Agree with this. Add to that the very late switching of lanes on the highway to take an exit, which I think is a very easy fix.
-
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
Yikes. I quit watching after the first attempt. FSD timing appears way off kilter. Maybe they need to better synchronize the new vehicle control nets or poor object kinematics estimation?
Apparently many more ADAS visits are forthcoming.
Last edited:
AlanSubie4Life
Efficiency Obsessed Member
Coming around to the idea that the lead engineer for this needs to be relieved of duty. It is not a difficult problem! Just simulate it. The problems will show up. If not, fix the simulation.
Last edited:
dachshund
Member
I tried disengaging by jerking the steering wheel slightly. The robot shrieked and gave errors about the speed being too high and said something about a steering problem. It refused to disengage, and continued to do loud beeping. The blue stripe finally turned gray and the beeping stopped.
Pretty strange. I’ll stop using the steering wheel method since it seems to be a problem.
Pretty strange. I’ll stop using the steering wheel method since it seems to be a problem.
12.3.4
Right tires off the road a couple of times and close to off several more. One time though clearly into the dirt intentionally to give some spaces for a large truck on the outside lane while my inside lane was curving to the right and they were crossing the midline.
Right tires off the road a couple of times and close to off several more. One time though clearly into the dirt intentionally to give some spaces for a large truck on the outside lane while my inside lane was curving to the right and they were crossing the midline.
Similar episode once when I tried to disengage using the yoke but corrective steering prevented me from turning manually, and there were lots of beeps. All happened very quickly while I was driving very slowly in a parking lot.
Daniel in SD
(supervised)
AlanSubie4Life
Efficiency Obsessed Member
I’m beginning to have concerns that they will not have the right approach for v12.4.Sep 30th 2022.
They appear to be leaning too heavily on real-world results and training. And it does not appear that there is any feedback loop.
Failure to solve this in 12.4.x could easily be a death blow to the entire program. Have to be able to solve the simple problems to have any hope of solving more difficult ones.
Any beer bet that 12.4 will be the one that finally gets the first 9?
No bets, but there is no such thing as a simple fix for ML. The challenge is that outcomes are probabilistic and not deterministic so the best you can hope for is that the inference is probably correct. Also, I think even the smartest minds have significantly underestimated self driving as a challenge - ML is a great way to do brute force statistical regression and solve certain problems - but there is no guarantee that as it stands that it is the right approach for this particular problem. If the scenario doesn’t exist in the training set, then it cannot be handled during inferencing. Just that small problem that your training set merely needs to have examples of every possible driving outcome that could ever occur in every possible environment that could possibly occur.
AlanSubie4Life
Efficiency Obsessed Member
Except that does not seem to be happening here.
Inference and understanding of the vehicles and their positions seems to be perfect.
And for being probabilistic, it is not like this is infrequent or somehow being tripped up by tricky situations!
Just simulate it! No need to drive! If it doesn’t fail in simulation, make it fail in simulation FIRST.
I would prefer to disengage with a gentle tap on the brakes and take control, rather than fidgeting with the steering wheel.
Chuck always test it right after installing a new version. My experience is that it takes a couple of days for a new version to “settle in”.
I find that FSD preforms much better after a couple of days of using it. (I don’t Know why or how, but it is just my observation)
Except when you need to use the steering wheel to avoid a pothole.
Correct me if I'm wrong, but haven't actually seen what degree of improvement a full retraining cycle can actually lead to. Most public testers at this point started on a 12.3.* release (and the various point releases that have come pretty quickly are not full retraining of the system, as far as I know).
12.4 will be our first real chance to see how much changes.
I personally don't care that much about unprotected lefts, specifically, but I'm am very interested to get a better picture of:
- how much change can they ship per iteration (meaning a full retraining of the system)
- are there noticeable regressions?
- how many iterations they can ship per year
Time to give up on Chuck's Left Turn. Left turns are dangerous, according to experts and NHTSA:
AlanSubie4Life
Efficiency Obsessed Member
Yep.
Yep, that is why it is one of the most consequential releases ever.
A lot more accidents at nearby traffic-light-controlled intersections.
Daniel in SD
(supervised)
Most collisions at intersections happen when there is a traffic signal.
I’m not sure it’s possible to exit Chuck’s neighborhood doing only right turns.
Last edited:
Similar threads
