I'm pretty sure that deriving position from acceleration data from an inertial sensor is not that cleanly done from first principles: it tends to drift a bunch.

I'm pretty sure that deriving position from acceleration data from an inertial sensor is not that cleanly done from first principles: it tends to drift a bunch. There's no reason the trained model can't emulate the dynamics accurately and do just as well as the traditional approach, but then also learn from additional clues.