Visual Servoing in Autoencoder Latent Space