SB-ObsGen: A Framework for Flyby Small Body Observation Plan Generation Using LLMs and Physics-Informed Tools

Deep space exploration missions may encounter opportunities to visit previously unplanned small bodies, which require timely and reliable observation planning under complex engineering constraints. However, traditional manual planning is labor-intensive and difficult to adapt to multi-objective flyby scenarios, while fully autonomous systems still face reliability limitations. To address these challenges, this study proposes SB-ObsGen, a framework for flyby small-body observation plan generation that integrates large language models (LLMs) with physics-informed tools. The framework follows a Generate-and-Optimize paradigm: an Observation Plan Generation Module decomposes tasks and invokes domain-specific tools to derive feasible observation information, and a Plan Optimization Module iteratively refines the initial plan through discriminator-guided feedback. Experiments on 80 mission scenarios show that SB-ObsGen achieves strong performance across different LLM backbones, with the best configurations reaching over 85% plan consistency with reference recommendation plans. Comparative experiments against classical planning methods further show that the proposed framework is competitive on structured planning tasks while offering additional advantages in tool use, flexible constraint handling, and end-to-end planning from heterogeneous inputs. Ablation studies confirm the critical role of the Plan Optimization Module in improving final plan quality.