-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla “can gather 1 billion miles of data per year”
- Thread starter strangecosmos
- Start date
32G in total, but the "snapshot" partition is 10G (10G for hd maps and then remainings are for firmware images, logs and other misc stuff)
im.thatoneguy
Member
Yeah, I believe Waymo has pretty much assumed that they've solved the identification questions ages ago (in clear dry weather). They're still just working through driving logic. And most of their "billions of miles virtually driven" aren't being run in a Grand Theft Auto video game Matrix-like construct, they're just running in a bounding box world.
Both problems are hard. It's hard to assign meaning to images and point clouds. It's also hard to act intelligently and safely on those meaning-rich objects. My argument is that annotating and guided training for assigning meaning is best done with engineer supervision. Learning how to drive and react to lots of very different situations is best done through massive data troves of raw data. And that if a driver acts particularly weird in a situation that is assumed to be 'well understood' that is a great hint for engineers to see what external variable wasn't assigned meaning that the driver is reacting to. Good feedback to automatically find areas that need more annotation.
Both problems are hard. It's hard to assign meaning to images and point clouds. It's also hard to act intelligently and safely on those meaning-rich objects. My argument is that annotating and guided training for assigning meaning is best done with engineer supervision. Learning how to drive and react to lots of very different situations is best done through massive data troves of raw data. And that if a driver acts particularly weird in a situation that is assumed to be 'well understood' that is a great hint for engineers to see what external variable wasn't assigned meaning that the driver is reacting to. Good feedback to automatically find areas that need more annotation.
