6D-poseschatting van niet-coöperatieve ruimtevaartuigen voor nabijheidsoperaties: Een Neural Radiance Fields-perspectief op datagedreven methoden

On-orbit servicing and active debris removal missions depend on rendezvous and proximity operations in which a chaser spacecraft must maneuver in close proximity to a target or dock with it. Autonomous operation around an uncooperative target requires monocular vision-based estimation of the target's 6D pose, i.e., position and orientation, relative to the chaser. Recent data-driven approaches show strong promise but suffer from several limitations that hamper their adoption. First, their scope of application is constrained to perfectly known targets because training hinges on an explicit geometry and appearance model such as a CAD model to produce a synthetic training set. Second, their dependence on synthetic data weakens out-of-domain generalization and creates a risk of performance degradation when models encounter real on-orbit imagery.Third, data-driven methods rely on opaque components, which lower transparency and affect confidence in safety-critical operations. Motivated by these limitations, this thesis addresses three research questions. (i) How can pose estimation be achieved when no explicit target model is available? (ii) How can we move beyond image-wise transforms to strengthen domain generalization? (iii) How can we visualize the target features exploited by pose estimators to understand their decision process and increase confidence? This work adopts a unified perspective based on novel view synthesis through Neural Radiance Fields (NeRFs).Novel view synthesis implements the inverse mapping of pose estimation by rendering an image from a specified camera viewpoint rather than inferring pose from an image. Within this perspective, NeRFs (i) replace explicit target models for dataset creation, (ii) enable appearance-augmented yet geometry-consistent training, and (iii) allow gradient-based inspection of the 3D cues that support pose estimation. This thesis advances vision-based spacecraft pose estimation through three core contributions. First, we introduce a NeRF-based 3D reconstruction method tailored to on-orbit imagery through additional degrees of freedom that compensate for illumination variability and pose uncertainty. Second, we develop an image synthesis method that exploits the decoupling of geometry and appearance in NeRFs to generate geometrically consistent images under diverse viewpoints and appearance conditions. This enables the learning of a pose estimation network without requiring access to a target CAD model and improves generalization beyond the training domain.Third, we introduce a visualization technique that learns a NeRF via gradients back-propagated through a pose estimator to reveal the target 3D cues on which the estimator relies. Together, these contributions (i) expand applicability to targets without explicit models, (ii) improve robustness under domain shift at deployment, and (iii) increase transparency to support adoption in future proximity operations.