Samuel Felton will defend his PhD titles “Deep latent representations for visual servoing” Tuesday, december 20h at 14h00
- Céline TEULIERE (Examinatrice), Maîtresse de Conférences, Université Clermont Auvergne
- François CHAUMETTE (Examinateur), Directeur de recherche, Inria
- Christian WOLF (Rapporteur), Principal research scientist (HDR), Naver Labs
- Guillaume CARON (Rapporteur), Maître de Conférences (HDR), Université de Picardie, Jules Verne
- Elisa FROMONT (Directrice) Professeure, Université de Rennes 1
- Eric MARCHAND (Co-directeur), Professeur, Université de Rennes 1
Visual servoing is used to control robotic systems using visual information. This framework can be used to solve many tasks, such as object grasping, navigation or target tracking. Classically, visual features are geometric or photometric in nature. However, geometric primitives can be hard to extract and track from raw images. Lately, deep learning has come to light as a potential tool for visual servoing, but its use remained limited to extracting features or estimating the pose from images. In this thesis, we propose to go further by removing the feature extraction step, directly linking the learned representations to the motion of the camera, mounted on the robot. This link can be learned, creating an end-to-end approach, or it can be analytically computed, based on the network structure. Moreover, both supervised and unsupervised learning can be employed to create useful servoing representations. Finally, framing visual servoing in a latent space allows us to leverage information from multiple modalities, such as poses and images, leading to a novel “hybrid” visual servoing scheme. We show that the latent space is ideal due to its flexibility, and can be leveraged to accomplish large motions, while retaining an excellent accuracy.