A quick update on that sharp 90 degree curve. I tried again today and failed a few times. My car tried to go through the curve at 25~30 mph, which is too fast. Now I guess the first day it dropped speed to 15 mph because there was a car from the opposite direction right when I entered the curve, on that day my car slowed down to 15 to avoid a potential collision.
Edit: I think 2019.8.5 is mainly about Navigate on Autopilot on highway without conformation. So this kind of sharp curve probably is not in this release. I'm just curious when it will be able to handle the curve.
anecdotally, there's an in-town curve that autosteer couldn't handle. To be fair, most traffic crosses over the line in it, even though the road is plenty wide enough.
However, the version
before 8.5 was able to reliably negotiate it, staying well on the correct side of the line. I was quite pleased. Then I got the latest update and it can no longer handle it reliably. Going the direction I normally do it crowds the line (like most human drivers) and may cross over it (hard to tell for sure just how far it goes), but yesterday in the opposite direction it went too fast and wide (towards oncoming traffic) before bailing. Despite my recovery of the situation my wife was not amused.
This is exactly the sort of problem that occurs with a probabilistic system and why I thought that Tesla was using conventional programming for the actual driving and using the neural net for it is good for (e.g., vision). In traditional programming this would be termed a regression bug, but in a probabilistic system it is a feature. I would expect even if Tesla had all the data they could not explain why the prior version handled the curve fine while the current handles it poorly or fails.
Despite my disappointment, having watched the video of Karpathy talk about autopilot I'm accepting that this is the best path forward and that, given that parts of AP hinge on probabilistic programming, it doesn't realistically gain that much determinative behavior by trying to segregate tasks. He didn't go into details, but indicated that as time went on the neural net gradually assumed more responsibility for the AP process.
So when you suggest that "...is mainly about Navigate on Autopilot on highway
without conformation," while that may be true for intent or even in fact, if they update the neural net portions of AP at all it can have unintended consequences. This is going to be true for any neural net-like system, but when you have something with as complex of a task as driving (it may only have two outputs: steering and speed, but the environment and interactions are quite complex) you just can't be sure how it will behave in any given circumstance until it has been demonstrated.
Put another way, Karpathy makes a point that -- unlike traditional programming with variable loads -- using the neural net has fixed costs in terms of memory and processing power for a set accuracy target. What he didn't explicitly say is that you can't predict before hand when/where/how those misses will occur or with what impact, just the probability.