Maya Harris, Massachusetts Institute of Technology ; Richard Linares, Massachusetts Institute of Technology
https://doi.org/10.64861/ZAKK5427
Keywords: Maneuver detection, maneuver classification, machine learning, semi-supervised
Abstract:
As the number of objects in low Earth orbit (LEO) continues to increase exponentially, tracking and characterizing Earth’s orbital population becomes ever more important. Collision avoidance predictions rely on an understanding of where space objects currently are and where they will be in the future. In addition to the natural forces that affect the behavior of space objects, many satellites also maneuver under their own power, for reasons ranging from routine stationkeeping to collision avoidance to mission objectives. However, operators are not required to publicly share when or why their satellites maneuver, leaving a significant gap in our understanding of the orbital environment. This work proposes a pipeline for detecting when satellite maneuvers occur and characterizing what type of maneuver each one is, using only publicly available data.
Maneuver detection is performed using the Response Estimation and Analysis using Correlated Trajectories (REACT) framework, which identifies maneuvers by comparing a target object’s behavior to the consensus behavior of nearby passive debris objects experiencing the same environmental forces. This work refines REACT’s background object selection method and demonstrates its ability to detect maneuvers across a wide variety of LEO objects using only Two-Line Element Sets (TLEs).
Once a maneuver is detected, physical and contextual features are extracted and used to characterize it. Because confirmed maneuver labels are scarce relative to the number of detected maneuvers, this work explores unsupervised and semi-supervised approaches to maneuver classification, including anomaly detection and label propagation. A feature separability analysis reveals that different maneuver types can have distinct, and sometimes object-specific, feature signatures. Results vary across objects, highlighting both the promise and the challenges of this approach. The end product of this work is a comprehensive pipeline where a user can input any tracked object and receive a labeled history of that object’s maneuvers throughout its lifetime.
Date of Conference: September 15-18, 2026
Track: Conjunction/RPO