Stanford's Jiajun Wu outlines de-rendering techniques to extract physical object intrinsics for robotic digital twins

The physical properties serve as biases for self-supervised learning.