Apple proposes a system of refining a simulator's output through SimGAN, a take on "Simulated+Unsupervised learning." The technique combines unlabeled real image data with annotated synthetic images using Generative Adversarial Networks (GANs), or competing neural networks.

Read about the paper here: [1612.07828] Learning from Simulated and Unsupervised Images through Adversarial Training