For deep-space missions operating at large distances from Earth, autonomous relative optical navigation is emerging as a key technique because communication delays and the need for instant, high-precision control make ground intervention impractical. For instance, NASA’s Double Asteroid Redirection Test (DART) planetary-defense mission relied on this technology to successfully impact Dimorphos, proving its operational feasibility. Similarly, the Japan Aerospace Exploration Agency’s (JAXA) Martian Moons eXploration (MMX) mission, scheduled for launch in 2026, depends on this autonomy, though for a different purpose, specifically to enable safe close-proximity operations around Phobos.
However, a major issue is that existing images of Phobos from previous spacecraft, even those with reconstructed poses and known camera parameters, do not cover the full interval of viewing geometries and illumination conditions expected during the MMX close-proximity operations. This lack of data limits the construction of realistic datasets for relative vision-based navigation, guidance, and control algorithms. To bridge this gap, this work utilizes CORTO, an open-source Blender-based rendering tool, coupled with an automatic data-driven optimization loop. In contrast to established tools such as PANGU and SurRender, which are often proprietary or restricted to specific organizations, this framework leverages CORTO’s open-source, modular architecture and pre-configured scenarios to generate synthetic templates with tunable shading and material parameters, thereby achieving the necessary “image equivalence” for rigorous validation.
In the present study, we refine both the evaluation metric and the optimization strategy to ensure reliable material estimation. At the core of the framework is a content-focused similarity metric designed to compare real and synthetic images within a navigation-relevant window. We implement this via a fully autonomous pipeline that reads raw PDS IMG products, applies automatic pre-processing (including hot-pixel filtering and brightness normalization), and then optimizes material and selected sensor parameters using particle swarm optimization (PSO). This approach is particularly suited to non-convex landscapes and exploits an efficient trade-off between population size, iteration count, and the computational budget required for rendering.
Using this framework, we constructed an initial Phobos-focused optimization dataset that samples viewing geometries relevant to MMX, including complex lighting scenarios where Mars contributes reflected light or where Phobos and Deimos are simultaneously visible. By jointly optimizing material parameters and the camera dark current threshold against historical mission data, the framework yields Structural Similarity Index Measure (SSIM) values between 0.82 and 0.90, with a normalized root mean square error (NRMSE) around 0.04. These quantitative results indicate a strong structural consistency between synthetic and real imagery, establishing a validated baseline for generating realistic navigation datasets.
To demonstrate operational utility, the optimized material parameters have been integrated into the Mars-Phobos-Deimos scenario, which utilizes literature-based geometry already implemented in CORTO. This integration enables the production of a high-fidelity labeled synthetic dataset along a reference Quasi-Satellite Orbit (QSO) around Phobos. Currently, the framework is being expanded to generate datasets covering a broader range of candidate MMX orbits from the literature, including complex scenarios where Mars enters the field of view. Furthermore, explicit sensor noise models are being integrated directly into the optimization loop to facilitate systematic comparisons of noise handling strategies.