How AI Is Used in Space, From Mars Rovers to Smart Satellites

AI is used in space for two main jobs. One is letting spacecraft make routine decisions on their own when they are too far away or moving too fast for people to steer. The other is sorting through more data than scientists could review by hand. Mars rovers pick paths around rocks. Small satellites judge which pictures are worth keeping. Models on the ground help map the Moon. Most of it is narrow, task-specific software that people still supervise.

Does NASA use AI?

Yes, and the agency says it has done so for decades. NASA's overview page states that 88% of the driving done by its Perseverance rover on Mars has been autonomous. It lists uses that range from sifting satellite imagery to searching telescope data for planets beyond the solar system.

As of October 2026 that page is titled Super Intelligence. NASA says it is updating its wording to follow Executive Order 14434 of September 29, 2026. The order tells federal agencies to use that term in place of artificial intelligence. For now, it defines the term as covering the same technologies. NASA also submits an inventory of its use cases to the Office of Management and Budget. NASA says many activities in its 2026 submission were still in formulation or development.

How is AI used on Mars rovers?

Distance is the reason. Mars is on average about 140 million miles (225 million kilometers) from Earth. Signals take too long to cross that gap for anyone to steer a rover live. Three kinds of onboard autonomy fill the gap:

• Self-driving. Perseverance photographs the ground ahead, and its onboard computer spots hazards in the images and steers around them.
• Choosing targets. Curiosity, the older rover, pioneered software that selects rocks for its laser by their shape and color. Perseverance can do the same.
• Deciding what to study. Software called adaptive sampling positions Perseverance's PIXL X-ray instrument and flags minerals worth a deeper look. NASA's Jet Propulsion Laboratory (JPL) described it in 2024 as the first use of AI on Mars to make autonomous decisions from real-time analysis of rock composition.

The newest step uses generative AI on the ground. On December 8 and 10, 2025, Perseverance completed the first drives on another world planned by AI, according to JPL. Vision-language models are a type of generative AI that works with images as well as text. Here they analyzed orbital photos and terrain-slope data and produced the waypoints that human route planners normally plot. JPL ran the demonstration with Anthropic, using its Claude models. The rover drove 689 feet (210 meters) and then 807 feet (246 meters). Before sending commands, engineers ran them through a virtual replica of the rover and verified more than 500,000 telemetry variables.

Do satellites use AI?

Some do, mostly to decide what to photograph and what to discard. Φ-sat-1 was launched on September 3, 2020. The European Space Agency (ESA) describes it as the first experiment to demonstrate AI for Earth observation. It screened out flawed images on board so that only usable data was sent down.

Its successor, Φsat-2, launched on August 16, 2024. The small cubesat carries six AI applications, including cloud detection, ship detection, wildfire detection and image compression.

NASA has tested a more active approach called Dynamic Targeting. In mid-July 2025, software on the commercial CogniSAT-6 satellite looked about 300 miles (500 kilometers) ahead along its orbit. It told clouds from clear sky and determined where to point its camera. It did all of this in less than 90 seconds with no human involvement. JPL says clouds can block the view of the surface as much as two-thirds of the time. Later tests are meant to hunt for storms, wildfires and volcanic eruptions.

How does AI help scientists on the ground?

Much of the work happens after data arrives. In September 2026, NASA and IBM released an open-source foundation model for lunar science. It was trained on roughly 2 million image tiles from the Lunar Reconnaissance Orbiter. A foundation model is trained once on a broad dataset and then adapted to specific tasks. Here, those tasks include mapping craters and estimating where ice may exist near the lunar poles. The lunar model joins earlier NASA-IBM models for Earth observation, named Prithvi, and for the Sun and space weather, named Surya.

NASA is also testing AI that coordinates several machines at once. In summer 2026, NASA scientists and industry partners ran a field test in Virginia of a drone and two rovers working as a team. The trial assessed whether the fleet could pick scientifically interesting spots by itself. It also assessed whether the fleet could decide which robot to send, balancing hazards against scientific return.

What to watch

JPL noted in 2024 that planning each day's commands for either Mars rover still took dozens of engineers and scientists. The AI-planned drives were a demonstration that engineers verified before transmission. JPL described Dynamic Targeting in July 2025 as still in testing. Two signals are worth following. One is whether generative planning becomes routine in rover operations. The other is whether onboard decision-making moves from small test satellites to major science missions.