Attitude estimation using AI-based hyperspectral technology for autonomous close-proximity operations

Tools

Cakir, Alper and Porter, Fiona and Campbell, Andrew and Savitski, Vasili and Hall, Iain and Longino, Jamie and Zabalza, Jaime and Vasile, Massimiliano and Murray, Paul and Feng, Jinglang; (2026) Attitude estimation using AI-based hyperspectral technology for autonomous close-proximity operations. In: IAF Astrodynamics Symposium - Held at the 76th International Astronautical Congress, IAC 2025. Proceedings of the International Astronautical Congress, IAC . International Astronautical Federation (IAF), AUS, pp. 194-208. ISBN 9798331329358 (https://doi.org/10.52202/083087-0019)

[thumbnail of Cakir-etal-IAC-2025-Attitude-estimation-using-AI-based-hyperspectral-technology]
Preview
 Text. Filename: Cakir-etal-IAC-2025-Attitude-estimation-using-AI-based-hyperspectral-technology.pdf
Preprint
License: All rights reserved
Download (5MB)| Preview

Abstract

Accurate attitude estimation of resident space objects (RSO) is essential for autonomous navigation and closeproximity operations such as in-orbit servicing and active debris removal. To address the challenges of using conventional cameras in terms of illumination conditions, this research developed a model that leverages hyperspectral imaging together with the machine learning approach. It identifies the features and components of the target spacecraft with a given spectral response and then estimates its attitude quaternions, rather than determining the exact geometric shape by using the conventional RGB cameras. The model utilizes sequences of spectral images or time-series data to determine the attitude of the target object over time, by employing the framework consisting of a convolutional neural network (CNN) and a recurrent neural network (RNN). To further enhance robustness, a Bayesian Neural Network was developed and integrated into the framework, allowing investigating the model uncertainties by generating sets of weights and biases of the neural networks rather than a single deterministic estimate. This approach provides a measure of uncertainty, which is crucial for applications where confidence levels in the estimated attitude are necessary for autonomous decision-making. The model was firstly trained and validated with synthetic data generated from open software Blender and then tested with real-world hyperspectral images that were generated in a controlled laboratory environment. This hardware in-the-loop (HIL) test is the first step in validating the performance of the model for real-word applications. The transition to real data required adaptations to the model architecture, including fine-tuning through a transfer learning approach. This method improved the model’s ability to generalize beyond synthetic datasets, and the result demonstrated that hyperspectral imaging can be effectively utilized for real-world attitude estimation tasks. Future work will focus on expanding both synthetic datasets and laboratory datasets to include a wider variety of objects and testing the model on more complex scenarios to further enhance its performance and applicability in real space missions.

ORCID iDs

Cakir, Alper, Porter, Fiona ORCID logoORCID: https://orcid.org/0000-0002-5695-0633, Campbell, Andrew ORCID logoORCID: https://orcid.org/0000-0002-4439-3630, Savitski, Vasili ORCID logoORCID: https://orcid.org/0000-0001-5261-1186, Hall, Iain, Longino, Jamie, Zabalza, Jaime ORCID logoORCID: https://orcid.org/0000-0002-0634-1725, Vasile, Massimiliano ORCID logoORCID: https://orcid.org/0000-0001-8302-6465, Murray, Paul ORCID logoORCID: https://orcid.org/0000-0002-6980-9276 and Feng, Jinglang ORCID logoORCID: https://orcid.org/0000-0003-0376-886X;
CORE (COnnecting REpositories)