Thanks. I was wondering if people with curbed rims were driving the Model S because it's about five inches wider than a Model 3. That's 2.5 inches on each side, and that's the difference between "uncomfortably close" and a curbed rim.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD beta v12.3 pulls the ultimate April Fools joke
- Thread starter AS9.9.
- Start date
I posted a similar thread from my experience last night. FSD curbed both passenger side rims and it wasn’t a matter of not being vigilant. The car was in FSD, driving great, there’s a turn near our house and the car, previously, never had a problem making it around. This time, it was going around it and decided to go straight during mid-turn! There was no time for me to turn the steering wheel to avoid it, either. My wife was with me and we were both shocked when it happened. I’m still a fan of v12 but I’m not sure how well the 12.3.3 rollout to everyone is going to go as FSD (supervised) seems to be having issues curbing wheels as well as the speed issues. As a shareholder, I’m a bit nervous about this rollout (as well as a possible class-action lawsuit being filed).
Here's another, fresh this morning.
FSD 12.3.3 '24 MX
Hit the curb when exiting traffic circle - no evasive maneuvers
Last edited:
I feel your pain.
FSD 12.3.3 on a brand new (400 miles) '24 MX
Hit the curb when exiting a traffic circle. No challenging traffic.
Interesting as the curb is painted red as well. I feel like the car really can't gauge its distance to curbs well enough.
I have a hypothesis. Are all the reports of curbing wheels from Model Y & X?
There is a major influence of Model 3 training data of course. The path charted by Model 3 from human drivers is not always acceptable for Y and X as they are wider, but the FSD system might not take that into account yet. It just does what its nets have seen as examples of human driving, and the point of view is from the center of the car.
This is a deficiency going to end-to-end nets (not insolvable but harder) instead of explicit physical space path planning and optimization and physics knowledge of sizes/widths, vehicle boundary and drivable area boundary.
FSD curbed my model 3, so I am unsure. But I also think S, X, and Y will have a higher curb rate considering their width.
I'm just mildly "amused" by the fact that folks let FSD beta - pardon "supervised" - curb their car and cause damage to the wheel/ potentially side wall of tire.... just so Elon can collect free data points. You would think they could have run FSD by now on plenty of 3 / Y so that doesn't happen when they offer a "free trial" to anyone. My "trial" FSD also nearly curbed my vehicle but I intervened and haven't used it since... not worth ruining a wheel ...
Also I wonder about the rate of hitting the curb in California vs outside California. Nets make assumptions all the time. I have a 3 in CA and little problem---but that's where the data set is concentrated.
KArnold
Active Member
Actually I don't get this part. Supposedly FSDs learns by watching millions upon millions of right hand turn videos to mimic an AI-generated precise route for the cirmstances.
How many of those videos curbed the wheels? Presumably none. So WHERE did it learn to do otherwise?
Feel free to insert Twilight Zone music if applies.
black knight 3
Member
black knight 3
Member
black knight 3
Member
It followed a path which would be OK (starting out) for a Model 3 in California. Maybe other states have sharper turns or build their intersections somewhat differently, or the curbs look different in some as yet not understood way.
It's still a very small minority of turns and so from statistical point of view the net is successful---that isn't enough for a deployed product.
In the ML sense I think they lack enough negative contrastive examples: "Don't drive like this" "Do drive like that" in similar or identical locations.. Maybe that could be done by simulation (present proper paths as positive examples and improper wheel-curbing but otherwise similar paths as negative examples so it sharply understands the boundary between them) but that may make the system work properly only with simulated data as it can probably distinguish simulated from real images implicitly.
This is the problem with full "nothing but nets"---its quite difficult to find all the weird behaviors. Just like with people.
black knight 3
Member
no, my friend. It hit the car hard and car shifted. You have to run video at 1/2 speed to see the dust-up and the car shift after scraping the curb. attached is the photo right after I came home and the dust and rim damage are visible..
Attachments
Similar threads
- Article