Do feel like your autopilot learns over time?

[timk225]

Is one of the functions of autopilot to learn over time, as in learn the roads you travel on often?

I finally jumped on the AP package for $2000 when it was on sale in early March. I've got things pretty well set, and it now does most of the driving when I go to work, which involves going into and out of the city, each way. I can get on the highway, put AP on, and I only have to take over for a mile or 2 when going around downtown Pittsburgh, to get me off one highway, off the correct off ramp, and onto the next one for the second highway of the trip to/from work.

When driving home from work, part of the highway has a rather tight curve on it, that I approach around 75 mph. And it also happens to have a big railroad bridge right over it, so lots of heavy steel there to mess with the radar. For the first week or two, AP would hit the brakes briefly when going under the bridge, but now, it doesn't seem to do it as much.

And there are other parts of the highways going to / from work, where it used to be more sensitive, but now it seems to handle it better.

I'm just wondering what degree of machine learning it is capable of doing. I presume it will never do some things totally right, but it can learn over time.

 

[Big Earl]

The system only seems to change when software updates are received. Autopilot "improves" with each software update, although sometimes its behavior can seem to regress due to expansions of the neural network expands the things it reacts to.

I have not noticed any sort of learning or behavior changes in between software updates.

 

[GoSharks]

No, it doesn't. Only with software updates from the mothership.

 

[woodisgood]

Every Model 3/AP should behave exactly the same on any given road. At least that’s what Tesla wants.

Therefore, it would run counter to that to have vehicles with differing abilities out there, even if it were feasible (and it is) for individual cars to do their own learning.

One area where I think it makes sense to start individualized vehicle memory/learning is with Summon and an owner’s home garage/driveway.

 

[StellarRat]

Well, sort of, I have noticed changes in its speed settings as Tesla has gotten more data about the speed limits around my town. At first it tended to speed A LOT. Now there far fewer of these mistakes in my area. However, it will still get you a ticket if you don't keep an eye on it.

One of the big problems is still getting good maps from government. I've had the nav system try to route me through a sheep pasture. There was probably a road there are one OR the planners draw in a road, but it was never built. I have to assume that speed limits in various areas are even less accurate. Tesla should probably crowd source those from the average speed that its cars are driving on streets.

 

[Enginerd]

My understanding is that all neural network updates are considered supervised learning based on fleet data, and those are periodically transmitted from HQ as software updates. I believe that initially with a new car, however, there is some camera calibration taking place. I had several phantom braking / false positives during my first few hundred miles. That behavior improved significantly, before I received my first software update.

 

[StellarRat]

My understanding is that all neural network updates are considered supervised learning based on fleet data, and those are periodically transmitted from HQ as software updates. I believe that initially with a new car, however, there is some camera calibration taking place. I had several phantom braking / false positives during my first few hundred miles. That behavior improved significantly, before I received my first software update.

There is a spot on one state highway I drive on that ALWAYS messes up on EAP. It thinks it needs to slow to 25 mph there even though the road is 55 all the way. Coming from the other direction it does not do this. Apparently, even with a bunch of Teslas crossing this area it still hasn't learned. I've learned to anticipate it and just put my foot on the accelerator at that spot. It's probably only 100' of road in the whole highway.

 

[JeffK]

From everything I've read the training takes places only at Tesla using data retrieved from the vehicles. Individual cars don't do training.

 

[spidey]

From everything I've read the training takes places only at Tesla using data retrieved from the vehicles. Individual cars don't do training.

A very basic example of individual cars ‘learning’ is the air suspension in s and x. If there’s a point in the road where u manually raise the car up it remembers that and does it automatically the next time. It also tells the hive mind so in future update other Tesla’s do same thing. Like I say on car learning is very basic compared to the software update knowledge but there are little bits there.

 

[NoMoGas]

OK some common misconceptions...

YES it does learn over time through "fleet learning" and NO it doesn't require a Firmware update. The Firmware (version software you have) is separate from AP learning and most updates (new features excepted) If you got a firmware update for every curve- correction you'd be installing every day.

For example, those that purchased AP after they took delivery should notice their AP just started working even though no software update was installed.

 

[Resist]

It definitely does not learn over time, as many believe. Each car doesn't have it's own ability to "learn". The software via updates, is the only thing that gets better.

 

[NoMoGas]

It definitely does not learn over time, as many believe. Each car doesn't have it's own ability to "learn". The software via updates, is the only thing that gets better.

Sorry, I should have been clearer. No, the car dosn't but the cloud it's attached to does. That is updated without software updates routinely.

 

[camalaio]

OK some common misconceptions...

YES it does learn over time through "fleet learning" and NO it doesn't require a Firmware update. The Firmware (version software you have) is separate from AP learning and most updates (new features excepted) If you got a firmware update for every curve- correction you'd be installing every day.

For example, those that purchased AP after they took delivery should notice their AP just started working even though no software update was installed.

Source?

AP starts working after post-delivery purchase because it's a software feature flag (on/off switch) and the car already has the hardware and software to perform AP tasks. Knowing companies employing various "learning" methodologies, no one in their right mind would be releasing daily tweaks. They're likely evaluated, assigned, grouped in a build, thoroughly tested for behavioural regressions, then finally released once tests have passed. Even then, releases are staged in groups (a small number of people receive the update first, and if all goes well a larger group receives the update).

The car cannot learn on its own as it has no concept of if the reason you cancelled AP was because AP did something wrong (and what it did wrong) or if you just wanted to turn it off or direct it somewhere else.

EDIT: To actually participate in the thread, I haven't noticed any change in AP behaviour in the months we've owned the Model 3. I have noticed it does much better in the dark with fewer vehicles on the road and higher contrast to see the lane lines with.

 

[NoMoGas]

Source?

AP starts working after post-delivery purchase because it's a software feature flag (on/off switch) and the car already has the hardware and software to perform AP tasks. Knowing companies employing various "learning" methodologies, no one in their right mind would be releasing daily tweaks. They're likely evaluated, assigned, grouped in a build, thoroughly tested for behavioural regressions, then finally released once tests have passed. Even then, releases are staged in groups (a small number of people receive the update first, and if all goes well a larger group receives the update).

The car cannot learn on its own as it has no concept of if the reason you cancelled AP was because AP did something wrong (and what it did wrong) or if you just wanted to turn it off or direct it somewhere else.

EDIT: To actually participate in the thread, I haven't noticed any change in AP behaviour in the months we've owned the Model 3. I have noticed it does much better in the dark with fewer vehicles on the road and higher contrast to see the lane lines with.

Point is that "flag" is done without a software update. It's remotely activated in the background much as the "tweaks" are (when x number of people correct the car in the same spot etc) Much like when the AP is polled for images of cut-ins, bicycles and the like. It's a two way communication that happens in the background as is corrections. The source is Elon explaining how the system works on the initial roll out in 2014 as well as the autonomy day event. He even pointed out a specific thing the car did on the 405 that it just stopped doing one day... no software update changed the behavior.

Here is an AP user who noticed his car learned to stop taking exits... no software update

 

[camalaio]

Point is that "flag" is done without a software update. It's remotely activated in the background much as the "tweaks" are (when x number of people correct the car in the same spot etc) Much like when the AP is polled for images of cut-ins, bicycles and the like. It's a two way communication that happens in the background as is corrections. The source is Elon explaining how the system works on the initial roll out in 2014 as well as the autonomy day event. He even pointed out a specific thing the car did on the 405 that it just stopped doing one day... no software update changed the behavior.

Here is an AP user who noticed his car learned to stop taking exits... no software update

A video from 2015 (is that AP1?) isn't relevant anymore, nor any statements regarding AP1. AP1 behaved very differently from what's on the Model 3 (in some ways better, some ways worse). Newer AP is likely (certainly anecdotally) making very "paranoid" decisions about how to navigate, and is likely being carefully controlled by Tesla now if it wasn't before. Think of the liability issues of a car "learning" behaviour on its own, completely unverified from the company. Who is at fault if a car crashed because of something it learned to do by itself? As a society we haven't even figured that out for intentionally programmed behaviours, at least not in a way everyone agrees on.

 

[NoMoGas]

A video from 2015 (is that AP1?) isn't relevant anymore, nor any statements regarding AP1. AP1 behaved very differently from what's on the Model 3 (in some ways better, some ways worse). Newer AP is likely (certainly anecdotally) making very "paranoid" decisions about how to navigate, and is likely being carefully controlled by Tesla now if it wasn't before. Think of the liability issues of a car "learning" behaviour on its own, completely unverified from the company. Who is at fault if a car crashed because of something it learned to do by itself? As a society we haven't even figured that out for intentionally programmed behaviours, at least not in a way everyone agrees on.

Thee autonomy day event was 2019. So no, not AP1 only. I never said it's not verified by the company. I said it's updated without a firmware update.

There are two very different systems on the car and the firmware version doesn't necessarily change AP behavior. It may add functionality, but how it handles a specific intersection based on error reports (driver take over events) can and are improved in the background without firmware updates.

 

[camalaio]

Thee autonomy day event was 2019. So no, not AP1 only. I never said it's not verified by the company. I said it's updated without a firmware update.

There are two very different systems on the car and the firmware version doesn't necessarily change AP behavior. It may add functionality, but how it handles a specific intersection based on error reports (driver take over events) can and are improved in the background without firmware updates.

If it is actually doing something aside from updates reliably, cool I guess? Though wasn't the whole goal of AP to be general driving autonomy, not based on map data like the other self-driving endeavours? If they're still following this goal, I don't see how it could collect info on specific locations to improve behaviour at that location and not impact how it drives elsewhere.

 

[NoMoGas]

If it is actually doing something aside from updates reliably, cool I guess? Though wasn't the whole goal of AP to be general driving autonomy, not based on map data like the other self-driving endeavours? If they're still following this goal, I don't see how it could collect info on specific locations to improve behaviour at that location and not impact how it drives elsewhere.

You are correct here, sort of. They are using GPS for certain things but not the primary at least thats what I got from the AD presentation. There are still some quirky things that require GPS as a band aid, but then if something changes there (road work etc) it cant be GPS dependent. So the idea is to train the car for "like events" . The point is that some of these don't require a firmware update, but when they do a firmware update they tend to update AP as well... Geez I hope that makes sense?

 

[camalaio]

You are correct here, sort of. They are using GPS for certain things but not the primary at least thats what I got from the AD presentation. There are still some quirky things that require GPS as a band aid, but then if something changes there (road work etc) it cant be GPS dependent. So the idea is to train the car for "like events" . The point is that some of these don't require a firmware update, but when they do a firmware update they tend to update AP as well... Geez I hope that makes sense?

Yeah I'm following. Some have suggested known overpasses prevent braking for an example of positional awareness. This is just the first time I've seen confident statement about updating/"learning" outside of updates. There are a lot of advantages to rolling it out via updates, but perhaps it's being version controlled separately (in a way we don't see) from other firmware.

 

[argon2018]

Is one of the functions of autopilot to learn over time, as in learn the roads you travel on often?

I finally jumped on the AP package for $2000 when it was on sale in early March. I've got things pretty well set, and it now does most of the driving when I go to work, which involves going into and out of the city, each way. I can get on the highway, put AP on, and I only have to take over for a mile or 2 when going around downtown Pittsburgh, to get me off one highway, off the correct off ramp, and onto the next one for the second highway of the trip to/from work.

When driving home from work, part of the highway has a rather tight curve on it, that I approach around 75 mph. And it also happens to have a big railroad bridge right over it, so lots of heavy steel there to mess with the radar. For the first week or two, AP would hit the brakes briefly when going under the bridge, but now, it doesn't seem to do it as much.

And there are other parts of the highways going to / from work, where it used to be more sensitive, but now it seems to handle it better.

I'm just wondering what degree of machine learning it is capable of doing. I presume it will never do some things totally right, but it can learn over time.

No it's not AI. Neural net (AP's brain) gets better with every software update which is the same for all cars. Neural net is manually programmed.