Deep learning for spacecraft pose estimation: from algorithms to hardware-in-the-loop testbed

The rapid growth of the New Space economy is increasing demand for in-orbit servicing and active debris removal missions that rely on rendezvous and proximity operations. In this context, relative pose estimation between a chaser and its target is a key enabling technology, achievable with various sensors (e.g., cameras, LIDAR). Monocular vision is attractive for its minimal system budget footprint. However, the complexity of space imagery and onboard deployment constraints still hinder its adoption in space. Monocular satellite pose estimation methods, especially when paired with deep learning for image processing, still suffer from domain gap, high computational cost, and limited uncertainty quantification. This work investigates these issues through an extensive set of experiments. First, a domain-adversarial, Vision Transformer-based pipeline is introduced to reduce the performance gap between synthetic and real imagery. This approach ranked fourth and fifth in ESA’s second Satellite Pose Estimation Challenge. Architectural and data-level strategies, such as generative models and hybrid neural network architectures, are then investigated to further mitigate domain gap while lowering computational cost. A parametric study is conducted to scale image processing on Edge TPUs, showing real-time performance with minimal energy consumption. Additionally, Sparse Variational Gaussian Processes (SVGPs) are explored to provide uncertainty estimates from neural network feature maps, showing a moderate-to-strong correlation between predicted uncertainty and actual errors on synthetic images. For unknown targets, the pipeline must begin with an inspection phase. To this aim, lightweight localization models are evaluated on previously unseen spacecraft at far range, followed by a lightweight deep-learning-based feature matcher that drives a frame-to-frame pose estimation pipeline to recover camera motion. Results highlight current dataset limitations for this task. Finally, the development, software integration, and calibration of a HIL testbed at the University of Bologna are presented, together with a dataset generation pipeline featuring a nearly symmetric target.