SLiG-Net: A joint pose optimization network for space robot grasping under low-light conditions in on-orbit operations

On-orbit space operations face the dual challenges of low illumination and microgravity environments, which adversely affect the accurate recognition and grasping of targets by space robots, thereby limiting their autonomous operational capabilities. To address grasping under low-light conditions in space, this paper proposes SLiG-Net, a joint optimization method for grasp pose estimation. SLiG-Net consists of two components: LIFE-Net, which enhances low-light image quality through multi-scale feature enhancement and attention mechanisms to effectively restore detail information; and HMPG-Net, which integrates RGB semantic features with point cloud geometric data to design a hybrid-metric grasping strategy based on local contact stability, surface flatness, and center-of-mass distribution, enabling multi-dimensional optimization of grasp poses for freely floating objects. To accommodate the computational constraints of space platforms, the network architecture is lightweight, ensuring real-time performance and deployment efficiency. Simulation experiments demonstrate that SLiG-Net outperforms RetinexNet and Restormer in image enhancement tasks, achieving higher PSNR and SSIM scores with fewer parameters. In grasp evaluation on the GraspNet-1Billion dataset, the method surpasses existing approaches in AP and various threshold metrics, while maintaining high grasp success rates under extreme low-light conditions, validating its practicality and robustness for space on-orbit maintenance tasks.