I’ll just leave this HERE
It’s one hell of an achievement. And I believe this isn’t counting shadow mode but actively driven.
Waymo and others have much lower numbers as they are using a different system entirely and I don’t believe one should think Tesla are massively ahead. But data is data. In this case, they have an advantage in terms of training.
Hopefully with HW 2.5 and HW 3.0 when more complicated neural nets are used, the training data will have resulted in major reliability. N
I was really surprised at this number - it’s about half of what I would have expected. It makes me think that either the take rate on AP is lower than I would have expected, or that common usage of AP is rather low. Given the media environment and comments that have been made by Tesla I’m inclined to the latter. It may be that everyday users are being dissuaded from employing the AP that they already bought because of hyperbolic criticism from mainstream media.
If so that’s a sad thing. It would mean that the media was effectively killing people by discouraging the use of a safety technology.
Lex Fridman’s model predicted about 1.5 billion miles, so that was off by about 50%.
I was just thinking along the same lines that Ohmman was thinking. Most of the average daily commute is actually a rather short distance from home to work and perhaps to other destinations all within a small radius. Which means mostly surface streets. Which is something autopilot isn’t optimized for and thus isn’t really used on. Yes. Tesla drivers do go on longer road trips, but this isn’t enough to really pull the numbers up.
I still think that hitting 1 billion overall is a really great achievement. And lets not discount the fact that the cars are also running in shadow mode which still makes the vast majority of the daily trips taken useful.
The only question I have is that in shadow mode, the car is seeing what it would have done if it were in control. But with autopilot limitations on surface streets, how useful is this data when it comes to training because when it comes to actually using the networks for inference then it wont have enough mileage to generally improve the fleet in a rapid fashion.
We’re not sure how “shadow mode” really works, or even if it’s running in HW2 Teslas right now. There are apparently (according to verygreen on TMC) triggers for stuff like suddenly changing the steering angle by a certain amount that trigger an upload of some kind of data to the cloud. Maybe this data could be used at Tesla headquarters to simulate the situation and see how Autopilot would have handled it.
At Tesla HQ, if the engineers have all the sensor data from a given moment on the road, they can just feed that sensor data into an instance of Autopilot and then see the output that Autopilot gives. Autopilot can’t tell the difference between real, live sensor data and replayed, pre-recorded sensor data. Autopilot would then output some actuation of the steering wheel, accelerator, and brake, which can be compared to what the human did in the same situation. This is just a hypothetical idea of how “shadow mode” might actually work in practice.
If this isn’t how shadow mode is supposed to work, and it’s really something that is supposed to — live, in the car — notice differences between its steering, accelerator, and brake output and what the human is doing with the controls, then perhaps shadow mode will have to wait until HW3 and Software V10 alpha.
It’s not clear what the term ‘shadow mode’ is supposed to refer to. But it’s certainly true that the fleet of AP2 tesla vehicles are all - whether using AP or not - capable of uploading data captured from excursions. More than one person with a rooted car has reported on this and a substantial amount of data about the process has leaked into public view.
My own car uploads megabytes of data after pretty much every outing.
Agree that AP usage varies widely. I know people who, like myself, use it for a great majority of their miles and I also know people who via either circumstance or attitude use it very little. But my completely baseless expectation had been that AP miles would be about 20% of fleet miles. It seems to be only 10%.
Do you have wifi monitoring, or any other way you can figure out the average daily upload over a period of weeks?
I don’t tend to monitor it all that closely, though there are people who do. Whenever I check the the vehicle’s network usage after a substantial drive I find that I has uploaded a few megabytes of data. I’m just back from a 3 mile errand and it doesn’t seem to have uploaded anything since, but when I come back from a long drive it seems to usually do an upload after it gets a wifi connection.