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

Autopilot is already improving.

This site may earn commission on affiliate links.
Unless the company claims it, it is only pure speculation that AP is learning. We did learn that TM was monitoring charge times and frequencies and used that information to address range issues and perceived range anxiety...

Well they have been claiming itfrom the beginning:

These mutually reinforcing systems offer realtime data feedback from the Tesla fleet, ensuring that the system is continually learning and improving upon itself.

Your Autopilot has arrived | Tesla Motors
 
I have noticed several articles over the past few days about how AP is designed to "learn."

Couple examples here:

How Tesla's autopilot learns - Fortune

Tesla is mapping out every lane on Earth to guide self-driving cars

Based on this info, it's reasonable to assume that AP will improve over time based on the fleet's driving habits in a particular area. I hope that's the case at least!! Regardless, the detailed mapping could mean a lot for not only the future of AP, but also the navigation system. And then perhaps the two merge over time.... :)
 
The only way this would be true is if

A.) Tesla can download new GPS map information to the car without doing an SW firmware update. It currently can download information (like a new supercharger network) so there is precedence.

Or

B.) Despite Elon's fear some Engineer snuck a deep learning network into the cars software.
 
There is indeed a deep knowledge network in the cars now. The Mobileye system that Tesla's auto steer is based on uses an extensive deep knowledge network to decode visual scene information. And there are reports out there that Tesla has 100+ engineers working on auto pilot. So, if this is true, those engineers have to be doing something, and part of that something could be putting in real time learning algorithms into the Mobileye system.

It isn't inconceivable, just more than I personally thought they were doing.
 
I wasn't going to start a post on it, but I've been noticing it too. On Day 1, I had to rescue the car a lot. For example, on my typical commute to my home office is 23.5 miles and includes mostly 4- to 5-lane divided freeway and congested city streets in SF. I had the following counts on the identical route, identical lane choices (I'm a creature of habit), and times of day:

Day 1: 6 rescues
Day 2: 5 rescues
Day 3: Weekend
Day 4: Weekend
Day 5: 3 rescues
Day 6 (today): 1 rescue

Today, in San Francisco on Bryant Street after the I-80 onramp, I was getting ready to rescue the AP because the lanes split as it freaked out every other time. Today, it didn't ... it stayed in the center lane and drove straight, despite the split. I was very puzzled ... is this car learning???

- K
 
I have noticed several articles over the past few days about how AP is designed to "learn."

Couple examples here:

How Tesla's autopilot learns - Fortune

Tesla is mapping out every lane on Earth to guide self-driving cars

From that Forbes article: When it comes to the autopilot software, Musk explained that each driver using the autopilot system essentially becomes an “expert trainer for how the autopilot should work.”

So, assuming this isn't hyperbole ******** from Elon (sorry, Elon, but you've tarnished your trustworthiness with way out there statements in the past), they are indeed using how humans drive to train autopilot.
 
And, I'm putting myself out there, for free, for any magazine, blog, or newspaper than actually wants a detailed article on how Autopilot really works. I can be your technical advisor and ask the right questions...
 
The only way this would be true is if

A.) Tesla can download new GPS map information to the car without doing an SW firmware update. It currently can download information (like a new supercharger network) so there is precedence.

Or

B.) Despite Elon's fear some Engineer snuck a deep learning network into the cars software.

I'm pretty sure they would make the GPS map information able to update without a firmware update. I've had various offline navigation apps on my phones for the past 5+ years. They've been able to update the maps without having to reinstall the app itself.
 
This is machine learning more than AI. The issue I have with these reports is that the Autopilot model can't be operating at HQ and steering from there. Whatever model is built from the data must be then sent again to the vehicle. If indeed there is noticeable day-to-day improvement, it implies that Tesla sends updates to the vehicle without actually doing a software update. That is, some sort of background model update.

When does this happen? Anyone want to throw a sniffer on their network and see if it's occurring overnight?
 
This is machine learning more than AI. The issue I have with these reports is that the Autopilot model can't be operating at HQ and steering from there. Whatever model is built from the data must be then sent again to the vehicle. If indeed there is noticeable day-to-day improvement, it implies that Tesla sends updates to the vehicle without actually doing a software update. That is, some sort of background model update.

When does this happen? Anyone want to throw a sniffer on their network and see if it's occurring overnight?

Paging wk057
 
Anyone want to throw a sniffer on their network and see if it's occurring overnight?

You already know this happens. Elon said it during the press release in the answer to the question, "What happens if there is no cell coverage?" Elon answered that the car would save the data from that portion then upload the data as soon as it was back in cell coverage. To me, this implies there is some kind of continuous upload of data, vs. once a month because the amount of data the car would have to hold until then could be quite significant.

- K
 
You already know this happens. Elon said it during the press release in the answer to the question, "What happens if there is no cell coverage?" Elon answered that the car would save the data from that portion then upload the data as soon as it was back in cell coverage. To me, this implies there is some kind of continuous upload of data, vs. once a month because the amount of data the car would have to hold until then could be quite significant.

There's a distinction to be made between the uploaded data (raw training data) and the downloaded model. Generally in this kind of problem, you take the training data and use it to build a predictive model. This model, in the case of Autopilot, controls the vehicle. The raw data is uploaded continuously, and that's fine. Models aren't built and downloaded continuously; that just doesn't make sense. So if it's being pushed silently, it's happening at a time that the vehicle is not in use. I'm guessing it's at nighttime, if at all.

Hopefully that clears it up a bit.
 
My day job is a porn star, why would I ever quit that!

TMI!!!

To get back on track, I also feel like I've had to override autopilot less since installing it, although I probably have a lot fewer autopilot drives and miles than many of the rest of you.

Wish there was a way to know for sure if this was happening, or if it's just a huge case of wishful thinking.