NASA has managed to make a satellite “understand” what it sees, without sending a single photo to Earth

The role of satellites is simple. They take data, send it to Earth and, once here, it is analyzed. So far, that is being viable. However, the process is getting complicated. There are more and more satellites (Hello, Elon Musk), which means more information flowing to Earth. There are not enough people, energy or time to analyze so much information. For this reason, NASA and the startup LoftOrbital have developed NAVI-Orbital, software for satellite observation that is capable of describing what it sees and even searching for what it is ordered to do.

A model of vision and language. NAVI-Orbital works with Gemma 3a series of vision and language models belonging to Google DeepMind. As their name suggests, they are capable of processing both images and text, so that, with proper training, they can describe images and respond to simple text commands. Before, a multitude of complex computer commands had to be used for a satellite to follow orders and even then the results had to be analyzed by humans. Now, all you have to do is ask him to obey and analyze the results.

Find what you ask for. A clear example of what NAVI-Orbital can do is search for objects, buildings or geographical features. If you know what a river is, you can search and display all the rivers in a given area. You can also locate all bridges or roads. But, in turn, it can identify anomalies in those places. For example, you may find rivers about to overflow or roads where there has been an accident.

much faster. Most artificial intelligence algorithms send information to data centerswhere the result is processed and returned. In this case, however, what is known as edge AI is used. That is to say, the entire model runs directly on the same device that captures the data. The satellite in this case.

Civil or military applications. In April, the first tests were carried out with the YAM-9 satellite, which already has the new software incorporated, with very good results. It was the first time that a satellite managed to describe what it “sees.” That’s why NASA and LoftOrbital are already excited about their applications. They consider that satellite constellations with NAVI-Orbital could be used to monitor the planet for civil or defense applications. They can search for oil spills, for example, but also study possible military attacks.

The difference with what we have so far with any of these purposes is that nothing has to be analyzed. That is, you don’t put a satellite to take photos of the sea and then here on Earth a team of people or a computer algorithm analyzes them for oil spills. It is the satellite itself that saves sending thousands of photos without spills and directly analyzes what it sees. The result is much faster, cheaper and more efficient. It may even be the future of satellite observation, although it is still too early. We will have to see how these models that have already begun to undergo the first tests evolve.

Image | LoftOrbital

In Xataka | Ukraine’s military has a problem almost as important as Russia: Starlink belongs to Elon Musk