Applying Machine Learning Techniques for Optical Relative Navigation in Planetary Missions

Artificial intelligence (AI) algorithms are playing an increasingly crucial role in onboard data processing to improve spacecraft autonomy and enhance the quality of scientific observations. While AI has found prominent use in Earth observation (EO) missions, it is also gaining importance in planetary missions. These missions require precise spacecraft navigation, especially in scenarios where GPS positioning is unavailable, such as current and future missions to the Moon or Mars. This article presents two novel elements: first, the utilization of onboard convolutional neural networks (CNNs) to provide real-time in-orbit navigation for planetary orbiters, and second, a new method for generating synthetic images to train the CNNs. The proposed solution consists on a set of regression CNNs, each responsible for estimating a different part of the pose solution. Furthermore, multiple data augmentation techniques have been implemented in the training process to extend during runtime the training set and successfully fill the gap between synthetic and real images. To illustrate the proposed solution, OPS-SAT, an Earth-orbiting spacecraft in a Sun-synchronous orbit has been selected as a use case to test and validate the CNNs, improving the current two-line element (TLE) derived geo-localization tag of its images. Although extensive datasets of Earth imaging exist, including a wide range of illumination conditions, weather conditions, or seasonal changes, a method for systematically generating synthetic image datasets for generic conditions is proposed. Earth albedo and terrain elevation datasets have been combined into a very high-definition Earth 3-D model, enabling photorealistic synthetic image generation.