>>AutoPilot has improved safety dramatically and that's the important point.<<
Says who?
The data.
In the 1st quarter, we registered one accident for every 4.68 million miles driven in which drivers had Autopilot engaged. For those driving without Autopilot but with our active safety features, we registered one accident for every 1.99 million miles driven.
Think about that statistic for a second. Thats 2.3x more distance between accidents on average
purely because of incremental AutoPilot usage on top of the active safety features (AEB, Lane Departure etc)
That's up from 1.6x compared to the same period last year
In the 1st quarter, we registered one accident for every 2.87 million miles driven in which drivers had Autopilot engaged. For those driving without Autopilot but with our active safety features, we registered one accident for every 1.76 million miles driven.
but the real numbers to look at are the miles travelled from A to B WITHOUT any disconnection.
No it isn't. The purpose of AutoPilot is to reduce the probability of being in accident not to reduce probability of disengagements.
That being said, phantom breaking in itself will increase the probability of causing an accident so therefore it should be minimised. What nobody seems to be talking about here is that the OP was in a 2015 Model S. AutoPilot is
far better now (as indicated in the above stats).
If you assume overrides are only taken when there's a serious risk of an accident the numbers change dramatically.
There's always a serious risk of an accident. Relatively speaking you are always closet to death the times you are in a vehicle.
Disengagements also aren't a corollary of an accident. For example statistically speaking how do you know that the time you turned on AutoPilot was the time it saved your life? When it works correctly you completely forget the whole experience and only the "bad" experiences survive in your memory (hence the survivor bias in personal anecdotes).
You could have had the worst ever A to B experience with AutoPilot (constant disengagements, phantom braking, etc), and it
still be the time it saved your life compared to if you drove it completely manually but you can never know this without undoing entropy. Therefore the measure of disengagements is certainly important but not the one we should be focusing on as a guiding metric of AutoPilot's net impact.
It may be OK for level 3 and a stretch at level 4, but 5 is a whole different ball game and I cannot see it happening with the present vehicles, if ever.
I'm talking about AutoPilot in the context of it being a level 2 driver assistance feature. Level 5 will happen but we will meet it halfway (basically everyone will understand the limitations of the vehicle and we'll design our society around it -- much like we have with non autonomous cars just more in our favour) but that is a different topic altogether.
But I will also acknowledge that the radar watching a couple of cars in front (not in AP) might be responsible for saving me from a significant crash.
This by far is the biggest contributor to the improved distance between accidents when using AutoPilot. I've covered this before but TACC
is AutoPilot just without the AutoSteer (Beta). For example TACC will stop at traffic lights (if you have the setting enabled) even with AutoSteer disabled and have the same propensity to phantom brake as when you have TACC+AutoSteer running.