These deterministic dice are rolled by duplicating all relevant physics objects into a new physics space (called the prediction space) then ticking that simulation until every object within it comes to rest. For every tick of the prediction space, each object within records it's transform, which the original object then uses to animate itself from initial to final transform. From the final and initial transforms of a die in the prediction space, one can calculate which face of said die will be face up when the die comes to rest - then simply place the textures on the dice object accordingly. (side note: setting the initial orientation of the die would require either a parent transform, or adjusting each reported transform accordingly, and i'm too lazy to do that.)