Gave NavOnAP some more behind-the-wheel time since my initial post. After trying it for ~2000 miles now I've finally decided............ to just disable it completely in the car's settings.
The last straw was last night when it almost got me and my passengers into a serious accident, despite hands-on-the-wheel and full attentiveness.
NavOnAP suggested a lane change to the right for an exit that was about a mile ahead. As what seems like par for the course, its suggestion was to change lanes directly into another car that it was fully aware of, so waited a moment to pass that car, then confirmed it.
Car started edging to the right, got 3/4ths of the way into the right lane.
About that point, an impatient car behind me that wasn't happy enough with my +14 MPH vs the speed limit driving started to approach to pass us. The Tesla was still moving right slowly, and the tailing vehicle hadn't really been tailgating or anything. They just saw us moving and started to speed up. As the car got to the point where the left tire was on the center line, the car approaching from the rear was nearly in line with my rear bumper.
Then, out of no where, the car pops up "Lane change cancelled" and jerks the wheel
hard to the left and gets back into the left lane. Was abrupt enough to wake my passengers. We also nearly collided with the rapidly approaching car behind us, to the point where they did a full panic brake to avoid me... quite narrowly it seems. A few more miliseconds of reaction time on their part and I'd likely be attaching an accident report to this post.
The most worrisome part is that it took my hand that was on the wheel with it as it jerked into the left lane, and didn't disengage based on my resistance until
after the maneuver, which only took a split second. (Technical side: My educated guess as to why this was the case is that whatever algo in AP2 determines user applied steering wheel torque didn't correctly cancel out the commanded steering's effect on torque, and thus was unable to properly register user applied torque during the abrupt maneuver and didn't disengage quickly as commanded. Likely due to some kind of internal sample-based or time-based smoothing of the values.)
I'm not one to try and blame anything on AP. Ultimately, the driver is responsible for the car's actions. If there was an accident here, I'd hope the person to the rear with their somewhat unsafe passing maneuver would end up being the at fault party... but honestly, there was no reason for the car to jump back to the left lane. Not for AP, not for a human. It's something that wouldn't happen, and I can't blame the driver behind for assuming the lane change would be quickly be completed. I couldn't blame any reasonable driver for believing that a car with their turn signal on, who's 90% completed a lane change, would just decided to abort. The problem here is that I just didn't have many options for a reaction to the car's action. I couldn't brake, since this would get me rear-ended. No way to reach for the gear stalk to disengage AP that quickly. More importantly, AP didn't disengage like it should during the maneuver based on human applied wheel torque and instead took my hand along for the maneuver.
In this case... this is the fault of AP. My hands were
on the wheel. AP should have disengaged instead of being able to turn sharply left at 79 MPH. (Beating a dead horse here probably, but this issue is completely unique to AP2. Based on first hand testing and experience, the same thing is impossible with AP1 as there is a safety decision to abort automated steering control based on torque on the wheel that is handled by the EPAS itself when human applied torque exceeds a threshold... doesn't appear to be the case with the newer EPAS setups since the Model 3, most AP2 S/X, and all AP2.5 S/X that have EPAS2 or 3.)
This is the first time I'm seriously considering reporting Tesla to someone like the NHTSA.
I've submitted a full accounting of this to Tesla as well along with date/time and such for log check purposes. Will see what they say before I do so.
Other notes:
Some people noted that the car suggesting lane changes directly into a vehicle it knows is there is intended behavior, since the human should be verifying before imitating the change. However, more testing suggests that this is just the result of poor coding, as the car will avoid suggesting lane changes into vehicles depending on which side camera is detecting the car. So this is definitely a bug and not a feature. Probably a huge part of the reason they pulled autonomous lane changes.
The car can NOT detect vehicles consistently to the sides and behind at night. Not at ALL. Especially motorcycles or cars with somewhat dim headlights. There is no way these cameras are going to work for night-time FSD without rear/side/corner radars... not on roads without overhead lighting (like most around me). This is a huge problem that I hadn't really considered the full ramifications of. An extreme example is that I was passing a truck with busted trailer lights (flickering off and on)... and the car completely lost it as soon as it was out of view of the front cam/radar.
Overall I continue to be disappointed with AP2.x.