Quote Originally Posted by edgesandnets
It's true that there may end up being developments faster than I anticipate, but I do not think what you described is possible using existing technology.
Sure there is. There is a lot of computing power for cheap now days. My graphic card has 2560 CUDA cores.
You can buy a cheap CUDA core development system for a few hundred $.

You can losslessly store the state of a chess or go board after any move that is both compact and interpretable both to machines and humans, which makes it suitable for machine learning.
You would need to store a sequence of trajectories for each rally. Memory and storage is cheap. To make searching an indexing faster use hash codes like the chess computers do.

For table tennis, I guess you could automatically learn a representation or learn directly from video pixels, but such a task sounds significantly harder.

Tracking human pose and robustly automatically knowing who won each point are both AI.

I would only use the video to compute the trajectory, speed and spin.
Chess programs think 30 moves ahead now. TT rallies rarely go that long. The tree would also be very narrow after a strokes.