Very nice overview.
Redundancy is necessary, which Tesla (Musk) has thrown out for simplicity and cost savings. Hopefully, Tesla will eventually realize their current approach (neural networks, code and only cameras for input) is not going to get them to level 4-5, and rethink their approach. Of course, the mapping s/w approach followed by some would work fine at Disney World, but is an idea that should have been aborted early on when you consider driving in vast America with near constant road work. The massive LiDAR sensors aren’t practical or economic unless they can miniaturize them while still retaining some usability. I do think radar would be really useful in rainstorms and fog conditions to avoid the massive vehicle crash looming ahead.
I own a Toyota with its Toyota Safety Sense (TSS) with a camera and radar, which is useful as adaptive cruise control on an interstate to keep a safe buffer distance to the car ahead. It’s also useful to remind you that you need to stop in case you are inattentive, & it applies the brakes. However, because its processing time is so slow it often reacts after the car ahead has turned off the road. It also suffers from phantom braking while on cruise if a car brakes in an adjacent lane. Still, it’s limited, but useful as long as the sensors are clean and one realizes its limitations.
Finally, I would like to add that I have ridden a road bike for close to 50,000mi for 8yrs, and I can’t wait for self-driving cars because humans are terrible drivers. However, I had become a lot better at seeing situations develop ahead of time and taking actions (getting off the road, or stopping at a green light as I watched a car go through a red light). Last year, I decided one day to switch to walking from road cycling because I was becoming a bit bored and the relentless risk of being run over by distracted, road-rage, sleepy, and basically incompetent drivers (passing in no passing corridors, not stopping for stoplights, etc.). I have no doubt that level 4/5 self-driving will be achieved in a few years. And today with simulations, it is straightforward how to test it under terrifying conditions to see if it can not just avoid deaths, but have the forethought, or would they call it a predictive processing algorithm (PPA)) to avoid the accident outright. When that day comes, cycling and resident deaths from cars will drop close to zero.