02 - Systems Engineering and Integration
MANEUVER- AND DELAY-ROBUST DISTRIBUTED SATELLITE SSA VIA SPATIO-TEMPORAL GRAPH FORECASTING
H. Rizk¹, H. Yamaguchi¹, S. Morioka², T. Inagawa²; ¹Osaka University, Japan ;²Interstellar Technologies Inc., Japan
The rapid growth of LEO constellations is turning Space Situational Awareness (SSA) into a real-time bottle- neck: conjunction screening must be refreshed frequently, yet catalog updates arrive with latency and gaps, and maneuvers can abruptly invalidate propagator-only predictions, destabilizing closest-approach estimates and risk ranking. We present a distributed SSA framework designed to remain reliable under delayed and incomplete information by combining interaction-aware trajectory forecasting with uncertainty-calibrated triage. Inputs are time-aligned inertial state sequences derived from TLE/SGP4 propagation, and outputs are multi- step forecasts of state means together with calibrated uncertainty. We model SSA as a dynamic spatio- temporal interaction graph whose edges encode relative motion and risk-relevant neighborhoods informed by CPA/TCA. A spatio-temporal graph forecaster predicts short-horizon trajectories, and candidate pairs are ranked using an uncertainty-aware score to produce a stable top-K triage list. Distributed execution exchanges only compact boundary summaries (forecast states, uncertainty, and timestamps) rather than raw catalogs, and robustness is evaluated under message delay and loss. In a 90-day catalog-driven test with 4,500 RSOs, our method reduces 60-min position RMSE by 33% (2.84→1.91 km) and improves triage quality by 27% in Re- call@50 (0.62→0.79), 27% in NDCG@50 (0.55→0.70), and 39% in PR-AUC (0.41→0.57). Under distributed delay/loss (τ = 3, p = 0.3), Recall@50 remains 0.73 with 7.6 kB/update/node.