Absolutely they do .. they are the main reason they HAVE the beta program.If you mean hitting the button to save the report, then yes. Does Tesla actually look at those? I've done a few dozen of them.
You can install our site as a web app on your iOS device by utilizing the Add to Home Screen feature in Safari. Please see this thread for more details on this.
Note: This feature may not be available in some browsers.
Absolutely they do .. they are the main reason they HAVE the beta program.If you mean hitting the button to save the report, then yes. Does Tesla actually look at those? I've done a few dozen of them.
Are you reporting these things using the FSD report button? That's how to get them fixed!Sure is interesting to see one driver's experience across versions, but 0.04 disengagements per mile is VERY different from my personal experience. I have FSD engaged for > 95% of my 30-minute all-non-highway commute every weekday, and I love using it, but I think my disengagement rate is roughly 1 disengagement/mile, and it's been pretty constant through the last several versions aside from a big increase on 10.8/10.8.1, which couldn't complete a right turn without crossing the yellow line and also wanted to cross solid lines to drive on the shoulder approaching a right turn (I haven't had the chance to drive 10.9 much yet to verify if these are fixed).
I think over 90% of my disengagements fall into one of these categories:
- Mis-classifies mailbox as a pedestrian and wanted to cross double yellow line to give the "pedestrian" appropriate space (usually it's those thick plastic mailboxes like this one)
- Tries to cross double yellow line to go around queue of cars in front waiting at a stop sign or traffic light that is out of sight (around a corner or over a hill)
- Poor lane selection (got into a turn lane when going straight, changed into a lane that ends in a few hundred feet, changed lanes away from an upcoming turn within 0.5 mile of the turn or sometimes within a few hundred feet of the turn)
- Doesn't stay right on roads with no yellow center line, especially at stop signs where it would prevent other cars from turning onto the road I'm leaving.
- Approaches sharp turns wayyy too fast (45 mph speed limit technically, but yellow warning signs indicate 15 mph is appropriate for this curve)
So I suppose if the roads in the area of that other driver don't have that big plastic mailbox type, or roads with stop signs / traffic lights can be out of sight when queueing, or unmarked roads, or medium speed roads with some extra sharp turns, I could believe 0.04 disengagements/mile, but it sure doesn't work that way where I drive.
How much did you get paid for risking your car, neck and family? So much for a relaxing week end drive to the mountains. But you have the honor of knowing you have served the Elon Land. I know, it's for the greater good.I just used FSD beta from my house to Big Bear Mountain Ski Resort and back. I was very surprised it handle it well. It did 2 major mistakes, 2 minors, and 2 slow decisions. Overall, I was very impressed.
2 Major mistakes were caused by pickup trucks not following traffic laws and drove like idiots. One was in the left turning lane to skip traffic and cut me off as he bullied his way into my lane which was going straight. As a human, I saw that one coming, but FSD reacted late and swerved to the right and slammed on the brakes. Another major ones, I was going straight and a pickup truck didn't stop at a stop sign and forced his way into my lane to make a left. As a human, I would have already slowed down, but FSD swerved to on-coming traffic, and I took over right before the truck hit me.
2 Minor ones where the car crossed double yellow doing a couple of tight turns where we were all going slower than the speed limit.
2 slow decisions were passing big rigs doing 20 mph, and I was leapped frogged by 20+ cars. I had to manually take over to pass the trucks.
Absolutely, and sending emails for most of the non-obvious ones, although I'm never sure exactly which of my many snapshots make it to the end of the day to upload over wifiAre you reporting these things using the FSD report button? That's how to get them fixed!
What is your source for this info?They have 10,000 labelers (not on contract, but apparently "internal"). They could potentially handle 300k clips a day - 30 per person. If each takes 5 minutes, still about 3 hours of the day. If labeling these are a large part of their job, they can do it ... but if its a small part (i.e. Tesla wants them to label mostly some other images/clips, it won't be possible).
I misremembered - it is actually 1,000 labelers.What is your source for this info?
More than 2 weeks after it began to roll out, it stopped (at ~40% of testers) and since he/they've been uncharacteristically silent.I'm still waiting for 10.9 to drop into my download queue, but the rollout is still at a dribble. There is still no feedback from Tesla as to why it has been limited to a smaller number of users.
I misremembered - it is actually 1,000 labelers.
I think that makes it 300 clips a day per labeler. AT 5 minutes, clearly not possible. They need some way to reduce automatically label those and for labelers to look at only critical ones.
I’m wondering if they use the auto labeler to label all videos and the 1k humans are used only to review auto-labeled clips and verify that they‘ve been labeled correctly as a kind of quality control?The Auto-labeler - which they use. Dojo will allow for higher efficiencies of the auto-labeler. Some labeling is done by humans and some are handled by an algorithm.
How much did you get paid for risking your car, neck and family? So much for a relaxing week end drive to the mountains. But you have the honor of knowing you have served the Elon Land. I know, it's for the greater good.
[conjecture]I bet they only check a few at random, like 1% just for quality control. Probably the Auto Labeling checks for common events like Right Turn on Red at a No Turn on Red intersection (bet they get 1000's of this event each day and easily labeled). It probably flags events that are unusual/edge, violent or that it can't identify and theses are sent to the humans to look over. Likely the identified clips (maybe <>95%) are sorted and aggregated into a "bucket" of similar events (like the Right at a No Turn) then analyzed as a group for improvements or changes that can be made.I’m wondering if they use the auto labeler to label all videos and the 1k humans are used only to review auto-labeled clips and verify that they‘ve been labeled correctly as a kind of quality control?
Or more likely the next release will be V11 with the FSD single stack. Total silence from Elon on anything 10.x.Well, either something is broken that they are focusing on, or Tesla is taking a 2 week team building exercise!
TL;DR: disengagements per mile go from 0.22 to 0.04.
What's the math to covert that into miles per disengagements? Seems like an easier number to understand.
If it's 0.25 that's 4 miles per, if it's 0.5 that's 2 miles per, so 0.05 would be 20 miles per
So went from more than 4 miles per to more than 20 miles per
10.0 blew his shoes off. 10.8 blew his socks off. Perhaps he is worried about more clothing items getting blown off by the next release.Or more likely the next release will be V11 with the FSD single stack. Total silence from Elon on anything 10.x.