Harnessing Large Language Models for Satellite Ground Tests
Brian J. Connolly, Kristen M. Anderson · 2024
This study examines the potential of Large Language Models (LLMs) in augmenting satellite ground testing tasks. The focus is on a Coarse Attitude Sensor (CAS), a component of a satellite system responsible for attitude determination. The capabilities of LLMs are evaluated across three primary tasks: refining component requirements, generating code documentation, and assisting in satellite test engineering tasks. The LLM employed in this study, GPT-4-32k-0613, demonstrated promising results in the initial stages of requirements analysis and code documentation. However, the generation of test scripts revealed severe limitations, particularly in the context of fully autonomous operation. LLMs used in an unsupervised manner can make dangerous oversights that can be difficult for even expert humans to spot. The study concludes that while LLMs hold significant potential in aiding complex engineering tasks, the current level of technology is most effectively used to supplement human expertise, not replace it. The study also highlights the importance of expert supervision and careful, iterative prompting to avoid critical errors in task execution.