I don't think there's any fleet learning going on. What's happening now is basic lane line recognition, general object avoidance, and probably not much else.
Computer vision is a funny game. The Mobileye guys built a vision library that was obviously very robust for a very specific task. Where a huge dataset, machine learned vision net wins out is on "general" recognition - but it's still a bit unpredictable. For example, you can show it 100 images of a cat, and it'll tell you it's a cat 98 times. But, twice, it'll tell you it's a banana or a flowerpot. No one is entirely sure why.
For basic "just drive in the middle of the lane" logic, there's no need for machine learning. You'll get better results by just hardcoding some specific rules. Nvidia pushed the machine learning line VERY hard, to sell a lot of GPUs, but it's not the silver bullet in this case.
To be honest, I don't think we'll see much improvement on the current phase of AP2. It might get slightly better at stop and go traffic, and the basic driving rules will get tweaked and so on. We may see the odd minor feature, like speed limit detection or something. But, nothing crazy interesting.
Tesla are probably more interested in developing the high definition map option... I would be if I were Tesla. It basically means you don't need to worry about working out
where to drive - instead, just focus on what not to hit. Everything else is already determined by the map. Speed limits, complex junctions... roadwork areas, roundabouts? These are a nightmare to work out with vision alone - but instantly solved with a good enough map. This is a much easier problem! For example:
Note that this is
not an overlay of a vision recognition system - you're seeing the road map here, so it knows exactly where it should drive.
I was hoping they might shed some light on this with the release of Model 3, but nothing was said.
I don't think we'll see any real improvement in Autopilot with the current methods, or even any new features. It may one day actually reach parity (UX parity rather than feature parity) with AP1, but I doubt it'll surpass it. That is, until they start activating the high definition maps. This will probably happen just around California to begin with. Then, major highways, cities... and eventually more local roads and rural areas... think of it like the rollout of high speed internet!
