It certainly happens just when the car in front is slowing noticeably faster than the current flow of traffic, I'm pretty sure it is calculating closing distance regardless of acceleration. I know in my case having an FCW it was a car pulling in from the side, but i've seen it in the past where a car pulls a crazy ivan and slams on the brakes say deciding late to turn or pull into a spot, etc.. of COURSE I lift off and regen is usually enough for sure - so I'm CLEARLY at a safe enough distance for human intervention and physics to avoid a front end collision, but the warning has gone off NTL.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Safety Score
- Thread starter ScottNY
- Start date
HighZ
Member
AlanSubie4Life
Efficiency Obsessed Member
I racked up some electrifying Safety Score miles last night (ended up perfectly clean). Driving carefully around the neighborhood to avoid FCWs was way more terrifying than the storm; today was garbage day. 
.
.1) Get this app:
B) An imperfect braking score is most expediantly fixed with lots of "good" braking. So find a nice side street with lots of stop signs and little traffic. Then do lots of long braking all the way to a stop trying to keep your G-force at 0.2 - "good" braking being more than 0.1 and LESS than 0.3
Next) Fix your following score by getting on the highway (at least 50 mph) and hanging about 2.5 seconds behind a nice steady slow driver, like a semi. "Good" following is more than 1 second to 3 seconds of following distance. It sounds like you and I both usually follow much farther back. My first time on the highway I got a following score of 17%! But that was ALL from one car cutting in front of me at an exit. I just didn't get any "good" following to cushon the, like 2 seconds of "bad". If someone seems like they are going to cut me off I just turn on autopilot, cause that doesn't count"good", or "bad".
EVNow
Well-Known Member
Yesterday I took a slightly different route with quite a few small ups & downs - didn't get any dings. One thing I noticed was that on "low" regen while going downn the speed wasn't increasing. That tells me it is safe for me to attempt going to the low-lands. Just have to be careful at the end to brake gradually. Also used an accelerometer app to get a feel for the g-forces at play.I racked up some electrifying Safety Score miles last night (ended up perfectly clean). Driving carefully around the neighborhood to avoid FCWs was way more terrifying than the storm; today was garbage day.
View attachment 718132
Last edited:
As a solid 100 on one of my two Teslas I feel pity for all of the little people who have given up on achieving the Tesla Gold FSD Ring. Now if only a few of the other many thousand 100’s would drop out so I would have a chance to get the totally not ready for prime time software this weekend. After watching some of the not beloved YouTube egomaniacs on 10.1 I can only thank my lucky stars that I have an up to date will.
jebinc
Well-Known Member
mackgriggs
Member
1) Get this app:
B) An imperfect braking score is most expediantly fixed with lots of "good" braking. So find a nice side street with lots of stop signs and little traffic. Then do lots of long braking all the way to a stop trying to keep your G-force at 0.2 - "good" braking being more than 0.1 and LESS than 0.3
Next) Fix your following score by getting on the highway (at least 50 mph) and hanging about 2.5 seconds behind a nice steady slow driver, like a semi. "Good" following is more than 1 second to 3 seconds of following distance. It sounds like you and I both usually follow much farther back. My first time on the highway I got a following score of 17%! But that was ALL from one car cutting in front of me at an exit. I just didn't get any "good" following to cushon the, like 2 seconds of "bad". If someone seems like they are going to cut me off I just turn on autopilot, cause that doesn't count"good", or "bad".
Finally downloaded this app. Hope that my phone's sensors are reporting like the Tesla's. Learned that I don't have to take every turn as slow as I have been. To get to .4g on a turn...its a fast turn. With the way I have been handling turns, probably 0 of them have even registered .2g. Braking though.....probably 100% of my braking has hit .1g. On full regen....is gets pretty close to .3g. Close enough that I'll still keep feathering my braking. If you do full regen and add some brake pedal, you are probably going over .3g.
BitJam
Active Member
TL;DR: Tesla is automating their AI training process in order to leverage their large fleet. This will require a large number of beta testers.
They need vast amounts of high quality data and this is what beta testers provide. Tesla has a team of 1,000 people doing manual labeling for FSD. I imagine this team has been able to keep pace with the 2,000 beta testers. But Tesla has been working on and improving automatic labeling. This is an essential part of their path forward. When machines handle half the labeling, Tesla can double the beta pool. When machines handle 90% of the labeling then the pool can increase by a factor of ten. And so on.
The idea is that FSD will improve and become safer with more training. This will allow Tesla to safely increase the pool of testers which will further improve FSD safety and so on. Tesla plans to increase the beta pool by one or two orders of magnitude in order to speed up training by one or two orders of magnitude.
FWIW, here is an article on data labeling. It says:
In addition, Tesla is doing the labeling in vector (3D) space instead of image space. This is tricky because you need to transform image space (what is seen by the cameras) to 3D vector space before doing the labeling. The most useful training data will come from cars running the most recent FSD software.
That is not how it works. They have the ability to collect data from the fleet and sometimes use it to get data from certain situations (like ice falling off a truck in front of the car) but Tesla is not collecting all the video from the entire fleet. Even if they were it's nearly useless for training. Whittling it down to what is useful would be a massive computational task, looking for needles in a haystack the size of a planet.
They need vast amounts of high quality data and this is what beta testers provide. Tesla has a team of 1,000 people doing manual labeling for FSD. I imagine this team has been able to keep pace with the 2,000 beta testers. But Tesla has been working on and improving automatic labeling. This is an essential part of their path forward. When machines handle half the labeling, Tesla can double the beta pool. When machines handle 90% of the labeling then the pool can increase by a factor of ten. And so on.
The idea is that FSD will improve and become safer with more training. This will allow Tesla to safely increase the pool of testers which will further improve FSD safety and so on. Tesla plans to increase the beta pool by one or two orders of magnitude in order to speed up training by one or two orders of magnitude.
FWIW, here is an article on data labeling. It says:
In addition, Tesla is doing the labeling in vector (3D) space instead of image space. This is tricky because you need to transform image space (what is seen by the cameras) to 3D vector space before doing the labeling. The most useful training data will come from cars running the most recent FSD software.
BitJam
Active Member
As I understand it the stacks will be merged in 10.2; adding new beta testers is waiting on the release of 10.2 so you and Tesla are on the same page. As I recently said, Tesla wants to increase the size of the beta pool as they automate the labeling process allowing them to handle more testers. The better their automation, the larger the pool they can handle.
AlanSubie4Life
Efficiency Obsessed Member
What's different about the stream of data from the beta testers vs. the stream of data from a few thousand carefully selected cars from the fleet? If the difference is that that stream of data from the beta testers will have information on where interventions took place, and bug reports from users tied to video clips...how is that not just a good way to find out what FSD bugs (limitations due to labeling, path planning, etc.) to focus work on (and work on labeling in those specific situations, etc.)?
Now I know why my wife has been asking me if my life insurance policy is paid up for the last 2 weeks
EVNow
Well-Known Member
To get more varied cases, you need more testers. To get more variations of the same case, you need more testers.
And ofcourse, there is this inconvenient fact that the ceo had promised "feature complete" two years back.
EVNow
Well-Known Member
I wonder what is the higher risk in AZ. FSD Beta or all the anti-vaxxers / anti-maskers
AlanSubie4Life
Efficiency Obsessed Member
When I say "stream of data from the beta testers," I mean including the several thousand (or over 10k) new additional testers.
EVNow
Well-Known Member
I read your post wrong.
They need the FSD beta software in those cars to see how FSD would actually react and what the person thinks is the correct action. Simulation can go only so far ...
BTW, the funny thing about simulation is that Tesla has argued they have a huge upperhand over Waymo in the battle for FSD because they can get a lot more real miles data than Waymo that relies on simulation. Its high time they started using the fleet to test fsd beta.
ok now I'm curious - how many of us who are obsessed with the mechanics and gaming the system are also of the generation that grew up on video games? Not saying that non gamers wouldn't game the system, but I expect there to be a pretty solid correlation between video gamers and safety score gamers... The mentality of completionism and perfectionism applies so strongly to both camps.
EVNow
Well-Known Member
Yes… they need to grab the data when the driver hits the breaks and disengages it. Or when overriding the steering wheel. Speaking of which… I sort of wish the car would drop fully out of control when I correct its steering. Not anymore, but when I first got the car it surprised me a couple of times by speeding up after I had (I thought) taken control.
Similar threads
