Evaluating Depth-Augmented Visual Features for Cross-Dataset Lunar Traversability Estimation

This research investigates whether augmenting visual features with depth-derived geometric information improves lunar traversability estimation. Using a weakly supervised framework trained on sparse pairwise annotations, the authors evaluate models in both terrestrial analogue and real lunar settings using Chang'E PCAM data. Contrary to expectations, naive geometric augmentation yields no consistent improvement over a visual-only baseline; instead, the choice of training source is the dominant factor for transfer quality. The evaluation runs on a resource-constrained edge device, providing a controlled baseline for future multimodal fusion architectures.