Intent Inference of Non-Cooperative Rendezvous and Proximity Operations

Abstract

Rendezvous and proximity operations (RPOs) are a standard class of space operation, increasingly characterised by dual-use and adversarial potential. When conducted non-cooperatively, the intent behind an RPO is unknown, complicating threat assessment for operators. This thesis presents a methodology for inferring RPO intent from observable trajectory data. Five intent classes are defined from representative mission objectives, and synthetic trajectories are generated for each via multi-objective optimisation, spanning both low Earth orbit and geosynchronous orbit regimes. An XGBoost classifier is trained on manoeuvre features extracted from this synthetic dataset to distinguish between intent classes, and is benchmarked against Random Forest and SVM alternatives. On synthetic data, the classifier reliably discriminates between intent classes, with accuracy improving as more manoeuvres, and thus more trajectory context, become available. Validation against real-world tracking data shows the same trend but at lower absolute accuracy, revealing a sim-to-real gap that is traceable to training data. This gap indicates that the synthetic training data does not fully capture the manoeuvre behaviour of real-world, potentially adversarial chasers, and that closing it will require closer alignment between simulated and observed RPO dynamics.