Post-Launch Capability Expansion of Vision-Language Models via Prompting for On-Orbit Spacecraft Inspection

The study investigates the use of prompt-driven vision-language models for post-launch semantic expansion in spaceborne inspection systems, allowing new spacecraft components to be identified via natural language without altering onboard model weights. The model, SAM3, evaluated on a dataset of 129 images, achieved zero-shot instance segmentation with a mean Average Precision (mAP) of 0.385 at IoU 0.5, showing significant scale-dependent performance, particularly for larger components. This approach enables practical on-orbit updates for spacecraft inspections, highlighting the potential of prompt-based methods while also noting challenges in localizing smaller components due to orbital domain shifts.