Some thoughts on my FSD beta experiences....
I'm one of those who finds reasons to disengage about once per mile. There are a lot of reasons for that. One may be the amount of snow that we have here in Ohio. The roads are clear, but instead of curbs and edge lines, the car sees snow banks. I don't think the NNs have been well trained for this environment.
I think Tesla could learn a lot by recording 2-3 seconds of audio whenever someone uses the wheel or the brake to disengage. (Note that I don't really suggest this because of obvious privacy issues.) They would get comments from me like "What the hell was that for?" or "Slow down, damnit!" or "No, let's not drive into the snow bank".
The Insurance Institute recently posted some guidelines for self-driving cars that would allow an automaker to get its seal of approval (which most automakers want). I don't agree with all of them. For example, even though I drive NOA with lane change confirmation, I don't want that to be a fixed requirement. I think automatic lane changes will be very useful as the software improves - I'm just not ready to trust them yet.
But they also say that the car should let the driver steer when he/she wants to. Don't disengage, just let the driver take precedence over the software. I agree with that completely. I would like it very much if the car would steer for me when I let it, but would defer to me (without disengaging) whenever i took control. In fact, I'd take it one step further - let the car learn my preferences by monitoring my driving style. But we don't seem to be anywhere near the level of intelligence that it would require to accomplish that.
In fact, based on what I've seen so far, I suspect that the current hardware in most of our cars does not have enough computing power to properly handle FSD. FSD seems to look no more than a short distance ahead. It doesn't seem to be doing much planning. It accelerates and brakes somewhat aggressively even in chill mode. It drives quickly up to a line of stopped cars and then hits the brakes hard to stop. It doesn't seem to understand traffic patterns very well. Sometimes it wants to go around cars that are stopped at a red light. It changes lanes away from an upcoming turn even within the last half mile or quarter mile. In general, it is long on tactics and short on strategy.
I agree with those who say FSD may need better cameras to be able to discern the situation at a greater distance. But that also means analyzing a lot more data. And that means faster processors will probably be needed.
I think I understand the difference between those who disengage often and those who can go long distances without it. I recently had a trip where I drove over 100 miles without disengaging. About a third of it was FSD, the rest was NOA on a freeway. The non-freeway part of the drive was on two-lane country roads, mostly 55MPH, but passing through several small towns where there were traffic lights, lower speeds, and more traffic. It handled all of this very well. In fact, I only had one disengagement, and that was in an unusual situation. The car was on a 35MPH road that was marginal to the freeway. There was a fork in the road where the left side was a ramp to the freeway and the right was a 35MPH business district. The car correctly took the right fork, but it set the speed limit as if it were heading onto the freeway. Even when I dialed it down to 35, it stayed around 55 and showed no sign of slowing. I disengaged quickly because I know this area to be a frequent speed trap. But other than that, it took me all the way home.
Yet if I ask it to drive me a few miles to the local supermarket or Costco, it makes mistakes often enough that I need to disengage at almost every turn, and sometimes in between them.
There is one thing I've observed that I haven't seen anyone mention. Sometimes when I am driving on a two lane road in the suburbs, even ones with well-marked lanes, it will simply refuse to engage FSD when I double click the stalk. Prior to the click, it will be showing me the FSD visualization, but when I try to engage FSD it will drop back to the old-style visualization and put me in AP instead. I have not figured out why this happens. If I try it again on the same road a short time later, it may work correctly.
Another thing I noticed is how the car responds to cross-traffic approaching from the right. Whenever there is a car in a position to enter the road from the right, whether it is moving toward the intersection or stopped at the corner waiting to turn, FSD will slow down apparently out of caution. This would not be a bad tactic - except that it consistently does it about 50 feet PAST the intersection. It reacts too late to matter. This may be another example of simply not being able to analyze situations quickly enough.
I'm on 10.8.1 now, hoping to be in the first batch of 10.10s when they start rolling out to non-employees. 10.10 will probably make some small improvements. Maybe version 11 will take some big steps forward. Maybe it will take some big steps backward first. I don't think we will have a trustworthy FSD for quite a long time.
www.teslarati.com
/cloudfront-us-east-2.images.arcpublishing.com/reuters/UMYS3CKKAFOXLMJEAVLQ6IVUXU.jpg)
