Prompting Enables Post-Launch AI Expansion for Spacecraft Inspection

Spaceborne inspection systems typically deploy perception models before launch, making subsequent updates to model weights or expansion of fixed label sets operationally challenging. This paper investigates a novel approach: using prompt-driven vision-language models to enable post-launch semantic expansion. The goal is to allow new spacecraft components to be identified and specified through natural language prompts, eliminating the need to modify or re-upload onboard model parameters. The study evaluates the zero-shot instance segmentation capabilities of a model (SAM3) on a test set of 129 images of previously unseen satellites, adhering to a strictly frozen, single-pass inference protocol. Under fixed global thresholds and without post-processing, SAM3 achieved a mean Average Precision (mAP) of 0.385 at an Intersection over Union (IoU) threshold of 0.5, and 0.267 across IoU thresholds from 0.5 to 0.95. Performance was found to be strongly dependent on the scale of the objects; large structural elements like spacecraft bodies and solar arrays were reliably localized, while smaller appendages such as antennas and thrusters remained more difficult. Crucially, prompt formulation significantly influenced performance, with structured prompts incorporating spatial and geometric descriptors yielding up to an 82% improvement over simple category-name prompts. The model operates within the memory and compute constraints of contemporary embedded GPUs, suggesting that prompt-driven grounding offers a practical mechanism for extending the semantic capabilities of spacecraft inspection systems post-launch, while also highlighting current limitations for fine-scale component localization under orbital domain shift.