Machine Learning Spacecraft Pose Estimation from Synthetic Data to Flight Constraints

Andrew Price

École Polytechnique Fédérale de Lausanne

Kinematic state estimation (i.e., attitude, position, velocity, and angular rate) is a foundational capability for space applications such as rendezvous and servicing, proximity operations, and small body science. Optical methods for estimating spacecraft and resident space object states date back more than a century, with geometric feature matching and comparator table approaches demonstrating proven reliability and robustness in well structured scenarios.

In the past five years, machine learning based methods have achieved state of the art performance in challenging regimes where classical techniques struggle, including partial observability, uncooperative targets, and extreme illumination conditions. At the same time, learning based state estimation introduces significant challenges relevant to space missions, including limited flight data, large domain gaps between synthetic and real imagery, stringent onboard compute and power constraints, and the need for transparency, validation, and stability guarantees.

This talk presents a computer science perspective on vision based state estimation, focusing on how modern learning methods can be adapted to the realities of the space environment rather than directly transferred from terrestrial applications. I discuss advances in synthetic data generation and domain gap mitigation, and show methods to adjust an architecture for space environment scenes. While recent trends emphasize large foundation models, I argue that carefully designed, tuned, and compressed models enable onboard deployment solutions.

The talk concludes with insights from our ongoing research aimed at improving the interpretability, reliability, and validation of learning based state estimation systems for flight critical space applications.