With the continuous advancement of technology, the application of artificial intelligence (AI) is becoming increasingly widespread in various fields, and the field of space exploration has also welcomed its revolutionary application—generative artificial intelligence (GenAI).
We have gathered some practical applications of generative AI/large models in the space sector from abroad to see what directions the industry is taking.
Four Major Advantages of GenAI in Space
-
Natural Language Processing (NLP) Astronauts can access and analyze data generated from space operations and experiments through an intuitive and user-friendly language interface.
-
Reduction of Data Transmission/Latency Issues Preliminary processing and analysis are completed on the spacecraft, only requiring the transmission of results back to Earth, significantly reducing the amount of data transmitted and latency.
-
Real-Time Decision Making Astronauts can access findings from experiments and other analyses more quickly, and Earth stakeholders can also access satellite sensor data analysis and other findings faster, which is crucial in emergencies or time-sensitive situations.
-
Adaptive Models LLMs can be updated and uploaded based on changing conditions and work requirements to reflect the latest space environment.
First LLM Application on the International Space Station
Booz Allen Hamilton deployed an LLM application on the International Space Station (ISS). Since mid-July 2024, this application has been part of an experiment aimed at testing the feasibility of using GenAI applications in a resource-constrained computing environment on a spacecraft. This is also the first actual deployment of an LLM in space.
Booz Allen collaborated with Hewlett Packard Enterprise (HPE) to develop this application in eight weeks. It was successfully uploaded to the ISS as part of a proof of concept (POC) experiment, aimed at assisting astronauts in conducting laboratory experiments using GenAI without relying on Earth’s internet.
Main features include providing remote data acquisition and retrieval enhanced generation (RAG), enabling astronauts to effectively access information and accurately interpret and resolve complex issues using natural language queries.
Φsat-2: ESA’s AI Small Satellite
The European Space Agency (ESA)’s Φsat-2 is a CubeSat designed to demonstrate how AI technology can optimize Earth observation from space.
Φsat-2 will carry six AI applications, which are
Multispectral Imaging Tools and Systems
Image De-clouding
Detection and Classification of Maritime Vessels
Real-Time Map Generation Based on Satellite Imagery
Discovery of Anomalies in Marine Ecosystems and Detection of Wildfires
Onboard Image Compression and Reconstruction Tools
Notably, the *Sat2Map* application uses CycleGAN to convert satellite images into electronic maps, achieving near-real-time map data generation and delivery. CycleGAN is a type of GenAI that falls under the category of unsupervised image-to-image translation models. By learning the mapping between two domains, CycleGAN facilitates creative transformations between different image styles or domains.
Image-to-image translation involves converting an input image from one domain (e.g., a photo) into an output image in another domain (e.g., a painting, map). This real-time mapping capability is particularly important for emergency services (e.g., quickly identifying passable roads during disasters such as earthquakes or floods).
ESA plans to first test the Sat2Map application in Southeast Asia to demonstrate its potential in crisis management. The Φsat-2 project is the result of collaboration between ESA and a consortium led by Open Cosmos, including Ubotica, CGI, CEiiA, GEO-K, KP Labs, and SIMERA.
Some Thoughts
GenAI in the space sector requires advanced AI edge computing capabilities. Although still in the experimental stage, the applications from Booz Allen Hamilton and ESA seek to demonstrate the possibility of directly integrating GenAI capabilities into onboard/spacecraft systems, promoting natural language-based data retrieval and analysis, improving data transmission efficiency, and enabling real-time decision-making.
There has been no public disclosure regarding the application of large models in space domestically; however, discussing large models inevitably brings up computing power. First, we need to solve the edge-side computing power issue before demonstrating the application of large models on the edge side.
In fact, there have already been several cases of onboard/airborne computing domestically.
-
According to public reports, as early as 2021, domestic satellite manufacturers had achieved edge computing, completing onboard de-clouding:
“Tiansuan Constellation” plan, jointly initiated by the Shenzhen Graduate School of Beijing University of Posts and Telecommunications and Tianyi Research Institute, aims to serve major national strategic needs and explore the forefront of international technology. By designing satellites to be intelligent, service-oriented, and open, it builds an integrated platform for space computing on-orbit experiments, integrating edge computing with satellite computing platforms, allowing satellites to simultaneously utilize edge and central cloud resources in space and possess AI capabilities and multi-task processing capabilities. Testing data shows that through collaborative reasoning between satellites and ground stations, the computation accuracy has improved by over 50%, and the data returned by satellites can be reduced by 90%, marking the entry of satellites into the cloud-native era.
-
Currently, there are also leading companies in China specializing in the research and development of onboard edge computing technology, such as Beijing Xingce Future, which has participated in multiple satellite R&D projects with a good level of product maturity.
It can be seen that by providing computing power support for large models on the edge side, there are already independent solutions available. The next step is to conduct pilot studies and validations. The tasks to be undertaken in the future roughly include the following:
Demonstrating Large Model Application Scenarios: In combination with the AI Agent concept, select some important value scenarios in satellite network security, collision protection, real-time image interpretation, and alerting Hardware Adaptation: Preferably select suitable large models, possibly involving multi-modal large models, with hardware manufacturers leading the adaptation of large models on the edge Model Training: For the target large model, conduct specific fine-tuning and training according to scenario tasks, such as completing a “satellite intelligence analysis large model” Agent Construction: Based on scenario requirements, construct a complete onboard AI agent for processing various tasks in a streamlined manner, such as a “satellite threat handling large model”
Of course, in the aerospace field, which has always emphasized stable operation, large model technology still seems not so rigorous, such as the hallucination problem; it should undergo sufficient technological research and the improvement of various safety assurance measures before being allowed to operate in space.
