I was thinking of path planning beyond the visual perception of the cameras. The car is in the right lane, preparing for a right turn, and the vision is occluded due to large amount of traffic. Currently the planner sometimes attempts to go around the traffic to the left, which then leads it to miss the turn. If the cloud had memory of the intersection, and previous drives, perhaps that behavior would be reduced (though not eliminated). I admit my knowledge of how the NN is handling decision making is limited, so I may be way off base.
I don't think the nets would be trained to remember that specifically, and if they did, it would be overfitting or over capacity nets which could have undesirable effects.
There appears much to be done on the driving policy side of the product, as opposed to the perception side. The technology of the perception side, with the latest major update, seems to be high quality machine learning with state of the art algorithms, though I would prefer higher resolution cameras, and in stereo.
But planning is a different task. To me this requires significant semantic mapping, which could be gained from fleet driving. Then, those semantic hints would go into a neural network/ML based planner. I don't think Tesla is there yet, and there's a big gap.
Humans could have made a mistake like you describe if they had no prior knowledge of the intersection (e.g. where can I turn from to get to where I'm going). There is some basic mapping used but not at a really deep level. Humans though usually drive places they've been before and intuitively understand the dynamics and optimal routes for their intersections.
The fleet should be generating routing information from human driving, and use that to build semantic maps of "what do people do to get from X to Y", not just allowed behavior, but average preferred behavior. The car might not be able to follow that on average, but could use it as a target. Collecting large datasets of routing under human driven conditions would also generate targets (presumably using reliable safe drivers only) for the driving policy training of neural networks.
This would require a major upgrade to back end and front end technology and an expansion of the capabilities of the mapping side to bring in semantic hints for use on the planning side. The non-neural network mapping side would be use usual geometric database lookups to provide the appropraite 'semantic tags' for the routes and intersections (getting this representation right is not at all easy) which would be fed to a primarily neural network based planner and optimizer.
