gottagofast
2022 Model 3 Performance
Is there a certain speed limit where you don’t get nags with FSD? The car never seems to nag me in stop and go traffic when I’m going under 10 mph
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.
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 has decision wobble fsd jerky driving fsd late deceleration fsd sometimes may not detect train crossings fsd stops too early too quickly fsd stops too far behind line fsd tailgating in city streets
Not sure V12 has correct behavior:
Last night FSD made a left turn from street to a restaurant parking lot. When the car head was about 7 ft to the entrance I disengaged because I saw 2 people walking on the sidewalk to the parking entrance (they were about 10 ft to the entrance opening). I waved them to go and they waved me to go too. Both sides were courteous
Should I disengage or should I ignore the pedestrians and let FSD continue to go to the parking lot?
Last night FSD made a left turn from street to a restaurant parking lot. When the car head was about 7 ft to the entrance I disengaged because I saw 2 people walking on the sidewalk to the parking entrance (they were about 10 ft to the entrance opening). I waved them to go and they waved me to go too. Both sides were courteous
Should I disengage or should I ignore the pedestrians and let FSD continue to go to the parking lot?
Mow them down and have lunch.
I can't. They are nice people. But I am willing mow down some impatient non-FSD Tesla owners cutting me off when I am using FSD.
Yeah for city streets driving, there wasn't as much noticeable improvements with V11 focusing on replacing the even older Autopilot behavior for highways such as fixing the strict lane centering even when widening for on/off-ramps. Sounds like the lane positioning is much improved with your experience with 12.x so far such as smoothly passing parked cars without weaving or approaching stop signs with enough space to fit 2 vehicles side-by-side?
AlanSubie4Life
Efficiency Obsessed Member
Yes, this is a good point. Seems like HW4 will be plenty capable, so it may be beneficial for Tesla to focus on that for now and get the robotaxis working. Then they can build up retrofit modules. I too am doubtful of the geofenced service.
Last edited:
gottagofast
2022 Model 3 Performance
diplomat33
Average guy who loves autonomous vehicles
Edge Case Summary after 2 weeks of testing V12:
Positive takeaways:
- Less interventions and disengagements. Most of these used to be due to bad planning/erratic lane changes
- No more weird counter intuitive “changing lanes to follow route”
- Turns are butter smooth
- In general, less unnecessary slow downs due to pedestrians or nearby objects, makes for a much smoother ride.
- More confident in unprotected left turns from the center
- Major improvements in speed bumps and dips
- Ability to do u-turns as long as road is wide enough without requiring a 3 point turn
- Phantom turn signaling is gone
- Lane changes (city streets) have gotten so good that even if it’s not in an optimal lane before a turn it can easily and safely squeeze in on time.
- Improved behavior at right turn on red where ego has to yield to opposing traffic making a protected left turn before safely proceeding
- Recognizes blocked lanes in a timely manner and moving away from them naturally and safely
- Recognizes keep clear road signs and keeps the box clear in heavy traffic scenarios (not perfect yet)
- It can navigate large open parking lots with heavy foot traffic very well (eg Costco). No jerky steering and proper speed to account for pedestrians
Negative takeaways:
- The stopping/starting speed curves need adjustment. It’s not natural when coming to a stop and passengers are either pushed back into their seat accelerating out of a turn or jerked forward due to sudden stops at yellow to red light transitions
- Still has trouble with no right turn on red intersections.
- Chuck Cook style unprotected left turns are less confident from initial start (requiring an accelerator press to get going). Need to do more testing
No comment:
- Have not tested yet some of the less common things on my common routes like roundabouts, yields, construction, etc
- Highway behavior is the same as V11 since it is not on the E2E stack. Actually slight regression with missing exits off a highway
Positive takeaways:
- Less interventions and disengagements. Most of these used to be due to bad planning/erratic lane changes
- No more weird counter intuitive “changing lanes to follow route”
- Turns are butter smooth
- In general, less unnecessary slow downs due to pedestrians or nearby objects, makes for a much smoother ride.
- More confident in unprotected left turns from the center
- Major improvements in speed bumps and dips
- Ability to do u-turns as long as road is wide enough without requiring a 3 point turn
- Phantom turn signaling is gone
- Lane changes (city streets) have gotten so good that even if it’s not in an optimal lane before a turn it can easily and safely squeeze in on time.
- Improved behavior at right turn on red where ego has to yield to opposing traffic making a protected left turn before safely proceeding
- Recognizes blocked lanes in a timely manner and moving away from them naturally and safely
- Recognizes keep clear road signs and keeps the box clear in heavy traffic scenarios (not perfect yet)
- It can navigate large open parking lots with heavy foot traffic very well (eg Costco). No jerky steering and proper speed to account for pedestrians
Negative takeaways:
- The stopping/starting speed curves need adjustment. It’s not natural when coming to a stop and passengers are either pushed back into their seat accelerating out of a turn or jerked forward due to sudden stops at yellow to red light transitions
- Still has trouble with no right turn on red intersections.
- Chuck Cook style unprotected left turns are less confident from initial start (requiring an accelerator press to get going). Need to do more testing
No comment:
- Have not tested yet some of the less common things on my common routes like roundabouts, yields, construction, etc
- Highway behavior is the same as V11 since it is not on the E2E stack. Actually slight regression with missing exits off a highway
AlanSubie4Life
Efficiency Obsessed Member
Unfortunate that Chuck didn't get the first edition. Might have been able to score twice.
OK, now I'm getting excited. Seriously, that issue alone was enough to keep from using V11 off the highway at all.
AlanSubie4Life
Efficiency Obsessed Member
Maybe this is the true reason they chose not to roll it to Chuck in the first wave (it just wasn't going to be good). Perhaps it was not his ignorance about the clear and present dangers to western civilization.
Unfortunate that Chuck didn't get the first edition. Might have been able to score twice.
I have to say the videos of simple cases of unprotected lefts with some traffic have not looked good so far. It's possible it is too confusing for end-to-end. Or maybe they'll have it trained up to superhuman levels for the next release.
Welcome to the Debate Party.
There are semi, uncomfirmed noises from Tesla that they think HW3 can do the job. Their pronouncements on this subject are as clear as mud.. because they haven't quite gotten FSD-b running whatever level they'd like and it's a research project (as in, nobody's ever done this before).
The tongue-in-cheek definition of research is, "The process of running up alleys to determine if they're blind."
There are a variety of posters of various levels of technical acumen, none of whom are actually working on an ADAS system, who claim, variously, that
- It's never coming, no matter how much hardware/software one throws at the problem
- It's coming, but it requires HW4/5/6/...
- It's coming, in HW3 or whatever, but the amount of time is much longer than anyone will expects. Years, decades, heat death of the universe
- And, just to keep things boiling, there's people who sometimes show up and say It's Beautiful, it'll be next month!
I happen to think positively of Tesla's efforts and think that they'll get there eventually. Just like everyone else around here, that opinion is somewhat fact-free (since nobody's actually looking at the code/hardware in Tesla's development labs). My gut feeling on this has two parts:
- Back when they started all this in 2017 or thereabouts somebody with some serious technical chops did the equivalent of a back-of-the-envelope calculation, thus generating the specs for HW3 that would nominally be able to do the job. (Or earlier, thus giving us HW1/HW2/HW2.5.. but HW4 has only come out during the 4th Quarter, and might just be the usual reduce costs/improve performance that thinner silicon line widths give people)
- A lot of people have paid for HW3 and, even with Tesla messing around with the specs of exactly what FSD-b is going to be able to do, there'd likely be some enormous financial hit involved with retrofitting all those older cars. They don't want to do that, so they have every reason to make it happen.
The proper answer as to what it's all going to take and how long is, "Nobody knows".. and that probably includes the people at Tesla. But since they don't talk, None of Us Know.
The debates are fun, though.
gottagofast
2022 Model 3 Performance
If hw4 is needed they will upgrade all the hw3 cars that have FSD installed for free?
Heat death of the universe - Dude, I did a spit-take that was so f'in hilarious. Thanks for the chuckle.
gottagofast
2022 Model 3 Performance
We're back to "Nobody knows".
What's the phrase: $DIETY doesn't know, and the Devil's not telling..
Had a HW2.5 car upgraded to HW3. Far as I know, the CPU box was swapped.
HW4 as presently built apparently involves different cameras and wiring harnesses.
But, in engineering, "Never say Never". I've seen some upgrade methods that involve craziness in HW & SW that are hard to describe, and have even been responsible for a few.
First off: Tesla has never said that HW4 is going to be required for FSD. Since they're the people actually designing and building the cars, their opinion rules, but "never said" covers a lot of ground.
Those around here who say, "It can't be done with HW3!" are expressing opinions, not facts. They may be well-researched opinions, but they're still opinions.
The design differences between HW2 and HW3 allowed the electronics to be swapped out with minimal effort. The design differences between HW3 and HW4 is significant (based on tear-downs), including new wiring, ports on different sides, different cooling needs, and overall size increases of the boards.
