A Physically Accurate Simulation Framework for Training and Validation of AI-Based Algorithms for Space Exploration

Space robotic missions demand high levels of onboard autonomy beyond those offered by standard guidance, navigation and control methods. The integration of artificial intelligence within these systems provides a framework to enhance onboard autonomy, yet the performance of machine learning algorithms remains strongly dependent on the quality and quantity of training data. In the space domain, the limited availability of annotated real-world imagery has prompted the growing use of synthetic datasets, enabled by advances in photorealistic simulation. Nonetheless, networks trained on synthetic imagery exhibit discernible differences in performance when evaluated on real data. To address these limitations, we developed a photorealistic simulation environment for lunar surface scenarios using Blender. The framework produces high-resolution digital elevation models through fractal expansion algorithms, with craters and rocks inserted according to lunar size-frequency distribution models. Illumination is simulated using the Hapke bidirectional reflectance function, and Blender's rendering pipeline generates images under varying observation geometries. A custom add-on yields the corresponding annotations, including semantic segmentation masks, object-detection bounding boxes, and depth maps. The simulator is assessed in representative tasks for planetary surface exploration, such as terrain-slope estimation and rock detection. The results indicate that the tool can serve both as a high-quality dataset generator and as a testbed for validating navigation pipelines, supporting the development of autonomous systems that integrates AI-based and conventional methods.