I am on the beta release and for the most part am amazed at what it does. However, what I would have thought super easy seems to be kryptonite for the FSD software: know which lane it should be in. This is, no exaggeration, the reason I have to take over 90% of the time. Examples: 1) I am in the right lane coming up on a right turn in 100 feet. Suddenly the Tesla thinks it needs to get in the left lane which would cause it to miss the turn. 2) Similarly, I have an upcoming left turn on a three lane road and the car wants to stay in the right lane until I practically arrive at the intersection and then try to make three lane shifts to get into the left turn lane. 3) Some times I am cruising along on the freeway and it will occasionally turn on the blinker and start a lane change and then change its mind. I could go on but basically more of the same. Thing is it isn't random but very repeatable - at least examples 1 and 2 are and occur every time I take a particular route. Some times, if I initiate a lane change to get in the proper lane it will stay there but some times it tries to go right back to the wrong lane. Seems like it should be pretty simple to tell the computer to get in the correct lane for an upcoming turn *and stay there* barring a disabled vehicle. I searched and have not found anyone else mention this so is it just me or have others observed the same?
Most of them, other than the freeway lane change, are all evidence of what I've been saying for months. AP + FSDb needs better routing maps and to use them. Its attempting to use AI/ML models for everything, but ones which have to figure out everything from a second of visual input at a time and no memory.
Every other autodriving company will use maps heavily because they know all relevant information isn't in visual input, and even that information which is there humans can better extract it, humans understand traffic and road design better, and humans usually drive confidently on routes they've used before and know intuitively the best lanes/routes to be in. Humans have maps in their brain.
The use of maps would be to measure what humans actually do to drive routes to go to the same place that you're going to. Tesla only needs to instrument the human driven fleet. It would know that 98% of people who needed to make that left turn would be in the left most lane by a certain distance. The maps would give a 'target route' to take which the autodriver would attempt to follow, but of course prioritizing safety above all.
Of course it's necessary for the AI to work reasonably well when no maps are available, or they're obsolete, but that's a different issue. 99% of the time, they would cover the routes 99% of humans want to go to.
I suspect the lack of maps is an ideological opposition from the CEO given what he has tweeted about them.
The lane change being aborted on the freeway is most likely because it detected a car behind you which looked like it might be moving into the space or accelerating into it. The code is very conservative, overly so, and aborts too easily in my opinion. The Navigate On Autopilot (which does this, not FSDb) doesn't understand the concept of changing lanes while accelerating out of the way. It is either change lanes at exact speed, or accelerate in lane, not simultaneously, which is what humans would often do.