Satellites send back information of one kind or another to Earth. They can do their jobs better when they operate independently, free from terrestrial tethers that restrict performance, keeping course and conducting vital tracking and analysis on their own.
Key to accomplishing this is the autonomy afforded by artificial intelligence and machine learning, or AI/ML — the same technology sweeping through everyday life on Earth. Nowhere does AI promise greater transformation than in space defense, where it can inform and speed the decision-making, first moves and reactions that spell the difference in conflict.
“There is no doubt that artificial intelligence will make satellites more effective for national security in various ways,” said Simone D’Amico, a professor in the Stanford University School of Engineering who has worked more than 20 years in autonomous spacecraft research and development. One example, D’Amico told Apogee, is the handoff between orbiting satellites known as tip and cue. “Think about detecting a missile launch plan very quickly. Autonomous decision-making will be highly boosted with AI, especially when the interpretation of a lot of complex data is needed in a short amount of time.”
Researchers are looking at AI/ML to automate spacecraft functions across the board. Among them: maneuvering, controlling megaconstellations, spotting orbiting objects, tracking debris, identifying and countering anti-satellite threats, detecting targets, and analyzing data. Instead of waiting until it passes a ground station to send down a slug of Earth observation data so someone can make sense of it back here, autonomous satellites can process the data themselves using AI-enhanced onboard computers and just send down what’s needed quickly and cheaply.
“The Pentagon is feeling they’re getting trucks pull up and just unload all of these haystacks, and they’re like, ‘My goodness, how do we get the needles out of these things?’” Michael Bartholomeusz, CEO of satellite maker NOVI Space, told Apogee. “AI and machine learning cut through that noise. They transform vast streams of global data into clear, actionable intelligence that can be understood and acted on much faster.”
The term AI broadly describes a process meant to mimic human intelligence. Machine learning is a way of achieving AI by identifying patterns. How well it works is determined, in part, by the volume and quality of data used to teach or train the machine. This is a special challenge in space because the most powerful computers don’t operate there and because overall human activity and the resulting data are limited compared to terrestrial realms.
United States Space Force Gen. Michael Guetlein, then vice chief of space operations, described this revolution as a military priority during an AI conference in July 2024 sponsored by the Space Force and the U.S. Space Command. “Clearly, artificial intelligence and machine learning can, should, and will play a powerful role in the coming years,” Guetlein said, “including in our efforts to counter emerging threats and preserve the safety and security of our nation, its assets and our allies.”
As with many space capabilities, the U.S. is in a race with the Chinese Communist Party to develop AI/ML for space superiority — a contest that extends beyond technology to questions of strategic dominance. Other countries also are pushing ahead with the technology. European countries are investing heavily in on-orbit data processing and AI-driven operations, Japan focuses on disaster monitoring and autonomous satellites, India is applying the technology to resource mapping and climate monitoring, Canada is advancing AI for maritime and satellite tasking, and Israel is applying AI to defense.
Needles from haystacks
Based in Arlington, Virginia, NOVI Space was founded in 2017 and develops AI/ML-driven systems for both defense and commercial customers. The company’s GENIE constellation platform will deliver autonomy alongside high-resolution geospatial intelligence, enabling satellites to process and interpret data directly on orbit. The first GENIE satellites were scheduled to launch in early 2026. NOVI already has validated its technology through successful mission operations, achieving Technology Readiness Level 9 — the highest standard for operational maturity, according to NASA’s “Small Spacecraft Technology” report issued in February 2025.
In commercial applications, the technology can turn data into decisions, Bartholomeusz said. A satellite could provide a hedge fund or logistics company with daily ship counts in Baltimore Harbor without transmitting gigabytes of imagery. Oil and gas operators could focus on just the perimeter coordinates of a pipeline leak, and farmers could be alerted when specific plots require irrigation — all without reviewing every image frame.
AI/ML also promises to open these capabilities to a much wider market, Bartholomeusz said. “When we developed onboard compute to solve latency challenges for defense missions, we also broke the back of the economic problem,” he said. “If you’re transmitting bits and bytes of intelligence instead of gigabytes of raw data, you reduce cost by an order of magnitude or more. That opens the door for small and midsized enterprises that previously couldn’t afford to use space data. It truly liberates and democratizes access for applications across precision agriculture, food security, safety, maritime operations and infrastructure monitoring.”
The technologies that make onboard computing possible, Bartholomeusz said, include radiation-tolerant miniaturized computer chips such as those made by Silicon Valley’s AMD, paired with optimized AI/ML algorithms made for environments where size, weight and power are at a premium. NOVI’s SP240 onboard computer, for example, weighs about 350 grams, the same as a can of soup, yet vastly outperforms the 26-kilogram guidance computer aboard NASA’s 1966 Gemini 8 spacecraft. That computer helped guide the first space docking mission and weighed about as much as a 7-year-old child, yet it processed the equivalent of just 4,000 words at a time — about 8 kilobytes of data.
Sifting valuable data from the haystacks also is the goal of a three-year demonstration mission orbited in November 2025 and carrying a computer 100 times more powerful than anything yet launched into space, according to the technology website IEEE Spectrum. The powerful H100 GPU from AI chip giant Nvidia of Santa Clara, California, is traveling aboard the Starcloud-1 satellite and will process data from a fleet of Earth-observing satellites operated by Capella Space of San Francisco, California. A GPU, or graphics processing unit, is an integrated circuit designed to rapidly process large amounts of data in parallel, by breaking it into parts. Starcloud, based in Redmond, Washington, is testing the placement of large data centers on orbit to help limit the acreage and resources they gobble up on Earth as the energy-hungry AI/ML revolution spreads.
Bartholomeusz grouped AI/ML satellite developers into two camps, drawing a comparison to the evolution of mobile phones. “Many companies treat onboard computing as just another feature; NOVI treats it as the foundation,” he said. “It’s the difference between early 2000s cellphones — which offered limited computing power — and the iPhone or Android smartphones that redefined the device as a true computing platform.”
He added, “An iPhone is essentially a compact bundle of sensors and compute, made widely available and affordable, unlocking an entire ecosystem of apps and users. Smart satellites follow that same model — integrating sensors with onboard processing to create an open platform for innovation.”
Self-driving spacecraft
Another terrestrial comparison that helps explain the value of AI/ML in space is automobiles. Specifically, the self-driving kind. Widely accepted standards differentiate the capabilities of autonomous cars and trucks, and these designators — from Level 1 to Level 5 — are working their way into space as well. “It borrows some of the terminology from self-driving cars because we want to do the self-driving spacecraft,” said D’Amico, who wears many hats through his work with Stanford. One of them is co-founder and chief science officer with the Stanford spinoff EraDrive Inc. Another is founding co-director of the Center for Aerospace Autonomy Research (CAESAR), a collaboration between D’Amico’s Space Rendezvous Lab and the Autonomous Systems Lab directed by fellow Stanford professor Marco Pavone. CAESAR aims to “endow space vehicles with autonomous reasoning capabilities.”
Most satellites in use today would be rated Level 3, D’Amico said — operating from a script that’s written on the ground and played onboard. Highly micromanaged, Level 3 satellites have limited ability to react to events. At Level 4 and 5, on the other hand, “the AI brain, so to say, is able to learn.”
Check out that Gemini docking mission, when a crewed spacecraft connected with a larger Agena rocket, to see how far space autonomy has evolved, he said. “It’s really very instructive to see how those operations were conducted back in 1966. Pure manual control by the crew, through the human eye. A probe in a cone, basically, male and female. There was an index bar to align with the Agena target vehicle. Like using a joystick.”
In a modern maneuver, as when a SpaceX Crew Dragon docks with the International Space Station, the operation is largely autonomous, a Level 4 with a manual backup. It relies in part on LIDAR, or light detection and ranging — a remote sensing method that sends laser pulses to Earth to measure distances. “The spacecraft uses a number of sensors, from LIDAR to cameras to infrared sensors, for the navigation and the alignment,” D’Amico said. “The process is actually controlled by software and algorithms. There are waypoints that the spacecraft has to acquire and maintain in order to proceed with the docking, preprogrammed before approaching the International Space Station.”
Today, the docking process also follows international standards, “where any port can dock with any other port … instead of being a manual one-time event, largely reactive.”
One project driving the work of CAESAR is known as ART, or Autonomous Rendezvous Transformer, which is using AI/ML to make satellites maneuver smarter. A goal of the center’s work is what D’Amico calls “cognitive spacecraft,” capable of taking direction in simple, common language. “Operators will be able to write prompts, so instead of the standard telecommunications that is used in space,” he said, “they describe what they would like to see the satellite do — for example, ‘Try to approach and repair that unfamiliar tumbling object’ — and all the rest is done by AI with minimal supervision.”
CAESAR was established with support from private sector affiliates Redwire, a space infrastructure company based in Jacksonville, Florida, and Blue Origin, the launch company based in the Seattle area and founded by Amazon’s Jeff Bezos. The Stanford center is conducting research for both companies — in the case of Redwire, to enable autonomous shape characterization and navigation for unknown space objects, and for Blue Origin, autonomous cooperative rendezvous, proximity operations and docking.
Space computing
Progress on autonomy in space hinges on the performance of the integrated circuits powering onboard computing. One type, known as GPU, like the device aboard the Starcloud satellite, originally was designed to render video and graphics for uses such as games and is prized for its capacity to handle massive workloads. Another type, called field programmable gate arrays, or FPGAs, features a programmable hardware fabric that allows it to be reconfigured for a specific ML algorithm’s needs — an advantage over hard-etched circuits.
D’Amico sees promise in FPGAs, with their lower power demands, programmable logic and capacity to survive longer missions across orbits in a demanding space environment. Still, he said, the future might lie in the broad category known as neuromorphic computing. Unlike traditional chips that use a constant clock and separate processing and memory units, neuromorphic chips integrate these units and use an event-driven approach, firing “spikes” only when necessary to save power and increase efficiency. “They work similar to how our brains work, where basically we have neurons that can be activated depending on conditions and the capability to have a level of dynamic control,” D’Amico said.
Like the human brain, AI gets better as it’s trained to the task — a process that involves feeding curated data to selected algorithms to help the system refine itself and produce accurate responses to queries, according to a primer from cloud database company Oracle. Typically, when trained AI models deliver consistent results using training and test datasets, the process moves on to testing with real-world data before going live.
Coming up with enough real-world data is a challenge in space, still a frontier of human activity, and especially in space defense. AI training for autonomous car operations, by comparison, benefits from an operational environment that is easily accessible, D’Amico said. When it comes to AI training for space, he said, even the limited data available can be restricted by secrecy classifications. Data is lacking for training on orbital warfare because — for the most part — there isn’t any, noted an article in the August 2025 edition of Momentum, a quarterly magazine of the U.S. Space Force. “We don’t yet have enough contested environment data in the real world,” the magazine reported. “We’re not seeing regular adversarial maneuvers, lost sensors, or spoofed signals at scale, but we will.”
Synthetic data generated through modeling helps fill the gap in AI training. At CAESAR, this is accomplished, in part, through test beds that function as a surrogate for the orbital environment. “We use robotics, with robotic arms, to move the objects in a scene — an observer satellite, a target satellite,” D’Amico said. “It’s similar to what happens during the shooting of a movie, where we produce the illumination condition of space using special lamps for direct light from the sun, and for diffuse light that’s reflected by the Earth. We can evaluate how well the model that has been trained on pure synthetic data does when deployed in the actual environment. That’s the step before deploying it in space.”
Another major challenge in training autonomous satellites is that the high-powered computers needed for the job exist only on Earth. “Training models in space is impractical due to server requirements,” Bartholomeusz said. “Most models are trained on the ground and then uploaded to the spacecraft. While limited onboard learning is possible, the true focus is on running optimized, power-efficient models within the constraints of space hardware.”
Operator trust
One factor in applying AI/ML to space defense will be the relationship between the new technology and its human operators. Finding ways to leverage the advantages each brings to the task likely will provide the best results, Space Force Lt. Col. Tanisha J. Saunders wrote in a March 2024 research paper for the Air War College in Montgomery, Alabama.
Saunders said, “Machines could cover the areas humans show some weaknesses in, like in rapid data assimilation and analysis, and still enable human counterparts to continue to be successful at processing contextual understanding of dynamic environments to apply military effects.” An important first step, she said, is showing operators that the new technology can be trusted, in part by involving them in testing.
It’s the path all new technology must travel, Bartholomeusz said. “As with any new tool, trust takes time,” he said. “It comes through validation, experience, and the natural ebb and flow of learning.” Added D’Amico, “I think the human should retain the responsibility for strategy, for the high-level objective … exercise ethical judgment, and delegate the routine, the low risk, very quick control loops, especially when it’s a split-second task like collision avoidance.”
Autonomous satellites will create a host of new possibilities in space, D’Amico said. “Drones in very hostile, unknown environments, far away from us, with huge communication delays. How do you get an underwater vehicle to explore underneath the oceans of, I don’t know, Europa? Even closer to us, how do you get a rover to safely explore deep craters or caves under the surface of the moon?”
Another possibility: “A full, autonomous, space-based orbital infrastructure where we provide a kind of intelligence in terms of space traffic monitoring and management. That’s something that my company is working on in making the self-driving spacecraft — not only autonomous, but aware of their surroundings so that the satellites can see what’s happening around them and react. And share that information with others globally through, literally, a space cloud.”
Bartholomeusz sees virtually limitless potential ahead. “AI and machine learning don’t just enhance today’s missions — they enable entirely new ones,” he said. “Satellites could autonomously retask during crises, operate in coordinated swarms, or detect patterns in Earth and space environments that humans would never think to look for. The real promise of AI and ML in space isn’t only about efficiency — it’s about unlocking missions we haven’t yet imagined.”