Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Does FSD learn?

This site may earn commission on affiliate links.
Does the FSD Beta system learn as it goes? Since the 2022.45.15 upgrade my car has consistently missed a rural exit off the interstate. On the nav display it showed that the exit was 200 feet further down the road than it is in reality. I would disengage, take the exit and report. I did this every day for weeks. Then a few days ago it started taking the exit flawlessly every time. AFAIK the car hadn't received any updates.
 
  • Informative
Reactions: zoomer0056
Does the FSD Beta system learn as it goes? Since the 2022.45.15 upgrade my car has consistently missed a rural exit off the interstate. On the nav display it showed that the exit was 200 feet further down the road than it is in reality. I would disengage, take the exit and report. I did this every day for weeks. Then a few days ago it started taking the exit flawlessly every time. AFAIK the car hadn't received any updates.
Some people swear that it does!

In my experience, things only improve after a software update, not instant learning from a current version.

That means it couldn't competently take a particular curve previously (curves of death), but many months or years after an update, it now can.

On the other hand, there's a stop sign that my FSD consistently ignores since I bought FSD in 2017. I gave up "teaching" it. You are talking about weeks, and I am talking about years!
 
The car isn't learning. When we enter a destination, the planning actually comes from Tesla, and that includes lots of details about the route. So Tesla can be updating the information that they have at that central location and our cars just start doing something new as a result.
 
My problem with stop signs is that it trys to stop at two different stop signs on a rural road that are meant for the cross traffic. The roads cross at an extremely acute angle so the sign is angled toward the oncoming lanes.

To be honest, even a rookie human driver would be confused. I'm tempted to go out a night with a big wrench and a cheater and twist those sign posts.
 
I've had similar experiences where, for example, my MY would try to turn into a shopping center entrance cut in the median that was several hundred feet short of the actual turn. Magically, after 4 or 5 times of my correcting it--it now does it perfectly. I strongly suspect there is some element of local learning. Perhaps part of Tesla's "secret sauce."
 
Does the FSD Beta system learn as it goes? Since the 2022.45.15 upgrade my car has consistently missed a rural exit off the interstate. On the nav display it showed that the exit was 200 feet further down the road than it is in reality. I would disengage, take the exit and report. I did this every day for weeks. Then a few days ago it started taking the exit flawlessly every time. AFAIK the car hadn't received any updates.
No updates leads me to believe your reporting did something. Either the car learned or some sort of stealth update happened. We can only speculate but there is an answer we don't know.
 
It's the map data that's changing, not the car learning.
That makes sense. These are issues with the map and routing, and should be fixed for all users, not just the one car which reports it. So the changes should be on map and routing servers.

I don't know how detailed the in-car map is. I don't know where the routing is calculated. In any case it would make sense to have a base map on board, and to download details and changes over the air as needed. This is how Google Maps on smart phones works. There is a low resolution world map which is available without cellular or WiFi, and GPS can locate the position on that map. With a data connection or a pre-downloaded detail map file, all the details are available. Perhaps Tesla follows a similar approach.

This would explain the improvement posters here have described, without contradicting the general understanding that our cars do not learn.

I, for one, wish the car did learn, or at least remember stuff about my common routes. But that will have to wait for some future update. ;-)
 
Has there been any official statement from Tesla or otherwise as to whether the car actually "learns" anything while we drive?

I suspect that the only learning (changes to AI network) is done by Tesla corporate computers, and the only way it gets back to our cars is by SW update, but I see many comments where people seem to think their car is getting better as it experiences certain situations over again.

I think the advertising phrase "End to end AI" gave the impression the cars were learning on their own.
 
I think the advertising phrase "End to end AI" gave the impression the cars were learning on their own.
I think the phrase was "end to end neural networks" which replaced 300,000 lines of hand written software code. As you say, the NN training happens in huge data processing centers and is downloaded to the cars in the periodic software updates. The cars do not update the NN themselves.

There is, however, feedback from the cars to Tesla's data centers, and this is used both in NN training and map updates. So the cars are leaning from their experiences, but only in a collective sense.

As discussed earlier in this TMC thread, some map data is downloaded to the car while driving as part of the navigation process, to a car's behavior can change from one day to the next without a full software update. But this does not imply the car learned anything by itself.

Also as I have said before, I hope Tesla does eventually let the cars learn. For example, mine has some favorite potholes it always bumps through, so I have to disengage AP for a couple seconds to manually steer around them. This happens every day till the city comes out and patches the road. The car could learn, like we do, from the first hard bounce.
 
  • Informative
Reactions: Ben W and rjpjnk
Between firmware updates 2024.8.7 and 2024.14.6 my car finally figured out how to take me to the grocery store. I had long ago given up on this. I tried it again because some new roads showed up on a route I take once a week. I was very excited that FSD might finally take me to the grocery store! Unfortunately, the car forgot the correct route during or before the 2024.14.6 update.

My conclusions:
  1. The car will update navigation data between firmware updates
  2. This is not related to learning from the disengagements
 
As you say, the NN training happens in huge data processing centers and is downloaded to the cars in the periodic software updates. The cars do not update the NN themselves.
But do we know this for a fact or is it just our best working assumption to date? Has anyone with access to the software confirmed there is no local behavioral learning or adaptation? (Not counting map updates)
 
But do we know this for a fact or is it just our best working assumption to date? Has anyone with access to the software confirmed there is no local behavioral learning or adaptation? (Not counting map updates)
NEVER has been. Consider this. If your car "learns" more it could become a smarter Tesla and more valuable. City driven FSD Tesla's would have more value/smarts than country driven Tesla's. All Teslas drive the same and every time you pull out of your driveway it its the "first time" your car has ever seen that street.
 
  • Like
Reactions: Ben W and ucmndd