UArizona Researchers Develop AI-Powered Methods to Improve Safety in Space

Researchers in the University of Arizona’s Space4 Center are using novel artificial intelligence techniques to improve space traffic management and automation in the increasingly congested space between the Earth and the moon.

Postdoctoral research associates Andrea D'Ambrosio and Lorenzo Federici and graduate research assistant Andrea Scorsoglio are designing algorithms to better track and guide space objects between the Earth and the moon and predict their orbits using ground-based telescopes.

They presented their work at the 2023 AAS/AIAA Astrodynamics Specialist Conference in August. Engineers and researchers from around the world attended, presenting papers on astrodynamics, guidance, navigation and control, space domain awareness, orbit determination and estimation and other topics. 

They also plan to present their work at the Advanced Maui Optical and Space Surveillance Technologies (AMOS) Conference September 19-22. The conference is the premier technical conference in the nation devoted to space situational awareness/space domain awareness.

Managing space traffic – avoiding collisions between objects in space – is becoming more important. As of November 2022, the US Space Surveillance Network reported nearly 26,000 artificial objects in orbit above the Earth, including more than 5,400 active satellites.

“Tens of thousands more are expected to be launched in the next decade as mega-constellations like Starlink expand, presenting challenges to satellite safety and national security,” said Roberto Furfaro, systems and industrial engineering professor and Space4 Center deputy director.

With a focus on cislunar space—the space between and around the Earth and the moon—the center develops and deploys research and education solutions that ensure space remains safe, secure and sustainable.

D'Ambrosio focuses on improving the orbit determination of space objects – estimating where they are and where they are going – using Space4 Center telescopes. Tracking certain kinds of space objects and predicting where they will be is complex but critical to avoiding collisions. The ground-based telescope observations indicate an object’s position in the sky, but not how far away it is. Some space objects like satellites can maneuver using engine thrust, while others move naturally through space. 

Existing methods can estimate where non-maneuvering space objects are in orbit. Using a new method he developed, D’Ambrosio can estimate the orbit of maneuvering and non-maneuvering satellites in the geostationary belt, which circles about 22,000 miles above the Earth’s equator, and cislunar space. That His work applies a novel framework that employs a technique called physics-informed machine learning.

“His research is focused on more complex scenarios in which objects are closer to the Moon, in a more chaotic dynamical environment where common techniques generally struggle,” Furfaro said.

Federici’s work involves scheduling and prioritizing what satellites are observed and imaged through a telescope within a certain amount of time.

“We use telescopes to take images of space objects to realize their orbit determination. But if you want to observe multiple objects in one night, which do you observe first, and which later?” he said.

Federici developed an artificial intelligence-based method to optimally schedule observations of satellites in low Earth orbit and geostationary orbit using RAPTORS-2, a telescope at the University of Arizona’s Space Domain Awareness Observatory. Low Earth orbit encompasses Earth-centered orbits with an altitude of 1,243 miles or less , where the International Space Station orbits. An object in geostationary orbit is in a very high orbit that is ideal for certain kinds of communication satellites and meteorological satellites.

“This way, we know in a single night that we observed the most objects possible, or the most significant ones,” he said. “We can determine what to observe within a pool of objects, and order them.”

While his colleagues focus on space traffic management, Scorsoglio aims to automate spacecraft control and navigation during the final phases of an in-orbit docking maneuver—a  high-precision process whereby a spacecraft moves within a close distance of another object.

“Sometimes in space, objects must maneuver quickly, but the distances are so huge that feedback in real-time from Earth is impossible,” Scorsoglio said. “We want spacecraft to be autonomous and to be able to self-correct.”

To increase autonomy, Scorsoglio uses artificial intelligence methods that leverage a simulated dynamic environment with three-dimensional objects and realistic imaging capabilities and inputs such as the spacecraft’s location, velocity and orientation in space.

Within this environment, he runs simulations of spacecraft nearing one another and learning to solve problems autonomously, through trial and error, using a cutting-edge machine learning technique called Meta-Reinforcement Learning, or MRL. Scorsoglio found that using  MRL can lead to intelligent systems that learn from a simulated environment how to optimally perform a certain task, like a spacecraft guiding itself within close proximity of another object, with no human interaction.