Multimodal occlusion-aware perception for robotic manipulation of biological samples in space station refrigerators

Abstract The on-orbit transfer of biological samples in space station cryogenic refrigerators is currently performed mainly through manual astronaut operations. However, the extremely low-temperature environment (−80 °C) increases operational risks and limits efficiency. Robotic autonomous grasping and transfer provide a promising alternative, in which reliable target perception is critical for safe and effective manipulation. In microgravity environments, samples inside the refrigerator are freely floating, resulting in complex and random occlusion relationships that challenge conventional object detection methods. To address this issue, this paper proposes an end-to-end occlusion-aware object detection method based on YOLO11. The proposed approach introduces an extended multi-channel RGB-D input, expanding the conventional four-channel representation to a six-channel structure to enhance depth and spatial feature representation. Furthermore, a spatial relationship perception module is incorporated before the detection head to model inter-object spatial configurations and occlusion relationships, enabling occlusion-aware detection without relying on explicit segmentation masks. Consequently, the detector produces outputs that are more informative for robotic decision-making. Experiments conducted on a self-built microgravity simulation dataset demonstrate that the proposed method significantly improves detection performance under complex occlusion conditions, achieving a 12.4% increase in mAP compared with the original YOLO11. These results indicate that the proposed method effectively models occlusion relationships while preserving the simplicity of an end-to-end detection framework, providing reliable perceptual support for robotic manipulation of biological samples in space station refrigerators.