ETD

• autonomous driving
• autonomous navigation
• curriculum learning
• depth
• EAC
• ESA
• European Astronaut Centre
• European Space Agency
• hybrid dataset
• LUNA Analog Facility
• lunar surface
• moon
• moon's surface
• multimodal
• object detection
• obstacle detection
• planetary exploration
• rgbd
• robot
• robotic
• rock detection
• rover
• sim-to-real
• space exploration
• style transfer
• YOLO
• YOLO-based

This work addresses the problem of obstacle detection in lunar scenarios characterized by extreme illumination conditions, investigating the combined role of sim-to-real domain adaptation and multimodal RGB–Depth perception. The study first establishes strong RGB-only baselines trained under controlled illumination, analyzing the impact of dataset composition on generalization performance. Real images, synthetic data, and sim-to-real stylized samples are evaluated to quantify their contribution to detection accuracy. Results show that, while style-transferred synthetic data improves generalization under favorable lighting, RGB-only detectors suffer significant performance degradation under strong illumination variations and low-visibility scenarios.

To address these limitations, a mid-fusion RGB–Depth detection architecture is proposed, integrating geometric information into a YOLO-based framework. A progressive fine-tuning strategy is introduced to stabilize training and mitigate catastrophic forgetting when extending RGB-pretrained models to multimodal inputs. Extensive experiments demonstrate that the proposed RGB–Depth approach significantly improves detection robustness compared to RGB-only baselines. In particular, depth information enables reliable object detection in scenarios where photometric cues become unreliable, without degrading performance under standard illumination conditions.