Unsupervised Learning of Visual 3D Keypoints for Control

                   Boyuan Chen 1 Pieter Abbeel 1 Deepak Pathak 2

             Abstract
    Learning sensorimotor control policies from high-
    dimensional images crucially relies on the quality
    of the underlying visual representations. Prior
    works show that structured latent space such as
    visual keypoints often outperforms unstructured
    representations for robotic control. However,
    most of these representations, whet ...