Using LLMs to Generate Data Benchmarks for LEO Constellation Task Scheduling
Abstract
Low Earth Orbit (LEO) satellite constellations require robust scheduling mechanisms to manage multiple observation requests under limited resources. However, a significant semantic gap remains between high-level human requests and the structured planning requirements of scheduling systems. Existing benchmarks primarily focus on the operational scheduling layer and cover only a limited range of scenarios. To address this limitation, we present the KEPLER Data Factory (KDF), a framework that bridges mission-level intent and satellite scheduling by combining deterministic orbital physics with semantic text generation using Large Language Models. We evaluate KDF using multiple large language models. The results show high accuracy in generating mission attributes such as sensor types and priority levels, while highlighting lower performance in producing fine-grained temporal information, including specific days and times. To facilitate reproducibility and future research, both the KDF framework and the generated datasets are publicly available.
The KEPLER Data Factory Subsystems
Data collector (DC)
The pipeline orchestrator reads a parameterized specification configuration file to establish a centralized scenario profile:
- Satellite Fleet Context Initialization: Incorporates empirical satellite orbital elements by mapping specified NORAD tracking numbers or fetching active LEO asset categories directly from the Celestrak API. Every profile is thoroughly parsed to ensure compliance with a strict LEO orbit limit mean motion constraint of 11.0 or more revolutions per day.
- Stochastic Payload Subsystem Mapping: Platform assets are dynamically bound to strict physical limits using user-defined probability density distributions, mapping discrete RF communication bands (S, X, or Ka), solid-state recorder memory storage profiles, nominal data downlink transmission rates, and specific payload sensor allocations.
- Procedural Task Generation: Synthesizes target configurations randomly sampled from localized bounding envelopes. Tasks are designated as distinct geodetic points or multi-point polygonal scan areas, dynamically assigned a strategic mission priority (levels 1 to 3), and parameterized with strict availability release time offsets and deadline constraints anchored relative to
t₀.
Physics engine (PE)
Operating as the high-fidelity Physical Simulation Layer, this module propagates orbital segments over a discretized horizon step sequence to evaluate exact visibility intervals:
- Deterministic Pass Window Calculation: Continuously computes satellite state position and velocity vectors relative to the geodetic grid using standard SGP4 analytical propagation logic. The engine isolates valid topocentric access windows bounded strictly by frequency band-specific tracking elevation baselines and slant range propagation loss boundaries.
- Payload Geometry and Attitude Filtering: Filters and pairs computed look-ahead passes according to required sensor capability matching rules. Spacecraft line-of-sight tracking geometry is continually modeled in the Local-Vertical Local-Horizontal (LVLH) reference frame to calculate precise roll and pitch limits, discarding passes that violate operational sensor look angle thresholds.
Request generator
Operating as the Mission Strategy Layer, this component interfaces with localized inference models to bridge the semantic domain gap:
- Linguistic Strategic Translation: Converts exact geodetic coordinate sets into human-digestible geopolitical descriptions (country and city) via a reverse geocoding pipeline using the Nominatim API. Mathematical epoch deadlines are contextualized into intuitive, natural language timeframes representing localized relative planning horizons and diurnal phases.
- Automated Evaluation and Validation: Mitigates model hallucinations by validating output requests directly against simulation parameters. Regular expressions extract time tokens into isolated temporal fragments to preserve vector directional integrity. These structures are transformed via a local embeddings model and compared using vector cosine similarity to evaluate geometric alignment against reference control vocabularies.