The Defense Department has been warning of the growing risks of weapons in space, but the massive increase in objects about to enter low-Earth orbit will make tracking potential weapons even harder. A California company called Slingshot Aerospace has created an AI system to address that challenge, by identifying strangely behaving satellites within large satellite constellations—which could indicate space weapons or other threats.
On Wednesday, Slingshot Aerospace announced a new contract with the Defense Advanced Research Projects Agency for the system, called Agatha.
Dylan Kesler, the director of data science and AI at Slingshot Aerospace, told Defense One that detecting the strange movements of satellites in geosynchronous orbit—like the Russian Luch satellite detected diverting from its predicted path in 2015 to sidle up next to a pair of Intelsat communications satellites in a possible hacking attempt—is simple compared to the challenge of detecting space weapons that could be hidden among new satellite constellations in the much more crowded low-Earth orbit.
“The way that these sorts of analyses have been done throughout history up until recently was literally to say ‘We know that this spacecraft is is maneuvering in a way we wouldn’t expect and it’s getting close to another spacecraft and the GEO belt, and we should be analyzing that,’” Kesler said.
Now, as the number of satellites in LEO increases, it’s harder to spot potential outliers. And LEO is only becoming more congested. Just last week, China announced a new plan to launch 10,000 new satellites into LEO. That’s on top of a January announcement of 26,000 future Chinese satellites into low-Earth orbit, which itself is just part of a much larger wave.
Tracking potential nefarious behavior in constellations of that size, which are orbiting the Earth far faster than satellites in GEO, requires the ability to process more data from more places, Kesler said. Instead of using the machine learning tools analysts used to track satellites like Luch, the new normal will require more sophisticated approaches, like inverse reinforcement learning.
Strange and possibly illicit payloads, or a host of other factors, could cause a single satellite out of a large constellation to behave unusually, Kesler said. The goal is to detect that behavior as soon as possible so that relevant parties, like Space Command, can use intelligence resources to determine what is causing that behavior.
But with little if any data on actual space weapons available, how can Slingshot Aerospace test whether the AI system is detecting them accurately?
Said Kesler: “We challenged it with both the simulated data and then we crossed the sim into real.” Specifically, the company ran a series of scenarios to show how well Agatha could detect anomalies in virtual settings, and then applied the AI to satellite constellations already in space.
“It functioned in the real world as well, and identified satellites within those real world constellations that were suffering faults or that were changing their mission. And we talked to the owner-operators and indeed validated that the satellites that Agatha identified were in fact different from the others.”
In at least a couple of instances, the satellite operators only learned of the potential malfunctions from the company.
