New AI system hunts for satellites behaving oddly in space
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Sarah Grillo/Axios
A new AI system can spot satellites acting strangely in the increasingly congested orbit around Earth and that could be malfunctioning or have more nefarious purposes.
Why it matters: Companies are announcing plans to launch large constellations in the coming years, potentially putting more than one million satellites in space. And governments are increasingly investing in space-based electronic warfare tools, anti-satellite weapons and other spacecraft for national security that could be concealed in the mix.
- "This isn’t some hypothetical problem," says Audrey Schaffer, vice president of strategy and policy at Slingshot Aerospace, which developed the new AI tool with funding from the Defense Advanced Research Project Agency (DARPA).
- "We think this tool is very powerful for addressing near-term security challenges," Schaffer says.
How it works: The system — called Agatha — identifies anomalous behavior in a constellation and then deduces the strategies and intentions of suspect satellites, the company says.
- Instead of looking at a single satellite and determining how its behavior is different from what would be expected of it, Agatha identifies small differences in how a satellite is behaving compared to large numbers of other satellites in a constellation.
- That includes a satellite that's heavier or lighter than the rest with different maneuvering characteristics, which could indicate it is carrying a different payload, Schaffer said.
- Agatha can also detect unusual communications patterns. Rather than having uplinks and downlinks around the world, an espionage satellite might only be communicating with a handful of ground stations in the host country or allied countries.
- "We are rapidly approaching a moment where no human or team of humans would be able to monitor all activity in space let alone these minute differences," Schaffer says.
Zoom in: Constellations are relative newcomers to space, which meant the team at Slingshot didn't have massive amounts of data with which to train the system's algorithms.
- They simulated 60 years-worth of constellation data by creating artificial satellite environments, then planting simulated bad actors and training the AI on that data, says Dylan Kesler, director of data science and AI at Slingshot.
- The system was then tested on real-world data to see if it could identify unusual behavior from satellites that were then verified by their operators to be in the midst of changing missions or malfunctioning.
- One technique Agatha uses is inverse reinforcement learning (IRL), an AI approach related to reinforcement learning (RL) that plays a role in natural language processing, generative AI and other tasks. But unlike RL that makes "an actor that is governed by a policy," IRL observes actors (in this case, satellites) and derives a policy, Kesler says.
- It looks not just for single maneuvers but for patterns of multiple maneuvers.
What to watch: Schaffer says U.S. Space Command would be an ideal user of Agatha.
- It has "tremendous potential to enable national security space operators to sift through increasingly large amounts of data on what's happening in space to identify the needle in the haystack of these potentially hidden satellites," Shaffer adds.
The big picture: Because Agatha is "data-agnostic," it could be turned on genomic, biomedical and other data, says Kesler, who formerly worked in the biotech industry.
