New AI framework learns to track and predict the behaviour of satellites and space debris

As the number of satellites and debris in orbit continues to grow, automated tools for monitoring their behaviour have become essential for space safety and sustainability. A new self-supervised learning framework, presented in a research paper on arXiv, offers a step toward building a comprehensive foundation model for space object analysis.

How the system works#

The researchers built their model around a Perceiver-Variational Autoencoder (VAE), a neural network architecture that learns to reconstruct and understand patterns in data without requiring human-annotated labels. The team pre-trained this model on approximately 227,000 light curves—brightness measurements over time—collected from the MMT-9 observatory. This self-supervised pre-training taught the model to recognize normal orbital behaviour by learning to reconstruct the light patterns it observed.

The pre-trained model achieved a reconstruction error of just 0.0012, meaning it could accurately reproduce the brightness patterns it had seen. Objects that produced light curves difficult to reconstruct were flagged as potentially anomalous—a simple but effective detection method.

Testing and refinement#

To validate and improve the system, the researchers fine-tuned it using two independent light curve simulators (CASSANDRA and GRIAL) paired with 3D computer-aided design (CAD) models of real spacecraft: boxwing platforms, Sentinel-3, SMOS, and Starlink satellites. After this refinement, the model achieved 85% to 82% accuracy in identifying anomalies and predicting motion modes—such as distinguishing between sun-pointing orientation, spinning, and tumbling. The model also scored high on ROC-AUC metrics (0.92 to 0.95), indicating strong discriminative ability.

Real-world insights#

When applied to actual observational data, the model revealed recognizable patterns: distinct brightness signatures for different object types, and the characteristic flashing of sunlight reflecting off satellite surfaces (satellite glinting). These findings suggest the model had learned meaningful representations of how real objects behave in orbit.

Broader implications#

The framework simultaneously enables three critical capabilities: (spotting unusual objects or behaviours), (determining how an object is moving), and (creating realistic simulated light curves). This multi-tasking ability, emerging from a single pre-trained representation, demonstrates how self-supervised learning can support automated space monitoring and help sustain a safer orbital environment as traffic increases.