2D and 3D Data Generation and Workflow for AI-based Navigation on Unstructured Planetary Surfaces

Abstract

Artificial Neural Networks (ANNs) promise improvements in autonomous navigation performances for future robotic space missions in unstructured and irregular planetary environments. This paper describes an AI-Development-Framework (AIDF) based on workflows for developing, training, and validating ANNs to improve the performance of Guidance, Navigation, and Control (GNC)-algorithms by corresponding ANNs. Further, the AIDF includes a high-fidelity dataset generator for 2D image data, 3D point cloud, and 3D distance data (depth images) inspired by 67P/Churyumov-Gerasimenko comet's surface. For tests of the AIDF and the simulated synthetic high-fidelity dataset, Convolutional Neural Networks (CNNs) for semantic segmentation tasks in grayscale 2D and corresponding depth images were trained and validated, which follows mostly an encoder-decoder structure, such as the YOLOv8 and the Deeplabv3+ with different backbone nets, such as ResNet, ResNet-RS, and the YOLO-NAS. This paper also includes the results of investigations of particular CNN's behavior against Perlin noise, added to the artificial comet surface generated by the dataset generator.