4DCodeBench

Can a coding agent watch a video of a physical event and write the code that reconstructs it: the geometry, the dynamics and the render?