Lecture summary: Simulations also used to train machine learning models. Example: Renderer could render images at 90,000 frames/sec. Can present robot with 90,000 training examples/sec so machine can learn from years of human experience overnight

I hope there are some industry expert who can give some insights about the effectiveness of training real-world ML models from simulated 3D scenes. From my experience in 2D, models could quickly over-fit when merely using simulated data and doesn't generalize well when dealing with real-world input.