Satellite AI involves different computing tasks: learning from observations, applying a trained model, and supporting mission decisions. An October 26, 2026 survey by Zheng Lin and 11 coauthors examines how these tasks can be divided among individual spacecraft, neighboring satellites, and ground facilities.
The preprint organizes its analysis around limited memory, power, computing capacity, and intermittent connections. It reviews distributed training, collaborative inference, and large language models. It is a research synthesis, rather than a demonstration of an operational constellation.
Applying Models Near the Sensor
Inference means applying an already trained model to new information. For an imaging satellite, that can mean identifying clouds or detecting ships before deciding what to transmit.
The European Space Agency’s Φsat-2 science-phase announcement provides a concrete example. In July 2025, ESA reported that the spacecraft had completed commissioning and begun delivering science data. Its onboard applications process and compress Earth observation images, select usable observations, and support detection tasks involving wildfires, ships, and marine pollution.
ESA also changed the data products made available to researchers. Alongside processed outputs, it provided earlier-stage image products to support application development and training. Processing in orbit and improving models on the ground can serve different parts of the same mission.
New Space Economy’s discussion of space-originated AI workloads makes the corresponding service distinction: processing near an orbital sensor can reduce the amount of information that must be returned to Earth.
Learning Without Collecting Every Observation Centrally
Training changes a model using examples. Federated learning provides one way to do this when observations remain distributed across devices.
In the original federated-learning research, H. Brendan McMahan and colleagues described devices computing local model updates and a coordinating system combining those updates into a shared model. The training examples remain on the devices rather than being collected in one central repository. This exchanges model updates instead of the complete underlying datasets.
The satellite survey examines how such exchanges must accommodate orbital contact windows. A neighboring spacecraft can relay an update when its source cannot reach a ground station. The coordinating system must also account for delayed updates, which may describe an older model state.
Connecting Language Models to Flight Software
A large language model requires more than enough storage for its parameters. The survey identifies additional memory demands during execution and the need to coordinate computing with changing spacecraft resources.
Its proposed operational approach connects reasoning to structured software tools, command verification, and feedback. Mission information needs timestamps and version tracking so that delayed instructions do not overwrite newer decisions. Validation must reproduce orbital connectivity and hardware constraints before these methods can be assessed as part of a spacecraft control system.
