Say you have some state being rendered into pixels at 60FPS.

If the state-space/domain was constrained enough, could you train a model that has high performance and looks intelligible?