A satellite just learned to find things on its own — here’s what that means

In a groundbreaking achievement, an Earth observation satellite has successfully located specific targets autonomously, marking a significant advancement in space technology. This milestone, achieved in April, represents the first documented application of a vision-language model (VLM) in orbit, showcasing the potential of artificial intelligence to enhance the capabilities of space-based sensors and their overall value. Traditionally, satellites transmit extensive data back to Earth, where analysts employ machine learning algorithms or visual inspection to interpret the information. However, onboard the Yam-9 satellite, developed by Loft Orbital, a specialized software from NASA's Jet Propulsion Laboratory was able to identify areas of interest based on natural language queries. The VLM, named Gemma 3 and created by Google DeepMind, is designed for edge computing, allowing it to operate effectively on limited hardware located far from centralized data centers. VLMs integrate the contextual understanding of large language models with image analysis capabilities. Researchers utilized the model to classify sensor data, such as distinguishing between natural environments and human infrastructures, or identifying structures near railway stations. This demonstration is pivotal for two main reasons: it enhances the utility of space sensors by performing initial data triage in orbit, thereby alleviating the overwhelming amount of raw data that analysts must sift through, and it serves as a proof of concept for deploying larger-scale AI systems in space. Paul Lasserre, Loft's head of AI, emphasized the transformative potential of this technology, stating, "It opens the door to always-on, patrol layers in space." With a VLM, satellites can be programmed for specific tasks, such as monitoring borders and alerting operators about suspicious activities. Loft's spacecraft function as platforms for third-party clients, operating under a business model akin to infrastructure-as-a-service rather than conventional satellite production. Recently, the company secured a contract to build, launch, and manage six new satellites for EarthDaily, which will utilize the satellites to analyze and commercialize the data gathered. Launched in the fall of 2025, the Yam-9 serves as a prototype for Loft's orbital AI initiatives and is equipped with a powerful Nvidia Jetson Orin AGX GPU, one of the top chips for space computing. Juan Delfa Victoria, a technical leader at NASA JPL, spearheaded the development of NAVI-Orbital, the software platform that enabled the Gemma 3 VLM's functionality. While Gemma 3 is commercially available, the engineering team adapted the software to optimize it for space by minimizing its libraries and memory requirements. Although this is the inaugural use of a VLM in orbit, other companies are expected to follow suit. Planet Labs, which operates satellites equipped with Jetson Orin processors, currently employs these for basic object detection but is exploring additional AI applications, including VLMs. Kepler Communications, known for hosting the largest fleet of GPUs in space, has not confirmed the deployment of VLMs due to confidentiality agreements but has mentioned several undisclosed use cases for its computing resources since their launch in January. "Now that we’ve proven the concept, that’s really the direction of travel," Lasserre noted. The ambition is to expand the satellite constellation to ensure real-time monitoring across the globe, which he estimates would require between 50 and 100 Yam-9 satellites. Currently, Loft operates 12 satellites in orbit. Insights gained from deploying these smaller AI models will guide future efforts to implement extensive computing infrastructures in space, especially in crucial areas such as power and memory management. This progress could also lead to the development of innovative scientific tools. The concept for NAVI-Space, for instance, originated from JPL researcher Taran Cyriac John, who envisioned digital assistants for astronauts on lunar or Martian explorations. Delfa Victoria remarked, "We’re considering how to provide an interactive assistant, similar to those seen in video games and films, to aid astronauts who are unable to use keyboards while in pressurized suits." One thing is for sure: this technology is far from the fictional HAL 9000.