Software Update 2018.42.x

[caligula666]

Bladerskb’s point is that ML is not happening on the hardware in the car AND given how little data is sent back to the mothership (see other threads in Autonmous Car section), not much can really happen either. So, if that’s true, all of the marketing dance around how many giga-trillions cumulative miles the Tesla fleet is driving is a tad irrelevant.

Bladerskb’s unique argumentation style might be a repellent to some but, unless proven otherwise by real data/analysis/observation, the point (s?)he makes is more than fair IMO.

 

[f205v]

You could just make a Hotspot with your Mobile Phone while Parked somewhere (15-30 Minutes is enough for an ordanary update, not for the Maps!).
This helps a lot! Or park near a public WiFi (MacD for ex.).

I just spent the night at a friend's house, parked in his yard with a wonderful WiFi connection. 12 hours connected. Nothing downloaded.
I really need to contact SC

 

[Drlev]

I just spent the night at a friend's house, parked in his yard with a wonderful WiFi connection. 12 hours connected. Nothing downloaded.
I really need to contact SC

Alternately, the update simply hasn’t been pushed to your VIN range yet.

Though on occasion, updates get hung up to fail repeatedly, and you get stuck in a sort of update limbo.

At any rate, probably reasonable to call your SC.

 

[MacGreiner]

Looks like rollout has slowed to a trickle. Possible revision?

 

[rnortman]

At the end, machine learning is not based on what you do. It is based on what vast majority of people would do under a specific scenario.

Machine learning works best when given an expert-curated corpus of known-good choices or classifications to imitate. Training your model on bad data risks having the model learn bad behavior if that bad behavior is prevalent in the corpus. The "experts" in this case are not machine-learning experts but domain experts. And you may say that everybody is an expert in the domain of driving but most of us are terrible drivers, in a hurry, distracted, and doing unsafe things constantly.

 

[Saimaannorppa]

Bladerskb’s point is that ML is not happening on the hardware in the car AND given how little data is sent back to the mothership (see other threads in Autonmous Car section), not much can really happen either. So, if that’s true, all of the marketing dance around how many giga-trillions cumulative miles the Tesla fleet is driving is a tad irrelevant.

I’m curious what’s the theory on data transfers?

I have reason to believe only selected cars participate. My car was sending 30Mb per hour driven for the first 9 months, and sometimes more. Then I kept the car on energy saving and always connected off during a few weeks of vacation, and it has barely sent anything anymore ever since.

Btw. Larry Ellison hinted they have a massive autonomous driving database in the latest Oracle analyst call, while talking about Tesla in same context.

 

[Anner J. Bonilla]

I’m curious what’s the theory on data transfers?

I have reason to believe only selected cars participate. My car was sending 30Mb per hour driven for the first 9 months, and sometimes more. Then I kept the car on energy saving and always connected off during a few weeks of vacation, and it has barely sent anything anymore ever since.

Btw. Larry Ellison hinted they have a massive autonomous driving database in the latest Oracle analyst call, while talking about Tesla in same context.

Do we have a thread dedicated to data uploads? My car uploads about 100-400mb every day or two. I drive about 45 miles every day.

 

[Pale_Rider]

No, but @verygreen has shown that the triggers they use to determine what is captured and uploaded do often have time limits, so if your car doesn’t see something interesting while that trigger is active you won’t be uploading very much.

 

[TicoMcC]

I got 2018.42.2 last night and can confirm that you can have the camera open on top (albeit in a floating window, not the top half) over the music app. Makes me a happy camper again.

 

[NewTMSMan]

Machine learning works best when given an expert-curated corpus of known-good choices or classifications to imitate. Training your model on bad data risks having the model learn bad behavior if that bad behavior is prevalent in the corpus. The "experts" in this case are not machine-learning experts but domain experts. And you may say that everybody is an expert in the domain of driving but most of us are terrible drivers, in a hurry, distracted, and doing unsafe things constantly.

Yes this, exactly! Alternatively measurement of outcome quality can be used as well, but same issue exists in this scenario.

 

[Kruegmeister]

My 42.2 impressions.
It is much smoother over all. It seems like it improved exponentially to me.
Cars & Trucks on the IC are much more Stable, not Jumpy and hitting my Virtual Car.
Less overreaction to a Car moving into your lane. Less SemiTruck Love.
Drive on Navigation needs to plan to get over to exit much earlier however.
I waited until it hit 1.9 miles before the exit, traffic was too busy for it to get over in time in my opinion & it hadn't suggested that it should so I had to intervine & suggest it on my own...

It was Rainy, bunch of Construction & it handled all of it very well.

 

[SeBsZ]

Why vertical video? Thanks for the info though.

 

[Kruegmeister]

Why vertical video? Thanks for the info though.

Easier to do & pay attention to the Road.
I have the Note 8 & the darn thingbis all screen.
I had another 1 that is sideways, but YouTube is choking on the Processing Step for some reason... hung at 95%

 

[SeBsZ]

Best to get a mount.

 

[MarcusMaximus]

Machine learning works best when given an expert-curated corpus of known-good choices or classifications to imitate. Training your model on bad data risks having the model learn bad behavior if that bad behavior is prevalent in the corpus. The "experts" in this case are not machine-learning experts but domain experts. And you may say that everybody is an expert in the domain of driving but most of us are terrible drivers, in a hurry, distracted, and doing unsafe things constantly.

Caveat: it works best when you have that and/or when you have labels inherent in the data. For any non-trivial problem, you’re going to have FAR more data than you can possibly have labeled by domain experts. So you find ways to flag select data where their expertise is required. That can mean having non-experts do a first set of passes and only send to experts those where you get significant disagreement. It can also mean, for example, picking out anomalous sensor data(defined as some significant deviation from the mean).

In some cases, the inherent labels are all that’s needed. AlphaGo and AlphaZero were both trained without any labeling of the data, because the end result of a game could be used as a label in itself. Similarly, here, any crash/near miss is an excellent label. Taking situations where a person took over and running it through a simulator can also provide an excellent label.

All this aside, we *know* they’re using neural networks for object detection and classification. And we know they’re doing a relatively good job at that task and improving over time. They must be getting the data for that from *somewhere*. And, of course, we also know that for a (changing) set of certain select situations and also at random intervals, each car occasionally gathers a short set of videos and sensor data which is sent to the mothership. Either they’re using that data for ML, or they’re lying to all of us and secretly using magic.

 

[BinaryField]

In addition the NEO-M8L does not have integrated RTK like their newer models this would give it a 1.5CM precision. (Including a long list of caveats at the bottom of the datasheet)
Tesla can implement RTK on their end since they have the required sensors mostly accelerometer and gyroscope/IMU.

I don't really follow what you're saying here. What is "integrated RTK", and what does it have anything to do with the presence of an IMU? RTK is the process of determining, in real time, the relative baseline between a rover's GPS reciever (e.g. a car) and a reference GPS receiver (e.g. a fixed station on top of a building of known location). It seeks to use the precise carrier phase measurements to obtain very high accuracy positioning.

How can Tesla implement RTK techniques on their own? They would need an extensive network of accurately-surveyed reference stations as well as a low-latency medium to transmit their measurements to users. Even then, accuracy is not guaranteed because of the type of environments cars operate in. For example, a car driving in an urban canyon would be subject to severe signal multipath and blockage. The results would likely not be good enough for activities such as driving lane resolution.

 

[conman]

Looks like rollout has slowed to a trickle. Possible revision?

Where "wide" means a handful of cars in North America, followed by a trickle throughout the rest of North America over the next 4 weeks, followed by a handful of cars outside North America and so on...

 

[Anner J. Bonilla]

I don't really follow what you're saying here. What is "integrated RTK", and what does it have anything to do with the presence of an IMU? RTK is the process of determining, in real time, the relative baseline between a rover's GPS reciever (e.g. a car) and a reference GPS receiver (e.g. a fixed station on top of a building of known location). It seeks to use the precise carrier phase measurements to obtain very high accuracy positioning.

How can Tesla implement RTK techniques on their own? They would need an extensive network of accurately-surveyed reference stations as well as a low-latency medium to transmit their measurements to users. Even then, accuracy is not guaranteed because of the type of environments cars operate in. For example, a car driving in an urban canyon would be subject to severe signal multipath and blockage. The results would likely not be good enough for activities such as driving lane resolution.

My bad I mixed up a few things.
(Was probably reading the wrong datasheet didn't notice they had UDR support inside the chip and proceeded to explain Tesla doing it themselves as our phones do.)

The original post should have said dead reckoning/UDR which the NEO M8L already has integrated.
They call it UDR it uses the sensor reading from the car including wheel rotation, Gyroscope, and accelerometer.
https://www.u-blox.com/en/system/fi...h=PjqS9ARqPmLSfTClgeotRmweEE2W_vZ5UvDl_xVs3P0

It seems the NEO-M8L is a 1.5M when using SBAS augmentation otherwise a 2.5M CEP
Can't find a mention if UDR would improve precision lower than the datasheets 1.5 Meters with SBAS.

The newer M8P is able to do "0.025 m + 1 ppm CEP" but that is using RTK without RTK is same as the M8L which is 2.5M .
As you mentioned RTK would require a base station for reference readings which I doubt Tesla would get into themselves or using an NTRIP provider to deliver RTCM messages over the internet this supposedly allows a precision of 1cm if Wikipedia is accurate.

 

[BinaryField]

My bad I mixed up a few things.
(Was probably reading the wrong datasheet didn't notice they had UDR support inside the chip and proceeded to explain Tesla doing it themselves as our phones do.)

The original post should have said dead reckoning/UDR which the NEO M8L already has integrated.
They call it UDR it uses the sensor reading from the car including wheel rotation, Gyroscope, and accelerometer.
https://www.u-blox.com/en/system/files/UDR_whitepaper_(UBX-16000376)_0.pdf?hash=PjqS9ARqPmLSfTClgeotRmweEE2W_vZ5UvDl_xVs3P0

It seems the NEO-M8L is a 1.5M when using SBAS augmentation otherwise a 2.5M CEP
Can't find a mention if UDR would improve precision lower than the datasheets 1.5 Meters with SBAS.

The newer M8P is able to do "0.025 m + 1 ppm CEP" but that is using RTK without RTK is same as the M8L which is 2.5M .
As you mentioned RTK would require a base station for reference readings which I doubt Tesla would get into themselves or using an NTRIP provider to deliver RTCM messages over the internet this supposedly allows a precision of 1cm if Wikipedia is accurate.

No problem. Appreciate the clarification!

I'm not sure that applying the UDR/IMU measurements would necessarily improve precision. My expectation is that they would be blended in with the noisy or occasionally dropped GPS solutions to provide a smooth and consistent overall navigation solution.

I'm still curious about how the vehicle is determining its lane position. If it is using GPS only, assuming that u-blox's accuracy estimates are correct (i.e. not overly-conservative / pessimistic), then we should sometimes see bad lane readings. I haven't experienced that in my car yet, so does this mean they are doing something else in combination?

 

[VValleyEV]

My car pulled 7.5GB, then it stopped. Do you think it’s a maps update?

How do you view such download activity?

I have a 2018 Model S running 2018.38.6, first wide release of V9. Real interested in getting the new 42.2 V9.

Thx for your help.