Onboard Latent Kalman Filtering for Robust Spacecraft Pose Estimation

Junghwan Park, Woojin Cho

Keywords: Satellite Pose Estimation, Kalman Filtering, Latent Dynamics, Onboard

TL;DR: Onboard latent Kalman filtering for sim2real-robust spacecraft pose estimation without online retraining.

Abstract: Vision-based satellite pose estimation is critical for future space missions such as docking and debris removal. Although deep learning has advanced this field, the scarcity of real-world data forces reliance on synthetic imagery, leading to a persistent sim2real gap. Prior work mitigates this gap through computationally demanding online adaptation. In contrast, we address it without any online learning by leveraging temporal dynamics during rendezvous. We propose Dynamics-aware Robust Inference with latent Filtering for Temporal satellite pose estimation (DRIFT), a latent-space filtering framework that combines dynamics prediction with current observations. We first train a single-image encoder–regressor to extract pose latents, then freeze it and learn a lightweight Neural Kalman Gain module that produces adaptive filtering gains from innovations and recent latents. By explicitly modeling temporal dynamics in latent space, DRIFT enables temporally coherent and robust pose estimation in challenging sequential scenarios without onboard retraining.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 5