Space
Last reviewed
May 13, 2026
Sources
40 citations
Review status
Source-backed
Revision
v2 · 4,698 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 13, 2026
Sources
40 citations
Review status
Source-backed
Revision
v2 · 4,698 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: Space ChatGPT Plugins
Artificial intelligence in space refers to the use of machine learning, deep learning, autonomous planning systems, and other AI techniques across the space industry. Applications span satellite Earth observation analytics, autonomous spacecraft operations, robotic exploration of other worlds, exoplanet discovery, telescope data analysis, space weather forecasting, debris tracking, and large constellation management. The combination of huge data volumes, limited communication windows with ground stations, and the impossibility of real time human control beyond low Earth orbit has made AI one of the most consequential enabling technologies in modern spaceflight.
Space missions generate data at a scale that long ago outstripped the capacity of human analysts. A single Earth observation satellite can capture terabytes of imagery per day. The James Webb Space Telescope (JWST) downlinks roughly 235 gigabytes of science data daily. The Vera C. Rubin Observatory's LSST camera, once survey operations ramp up, is expected to produce around 20 terabytes of raw imagery per night and create roughly 10 million transient alerts per night, far more than any human team could vet. Spacecraft operating beyond Earth orbit also face round trip light delays that make real time joystick control impossible, with one way signals taking between about 4 and 24 minutes to reach Mars depending on planetary geometry.
These constraints push mission designers toward onboard autonomy and ground side machine learning pipelines. Commercial Earth observation firms train neural networks on labeled imagery to detect ships, aircraft, vehicles, oil tanks, deforestation, flood extents, and structural damage. Space agencies use AI planners to schedule rover activities, select science targets, avoid hazards during descent and landing, and recover from anomalies. Astronomers use deep learning to classify galaxies, identify gravitational lenses, vet exoplanet candidates, and search for fast radio bursts. Satellite operators rely on machine learning to predict conjunctions, decide when to maneuver, and characterize unknown objects in orbit.
The market has grown quickly. By 2025 there were roughly 12,000 active satellites in orbit, with SpaceX's Starlink alone executing close to 50,000 autonomous collision avoidance maneuvers in the six months from December 2023 to May 2024, a level of operational tempo that would be impossible without on board decision logic. Industry analysts expect the number of active spacecraft to push past 100,000 by 2030, which essentially forces every major operator into some form of AI assisted traffic management.
AI in space is older than most people think. The Jet Propulsion Laboratory and NASA Ames Research Center began experimenting with autonomous control software in the early 1990s as part of the New Millennium Program. The breakthrough came on May 17, 1999, when the Remote Agent Experiment took command of NASA's Deep Space 1 probe for roughly two days during its cruise to asteroid 9969 Braille. Remote Agent was the first artificial intelligence system to control a spacecraft without direct human supervision, handling three simulated failures including a stuck attitude control thruster. The system combined a planner and scheduler, an executive component, and a model based fault detection and recovery engine. It won NASA's 1999 Software of the Year award.
The NASA Mars Exploration Rovers Spirit and Opportunity, which landed in January 2004, introduced limited autonomous navigation, basic hazard detection from stereo imagery, and the AEGIS autonomous targeting system. AEGIS, short for Autonomous Exploration for Gathering Increased Science, was first uploaded to Opportunity and later ported to Curiosity, where it entered routine use on the ChemCam instrument in May 2016. AEGIS lets the rover identify rock targets in navigation camera imagery and command its laser spectrometer to analyze them without waiting for a ground command cycle. Performance reports show better than 93 percent success at selecting the requested geological material.
Curiosity itself debuted full AutoNav on August 27, 2013, becoming the first NASA rover able to drive across terrain that had not been cleared by Earth based operators. AutoNav builds a local hazard map from stereo cameras, picks a path, and drives to it. Curiosity ended up using AutoNav for roughly 6 percent of its odometry.
The Perseverance rover, which landed in Jezero Crater on February 18, 2021, brought several major upgrades. Its Vision Compute Element, a dedicated FPGA based computer separate from the main rover brain, lets path planning run while the wheels are turning. By late 2024 Perseverance had completed about 88 to 90 percent of its 17.7 kilometers of driving autonomously, an enormous jump from prior missions. The same mission introduced Terrain Relative Navigation during entry, descent, and landing, with a Lander Vision System matching descent imagery against an onboard map of the landing zone. The system delivered the rover within about five meters of the target, an order of magnitude better than the requirement.
The Ingenuity helicopter, which made its first powered flight on Mars on April 19, 2021, ran a fully autonomous control loop on a Snapdragon 801 mobile processor, using visual odometry from a downward facing camera, an inertial measurement unit, and a laser altimeter. Its flight software used JPL's open source F Prime framework and ran on Linux, a first for an off Earth flight system. By the time of its final flight on January 18, 2024, Ingenuity had completed 72 flights, vastly exceeding the planned five flight technology demonstration.
ESA's Phi-Sat-1 launched on a Vega rocket on September 3, 2020 as part of the FSSCat mission and became the first European Earth observation satellite to run a convolutional neural network onboard. The cloud detection model, running on an Intel Movidius Myriad 2 vision processor, filtered cloudy hyperspectral imagery before downlink, saving bandwidth. First successful onboard AI inference was confirmed on September 28, 2020.
On October 14, 2024, NASA's Europa Clipper launched on a Falcon Heavy from Kennedy Space Center, beginning a 1.8 billion mile cruise to Jupiter. The 7,145 pound spacecraft will perform 49 close flybys of Europa after arrival in April 2030, relying on a substantial onboard fault management and autonomy stack to survive Jupiter's intense radiation environment.
Earth observation has been the most commercially active area for AI in space. A dozen or so private operators run constellations of optical and synthetic aperture radar (SAR) satellites and use machine learning to turn pixels into structured information. The table below summarizes the major players.
| Company | Sensor type | Notable AI capability |
|---|---|---|
| Planet Labs | Optical (Dove, SkySat, Pelican) | Planet Insights Platform combines Analytic Feeds and Planetary Variables; the Pelican-4 satellite, launched in 2025, performs onboard inference on an NVIDIA Jetson Orin module |
| Maxar (now Maxar Intelligence) | Very high resolution optical (WorldView, Legion) | Precision3D building footprints, automated change detection, classifier driven object extraction used heavily for Ukraine damage assessment and the 2023 Türkiye earthquake response |
| Capella Space | X-band SAR | Vessel detection model and the Analytics Partner Program for ML based SAR analytics |
| ICEYE | X-band SAR | Flood Rapid Impact product delivers automated flood extents within 6 to 12 hours; partnership with SATIM for Detect and Classify model targeting vessels, aircraft, and vehicles at better than 90 percent accuracy |
| BlackSky | Optical (Gen-2 and Gen-3) | Spectra AI tasking and analytics platform with around 90 minute end to end delivery |
| Satellogic | Optical (NewSat, NextGen) | AI-first constellation with onboard inference; released about 6 million open training images via the EarthView dataset in 2024 |
| Umbra | High resolution SAR | Commercial SAR feeds used widely by third party ML developers |
Larger software platforms aggregate these feeds. Microsoft's Planetary Computer hosts more than 120 public geospatial datasets totaling over 50 petabytes and has rolled into the production Planetary Computer Pro service that integrates with Azure AI Foundry and Microsoft Fabric. Google Earth Engine offers a managed cloud platform with built in ML APIs for supervised and unsupervised classification and regression, and through ee.Model.fromVertexAi users can serve TensorFlow or PyTorch models trained outside the platform.
NVIDIA's Earth-2 platform, announced as a digital twin and then expanded in 2025 into an open family of weather and climate AI models including FourCastNet3, CorrDiff, and a global data assimilation stack, illustrates how the same pixel feeds are being used to train physics aware AI models for forecasting. NVIDIA claims CorrDiff produces images at roughly 12.5 times higher resolution than current numerical weather models, around 1,000 times faster, and about 3,000 times more energy efficient than equivalent CPU runs.
Commercial applications cluster in a few areas. Precision agriculture uses ML over multispectral and SAR imagery to compute NDVI time series and crop yield forecasts. Insurance and reinsurance carriers buy automated flood, wildfire, and hail damage assessments. Defense and intelligence customers consume vessel tracking, aircraft activity reports, and infrastructure change alerts. Disaster response agencies use SAR derived flood masks because radar penetrates cloud cover, which is usually thick over flooded areas. After the February 2023 earthquake in Türkiye and Syria, providers including Planet and Maxar released annotated damage assessments within hours.
The Mars rovers remain the most visible demonstration of autonomy in deep space. Beyond AutoNav and AEGIS already described, the rovers also use onboard schedulers. Curiosity's mission has used a system called MSLICE for command sequencing and a ground side activity planner, while Perseverance's autonomy stack includes onboard ENav planning that builds a local cost map and chooses arcs through it. Perseverance has driven more than ten times the autonomous distance evaluated by Opportunity over its entire 14 year mission, in a fraction of the time.
ESA's Phi-Sat-1 set the template for AI on orbit in Earth observation. Its successor, Phi-Sat-2, launched on a Falcon 9 rideshare on August 16, 2024, and runs a larger suite of onboard apps for ship detection, cloud removal, marine anomaly detection, and on demand image compression. ESA's broader Phi lab continues to push AI onto commercial cubesats in cooperation with vendors like Ubotica, which provides the CogniSAT hardware platform.
OpenAI free domain projects aside, the major space agencies are also investing in onboard fault detection. NASA's Autonomous Sciencecraft Experiment, originally flown on the Earth Observing-1 satellite in 2003, applied onboard machine learning to detect volcanic eruptions, floods, and ice changes and to retask the satellite without ground intervention. That heritage informs current onboard autonomy on missions like Europa Clipper, which must survive Jupiter radiation and operate during long communication gaps.
AEGIS has now been deployed to Perseverance's SuperCam instrument, with the first autonomous SuperCam target acquired on May 18, 2022. The system has the same role as on Curiosity: pick rocks that look interesting to the laser spectrometer without waiting for the next uplink. Mission operators report that AEGIS routinely beats unguided pointing by a large margin, which would otherwise waste laser shots on soil or sky.
The Kepler space telescope ran from 2009 to 2018, the K2 extended mission ran from 2014 to 2018, and the Transiting Exoplanet Survey Satellite (TESS) has been operating since 2018. Each one produced light curves for hundreds of thousands of stars, and the false positive rate for transit candidates is high. Machine learning has become the standard tool for sorting through them.
The first widely cited deep learning paper in this area was Christopher Shallue and Andrew Vanderburg's AstroNet, published in The Astronomical Journal in 2018. AstroNet uses a one dimensional convolutional neural network trained on phase folded Kepler light curves. The model recovered Kepler-90i, the eighth planet around Kepler-90, and confirmed an additional planet in the Kepler-80 resonant chain. The paper is one of the most cited applications of CNNs to astronomy.
ExoMiner, led by Hamed Valizadegan at the Universities Space Research Association and NASA Ames, took the approach further. The 2021 paper validated 301 new planets from the Kepler archive in one shot using a deep neural network trained on physics motivated diagnostics that mirror what human vetters look at. The model achieved roughly 93.6 percent recall on a held out test set. ExoMiner has since been extended to TESS data, and by 2025 had been credited with the validation of more than 370 planets.
Deep learning also underpins the analysis pipeline for the JWST. The Morpheus framework, originally developed at UC Santa Cruz, performs pixelwise galaxy morphology classification and is used in the COSMOS-Webb program. Researchers report that AI driven pipelines compressed analysis windows that would have taken years into days.
The Vera C. Rubin Observatory in Chile reached first light through its full optics with the LSST Camera on April 15, 2025, and released its first publicly imaged data on June 23, 2025. The observatory's automated alert system pushes images from the mountain to the SLAC data facility within seven seconds, compares them to a reference template, and generates difference image alerts. The official LSST Science Pipelines, available as open source on GitHub, lean heavily on classical computer vision plus ML based classifiers. The Rubin LSST Dark Energy Science Collaboration's 2026 white paper on AI and ML opportunities lays out the case that no human team could touch the tens of billions of sources Rubin will monitor.
The Euclid space telescope, launched by ESA on a Falcon 9 on July 1, 2023, started routine science observations on February 14, 2024. Euclid is expected to image about 10 billion sources, of which about 1 billion will have gravitational shear measurements for weak lensing cosmology. ESA's first Quick Data Release in 2025 was paired with a Euclid Space Warps citizen science project, in which volunteers and ML algorithms together identified more than 500 strong gravitational lens candidates in the Q1 fields, with projections of more than 10,000 lens discoveries to come.
Machine learning has also driven a series of results in gravitational wave astronomy and radio transient detection. Daniel George and E. A. Huerta's 2017 work demonstrated that deep convolutional neural networks could detect LIGO binary black hole merger signals and estimate parameters orders of magnitude faster than matched filtering. The Gravity Spy project combines machine learning with citizen science to classify glitches in LIGO data. Liam Connor and Joeri van Leeuwen's 2018 paper in The Astronomical Journal applied a hierarchical deep CNN to classify fast radio burst (FRB) candidates from CHIME Pathfinder and Westerbork Apertif data, and it remains one of the foundation references for ML based FRB pipelines.
The COSMIC system at the Karl G. Jansky Very Large Array, run by the SETI Institute and Breakthrough Listen, commensally taps VLA data and feeds it through an ML based candidate filter. Breakthrough Listen researchers reported in Nature Astronomy in 2023 that a deep learning pipeline applied to previously studied Green Bank Telescope data uncovered eight new signals of interest while running roughly 100 times faster than the legacy turboSETI pipeline. None of those candidates has held up under follow up, but the technique has changed how SETI searches are run.
The Sun drives a long list of operational risks: geomagnetic storms that knock out HF radio and damage power grids, energetic particle events that endanger astronauts, and atmospheric drag changes that pull low Earth orbit satellites down faster. NOAA's Space Weather Prediction Center has issued daily M and X class flare probabilities since the late 1990s, and a 2025 verification study published in Space Weather, by Camporeale and colleagues, found that the SWPC operational forecasts perform comparably to, or in some cases worse than, simple statistical baselines and lightweight ML classifiers. That has helped motivate a wave of new ML work.
NASA's Frontier Development Lab, run by the SETI Institute in partnership with NASA, ESA, Google Cloud, NVIDIA, and Intel, has produced a long list of heliophysics tools. DAGGER (Deep leArninG Geomagnetic pErtuRbation) predicts ground level geomagnetic disturbance up to half an hour in advance using solar wind observations. The SHEATH model targets the sheath region ahead of coronal mass ejections. The Instrument to Instrument Translation (ITI) model from FDL Heliolab projects 2024 fuses data from multiple solar observatories to produce a 3D view of the Sun, and the MEGS-AI neural network forecasts extreme ultraviolet irradiance.
Other teams have shown that vision transformers can classify active region morphology in NOAA imagery and that ML methods could have predicted essentially all of the major space weather events around the May 2024 G5 superstorm with usable lead time, an arXiv preprint by Camporeale and collaborators argues. Solar flare prediction at the rare X class end remains hard because the training data is so unbalanced.
The US 18th Space Defense Squadron maintains the standard public catalog of tracked objects, currently around 47,000 objects larger than 10 cm in low Earth orbit, but the commercial sector has built its own networks. LeoLabs operates phased array radars in Texas, Alaska, New Zealand, Costa Rica, Portugal, and Australia. The company's processing pipeline trains XGBoost classifiers on historical pass data to identify object type and maneuver status within 72 hours of first observation, reportedly hitting about 95 percent agreement with the public catalog and with expert hand labels. LeoLabs has also published work using deep learning for radar signature based debris characterization.
Slingshot Aerospace operates around 200 optical sensors across 21 sites on five continents. In June 2024 the company and DARPA announced Agatha, a system that uses inverse reinforcement learning to flag anomalous satellites inside large constellations, a useful tool when an adversary tries to hide a maneuvering inspector inside a swarm of commercial broadband satellites. Slingshot also won a NOAA contract in late 2024 to provide the presentation layer for the Traffic Coordination System for Space (TraCSS).
Privateer, the space data company co-founded by Steve Wozniak, Alex Fielding, and astrodynamicist Moriba Jah in 2021, released the free Wayfinder visualization tool and the Crow's Nest collision risk product in 2022. In 2024 Privateer acquired Orbital Insight and raised $56.5 million, broadening from pure space situational awareness into Earth observation analytics.
In Europe, the EU Space Surveillance and Tracking partnership, set up under the EU Space Programme and the Galileo regulation in 2021, links a network of 15 member state radars and telescopes, with EUSPA acting as the EU SST Front Desk. EU SST issues collision avoidance warnings, re-entry analyses, and fragmentation analyses for European spacecraft including Galileo, Copernicus, and GOVSATCOM.
SpaceX's Starlink is the canonical example of an autonomous constellation. Each satellite ingests US Department of Defense conjunction data screens (provided by what is now called the 18th Space Defense Squadron and the new TraCSS feed) every 30 minutes, computes its own conjunction probabilities, and fires krypton ion thrusters when the joint probability rises above the company's threshold of about 1 in 1,000,000. That threshold is roughly two orders of magnitude stricter than the typical industry standard of 1 in 10,000. SpaceX has documented around 50,000 maneuvers in the six months between December 1, 2023 and May 31, 2024 in filings with the FCC, averaging about 14 maneuvers per satellite over that window. Starlink also auto deconflicts against crewed stations like the ISS and Tiangong, with the SpaceX satellite always being the one to move.
Eutelsat OneWeb, formed in September 2023 when Eutelsat completed its acquisition of OneWeb, runs a 633 satellite operational constellation in 12 orbital planes at 1,200 km. The combined operator is moving toward digital channelizer payloads from Ramon.Space and other suppliers so future OneWeb satellites can be software defined and reconfigurable in orbit. That sets up AI driven traffic shaping, beam hopping, and interference management as a core operational pattern.
Amazon's Project Kuiper, Telesat Lightspeed, and the Chinese Guowang and SpaceSail constellations are each pursuing similar autonomous flight and traffic management stacks. The combinatorial challenge of routing tens of thousands of inter satellite laser links in real time pushes most operators toward reinforcement learning and graph neural network approaches.
NASA publishes an annual AI Use Case Inventory under OMB direction, listing dozens of active machine learning and AI projects across science directorates and mission directorates. The 2024 and 2025 inventories include rover autonomy on Mars 2020, ExoMiner exoplanet validation, on board fault management, wildfire detection from satellite imagery, anomaly detection in deep space network telemetry, and an internal AI assistant for engineering documentation. The agency operates under the Responsible AI principles articulated in Executive Order 13960.
NASA's Frontier Development Lab, already discussed, runs an eight to nine week summer research sprint each year, pairing PhD students and postdocs with industry mentors on space science problems. Past projects include lunar resource mapping, asteroid orbit determination from sparse observations, atmospheric escape modeling for exoplanets, and astronaut health monitoring.
The Jet Propulsion Laboratory runs a long standing autonomy group whose alumni built Remote Agent, the Earth Observing-1 onboard science, the CASPER ground side planner, AEGIS, and AutoNav. JPL's MEXEC executive software has been demonstrated in flight on M-Cubed and other small spacecraft.
ESA's Phi lab is the closest European equivalent of FDL. It runs flagship programmes like Phi-Sat that move ML inference onto Earth observation cubesats, the InCubed accelerator for commercial EO innovators, and partnerships with industry that have led to onboard hardware like Ubotica's CogniSAT and OBC stacks for AI workloads. ESA also runs the OPS-SAT in orbit lab, which from 2019 to 2024 hosted hundreds of AI and software experiments uploaded by external researchers.
The US Space Force formed in December 2019 and now manages most operational military space activity. The Space Development Agency, founded in 2019 and absorbed into Space Force in October 2022, is building the Proliferated Warfighter Space Architecture, a multi tranche constellation that aims for at least 1,000 satellites in LEO by the late 2020s for missile tracking and tactical communications. SDA's Tranche 4, expected to begin launches in 2030, is targeted as the first tranche with substantial autonomous operations including on board sensor tasking and downlink prioritization.
Space Force has run a series of pilot programs to apply machine learning to space domain awareness backlogs at the National Space Defense Center and to automate routine console operations. The Air Force Research Laboratory's POET (Prototype On orbit Experimental Testbed) demonstrated AI based edge processing on orbit in February 2022 in a series of containerized software uploads. The Defense Advanced Research Projects Agency's Hallmark and Sea Eagle programs have funded ML based threat detection prototypes. RAND has published a two volume study, Artificial Intelligence and Machine Learning for Space Domain Awareness, examining how AI changes mission effectiveness for both the US and adversaries.
Commercial vendors are deeply embedded in this work. Slingshot, Privateer, LeoLabs, BlackSky, Maxar Intelligence, and Capella Space all hold significant defense contracts. The 2025 AFWERX selection of Slingshot to track nefarious in space activities is one of many examples.
AI in space has limits that mission planners take seriously. Spacecraft hardware lags well behind terrestrial processors because radiation hardened chips run on older process nodes and have far less memory. Ingenuity flew a Snapdragon 801, a 2014 mobile chip, and even that was unprecedented; most planetary spacecraft fly variants of the BAE RAD750 or RAD5500 PowerPC processors that have orders of magnitude less throughput than a modern GPU. That gap forces aggressive model quantization, distillation, and feature engineering, and limits which architectures can run on board.
Validation and verification are also harder than in earthbound AI. NASA and ESA require flight software to be certifiable, which is awkward for large neural networks whose decision boundaries are opaque. The trend has been to use ML for advisory roles, with deterministic logic in the safety critical loop, although autonomous collision avoidance on Starlink is a notable case where the AI has the trigger.
Dual use concerns have grown. The same vessel detection model that helps Greenpeace find illegal fishing can be used to target ships. The same change detection pipeline that supports humanitarian damage assessment in Ukraine and Gaza supports kinetic targeting elsewhere. Commercial Earth observation operators have run into export control disputes and pressure to limit imagery over conflict zones, including the brief Maxar restriction on Ukrainian access in March 2025 that was later reversed.
False positive and bias issues recur. Solar flare prediction models are calibrated against rare events and tend to be poorly calibrated at the X class tail. SETI ML pipelines sometimes flag interference patterns that look almost like the artificial signals they are trained to detect, with the BLC1 candidate eventually traced to local radio interference at the Parkes telescope. Exoplanet vetting models inherit the biases of their training labels, and several confirmations have been quietly retracted as follow up observations failed.
Finally, the carbon and water cost of training ever larger Earth observation foundation models is becoming a part of the conversation. NVIDIA, IBM, and Hugging Face have all released satellite imagery foundation models in 2024 and 2025, and the same questions about compute footprint that apply to text large language models apply here as well.