Firmware Update 2018.12

[rnortman]

Bottom line it seems when there is a new up date, Give it a day.... The NN will sort itself out.

I am highly skeptical of this idea that the NN is "learning" anything in an individual car. I think all the evidence we have is that a fixed NN is delivered to the car from Tesla, and there is no ability for the NN to adjust its own weights or learn outside of Tesla shipping a new NN in a new software update. And everybody (on that hardware configuration) gets the same NN.

We do know that the vehicles are capable of calibrating their own cameras. It is possible that with a software update, Tesla changes the calibration procedure and the cars silently recalibrate themselves during the first few days of driving. I'm pretty skeptical of this also. The sorts of behaviors that might change due to calibration are persistently hugging one side of the lane or another, or seeing cars ahead in your lane as being in a different lane, or vice versa. Sensor calibration would not change the actual driving or detection algorithms, just make the car's estimation of where things are more accurate.

It is possible that in fact these new updates have shipped with two sets of algorithms, one that is disabled until a sensor recalibration is completed, and when then would kick in as soon as recalibration is complete. I don't think we have any evidence that this is happening, and frankly it would probably be a lot of work for Tesla to do something like that, so I really doubt it. More likely they would ship the new calibration procedure in one update and then ship the code that makes use of it in the next.

 

[conman]

I am highly skeptical of this idea that the NN is "learning" anything in an individual car. I think all the evidence we have is that a fixed NN is delivered to the car from Tesla, and there is no ability for the NN to adjust its own weights or learn outside of Tesla shipping a new NN in a new software update. And everybody (on that hardware configuration) gets the same NN.

I still believe this is the case - there is no suggestion that each car is adapting and learning as it goes along. Your driving isn't training your car's neural network; I'm inclined to believe your car's neural network is training your driving. You get used to the feel of it and it it doesn't seem so bad... just stick another driver in and see what they think. Putting my wife, who only infrequently drives my car, into the driver's seat is always a good sanity check.

 

[boonedocks]

Ok so I'm not alone. Sucks because AP in stop and go is usually solid.

You are definitely not alone. I have had the same behavior since 2018.10 making AP unusable for parts of my daily commute.

On a side note it’s coming up on the NEXT “NEXT WEEKEND” and still no New Nav or 2017.12 for either of our ‘17 S or X

 

[MacGreiner]

On a side note it’s coming up on the NEXT “NEXT WEEKEND” and still no New Nav or 2017.12 for either of our ‘17 S or X

Yup.

 

[croman]

I still believe this is the case - there is no suggestion that each car is adapting and learning as it goes along. Your driving isn't training your car's neural network; I'm inclined to believe your car's neural network is training your driving. You get used to the feel of it and it it doesn't seem so bad... just stick another driver in and see what they think. Putting my wife, who only infrequently drives my car, into the driver's seat is always a good sanity check.

I agree that the effect is likely human psychology but its a tough concept to internalize because I've had various builds that did *noticeably* worse and then improved. It was so stark (2017.50.2 comes to mind), that it couldn't be simply adjusted expectations...

 

[MorrisonHiker]

Has anyone noticed improved browser speed on 2018.12? Apparently one person with MCU2 have also gotten this version, on up until now MCU2 have been running a different browser versions. I am wondering if MCU1 and MCU2 are now running the same browser version, and if this has improved the browsing speed?

https://jsfiddle.net/qja8wjxc/95/show/

I didn't notice any change in the browser. I was able to test some websites that work on MCU2 but they still don't work on the original MCU with 2018.12

 

[boaterva]

You are definitely not alone. I have had the same behavior since 2018.10 making AP unusable for parts of my daily commute.

On a side note it’s coming up on the NEXT “NEXT WEEKEND” and still no New Nav or 2017.12 for either of our ‘17 S or X

Of course, it's always 'next weekend' ("tomorrow is always a DAY away....." (where's that music graphic..... ))

 

[ranma64]

Hi, i am sure AP 2.0 is now learning..or something similar. I tried many times on my commute to engage autopilot on a freeway ramp but was not able to keep the curve angle. After the fourth try it started to remain engaged until the first half...then 3/4 then finally is able to run all the ramp.
Made a video.

 

[Bebop]

Are the new maps a seperate update?

I got .12 at the service center but don’t have the maps. Does it magically pop up at some point? Or do you actually get a notification for a software update?

 

[MorrisonHiker]

Are the new maps a seperate update?

I got .12 at the service center but don’t have the maps. Does it magically pop up at some point? Or do you actually get a notification for a software update?

Yes. Yes. No.

 

[boaterva]

Yes. Yes. No.

You get a notification in the car. Time to start sleeping in the car like Bjorn!

 

[dcmetro]

Are the new maps a seperate update?

I got .12 at the service center but don’t have the maps. Does it magically pop up at some point? Or do you actually get a notification for a software update?

Yeah, that's exactly what my local SC rep told me. Said that new maps have some massive files that may hit ours cars at some point, after the .12 update. Had no idea when..

 

[MorrisonHiker]

You get a notification in the car. Time to start sleeping in the car like Bjorn!

Yeah, once you get back in the car, there will be a message indicating the maps were updated. You don't get a push notification in the app as you do with a firmware update.

 

[boaterva]

Yeah, that's exactly what my local SC rep told me. Said that new maps have some massive files that may hit ours cars at some point, after the .12 update. Had no idea when..

Been six days for me... mapless! So, don't start a stopwatch, it may take some time!

 

[MorrisonHiker]

Been six days for me... mapless! So, don't start a stopwatch, it may take some time!

4 days for me. I've been checking the router logs but nothing really happening. Only 150 kb downloaded in the last 2.5 hours...so no map downloads yet!

 

[bebbiXpress]

Hi, i am sure AP 2.0 is now learning..or something similar. I tried many times on my commute to engage autopilot on a freeway ramp but was not able to keep the curve angle. After the fourth try it started to remain engaged until the first half...then 3/4 then finally is able to run all the ramp.
Made a video.

I watched the video and despite the learning argue, you're going 30km/h or nearly 50% over the speed limit in a curve and then you complain that the car goes too much to the outside or even over the line of that curve? Sorry, but my experience says it would perfectly stay in the lane if you take that curve with the allowed 70km/h. Human drivers or Autopilot, for both it's more challenging to take such at 100km/h than by 70.
Because of the speedup of the video it's hard to see the 70 sign and so some folks seeing such a video not familiar with Autopilot could think it's not safe and should maybe even forbidden.

 

[luckyj]

Hi, i am sure AP 2.0 is now learning..or something similar. I tried many times on my commute to engage autopilot on a freeway ramp but was not able to keep the curve angle. After the fourth try it started to remain engaged until the first half...then 3/4 then finally is able to run all the ramp.
Made a video.

Be careful with calling AP handing the same stretch of road differently at different times as "learning". There are many variables that can affect it's performance. Speed, position in lane, objects in/near lane, weather, lighting, etc etc.

 

[rnortman]

Hi, i am sure AP 2.0 is now learning..or something similar. I tried many times on my commute to engage autopilot on a freeway ramp but was not able to keep the curve angle. After the fourth try it started to remain engaged until the first half...then 3/4 then finally is able to run all the ramp.
Made a video.

I'm sorry, but you're mistaken. This is not how machine learning works. What happens on your car in real time is inference, not learning. The learning happens offline. There is almost zero value to learning in an individual car. Why would Tesla learn in an individual car? They learn across the whole fleet, and everybody gets the same NN. And the NNs only update during software updates. Even if there is learning happening in each car, typical machine learning algorithms (excepting a few experimental and limited techniques) require hundreds or thousands of examples to learn from; it would not improve over the course of four attempts. Find me hard evidence otherwise; anecdotes don't cut it.

Except perhaps learning maps. Tesla has dabbled in that in the past. We still have no evidence that they're doing it in any sort of scale outside of specific areas (mostly in California). It is possible that a map update may have been (silently) delivered to your car that improved its performance on that specific route. This would probably happen overnight rather than while you were driving. But this is also highly unlikely, we have at best very tenuous evidence that the autopilot mechanisms use GPS data right now beyond categorizing the type of road you're on (limited access freeway, divided highway, local road) and determining the speed limits.

In your car, perhaps there were different weather, lighting, or traffic conditions. Maybe your windshield was more or less dirty. Very, very small differences that are not perceptible to you can have a large impact on the performance of the NN in any given situation. It is likely that these very small differences explain all of the anecdotal evidence that learning is happening in real time on vehicles.

 

[alcibiades]

I'm sorry, but you're mistaken. This is not how machine learning works. What happens on your car in real time is inference, not learning. The learning happens offline. There is almost zero value to learning in an individual car. Why would Tesla learn in an individual car? They learn across the whole fleet, and everybody gets the same NN. And the NNs only update during software updates. Even if there is learning happening in each car, typical machine learning algorithms (excepting a few experimental and limited techniques) require hundreds or thousands of examples to learn from; it would not improve over the course of four attempts. Find me hard evidence otherwise; anecdotes don't cut it.

Except perhaps learning maps. Tesla has dabbled in that in the past. We still have no evidence that they're doing it in any sort of scale outside of specific areas (mostly in California). It is possible that a map update may have been (silently) delivered to your car that improved its performance on that specific route. This would probably happen overnight rather than while you were driving. But this is also highly unlikely, we have at best very tenuous evidence that the autopilot mechanisms use GPS data right now beyond categorizing the type of road you're on (limited access freeway, divided highway, local road) and determining the speed limits.

In your car, perhaps there were different weather, lighting, or traffic conditions. Maybe your windshield was more or less dirty. Very, very small differences that are not perceptible to you can have a large impact on the performance of the NN in any given situation. It is likely that these very small differences explain all of the anecdotal evidence that learning is happening in real time on vehicles.

Sometimes my X fails to handle something well that it handled well in the past. The NN must be forgetting.

 

[suraj1194]

I haven't seen any low speed Tokyo Drift behavior from the car with 2018.12 . It's been very stable on my long (100+ mile) drives recently.

However, my 17" screen has rebooted TWICE since the new maps downloaded on Tuesday. First time, it rebooted on me just as I touched the screen after getting into the car and seeing the notice on screen saying I've the new maps. The second time was last night - I tried to access the car from the app, but there was no response. Went into the garage and the dash display says 'Touchscreen unresponsive. Press scroll wheels on steering wheel to restart'. This is pretty odd for me. I've pretty much never seen this happen before - especially the dash display last night, which is the first time I've seen that message.