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

Researchers have evaluated whether prompt-driven vision-language models can allow spacecraft inspection systems to recognize new components after launch without updating onboard software, a capability that remains operationally impractical with conventional supervised models. Spaceborne inspection systems typically deploy perception models before launch. Once in orbit, updating model weights or expanding fixed label sets becomes operationally impractical, limiting the system to the semantic categories defined on the ground [1]. The study investigates zero-shot instance segmentation of spacecraft components using a frozen, single-pass inference protocol on a test set of 129 images of previously unseen satellites [1]. Under fixed global thresholds and no post-processing, the SAM3 model achieved 0.385 [email protected] and 0.267 [email protected]:0.95 [1]. Performance varied sharply by component scale. Large structural elements such as spacecraft bodies reached 0.639 [email protected], and solar arrays reached 0.598 [email protected] [1]. Smaller appendages proved more difficult: antennas scored 0.221 [email protected], while thrusters registered only 0.081 [email protected] [1]. Prompt formulation significantly influenced results. Structured prompts incorporating spatial and geometric descriptors yielded up to 82% improvement over short category-name prompts [1]. The model operated within the memory and compute envelope of contemporary embedded GPUs, suggesting prompt-driven grounding can provide a practical mechanism for post-launch semantic extension of dominant spacecraft structures [1]. The challenge of maintaining and upgrading spaceborne hardware after deployment is not new. The Hubble Space Telescope, launched in 1990, remains the only telescope designed to be maintained in space by astronauts, requiring five Space Shuttle servicing missions to repair, upgrade, and replace systems [2]. For smaller spacecraft and inspection satellites, such physical servicing is not feasible, making software-based adaptability a critical research direction. Unmanned aerial vehicles have faced similar inspection challenges in terrestrial settings. Originally developed for military missions considered too "dull, dirty or dangerous" for humans, UAVs expanded into infrastructure inspections as control technologies improved and costs fell [3]. The pattern of deploying fixed perception models and later needing expanded capabilities mirrors the spacecraft inspection problem. The research appears on arXiv, a preprint server that has become central to rapid dissemination of machine learning research. Platforms such as Hugging Face now provide paper pages that aggregate related models, datasets, and demos, and have collaborated with arXiv to embed interactive demos directly alongside paper abstracts [5][6].