Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
I am still completely wowed by the recent map optimization (still on 11.3.6). Literally reduced my intervention rate by over 50%+. My disengagement free drive increased dramatically just from correct lane selection and zero phantom stopping! Tried some other routes today I know had problems before and today they were all fixed!
Yes - FSD is quite heavily tied to maps. Even humans drive much better on known routes than new. Tesla needs to put more effort and resources in getting good maps and more importantly correcting them in a timely manner. If I see an error I should be able to get them corrected in a matter of days rather than once in 6 months.
 
Drove 650 miles today. The biggest improvement is the cone behaviour. Previously if the cones were in the centreline where they could be hit, the car would get as close to them as possible. But if the cones were 2-3 metres from the road edge, where even a drunk couldn’t hit them the car would move over to get away from cones. I had zero cone disengagement today.
 
Despite the significant improvements in FSD up to and including 11.4.2 there are 2 problems that never seem to get corrected.

1) I live in a gated community and the "gate" is a single pole that lowers much like on a railroad crossing. FSD never reacts to the gate in any way at all. There is no question it would crash right through it every time if I didn't intervene. The display also has no visual indication that it sees anything at all. I can understand that it might not have been trained yet to recognize a pole across the road but I can't understand how it doesn't know that there is SOMETHING there that it shouldn't hit. I thought rule number 1 was "Never Hit ANYTHING."

2) On most 2-way roads that have no painted center line it still drives almost exactly in the center of the road unless a car is coming from the opposite direction at which time it moves over until the car passes. I know I know nothing but it just seems easy to conclude if you're on a 2-way road with no painted center line and you know exactly where the road edges are that the default location should be centered between the road edge and an imaginary center line. Strangely this problem doesn't happen all the time. Sometimes it gets it right. But usually not.

Am I the only one seeing these problems?
 
Yeah, that seems to match my experience with 11.4.2 now showing upcoming intersections in the visualization even though Vision alone shouldn't have been able to guess the lanes and geometry with such confidence and accuracy:

11.4.2 map lanes.jpg


This yard continues slightly uphill with more houses and fences blocking the view, but FSD Beta was able to visualize 2 intersections ahead correctly reflecting 2 lanes that merge after the intersection, double yellow lines next to the main road's left-turn lane vs dashed yellow for the center turn lane, and even the bike lanes on each side.

The example from Twitter is even more evidence of map data being used where in that case outdated maps result in incorrect visualizations from further away that Vision is able to correct when closer.

Compared to earlier 11.3.x experiences from even a month ago, I've noticed about half of our improvements were from updated navigation data (that can be updated every trip not even requiring a full map update) and other half requiring 11.4.2 software update. This definitely allows Tesla to make improvements to FSD driving behavior closer to real time.
 
Yeah, that seems to match my experience with 11.4.2 now showing upcoming intersections in the visualization even though Vision alone shouldn't have been able to guess the lanes and geometry with such confidence and accuracy:

View attachment 944053

This yard continues slightly uphill with more houses and fences blocking the view, but FSD Beta was able to visualize 2 intersections ahead correctly reflecting 2 lanes that merge after the intersection, double yellow lines next to the main road's left-turn lane vs dashed yellow for the center turn lane, and even the bike lanes on each side.

The example from Twitter is even more evidence of map data being used where in that case outdated maps result in incorrect visualizations from further away that Vision is able to correct when closer.

Compared to earlier 11.3.x experiences from even a month ago, I've noticed about half of our improvements were from updated navigation data (that can be updated every trip not even requiring a full map update) and other half requiring 11.4.2 software update. This definitely allows Tesla to make improvements to FSD driving behavior closer to real time.
Yeah likely in future releases we will see improved recall and precision for various tasks that partly are due to better use of map data. Likely Tesla is building up a dataset with labels for when to rely more on map data and when to rely less on map data so the neural network will learn to ignore incorrect map data but take advantage of correct map data. And likely they are setting up the entire mapping stack to automatically update the map whenever the car sees things that disagree with the map. Guess it's good that they will 100x their compute in the next 2 years.

There will be some pointless debate around "Elon said HD maps suck but now they realized that they need to use HD map and are years behind Waymo" etc. Imo this is not HD map, this is a very data compact representation that is used to augment what the neural network perceives. HD maps are perfect maps that the car intends to drive on that it localizes itself within and place all dynamics into. Maybe the difference is pretty small in practice... Hmm... :-/ *scratches beard*

You could argue that HD maps are a lot larger than semantical maps, but there are ways to compact them. And over time Tesla's compact map might grow in data size. Hmm...

You could argue that HD maps uses localization in the map, while Tesla just feeds the map into the neural network and let's it do the positioning magically. In the long run I expect localization to become done by neural networks rather than particle filters etc. So guess they are converging here also...

Intuitive I like Tesla's direction of reaching it, but need to think a lot more about the theoretical differences between the approaches.
 
I’m on a long road trip. Today in eastern Montana eastbound no cars near me for miles, FSD repeatedly executed phantom braking. Sometimes merely hard and once very hard. Sometimes accompanied by an unwarranted lane change. Otherwise, 11.4 is pretty good.
 
  • Informative
Reactions: JRP3
Despite the significant improvements in FSD up to and including 11.4.2 there are 2 problems that never seem to get corrected.

1) I live in a gated community and the "gate" is a single pole that lowers much like on a railroad crossing. FSD never reacts to the gate in any way at all. There is no question it would crash right through it every time if I didn't intervene. The display also has no visual indication that it sees anything at all. I can understand that it might not have been trained yet to recognize a pole across the road but I can't understand how it doesn't know that there is SOMETHING there that it shouldn't hit. I thought rule number 1 was "Never Hit ANYTHING."

2) On most 2-way roads that have no painted center line it still drives almost exactly in the center of the road unless a car is coming from the opposite direction at which time it moves over until the car passes. I know I know nothing but it just seems easy to conclude if you're on a 2-way road with no painted center line and you know exactly where the road edges are that the default location should be centered between the road edge and an imaginary center line. Strangely this problem doesn't happen all the time. Sometimes it gets it right. But usually not.

Am I the only one seeing these problems?
My gated community has a tall double swing gate at our entrance. It's on a slight corner. Even when the Homelink manages to open the gate my MY speeds up and will run into it if I don't hit the brakes. Occasionally it will brake hard just before I slam on the brakes. I'm on 11.4.2 and it almost seems worse than on some of the prior verions.
 
  • Informative
Reactions: JRP3
I’m on a long road trip. Today in eastern Montana eastbound no cars near me for miles, FSD repeatedly executed phantom braking. Sometimes merely hard and once very hard. Sometimes accompanied by an unwarranted lane change. Otherwise, 11.4 is pretty good.
My MY is on 11.4.2 an on an 82 mile trip today there were 3 instances of phantom braking. Two of them very hard braking. The prior version of FSD didn't have any phantom braking.
 
After driving more on 23.12.10 it’s a rough experience on busy streets in the city. Lots of phantom breaking and not good merging. it’s incredible on the freeway in stop and go traffic, but on the streets there’s just too much for it to function smoothly. I have enjoyed the gas/brake autonomy and I’ll steer the car. Sometimes it will lag in acceleration and I just have to push the pedal a little to get it going again.