It seems the creep line width is proportional to the speed of the road. I will check more on this.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD v12.x (end to end AI)
- Thread starter Buckminster
- Start date
-
- Tags
- #fixphantombraking autopilot chill is not chill elon is an ass fix fsd extreme late deceleration fsd does not commit to go fsd does not maintain proper speed fsd hard and late braking fsd jerky driving fsd late deceleration fsd stops too early too quickly fsd stops too far behind line fsd tailgating in city streets
GlmnAlyAirCar
Active Member
If you're on Auto Speed, turn that off.
I think that the car always chooses its speed based on conditions, but having an upper limit on the speed means that once it reaches that limit, the hunting around has no effect. If your setting is 50, then whether the car thinks the best speed is 50 or 53, it's going to do 50. In cases where conditions require the car to slow down, it'll slow down and hunt for the best speed, but if the appropriate speed for the road is 50, conditions should allow that again pretty quickly.
This means that if you want to shut down the hunting effect, lower your set speed to the point where the car always wants to go faster. If you're in a 25 zone, and the car is hunting around between 23 and 28, set your speed to 23. Unfortunately, that would mean that you're undoubtedly going too slow for other drivers, but that's the idea.
To return to Auto Speed, that seems to do only one thing: puts the set speed at something very high to allow the car to hunt for the appropriate speed all the time.
I think at this point, TeslaFi shows most states have at least one install, and the states without any detected might just have too small of a population on that service to detect the rollout. So potentially 12.x has been deployed to all states at least 1% rollout -- i.e., it's possible in any state.
Here's the states with at least one install for the initial validation rollout sorted with largest potential population first:
Code:
Washington (9) 4%
New York (2) 1%
Pennsylvania (4) 3%
Illinois (10) 8%
Maryland (4) 3%
Colorado (8) 6%
New Jersey (3) 2%
Massachusetts (3) 3%
Ohio (3) 3%
Minnesota (2) 2%
Michigan (3) 3%
Indiana (4) 5%
Oregon (4) 5%
Connecticut (1) 2%
Missouri (3) 6%
Utah (1) 2%
New Hampshire (1) 3%
Kansas (2) 9%
Alabama (2) 10%
Louisiana (2) 10%
Idaho (1) 6%
New Mexico (1) 6%
Rhode Island (1) 7%
Maine (2) 17%
Vermont (2) 17%
Mississippi (1) 10%
Alaska (1) 20%
North Dakota (1) 25%
South Dakota (1) 33%Whereas these have none detected but Wisconsin at the top of this list only has 70 vehicles potentially getting 12.x, so a 1% rollout would expect .7 vehicles:
Code:
Wisconsin (0)
Tennessee (0)
Oklahoma (0)
Iowa (0)
Kentucky (0)
Delaware (0)
Washington DC (0)
Nebraska (0)
Arkansas (0)
Hawaii (0)
Montana (0)
West Virginia (0)
Wyoming (0)And for reference, the states with wider rollout seems to be around 40% now:
Code:
California (264) 28%
Texas (123) 34%
Florida (92) 29%
Virginia (73) 41%
North Carolina (46) 39%
Arizona (41) 38%
Georgia (33) 31%
Nevada (22) 52%
South Carolina (11) 30%
sleepydoc
Well-Known Member
there was a separate thread with v12 driving experiences and for some reason the mods decided to merge it with the 2 year old rumor thread. Maybe they're trying to see how long they can make a thread before it crashes the server.
What is your view on the IV Drip style waves? It’s been looking like it goes down to ~40 and goes back to ~140 pending.
I received v12 on Sunday. I took it for a Trader Joe's run, which is a simple drive from my house, around 7 miles, almost entirely on one road.
On the way there, it did extremely well, much better than any v11 build I had previously tried. Driving was smooth, and I was particularly impressed with a couple of lane changes the car made to get around slow traffic; they were ones I probably would not have done myself, as I'm pretty patient, but they actually resulted in making a couple of green lights that I otherwise would have missed. On the downside, even though I have the car set to alert me with both a tone and steering wheel vibration before making a lane change, I don't think it did either.
The drive home was similarly good, though at one point it appeared to me that the car was waiting too long to get out of a lane that was about to become turn-only, and the lane I needed to be in had a motorcycle in it approaching from behind, so I went ahead and changed manually to avoid any possible issues with the car deciding to cut off a motorcycle at the last minute. It might have done just fine, but I really didn't want to wait to find out.
One thing that will prevent me from using FSD as it stands today with any regularity, however, is following distance. Even in Chill mode, it follows cars much too closely. I drive defensively and like to sit back several car lengths so I can see not only the car in front of me but also the cars around and in front of it. Since I still have to supervise FSD, I want it to do the same, but it was much too close to the car immediately in front of me for much of this drive - often within a car length or less at 45 mph. This is simply not safe behavior, particularly when it's clear that the car cannot see, based on what's on the display, that there are indeed other cars in front of the lead car. This tailgating behavior was especially bad when behind black and other dark cars - maybe just a coincidence, but I do wonder if the cameras perceive the distances to darker and lighter colored vehicles differently somehow.
This morning, I tested the system again on my interstate commute, but had to turn it off pretty early in my drive. The first hiccup was when I was stopped in a line of cars at a light. The car needed to continue straight and the map showed that it intended to do so, but while I was stopped, the wheel jerked pretty violently to the left several times, until it had turned ~180 degrees. This was quite unnerving. I cancelled FSD, righted the wheel, reengaged, and it did it again. The car righted the wheel as we began to move, with my hands ready to correct it if it failed to do so, but it was still disturbing and unexplainable behavior.
After it failed to start braking as I was an approaching a nearly stopped car in a line of traffic on the interstate, I took over at around 40 mph with maybe 100 feet to spare. Maybe it would have braked, but again, I would rather it not wait so long. I like gradual stops and gradual acceleration, particularly when there are other cars all around. Making a sudden stop like that is just asking to get rear-ended.
In short, this iteration of v12 represents progress, but not nearly enough for me to feel comfortable using it on a regular basis. Back to EAP on the highway for me.
On the way there, it did extremely well, much better than any v11 build I had previously tried. Driving was smooth, and I was particularly impressed with a couple of lane changes the car made to get around slow traffic; they were ones I probably would not have done myself, as I'm pretty patient, but they actually resulted in making a couple of green lights that I otherwise would have missed. On the downside, even though I have the car set to alert me with both a tone and steering wheel vibration before making a lane change, I don't think it did either.
The drive home was similarly good, though at one point it appeared to me that the car was waiting too long to get out of a lane that was about to become turn-only, and the lane I needed to be in had a motorcycle in it approaching from behind, so I went ahead and changed manually to avoid any possible issues with the car deciding to cut off a motorcycle at the last minute. It might have done just fine, but I really didn't want to wait to find out.
One thing that will prevent me from using FSD as it stands today with any regularity, however, is following distance. Even in Chill mode, it follows cars much too closely. I drive defensively and like to sit back several car lengths so I can see not only the car in front of me but also the cars around and in front of it. Since I still have to supervise FSD, I want it to do the same, but it was much too close to the car immediately in front of me for much of this drive - often within a car length or less at 45 mph. This is simply not safe behavior, particularly when it's clear that the car cannot see, based on what's on the display, that there are indeed other cars in front of the lead car. This tailgating behavior was especially bad when behind black and other dark cars - maybe just a coincidence, but I do wonder if the cameras perceive the distances to darker and lighter colored vehicles differently somehow.
This morning, I tested the system again on my interstate commute, but had to turn it off pretty early in my drive. The first hiccup was when I was stopped in a line of cars at a light. The car needed to continue straight and the map showed that it intended to do so, but while I was stopped, the wheel jerked pretty violently to the left several times, until it had turned ~180 degrees. This was quite unnerving. I cancelled FSD, righted the wheel, reengaged, and it did it again. The car righted the wheel as we began to move, with my hands ready to correct it if it failed to do so, but it was still disturbing and unexplainable behavior.
After it failed to start braking as I was an approaching a nearly stopped car in a line of traffic on the interstate, I took over at around 40 mph with maybe 100 feet to spare. Maybe it would have braked, but again, I would rather it not wait so long. I like gradual stops and gradual acceleration, particularly when there are other cars all around. Making a sudden stop like that is just asking to get rear-ended.
In short, this iteration of v12 represents progress, but not nearly enough for me to feel comfortable using it on a regular basis. Back to EAP on the highway for me.
Ben W
Chess Grandmaster (Supervised)
Yes, I meant controlled-access highway (freeway). I already find it meets the "useful L2" definition there, although it still regularly makes mistakes that require intervention, perhaps one or two per 100 miles. I still find it useful even though it's not L3 (eyes-off-road) yet.
Apparently the nags are context-dependent, so you should get them much less often in "easy" driving conditions. (On an uncrowded wide-open freeway, perhaps once every several minutes.) I do wonder how this will evolve as the system reliability improves. I have a 2017 M3 and 2022 MY, both with FSD, and plan to upgrade the M3 to a MS in a year or two, transfering FSD if that's still an option. I do expect Tesla to achieve controlled-access highway L3 (no nags or strikes unless prompted to take over) within the MS lifetime, and possibly controlled-access highway L4 (no in-motion nags or takeover prompts at all) towards the end of it.
I've found that the system is particularly sensitive to the first 10 seconds or so after enabling FSD; if I look away for more than a split-second during that time, it dings me. This is presumably to discourage drivers from enabling FSD just in order to do some immediate task that takes their eyes off the road. But agreed, even the first ding adds anxiety. (And I sometimes get dinged even when I'm looking straight ahead at the road; the camera system is far from perfect.) If this happens on a long drive, I'll sometimes pull over and put the car in park, just to reset the strikes. But I think that some system of nags/strikes is unfortunately necessary while FSD requires the level of supervision it does. Until it can drive a million miles between unforced safety-critical errors (such as merging into a turnout lane at speed on the highway, which I've seen it do several times), it won't be ready for L3.
As many have pointed out, V12 isn't used on the interstate. It reverts to the old stack.
sleepydoc
Well-Known Member
Which is precisely why we need to get V12 out to as many people as possible. Time is short - Opossums are dying, people!
sleepydoc
Well-Known Member
Seriously - if you can't read all 6290 posts in this thread, why are you pestering the rest of us? It's right there if you'd just look!
We'll need more examples to see but this is a good sign that they've fixed the 'school bus bug'
Maybe we should add an Elon curse quotient as a score of usability?
hey8you
Member
Agreed. Also, there are a few spots near me I want to try out. V11 does not handle them very well. 2 spots V11 acts dangerously.
Oh, I know, but the behavior is also clearly different from what it does when using EAP even on the same software build. When I switched back to EAP this morning, following distance immediately became more comfortable (longer), and stopping for cars ahead of me also became more gradual.
And Go Bucs !
I think his post reveals the direction he's given to focus the Autopilot team. In some sense, "Operation Vacation" is mostly there especially with end-to-end able to improve the neural networks with general data collection, but instead of going on vacation just waiting for data to come in, he wants the team to use the usual 2-week development cycle actively working on the top 3 issues.
AlanSubie4Life
Efficiency Obsessed Member
9-mile commute. Link provided yesterday here which is important for evaluating the scenario. Pretty awesome overall. Light traffic.
0 disengagements (a first, by far, fewest ever is like 3-4 or something probably). There were one or two lucky situations where it was drifting uncertainly for a second, and if there had been cars around I would have disengaged.
~7+ interventions, for speed and a couple lane changes I wanted to make, or cancel impending incorrect ones it wanted to do.
Absolutely ZERO nags. Not one. (9 and 3 FTW.)
(Seriously!)
This thing is good (for my application).
To me, more investigation needed, but it seems to make some attempts to drift around the lane to create buffers, and it may be avoiding blind spots. BUT, I am not sure about this; obviously could be chance. Wasn’t super busy this morning so not too many opportunities to observe. TBD
Will it make mistakes? Yes
Is the speed control broken? Yes
Is the stopping behavior stupid? Yes
Will it be pretty useless in some scenarios? Yes
Will it be a robotaxi? Absolutely not on current hardware, obviously.
So it appears fundamentally good for certain specific scenarios, but is also currently broken, as previously described. We will see if it can overcome these issues and be useful in slightly more complex scenarios.
0 disengagements (a first, by far, fewest ever is like 3-4 or something probably). There were one or two lucky situations where it was drifting uncertainly for a second, and if there had been cars around I would have disengaged.
~7+ interventions, for speed and a couple lane changes I wanted to make, or cancel impending incorrect ones it wanted to do.
Absolutely ZERO nags. Not one. (9 and 3 FTW.)
(Seriously!)This thing is good (for my application).
To me, more investigation needed, but it seems to make some attempts to drift around the lane to create buffers, and it may be avoiding blind spots. BUT, I am not sure about this; obviously could be chance. Wasn’t super busy this morning so not too many opportunities to observe. TBD
Will it make mistakes? Yes
Is the speed control broken? Yes
Is the stopping behavior stupid? Yes
Will it be pretty useless in some scenarios? Yes
Will it be a robotaxi? Absolutely not on current hardware, obviously.
So it appears fundamentally good for certain specific scenarios, but is also currently broken, as previously described. We will see if it can overcome these issues and be useful in slightly more complex scenarios.
Yes, both modes are broken.
Last edited:
I think that's pretty standard Tesla vehicle software deployment process of small rollout to evaluate knowing the typical observed install rate (e.g., some people are traveling or generally fast/slow to install updates). Tesla needs time to allow for the installation and usage and potentially long enough to avoid "first day" bias to have meaningful data to compare, and some evaluations can be done in parallel resulting in multiple rollouts. The new thing with 12.x is the state-by-state selection, but the deploying to various hardware/software configuration is pretty standard (although usually across different versions whereas here with 12.3 / 2023.44.30.25, it's the same version).
Most people getting 2024.x software skipping past 2023.44.30.25 are likely those outside of US/Canada or don't have FSD Capability or have incompatible hardware. Tesla is choosing not to deploy this "too new" software to vehicles that could get 12.3, so that's why US HW3 vehicles with FSD have been kept on 2023.44.30.8/.14. Although some reason it seems like Tesla decided 12.3 might not be ready for portrait screen Model S/X (without interior camera?) and pushed out 2024.2.7, but hopefully Tesla bumps up 12.x to some 2024.x version soon to include those vehicles as well as new subscribers and purchases.
I don’t have the update and I’m one of the OGs from back when we had to play the Safety Score game. (Still have PTSD from it
)
AlanSubie4Life
Efficiency Obsessed Member
I for one am curious about how the J Crew will react. But I hope Tesla fixes the stopping and speed issues before that wider release. You guys don’t deserve to be saddled with the current version; you’ve been through enough.I don’t have the update and I’m one of the OGs from back when we had to play the Safety Score game. (Still have PTSD from it
mtndrew1
Active Member
I’ve been on the J crew for a while now and I’d like to bring it up one notch. What’s one level up from J?
Similar threads
- Article
