I am going to try to help you. I realize the odds are what I am going to say if hype. Best I can tell you is it is not. And can easily be proven so.
It is not remotely possible, in time, money or lives, to use public shadow and safety driving, rely on deep learning and using gaming based simulation, to get to L4. The solution is to move most of the development and testing to DoD simulation technology which can produce a legitimate physics, not ray tracing, based all model type real-world digital twin. And yes I mean every possible complex and loaded scenario. For example — hundreds of exact model radars running faster than real-time in an actual city with actual objects, all in physics down to 5cm fidelity, with all direct and indirect interference modeled for every radar individually and cumulatively. And run with a full motion DIL with under 16msec delay. Not the debilitating 40msec or more everyone else does. And we make HD Maps by default and with point clouds with far more fidelity than any actual LiDAR can make. I know this sounds like hype. It’s not. DoD had to figure this out 25 years ago. They have significantly different architectures that permit this. And they know how to make extremely detailed sensor models. After all they deal with electronic warfare which is far more complex that what is needed here. Please take the shot at letting me show you.
Proposal for Successfully Creating an Autonomous Ground or Air Vehicle
Autonomous Vehicles Need to Have Accidents to Develop this Technology
Simulation can create a Complete Digital Twin of the Real World if DoD/Aerospace Technology is used
The Hype of Geofencing for Autonomous Vehicles