Here are the screenshots of an area I described before but without pictures. The first photo has a circle around the problem area and the second photo is zoomed in with path markers. When traveling north-east on highway 66 in the left lane, the car behaved perfectly until reaching the spot with the circle - at which it would immediately dive hard left toward oncoming lanes where the center divider disappeared. I have no idea why it never had problems with the three similar spots directly preceding it. However I drive each way on this road every day 2-5 times, sometimes more. So when I read about the video uploads I decided to see if I could help "train" the system. I let autopilot drive until the absolute last moment when I could "save myself" and then break away with the steering wheel. I am the only Tesla in the area AFAIK - have not spotted any others on this road in the 8 months I've been driving it. This "training" started when the car was shipped to me in, I believe late April/early May - it was on 17.17.4 8.1 when I got it and it was "misbehaving. 17.26.76 resulted in what I thought was a fix at first but later realized was not 100% - now the car would not go hard left at the divider but just go slightly jogged, but remain in lane - but unfortunately would not go right far enough when the median reappeared - so I had to break off every time to avoid what I thought would be a scraping of the concrete with my wheels - but I continued to use AP, stay in that troublesome lane - and break off each time at the last possible moment. Finally 17.28.c628869 was released a couple weeks ago and BAM - problem now completely gone - if anything the car actually jogs right maybe 2 inches when the median disappears - like it was spooked or something by all my "breaking away" at that spot. I've driven it dozens of times on 17.28.c628869 and the behavior is perfect each time.
Proof that video uploads are being used for reinforcement learning? No - but I never saw learning like this of any particular spot on AP1. Hopefully a good sign of things to come...
Proof that video uploads are being used for reinforcement learning? No - but I never saw learning like this of any particular spot on AP1. Hopefully a good sign of things to come...