Understandable reaction as you're getting used to it. The big difference is that you can have your hands and feet hovering close to the controls and take over whenever you feel uncomfortable, unlike the roller coaster where you're completely at its mercy. Still gives a little adrenaline rush each time though! I think you'll find that if you allow TACC to do its own thing, you will become less nervous. It does act consistently, even though it is too aggressive and not very smooth. There's lots of room for improvement, and I feel it's particularly bad for passengers - who are in no different situation than on that roller coaster.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
TACC jumpiness
- Thread starter senor_sausage
- Start date
It is my understanding that Tesla tracks "operator intervention" on AP and perhaps on TACC as well... so don't we need more "Nervous Nellie's" to hit the brakes when flying up to a stopped car, or slower moving traffic, or red lights to train the system to brake earlier and less aggressively?
Rather than using customers as Beta testers, I would much rather Tesla would put the car into "learning mode" where it makes choices but doesn't implement them while observing the choices a competent attentive driver would make. They could then review a driving session and "correct" its faulty choices. I guess what I am saying is that trained and paid human interface developers should train the cars?
Silly of me to think that the company that makes the car should make the systems work properly... so silly of me!
Keith
Rather than using customers as Beta testers, I would much rather Tesla would put the car into "learning mode" where it makes choices but doesn't implement them while observing the choices a competent attentive driver would make. They could then review a driving session and "correct" its faulty choices. I guess what I am saying is that trained and paid human interface developers should train the cars?
Silly of me to think that the company that makes the car should make the systems work properly... so silly of me!
Keith
What you don't include in your thought process is that AI is not trained with a handful of inputs, it requires a pretty high amount of (properly organized) inputs. You could use a couple of paid drivers if you wanted to devise a traditional algorithm, maybe. The other aspect is that AI requires multiple inputs of data from a wide variety of situations and locations so it is adapted to respond to all locations and situations as much as possible. You would need those paid drivers to roam around the world... 
It is essential that Tesla use data from the cars if they want to build a proper machine-learning dataset. That's not enough though, they also need to curate that data. I'm sure they do have many, many paid people to look at the data and label it or otherwise organize it so that it is in the right state for the machine to learn, eliminating bad data in the process. They most certainly do not just take all the data without any checks.
It is essential that Tesla use data from the cars if they want to build a proper machine-learning dataset. That's not enough though, they also need to curate that data. I'm sure they do have many, many paid people to look at the data and label it or otherwise organize it so that it is in the right state for the machine to learn, eliminating bad data in the process. They most certainly do not just take all the data without any checks.
What you don't include in your thought process is that AI is not trained with a handful of inputs, it requires a pretty high amount of (properly organized) inputs. You could use a couple of paid drivers if you wanted to devise a traditional algorithm, maybe. The other aspect is that AI requires multiple inputs of data from a wide variety of situations and locations so it is adapted to respond to all locations and situations as much as possible. You would need those paid drivers to roam around the world...
It is essential that Tesla use data from the cars if they want to build a proper machine-learning dataset. That's not enough though, they also need to curate that data. I'm sure they do have many, many paid people to look at the data and label it or otherwise organize it so that it is in the right state for the machine to learn, eliminating bad data in the process. They most certainly do not just take all the data without any checks.
In this model of machine learning (I have no problem with them doing this BTW) they need some method for drivers to alert the review teams when the car makes a mistake... and we need to mark every damn time the car charges up on stopped traffic / slow traffic / stop sign (or light) and slams on the brakes at the last moment as a mistake. Also, we need some way to mark slowing down to a crawl for sharp curves as a mistake...
In reality different people have different driving styles. There are flesh and blood humans out there that charge up to stopped traffic and slam on the brakes at the last moment, and that is how they want AP to behave. There are others that want to take sharp curves at 10 mph under the speed limit and that is how THEY want AP to behave. Personally I want AP to slow down well before reaching a stop using regen only if possible, and take sharp curves at 10 mph OVER the speed limit. So, if we have two people in the review team person one is a gradual stop, gradual start, drive fast on sharp curves type of person and we have person two in the review team that is a slam on brakes to stop, slam the accelerator to start, drive slow on sharp curves type of person what kind of algorithm will we end up with? The only solution I see is user selectability in the TACC and AP and FSD settings.
I would love to be able to help it learn out of my own selfishness, let alone the common good. If they implemented driver input to the review teams with lootbox rewards for good input it would speed up the learning process.
Keith
PS: The alert to the review team (if they ever implement it) can't be the horn honk used to save dash cam footage... I don't want people thinking I am nuts honking my horn in traffic for no reason
Except that is not true. the bug report command lets you put a note in the car's log, and it stays there. If you get a tech service your car for a problem, you can tell them that you entered a bug report at the right moment, it helps them pinpoint the interesting area of the logs. These things are huge.
I don't know if they get a summary of them sent versus the entire image/snapshot, but this is the proper way to report bugs.
from the manual:
Gpkarma
Member
So, I have following distance on 2021 MY set to 3-4 cars (currently I cannot set it fewer than 3) — and it typically adjusts and does not stil remain 3 cars away when cars are slowing down in front (except buggy situation of car ahead is just over the crest of a hill) … it just does not stop right behind the car in front of me … it is like a smaller size sedan distance, away from vehicle in front. I test drove on a radar car with FSD, with setting at 1 car and it stopped short right behind upcoming car (almost had heart attack btw… did not know about the setting!) at the last second, less than 1 car away… will have to compare when they feel TeslaVision / network has been trained enough to allow shorter stopping distance. I do think it would be nice to enable the cameras tos ee the car ahead through the car in front of you’s windows… like 2 cars in front CAN be seen with a human eye so I hope this will come back, or at least training/adjustment for when there is a hill crest in front of you & you are following another car,…… speeding up extremely faster than normal as then the car in front suddenly appears closer not fun.
I'm still getting used to my basic Autopilot, too, but I actually think my MY does better than the adaptive cruise control in my wife's BMW, which we've had for over 2 years.
I was just using it in stop and go traffic for about 20 minutes this week, and I though it was more "safely aggressive" than I would have. By that I mean, I felt safe, but I didn't get cut off by anyone changing lanes when just enough space opened up as the traffic started to move.
I have had the same thought about the AutoSteer being too perfect in lane center, and how that's not how I, or most human drivers, actually drive (e.g., when I pass someone, I now realize that I hug the left lane line). But I suppose if the engineers aren't comfortable with the system mimicking human idiosyncrasies like those yet, I prefer it split the different and center in the lane. I've already tried it in the middle lane of a 3-lane wide interstate, and while it was a little nerve-wracking, I think was actually very good and safe.
I was using TACC recently on a highway that had some grade crossings with traffic lights. After stopping at a red light, I could always tell if the car in front of me was a manual transmission when the light turned green . My Y exactly matched the speed of the car so I got to feel every gear shift. Starting up behind a tractor trailer was rough. Those things don't accelerate very smoothly. Even with that, I really like using Autopilot in stop and go traffic. It's much more relaxing than doing all the stops and starts manually. It's much more pleasant behind a car with an automatic transmission though.
RNHurt
2020 MY, White, Black
I just subscribed to FSD, which installed 2021.12.25.7 and I've noticed a more gentle TACC. I don't know if it's due to FSD or the software upgrade, or maybe I'm just imagining it, but it feels different to me. Previously, in heavy stop and go traffic, I was concerned that the person behind me was going to ram into me. My MY would accelerate hard to catch up to the traffic and then hit the brakes hard to stop. Now, it seems to ease into the acceleration and let the car glide to a stop more slowly; not quite how I do it but much better.
Does anyone notice this? I'm I just imagining it?
Does anyone notice this? I'm I just imagining it?
AlexHung
Member
TACC has been more smooth since 18.2. I’m on 25.7 and it’s really good.
Similar threads
