Tesla and Map Data

[sleepydoc]

There's been a lot of discussion about the effect of map data on FSD, how FSD handles inaccuracies, etc. The question of where Tesla gets its map data and how to correct inaccuracies also frequently comes up. Recently, Elon tweeted "10.13 smooths out intersection control, especially long lefts, and starts to handle roads with no map data at all."

Here's my question - Tesla has a fleet of several million vehicles with GPS sensors, all constantly scanning the road with cameras and sending data back to the Tesla Borg. Why do they need map data at all? They theoretically should be able to have the best map data in the country, and it doesn't seem like it should be a big deal for the system to flag an area where the map data differs from the camera or GPS data. For example, there's an area I drive that the map has the speed limit listed as 35 even though it's 55. If the car's driving along and sees a 55MPH speed limit sign and the map data says 35 it could trigger an alert to be reviewed and corrected. They could also easily partner with TomTom or another country to sell them data.

Am I missing something or is it just something Tesla hasn't focused on before?

 

[EVNow]

The question of where Tesla gets its map data

Search for my thread on maps.

The best we know - they get data from multiple sources and integrate.

 

[wknickless]

Why do they need map data at all?

Who are you going to trust: me or your lying eyes?

At the risk of overgeneralizing, the history of self driving technology started with the assumption that mapping data was available and reliable. This goes back to the original DARPA challenges and continues up to the present day with existing self driving taxis that navigate with high resolution LIDAR referencing their 3D surroundings.

Elon is challenging this orthodoxy by claiming L6 driving can be achieved purely through computer vision, because (after all) humans can drive without LIDAR or RADAR or even high resolution map memories.

He might be right, but in the meantime the external research they depend on assumes good mapping data is available. And in a civil court, one had better be able to defend a decision to rely more on a novel technology like computer vision than industry standard best practice: high resolution mapping data.

So whether you want to call it a crutch or training wheels, the FSD beta development process necessarily privileges mapping data. As the computer vision tech improves over time and (eventually?) measurably outperforms static mapping data, then we can expect to see computer vision data being used to correct the static mapping databases.

But as of this writing I don't think we're there yet.

 

[sleepydoc]

Who are you going to trust: me or your lying eyes?

At the risk of overgeneralizing, the history of self driving technology started with the assumption that mapping data was available and reliable. This goes back to the original DARPA challenges and continues up to the present day with existing self driving taxis that navigate with high resolution LIDAR referencing their 3D surroundings.

Elon is challenging this orthodoxy by claiming L6 driving can be achieved purely through computer vision, because (after all) humans can drive without LIDAR or RADAR or even high resolution map memories.

He might be right, but in the meantime the external research they depend on assumes good mapping data is available. And in a civil court, one had better be able to defend a decision to rely more on a novel technology like computer vision than industry standard best practice: high resolution mapping data.

So whether you want to call it a crutch or training wheels, the FSD beta development process necessarily privileges mapping data. As the computer vision tech improves over time and (eventually?) measurably outperforms static mapping data, then we can expect to see computer vision data being used to correct the static mapping databases.

But as of this writing I don't think we're there yet.

I Didn’t mean to say Tesla should use no map data, just that they don’t need externally sourced map data since they have the equivalent of a million google map cars roaming the streets constantly. They could conceivably continuously update their map database and have the best data in the industry.

 

[orangespoon]

I Didn’t mean to say Tesla should use no map data, just that they don’t need externally sourced map data since they have the equivalent of a million google map cars roaming the streets constantly. They could conceivably continuously update their map database and have the best data in the industry.

Converting visual car level environmental information to usable mapping information reliably is a difficult task, even for Tesla.

 

[Buckminster]

Read the whole thread from green:

https://twitter.com/x/status/1641474882520358913