental conditions, in simulation that driving around stumbling on the same combinations.
Same goes for “long tails” like the one Tesla showed at its industry day. OK you found that one. Now what about repeating it thousands of times or finding variations and repeating those?
Every AV maker and simulation company I am aware of in this industry uses simulation based on gaming architectures. Those systems have debilitating real-time and model fidelity issues as well as problems scaling or handling loads as math models and physics get more precise and there are more of them. Basically too much isn’t real enough. That will cause planning issues, false confidence and real-world tragedies. Aerospace/DoD simulation technology does not have these issues. Before someone says air travel is not that complex etc let me interject with DoD simulated urban war games. Those events have to deal with the same complexity the Av world has to deal with plus they can drive off the roads and shoot at each other. DoD simulation Technology can have far more precise models and far more of them including environmental physics than gaming by far.
This problem exists because the gaming industry chose to make architecture issues long ago that made sense for OS interoperability and games but not for the complexity and scale of what is needed here. No matter how many CPUs you throw at it.
The ONLY way to get close to L4 is to do 99.9% of the engineering and testing using aerospace/DoD simulation that is informed and validated by the real-world. (Which includes test tracks and various other real-world sources. Not just driving around.)