Star-Fusion: AI Model for Spacecraft Orientation Using Spherical Topology
A new multi-modal transformer architecture, named Star-Fusion, has been developed by researchers to enhance celestial attitude determination for the navigation of autonomous spacecraft. Conventional Lost-in-Space (LIS) algorithms are hindered by significant computational demands and vulnerability to sensor noise. Meanwhile, deep learning methods encounter challenges due to the celestial sphere's non-Euclidean topology and the periodic boundary conditions associated with Right Ascension (RA) and Declination (Dec). Star-Fusion addresses these issues by reinterpreting orientation estimation as a discrete topological classification problem, utilizing spherical K-Means clustering to divide the celestial sphere into K topologically coherent regions, thus reducing coordinate wrapping artifacts. This architecture features a tripartite fusion approach with a SwinV2-Tiny transformer backbone for extracting photometric features. The research can be accessed on arXiv with reference 2604.26582.
Key facts
- Star-Fusion is a multi-modal transformer architecture for celestial attitude determination.
- It addresses limitations of traditional Lost-in-Space algorithms and deep learning regression models.
- The approach uses spherical K-Means clustering to partition the celestial sphere into K regions.
- It reformulates orientation estimation as a discrete topological classification task.
- The architecture includes a SwinV2-Tiny transformer backbone for photometric features.
- The paper is published on arXiv with ID 2604.26582.
- The method mitigates coordinate wrapping artifacts from RA and Dec periodic boundaries.
- The work targets autonomous spacecraft navigation.
Entities
Institutions
- arXiv