MarkS22
Member
I believe part of the process is distilling the camera data for landmarks. For example, a mountain in the distance, a building, a lane marking, and a road sign. At night, you might look for the full moon or street lamps. Let's say every frame has 5 clear landmarks. All you'd need to send would be the boundary box, its coordinates, and GPS data. It my even include steering, accelerator, and braking levels. This is a trivial amount of data compared to video or photos. (If you look up Mobileye videos, you'll see how it generates boundary boxes around cars, signs, and pedestrians, and has the ability to classify them.)
Now, if we have another Tesla on a road at a similar GPS location, it should "see" the same landmarks at the same exact position (taking into account offsets, like air suspension). If 20 cars (or even the same car) drives the same location, the map can be continually refined. Any noise (margins of error) in the GPS gets evened out until the landmarks with GPS can become extremely accurate. Inches instead of feet.
Now take it a step further. There's a layer of snow on covering the lane markers, but other landmarks are still visible. Like celestial navigation, you can triangulate your position in the lane based on the location of the landmarks. The more the better.
I don't believe we're specifically teaching Autopilot as drivers. Our driving is more likely capturing data to continuously refine a 3D mapset and make them usable. It's more an automated like Waze on steroids.
(Of course, a version could look also for repeated braking or steering corrections at the same exact 3D location and either a human or algorithm would note future Autopilot cars should slow down or be prepared for the same action. I'm sure they're notified if the same "sudden braking" is happening at the same location above a preset threshold.)
Now, if we have another Tesla on a road at a similar GPS location, it should "see" the same landmarks at the same exact position (taking into account offsets, like air suspension). If 20 cars (or even the same car) drives the same location, the map can be continually refined. Any noise (margins of error) in the GPS gets evened out until the landmarks with GPS can become extremely accurate. Inches instead of feet.
Now take it a step further. There's a layer of snow on covering the lane markers, but other landmarks are still visible. Like celestial navigation, you can triangulate your position in the lane based on the location of the landmarks. The more the better.
I don't believe we're specifically teaching Autopilot as drivers. Our driving is more likely capturing data to continuously refine a 3D mapset and make them usable. It's more an automated like Waze on steroids.
(Of course, a version could look also for repeated braking or steering corrections at the same exact 3D location and either a human or algorithm would note future Autopilot cars should slow down or be prepared for the same action. I'm sure they're notified if the same "sudden braking" is happening at the same location above a preset threshold.)
