EVNow
Well-Known Member
Larger data set - crowd sourced. 1000s of miles on each version, almost 100 testers.
Last edited:
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Enginerd
Member
This is awesome! How can I learn more about this effort? I attempted a lesser version of something like this a while back on auto lane change, but with a sample size of 1.Larger data set - crowd sourced. 1000s of miles on each version, almost 100 testers.
View attachment 808989
AlanSubie4Life
Efficiency Obsessed Member
Yeah I'm curious how this data is actually gathered. Is it recall, or telemetry combined with user statements about driving style?
Last edited:
I'm sure Chris would agree that you need to be careful about reading too much into this. There are a huge number of variables outside of the different FSD version .. traffic, weather, even Chris nervousness when supervising FSD. To eliminate these you would have to do many test runs with each version on different days and times, a major project.
clydeiii
Member
Not sure that follows...perhaps he was driving those earlier drives with little to no traffic, for instance (due to time of day or COVID). Or he did a totally different route.
EVNow
Well-Known Member
See here. Anyone can register and submit data. They make sure you actually have FSD Beta (that also means you have to give them your VIN ...).
Also, the youtube video on the latest report, with lot of links in the description.
A lot of his earlier drives were testing unmarked dirt roads. Seasonality also makes a difference. I remember gravitating toward his videos because he did a lot of snow on dirt road testing with v8 and v9.
But yeah, unless he's controlling his variables, like route, season, time of day, and a bunch of other stuff, I don't think his intervention/disengagement tracker is very useful. But if he's consistent about testing all sorts of stuff, then longer term, a valid trend may appear.
AlanSubie4Life
Efficiency Obsessed Member
See here. Anyone can register and submit data. They make sure you actually have FSD Beta (that also means you have to give them your VIN ...).
Also, the youtube video on the latest report, with lot of links in the description.
I should sign up! Gotta get that disengagement rate up, haha. It does seem like it would be highly susceptible to selection bias though, and recall could be a problem (even though they provide the counter webpage).
EVNow
Well-Known Member
Its just "crowd source" data. Better than a single individual .... but more like the polls we run here on TMC. Obviously not a "statistically valid random sample".
Chuck Cook attempts unprotected left turns on 10.12.1. LOL.
Last edited:
AlanSubie4Life
Efficiency Obsessed Member
Tesla really took full advantage of the data from the test car they sent out to optimize this turn. Amazing! Not bad!
I'd turn right too. Very scary. The neural nets know best.
Last edited:
EVNow
Well-Known Member
Yes - the correct way to handle that road would be to turn right and have some U-turn opportunities at the next traffic light.
I've definitely had lots of occasions where I bail out on ULT and turn right because of heavy traffic.
Navigation needs to improve so that the car goes to an intersection with traffic lights using back roads.
AlanSubie4Life
Efficiency Obsessed Member
Haha, I was kidding. If it's busy enough, certainly a right would be advisable and often faster, but in a Tesla, there were plenty of opportunities. 0-60 in 3.1 seconds, man. If you're not using it, you're not living. But in this case it was possible even in a Corolla.
Like this opportunity...:
The other fix you propose (and support for u-turns) are just Band-Aids for a system with incomplete function (at the current time).
diplomat33
Average guy who loves autonomous vehicles
Yeah, there were lots of opportunities to safely do the unprotected left turn. FSD Beta should have been able to do it. The fact that it bails is because it is not good enough yet.
Release Note Engineering Description:
Upgraded decision-making framework for unprotected left turns with better modeling of objects’ response to ego’s actions by adding more features that shape the go/no-go decision. This increases robustness to noisy measurements while being more sticky to decisions within a safety margin. The framework also leverages median safe regions when necessary to maneuver across large turns and accelerating harder through maneuvers when required to safely exit the intersection.
English: Bail to the Right…LOL
Agree but I find it interesting that the most dangerous turn I take every day is a right hand turn since the B-Pillar camera cannot see properly and wants to launch into the intersection when cars are 90 feet away. I would have been hit many times by now if I didn't disengage.
jabloomf1230
Minister of Silly Walks
There are a number of possible responses to Chuck's most recent ULT video. First, Tesla may have changed the logic for such turns (poor visibility, high speed cross traffic, tee intersection with median). If it encounters such an intersection, it attempts to make the left turn, but if it can't safely do so quickly, it bales to the right instead.
The second possibility is that this 2022.12.x version of FSD beta is just buggy. The red steering wheel issue on unlined roads has been reported by at least one other tester. What also worries me is that people without FSD beta but on 2022.12.x are reporting AP components being randomly unavailable. There are two active threads here describing this issue, along with second hand information stating that Tesla says that it's a bug that will be fixed in a future version.
I hope that it's either the first possibility or maybe a third.
The second possibility is that this 2022.12.x version of FSD beta is just buggy. The red steering wheel issue on unlined roads has been reported by at least one other tester. What also worries me is that people without FSD beta but on 2022.12.x are reporting AP components being randomly unavailable. There are two active threads here describing this issue, along with second hand information stating that Tesla says that it's a bug that will be fixed in a future version.
I hope that it's either the first possibility or maybe a third.
Iain
Member
people are downloading 10.12.2 right now. It won't be long before we see if there are improvements.
Yea, if Tesla sent an engineer from Calf to Fla to test and gather data for an improvement then PLAESE don't send any to ATL since Beta and I have enough issues without those "improvements".
EDIT: I believe I have watched all his "left turn" and even the videos that evolved into to the left turn videos and I think this is the first 100% failure rate. What a terrible job 10.12.1 Beta did. Several EASY opportunities to make the left and it was a complete bail out to an infinite loop. Take that back since it couldn't even infinite loop with that STRANG repeatable Take Over Immediately bug.
Last edited:
Similar threads
