Exploring Stochastic Variational Gaussian Processes for Trustworthy Monocular Pose Estimation in Space

Estimating the relative pose of a non-cooperative target from monocular images is a key technology for enabling autonomous proximity operations in In-Orbit Servicing (IOS) and Active Debris Removal (ADR) missions. These systems must operate reliably under strict power constraints, making lightweight yet trustworthy algorithms essential. While compact Convolutional Neural Networks (CNNs) offer a feasible solution for onboard inference, the lack of reliable uncertainty estimates is a critical limitation, as it hinders the identification of failure cases, particularly under domain shifts. To address this, we propose an uncertainty-aware pose estimation pipeline that incorporates Stochastic Variational Gaussian Processes (SVGPs) acting on top of CNN feature maps. Correlations in the CNN feature space are used to identify CNN failure cases, enabling the rejection of unreliable predictions, and to estimate uncertainty. We demonstrate the feasibility of our approach with real-time deployment on a Coral Dev Board equipped with an Edge Tensor Processing Unit.