Limit on 90% of yesterday's drive was either 50, 55, or 60mph, and my default is limit +5. We have very windy twisty roads following the river and up mountain canyons, some well marked with the yellow advisory speed signs (lowest on yesterday's route is 45 IIRC), but most not. Yes, for huge cuts in speed from 60 to advisory of 35, I will manually run the TACC max down early for smoothness purposes, and because I do not trust TACC to look far enough ahead to perceive that deep of a cut is required and get there safely and smoothly.
Again my issue is not _that_ TACC slows for curves, but _how_ it does it. Too late, too deep, too inconsistently, and the nauseating and dangerous hunting up and down in the speed that it started doing sometime around 2023.30.6.
The pre-slowing is based on pre-mapped checkpoints (as per manual),
Perhaps that's how it used to work, but there's a lot of things about its current behavior that make me not believe that's still true. First and foremost, TACC will slow for the outside lanes of curve and _not_ slow for the exact same curve going the opposite direction, now on the inside lanes. The slowing for the outside lanes was again, too late, too deep, and super waffling in its decision of what speed to cut down to. The inside lanes it probably should have come down 2 or 3 mph, but did not come down by any from TACC max speed at all.
Second, if your / the manual's assertion that slowing is based on map data is true, why is it hunting up and down in the speed it wants to go? If this curve is map data designated as a "48 mph curve", why does it dive in at 53, stutter (hard on accel then hard on regen and repeat) down to 51, then stutter back up to 55, stutter back down to 49, and repeat?
Third and most damning piece of evidence for that theory is you can run the same curve in the same direction and get completely different results. -8 reduction one time, -2 the next, no slowing at all the next time you run the same curve.
I strongly believe the hard coding changed to an algorithm or machine learning subsystem for TACC to dynamically read the degree of bend in the curve sometime around or just before 2023.30.6 (and someone forgot to update the manual), and that algorithm/ML needs some serious tuning and especially smoothing work still.
I don't think AP will ever handle all slowing for curves by itself
Then how does the Full Self Driving system, that will supposedly one day be running robotaxis do things? Is it just going to run off the road if the map curve checkpoint data is wrong?
Yes, I know I'm on "basic autopilot", "basic TACC", etc and the stacks weren't the same in the past, but it's supposed to be moving towards merged stack for FSD, Auto steer, and TACC, right?
But let's say your assertion is correct, that it's going to get it wrong some percentage of time due to map data error, perception error, whatever. At that point I honestly would rather have an old dumb cruise control that I knew I had to intervene with 100% of the time, versus the current state of affairs where it works 90% of the time, doesn't do any slowing when it probably should 5% of the time, and makes me want to puke 5% of the time.
I imagine there is still a open split probably (like there isn't a separate lane line yet)? In general, AP is slow to speed back up after slowing for something
Nope, lanes were completely split with full solid line for the new lane many yards behind me, lead car had completely moved into the new lane quite a few seconds ago and was deliberately but gently slowing. And so was TACC. It was exactly the behavior I would expect if the lead car was still ahead of me in my lane, trying to hold following distance and everything. Except the lead car wasn't in my lane anymore. At all. And therefore should not be guiding the adaptive part of adaptive cruise control anymore, because it was no longer my lead car, but the system just never figured out to disregard it. There was nobody behind me so I just watched to see what it would do, down to maybe max -15 a across the next 10 seconds or so.