Picked up
@Gigapress and headed for Volunteer Park. Roads that are super steep, rough, tight and not marked well. FSD handled these well with only one missed occluded stop sign (wanted to try this again, but we didn't, but it missed another occluded stop sign so this is an issue). Did some other steep/tight accent/decent lefts, rights and straights, which it handled somewhat ok, but some jerky steering wheel movements are quite unnerving when traffic is coming head-on (grr,
Kalman filters at slow speeds is the root cause of this due to high jerk rate being allowed for any inputs to path planner being passed on to controls).
Not able to handle Seattle's hardest turn
here, but no big deal as this turn is ridiculous and from my time on Google Maps, it is the most missed turn in Seattle.
Did a whole bunch of UPLs and it handled all of them, but was too hesitant and had to intervene about 30% of the time on average (about 5% of the time it was an unsafe hesitation). I believe the new network(s) that are controlling this will continue to get better (does NOT feel like they are anywhere close to a local maxima in performance).
Several times after a left or right, FSD was not able to get into the correct lane (one lane change left or right) when it needed to immediately (it is not either using any sparse map 'look ahead' road segmentation data or the network is sometimes not present or is just too sparse and is somehow lacking for that particular segment).
My two lane roundy's were not handled well and I had to intervene each time (a few unsafe moments) where it was very hesitant, it chose the wrong lane, changed lanes in the turn or simply was going to hit the target car. I had such high hopes, but I'll test this more tomorrow, hoping for some better results.
In conclusion: I'll be driving a bunch more tomorrow and hope to achieve more statistical data on the same turns