Targeting on-orbit servicing scenarios and focusing on two key challenges of complex instruction parsing and lightweight deployment in space perception, this paper provides a systematic review of the collaborative parsing paradigms of vision foundation models driven by complex instructions, along with progress in their lightweight deployment on resource-constrained spaceborne platforms. The study compares the applicable scenarios and computational overhead of various collaborative segmentation paradigms under aerospace constraints. To address the \"memory wall\" bottleneck of spaceborne chips, it systematically elaborates on software-hardware co-designed lightweight architectures. Additionally, considering the high-risk characteristics of the space environment, it explores uncertainty-aware hierarchical fallback mechanisms and identifies the limitations of current evaluation benchmarks. Comprehensive analysis demonstrates that appropriately coordinating collaborative parsing paradigms and integrating software-hardware lightweight architectures with uncertainty-aware fallback mechanisms constitute a crucial pathway to addressing spaceborne resource constraints and environmental unpredictability. Meanwhile, embodied perception, multimodal collaboration, and high-fidelity synthetic data will emerge as pivotal directions for future research. This study closely adheres to aerospace physical constraints and systematically maps the critical technical links－from complex instruction parsing and spaceborne lightweight deployment to a closed-loop reliability assurance. It establishes a comprehensive evaluation perspective for on-orbit perception systems, providing insights and references for research on next-generation autonomous on-orbit perception systems.