Visual Servoing in Autoencoder Latent Space

Visual servoing (VS) is a common way in robotics to control robot motion usinginformation acquired by camera. This approach requires extract visual information from the image design law. The resulting servo loop built order minimize an error expressed space. We consider direct (DVS) whole images. propose new framework perform VS latent space learned convolutional autoencoder. show that this avoids explicit feature extraction and tracking issues provides good representation, smoothing cost function of process. Besides, our experiments unsupervised learning allows us obtain, without labelling cost, accurate end-positioning, often on par with best DVS methods terms accuracy but larger convergence area.