Last year, FSD had a low proba-closer-inspection-of-teslas-autopilot-safety-statistics-533eebe0869d)soon. Even more so after Tesla pulled it off the menu. Since then, some pieces of new information made me re-evaluate the topic. Individually, they have little relevance but for the first time, I see a path to autonomous driving with a fair chance of realization within a decade.
HW3
The upgrade of processing power by a factor of 10 enables implementation of more complex models with improved recognition quality at higher frame rates. More importantly, new features can be added without the need to first optimize performance of the existing SW to free up capacity. Recent announcements of upcoming features seem to depend more on rollout of the new HW than remaining SW implementation efforts.
AP + FSD
After moving some existing features into the FSD package, there is for the first time an actual proposition in this purchase. Until then, it was just the promise of future miracles to come and I suspect most orders were intended as a zero-interest loan to fund development. If Tesla can deliver according to the roadmap, the FSD package will actually be worth its money by the end of the year. Take rates will consequently improve and free upgrades for existing EAP owners after recent price cuts will add to the fleet of FSD - enabled cars.
Granted, Elon's interpretation will remain to be an assistance system for a while. It won't give me the ride home after a beer too many and it won't drop off the kids at school. Would I appreciate a car that is 99% of my commute hands - off while not meeting the mind - off requirements to meet L4 autonomy requirements? Definitely! Will it be worth an incremental 5k? Time will tell but I am confident that Tesla will get the pricing right in order to have take rates close to those of EAP.
Regulatory Approval
The SAE definitions of autonomy levels are a good definition of the requirements to the system but (to the extent of my knowledge) they don't specify how to prove that a system complies. Frankly, I have conveniently avoided the question as this all seemed to be pretty far out. By chance, I came across an article that discusses Tesla's EAP safety statistics. (
A Closer Inspection of Tesla’s Autopilot Safety Statistics).
My main takeaway was that it takes billions of miles to gather statistically relevant numbers to even tell whether the system is better than the human driver and even more to prove that the system is 10 times or 100 times less likely to cause an accident.
Not completely unlike the real - world test a human needs to pass for getting a driving license, an automated system will need to prove it statistically. This, however, means that only a significant large fleet can acquire the needed miles traveled. And in consequence, the Waymo fleet will likely not scale to reach a critical threshold.
Bottom line: Until recently, I gave a zero chance to L5 automated driving in the next 5 years and only slightly higher within a decade. Now for the first time, I can see how this could become feasible but only for OEMs with a critical mass of cars that constantly deliver evidence of correct behavior of the cars and Tesla is uniquely positioned if they manage to get away without expensive Lidar.
Edit: Completed the accidentally submitted post.