In an international collaboration with interdisciplinary groups from the Astrodynamics, Space Robotics, and Controls Lab (ARCLab) of the Massachusetts Institute of Technology (MIT) – with Professor Richard Linares – and the Aerospace Centre of Excellence of the University of Strathclyde (US) – with Professor Massimiliano Vasile –, UPM researchers are driving the evolution of the space sector with advances ranging from the discovery of unprecedented orbits to the prediction of solar storms and the control of autonomous spacecraft.
New Periodic Orbits
One of the most significant advances of this collaborative work is “generative astrodynamics,” where AI-based generative models have been used to discover new periodic orbits in the Earth-Moon system. Using a variational autoencoder generative model trained on NASA data, the team has been able to classify different families of orbits without prior information and generate novel trajectories that, when applied on a large scale, could revolutionize the design of space missions and the theoretical understanding of celestial dynamics.
Improvements in Space Weather Prediction
In parallel, space weather prediction has taken a leap forward with the application of transformer-based architectures, which have become very popular due to their use in language models such as ChatGPT. This approach has allowed for more accurate predictions of the impact of extreme solar events on Earth's atmosphere and satellites. Unlike traditional statistical methods, transformers identify complex temporal patterns and nonlinear relationships in the data, improving the ability to prevent uncontrolled satellite crashes and develop contingency plans for communications and electronic failures during periods of high solar activity.
Autonomous Spacecraft Control:
Finally, the AIDA has made an innovative contribution to autonomous spacecraft control, using language models such as LLaMA and ChatGPT. Through fine-tuning techniques and the design of specific prompts, they have surpassed traditional reinforcement learning-based methods, offering a data-efficient alternative that also facilitates knowledge transfer. This approach earned them second prize in the MIT Lincoln Lab's KSPDG competition and marks the beginning of the use of language models in Guidance, Navigation, and Control (GNC) in space, opening new possibilities for spacecraft automation in future missions.
In short, these advances—the result of collaboration between the UPM team, MIT, and the University of Strathclyde—not only optimize space operations and improve the accuracy of predictions, but also open new avenues for discovery and technological innovation. According to Víctor Rodríguez-Fernández, “The integration of AI into space research promises safer and more efficient space exploration, benefiting various industries and promoting technological development globally. As AI continues to evolve, its application in the space sector will transform our understanding of the universe and enhance our capabilities to explore beyond Earth.”
These teams recently presented their work at the ESA SPAICE Conference 2024, a conference focused on the application of AI to space.
The generative astrodynamics research is part of the project “OrbitGPT - Generative Astrodynamics,” European Space Agency Activities Portal, TEC-SF, Feb. 2024 - Feb. 2025. Contract number: 4000143521, coordinated by Professor Massimiliano Vasile of the University of Strathclyde.
Proceedings of SPAICE2024: the First Joint European Space Agency / IAA Conference on AI in and for Space https://zenodo.org/records/13889941
