# In a First For Science, A Satellite Has Identified What It's Seeing From Space > Source: > Published: 2026-07-01 11:17:08+00:00 The standard approach to [satellite imagery](https://www.sciencealert.com/striking-satellite-images-reveal-antarctica-is-10-times-greener-than-35-years-ago) is to snap huge batches of pictures and beam them back to Earth, where they can be sifted through by human operators and the best available algorithms. It's all worked well so far, but the time, transmission bandwidth, and energy required are starting to become bottlenecks. Modern satellites are simply [capturing more pixels](https://www.sciencealert.com/ai-combed-millions-of-images-of-the-arctic-it-found-an-alarming-pattern) than scientists have time to look at. However, the YAM-9 satellite has just done something different: It has identified and described features in its image scans without needing to check back with ground control. Not only that, but it can be instructed with natural prompts that you might use with [Google Gemini](https://www.sciencealert.com/ai-chatbots-are-bad-at-diagnosing-symptoms-for-a-surprising-reason-study-finds) or Siri, such as "find me all the railway hubs in this country". The advance comes via a NASA-built program called NAVI-Orbital, developed by researchers at the NASA Jet Propulsion Laboratory (JPL) and tech startup [Loft Orbital](https://www.forbes.com/sites/the-prototype/2026/03/07/startup-loft-orbital-is-launching-ai-powered-satellites-this-fall/). "Tasking a satellite to recognize a new feature has historically required writing command sequences, revalidating onboard software, and uplinking new binaries," [write](https://doi.org/10.48550/arXiv.2606.18271) the researchers in their arXiv preprint, which is yet to be peer-reviewed. "Under the NAVI-Orbital paradigm, re-targeting amounts to editing and uploading a new prompt. This shortens the re-tasking cycle and broadens the set of potential task authors beyond those with specialized command-sequence expertise." Most of the time, when you interact with an AI chatbot like Claude or ChatGPT, your queries are sent to [energy-intensive](https://www.sciencealert.com/ai-is-heading-for-an-energy-crisis-that-has-tech-giants-scrambling) [data centers](https://www.sciencealert.com/engineers-found-a-genius-way-to-slash-data-center-energy-use), where they are processed and a response is sent back. By housing AI models on a device instead, the processing is much faster without that back-and-forth, and if you're a satellite, that means less need for ground communications. "Usually, a user has to task the satellite using an API, wait for image collection and downlink, and then analyze the image using a pre-trained algorithm on the ground," Loft Orbital's senior marketing manager Sarah Preston told ScienceAlert. "This AI can actually 'see' what's in the image and identify what the analyst is looking for, such as bridges, highways, specific bodies of water, or signs of natural disasters like flooding and wildfires." With this satellite, the locally installed AI is Google DeepMind Gemma 3, a series of 'lightweight' models that are small enough to [run on laptops](https://huggingface.co/google/gemma-3-27b-it). It's a vision-language model (VLM), which means it processes both text and images. Crucially, it can run on a small satellite, where physical size, energy use, and computing power all have to be carefully managed. "This pipeline is orchestrated by a multi-agent architecture composed of three self-contained agents that hand off work to each other: an orchestrator that coordinates execution, a detector that analyzes, classifies, and summarizes images, and a dialog agent that enables operators to ask questions about the results," [write](https://doi.org/10.48550/arXiv.2606.18271) the researchers. In other words, technicians can ask the satellite's software a question, rather than programming it for each individual job. "This design makes NAVI adaptable to different missions without rebuilding from scratch." In on-ground, baseline tests, the system was able to broadly recognize what was in some 7,960 images with an accuracy of 88.2 percent, classifying them into categories such as residential areas, beach, agricultural zones, and mountains. Just two in-orbit live captures have been performed so far, with more planned. In future, this kind of technology could have applications far beyond [low Earth orbit](https://www.sciencealert.com/greenhouse-emissions-threaten-the-future-of-low-earth-orbit-scientists-warn). A simple and fast prompt-and-analysis approach like the one demonstrated here could be used with rovers exploring the surface of [the Moon](https://www.sciencealert.com/moon) [or Mars](https://www.sciencealert.com/perseverance-finds-complex-organic-compounds-in-strange-mars-rocks). "We're thinking, okay, you have astronauts with pressurized suits, and you know they cannot be tapping on a keyboard, whatever they want to do is complex," senior system engineer Juan Delfa Victoria told Tim Fernholz at [TechCrunch](https://techcrunch.com/2026/06/15/a-satellite-just-learned-to-find-things-on-its-own-heres-what-that-means/). "So, how about we provide an assistant, like in video games and in movies, where you see an AI which is interactive?" With around 100 satellites like YAM-9, real-time coverage could be set up across our entire planet, the researchers say. Loft Orbital certainly [has its sights set](https://www.forbes.com/sites/the-prototype/2026/03/07/startup-loft-orbital-is-launching-ai-powered-satellites-this-fall/) on providing such a service. "The vision is to have satellites working together for continuous and global real-time monitoring, aided by a marketplace of AI agents," Preston said. To a large extent, [satellite imagery](https://www.sciencealert.com/europe-is-now-experiencing-the-same-drop-in-pollution-as-china-thanks-to-the-lockdown) wouldn't need to be sent back to Earth to figure out what is pictured if this kind of technology were deployed. That could be useful for everything from tracking wildfire smoke to monitoring unusual activity at ports or borders in real-time. Of course, that's also potentially concerning in terms of the level of surveillance it would entail. "The company is working to scale those capabilities, supporting missions that need fast, on-the-spot decision-making, whether that's civil, commercial, or defense use cases," Preston told ScienceAlert. "The goal is for satellites to operate like a continuous lookout – you tell it to monitor a coastline for oil spills or flag new construction near a border. The satellite would evaluate what it sees and only report back when something meets that criteria." **Related: NASA Satellite Reveals Just How Fast Mexico City Is Sinking** There's still room for improvement in terms of accuracy and reliability, and huge questions around the ethics of outsourcing high-stakes image interpretation to AI. The researchers haven't examined what might happen should any "adversarial prompts" be entered, so their findings should "therefore be read as a feasibility result rather than as a robustness characterization", they [conclude](https://doi.org/10.48550/arXiv.2606.18271). Despite those hurdles, the researchers are confident that this breakthrough technology will soon become the norm. "It opens the door to always-on, patrol layers in space," Loft Orbital Head of AI Paul Lasserre told [TechCrunch](https://techcrunch.com/2026/06/15/a-satellite-just-learned-to-find-things-on-its-own-heres-what-that-means/). A report on the research is available on the preprint server * arXiv*. This article was fact-checked by [Jess Cockerill](https://www.sciencealert.com/michaelirving) and edited by [Clare Watson](https://www.sciencealert.com/clare-watson). While we pride ourselves on our process, we are only human. If you spot a mistake, [please let us know](https://www.sciencealert.com/contact-us).