Tesla Autopilot miles estimated at over 1.2 billion

[Lex_MIT]

Tesla Autopilot has driven over 1.2 billion miles (estimated). See visualization sand details in blog post: Tesla Autopilot Miles | MIT Human-Centered AI

This is a remarkable accomplishment in the history of applied AI research, development, and deployment.

 

[daktari]

I could not find the source for ratio of AP2+ cars with EAP option enabled?

 

[Lex_MIT]

I could not find the source for ratio of AP2+ cars with EAP option enabled?

I assumed that the ration of AP2 cars with EAP is the same as AP1 cars with AP. It's the most reasonable assumption given the lack of data on this point.

 

[conman]

There's the shadow mode claim again. I didn't think any of the people who'd hacked into their Tesla had any evidence of any shadow mode going on. I wonder if @verygreen could comment?

 

[Tezlanian]

There's the shadow mode claim again. I didn't think any of the people who'd hacked into their Tesla had any evidence of any shadow mode going on. I wonder if @verygreen could comment?

I think he's just saying the ratio is the same, not implying any shadow mode. That said, it'd be an awful waste to not run a shadow mode.

 

[verygreen]

I think he's just saying the ratio is the same, not implying any shadow mode. That said, it'd be an awful waste to not run a shadow mode.

In the linked blog post it's literally item #3:

Miles driven in “shadow mode” on hardware version 2+

Back to the original question. If we define "shadow mode" narrowly enough as "Tesla might produce a set of conditionals that when met would collect a bit of data from the car and transfer back to Tesla" then the shadow mode is real, BUT not every hw2+ enabled car gets the triggers (under 10% I think? I forgot how I estimated this) and not every car that does get the triggers would actually be able to trigger them (see hw2.5 triggers sent to hw 2.0 cars, model3 triggers sent to s/x cars and so on).

As such tracking the "shadow mode" miles is totally pointless in my opinion. On the other hand total miles driven on hw2+ is still important since they do gather valuable info from every trip taken, where the car was driven, what speeds were taken, what autopilot modes were available and so on: Data sharing and privacy - what's actually collected?

 

[Tezlanian]

In the linked blog post it's literally item #3:


Back to the original question. If we define "shadow mode" narrowly enough as "Tesla might produce a set of conditionals that when met would collect a bit of data from the car and transfer back to Tesla" then the shadow mode is real, BUT not every hw2+ enabled car gets the triggers (under 10% I think? I forgot how I estimated this) and not every car that does get the triggers would actually be able to trigger them (see hw2.5 triggers sent to hw 2.0 cars, model3 triggers sent to s/x cars and so on).

As such tracking the "shadow mode" miles is totally pointless in my opinion. On the other hand total miles driven on hw2+ is still important since they do gather valuable info from every trip taken, where the car was driven, what speeds were taken, what autopilot modes were available and so on: Data sharing and privacy - what's actually collected?

Didn't read blog post yet, just this thread and his ratio comment.

On the topic though, why would such a random sampling be done with the triggers? Why not run consistently whether explicitly enabled or not to gather data? If you only collect in places where people explicitly enable it you're probably also biasing for areas it already works reasonably well in. Why not take advantage of a larger data set in areas where it probably works less well? Seems silly to not take advantage of a resource that is essentially free if you can handle the volume.

 

[electracity]

In the linked blog post it's literally item #3:


Back to the original question. If we define "shadow mode" narrowly enough as "Tesla might produce a set of conditionals that when met would collect a bit of data from the car and transfer back to Tesla" then the shadow mode is real, BUT not every hw2+ enabled car gets the triggers (under 10% I think? I forgot how I estimated this) and not every car that does get the triggers would actually be able to trigger them (see hw2.5 triggers sent to hw 2.0 cars, model3 triggers sent to s/x cars and so on).

As such tracking the "shadow mode" miles is totally pointless in my opinion. On the other hand total miles driven on hw2+ is still important since they do gather valuable info from every trip taken, where the car was driven, what speeds were taken, what autopilot modes were available and so on: Data sharing and privacy - what's actually collected?

I agree, I also don't get how Tesla could train perception in shadow mode.

I do get how shadow mode could be useful once perception is trained.

 

[verygreen]

On the topic though, why would such a random sampling be done with the triggers? Why not run consistently whether explicitly enabled or not to gather data? If you only collect in places where people explicitly enable it you're probably also biasing for areas it already works reasonably well in. Why not take advantage of a larger data set in areas where it probably works less well?

I don't know why Tesla made this or that decision. You can try asking them.

I can speculate that the amount of data is so massive they just cannot afford to transfer and store it AND then to pay people to actually look at it all (probably a much bigger expense). It was quite telling after initial snapshots poured out en-mass and then died to a trickle very shortly after.

Also note it's not under control of people when to enable it, triggers absolutely get pushed by Tesla and you have no say to activate or deactivate them (other than opting out of the program, I guess), the triggers are also largely independent from whenever you use AP or not (unless the trigger asks for a specific autopilot state).

Currently majority of the triggers (from what I have seen, I do not have 100% visibility into it and have seen only a fraction of distributed triggers) appear to be debugging features. Things like "user asked for windshied wipe - send us 2 camera samples of that" or "steering wheel angle is X and speed is above Y - send us 5 camera samples of that". Initial ones were a stark departure - they were mostly "every once in a while collect a random video and upload it - do it 10-20 times". They still do a bit of random data collection now, but it's a lot more lightweight, like "random imu reading".

 

[Tezlanian]

I agree, I also don't get how Tesla could train perception in shadow mode.

I do get how shadow mode could be useful once perception is trained.

Doubt they're doing training here, probably just inference. The shadow mode would just be logging select raw data based on hand made heuristics to upload in the case of failures to further training. People with hacked cars would be able to tell you for sure though! I'm not taking mine apart until the warranty period is over.

 

[verygreen]

Here's an example of a trigger if you are too lazy to look in the old threads:

{"query":{"$and":[{"$eq":["@LegacyDebug.collision_warning.collision_warning_event",1]},{"$gt":["@TelemetryOutput.distance_travelled_m",2000]},{"$eq":["@MiscTriggerFields.vehicle_model",3]}]},"topics":["main-replay","fisheye","main","narrow","single-frame","latest-frames","can"],"requester":"img-collision-warning-0063","request-delay":1500,"max-requests":1,"timeout":1509764400,"min-period":60000,"request-time-to-live":432000,"request-wifi-hold-time":43200,"num-requests":0,"previous-request-time":0}

 

[Tezlanian]

Lol, thanks. It's not quite laziness as the search on this forum is just somewhat limiting. Need more compilation threads of a particular subject containing informative threads where such has been discussed before. Kind of surprised that isn't a thing with the volume of information here. As is I've tried just top informative threads and crawling particular users to some extent.

 

[BigD0g]

Shadow mode is a myth, there is no side driving comparing my driving to what it would have done, it’s simply not happening.

 

[Tezlanian]

Shadow mode is a myth, there is no side driving comparing my driving to what it would have done, it’s simply not happening.

That would largely seem to be the case. Waste of a competitive advantage others won't have for some time.

 

[Pyronx]

Wheelhack.com
Check this out! The Wheel Hack fixes the wheel nag. I like using it for when I’m stuck in traffic and It works perfectly when I drive to Las Vegas for 2 hours straight. With this in place and never had to touch the steering wheel once.

 

[Ludalicious]

Wheelhack.com
Check this out! The Wheel Hack fixes the wheel nag. I like using it for when I’m stuck in traffic and It works perfectly when I drive to Las Vegas for 2 hours straight. With this in place and never had to touch the steering wheel once.

Have fun explaining to your insurance company what's on the steering wheel if you get in an accident...

 

[conman]

Shadow mode is a myth, there is no side driving comparing my driving to what it would have done, it’s simply not happening.

This is precisely what I thought which is why I wondered where or why they even got figures for estimation of shadow driving. Snapshots I already understand and know about but that cannot in all conscience be called "shadow mode." My car uploads 50-100MB each day when I use autopilot in adversarial conditions - usually in busy urban areas; I do it to challenge it and test its limits (and occasionally laugh at it.)

 

[strangecosmos]

Lex, you are doing God’s work. Thank you.