Operating a spacecraft is usually a labor-intensive job. NASA’s OSIRIS-REx mission, which rendezvoused with an asteroid in 2018, had 100 operators on each eight-hour shift. Startups don’t have those resources, which is why asteroid-mining company AstroForge is turning to AI. It has built an in-house autonomous control stack called Solo, a transformer-based model, and plans to fly its first autonomous spacecraft in 2027 on the first rocket from Stoke Space, a NASA-backed mission expected to gather solar science data.

Most spacecraft autonomy today relies on traditional control algorithms because neural networks are seen as unreliable. The first use of a neural network to control a satellite’s positioning in orbit happened only last year. AstroForge, founded in 2022 with $56 million in venture funding, has already launched two prototype spacecraft, both of which suffered anomalies that prevented them from achieving most of their mission objectives. Its 2025 Odin spacecraft reached deep space but communication failed: there are few antennas on Earth large enough to reach spacecraft hundreds of thousands of miles away, and the usable time windows are small. AstroForge never gained control of Odin. That experience drove the company to consider putting enough intelligence onboard for a spacecraft to solve its own problems. CEO Matthew Gialich said, “Would that have been recoverable with all the data on the spacecraft? I don’t know, but I can tell you nothing onboard tried it, and I would love something onboard to try if the spacecraft is unrecoverable at launch.” He framed the decision as building a ground network that would cost around $200 million for five dishes worldwide versus trying to remove that need with a model.

Head of flight software Armand Awad said AstroForge decided to leverage transformer advances from frontier labs. The stack combines traditional control algorithms, models trained on test data for specific subsystems such as power generation or navigation, and an overall intelligence layer trained on about 2,500 spacecraft sensors. Gialich stressed the scope is deliberately narrow: “I’m not saying I’m going to make general spacecraft autonomy or general autonomy for the world. I’m making a constrained autonomy at a very low sensor input, following the basic training of a transformer model.” In theory, the agent would handle anomaly resolution: Awad imagines it realizing it has lost track of its position, correlating a power anomaly to an issue in its star tracker, and fixing the problem, “probably turning it on and off in this case.”

The company’s third vehicle, DeepSpace-2, is set to launch with Intuitive Machines‘ third moon mission, expected to head to space by the end of 2026. Solo will be onboard in “shadow mode” so engineers can test it before the Autonomy-1 mission. Gialich is committed to making that mission fully independent: “I don’t plan on flying radios that can receive from Earth on Autonomy-1. We have to go all in right now. The team’s probably going to talk me into it by the time we fly it. But right now, I’m telling them no radios.”