Ohio State Researchers Develop LUMINA Framework for Lunar Landing Site Analysis

Researchers at The Ohio State University have introduced LUMINA, an artificial intelligence framework designed to automate the assessment of potential lunar landing sites by integrating planetary imagery, geological observations, and geospatial datasets. This system represents a shift from passive image analysis toward an agentic model capable of executing multistep investigations to identify surface hazards and provide evidence-based recommendations for space mission planners.

Gokce Inal, a PhD candidate in civil, environmental, and geodetic engineering, leads the development of the framework, which builds upon previous university research into vision-language models trained specifically for lunar terrain interpretation. The team presented preliminary findings regarding the system’s capabilities at the 58th Annual Meeting of the American Astronomical Society’s Division for Planetary Sciences in October 2026. The framework aims to assist scientists in processing the substantial volumes of data required to select viable locations for future lunar exploration.

LUMINA functions by extending the utility of LLaVA-LE, a vision-language model previously developed by the research group to interpret lunar imagery. The underlying LLaVA-LE model was trained on the LUCID dataset, which includes approximately 96,000 high-resolution monochrome images paired with descriptive captions and 81,000 question-and-answer pairs. This dataset provides the model with a diverse range of lunar features, including crater distributions, boulder fields, and varying regolith textures, which are essential for accurate terrain classification.

The architecture of LLaVA-LE utilizes a specialized vision-language bridge that maps visual features from lunar imagery into the latent space of a large language model. By fine-tuning this bridge on the LUCID dataset, the researchers enabled the model to perform complex reasoning tasks, such as identifying geological hazards or estimating surface slopes from orbital photographs. In performance evaluations using GPT and Gemini models as benchmarks, LLaVA-LE demonstrated a performance score 3.3 times higher than the baseline LLaVA model in interpreting geological characteristics.

LUMINA takes this foundation further by introducing an agentic framework capable of organizing and carrying out a sequence of analytical operations. Instead of producing an isolated description of an image, it can examine a specific query about a potential landing area, undertake multiple assessment steps, and construct an explanation using the available information. This approach addresses the technical challenge of synthesizing observations that vary in resolution, coverage, and collection conditions across different lunar missions.

Assessing a landing site requires the integration of diverse variables, including surface slope, obstacle density, and operational constraints. The researchers emphasize that the system is designed as a decision-support tool rather than an autonomous authority, ensuring that human specialists retain final responsibility for mission-critical decisions. By providing a coherent scientific explanation for its findings, the framework allows experts to verify the evidence behind each recommendation.

The development of such tools is increasingly necessary as space agencies target the lunar south pole for future missions. This region presents significant engineering challenges, including rugged terrain, steep slopes, and extreme lighting conditions that create high-contrast environments. The Sun’s low position above the horizon produces long shadows, which complicates terrain interpretation and affects the placement of critical infrastructure like solar power arrays and communication systems.

LUMINA evaluates these variables to determine site suitability by examining terrain characteristics alongside mission-specific requirements. The framework provides a structured method for evaluating whether a location meets the necessary criteria for safety and operational success, effectively filtering out sites that do not align with the strict constraints of lunar surface operations. This systematic evaluation process helps reduce the cognitive load on human planners tasked with reviewing massive amounts of orbital data.

Future development phases will involve testing the framework against higher-quality lunar datasets to refine its accuracy and performance under operational conditions. While the preliminary results are promising, the research team acknowledges that further evaluation is required to quantify improvements in planning efficiency and decision-making accuracy. The project continues to focus on creating a specialized tool for automated terrain analysis that can collaborate with human experts to solve complex logistical problems in space exploration.