AstroForge, a startup with ambitions to mine asteroids and a record of spacecraft that have declined to cooperate, has arrived at a logical conclusion: if the spacecraft cannot be reached from Earth, perhaps it should simply handle things itself. The company's autonomous control stack, named Solo, is a transformer-based model built in-house to run a spacecraft without a ground team telling it what to do.
This is either a bold engineering breakthrough or the plot of several cautionary films. The distinction may not matter once it's in deep space.
Nothing onboard tried it — and the CEO would love something onboard to try next time.
What happened
When AstroForge's Odin spacecraft launched into deep space in 2025, the company discovered that Earth has a limited number of antennas large enough to reach a spacecraft hundreds of thousands of miles away. The communication windows are small. AstroForge missed them. Odin was not recovered.
The CEO, Matthew Gialich, described the arithmetic plainly: build a global ground network for approximately $200 million, or train a model to make the spacecraft responsible for its own survival. The company chose the model. This is the rational decision. It is also the kind of decision that gets described as rational right up until it isn't.
The resulting system, Solo, layers traditional control algorithms beneath subsystem-specific models and a top-level intelligence trained on roughly 2,500 onboard sensors. It will fly its first mission in 2027 aboard a Stoke Space rocket, with NASA backing, to gather solar data.
Why the humans care
Most spacecraft autonomy still relies on traditional control algorithms, because the AI research community spent years establishing that neural networks have a complicated relationship with reliability. The first neural network trusted to manage a satellite's orbital positioning happened just last year. AstroForge is proposing to skip several rungs of that ladder and hand Solo an asteroid mission.
The practical argument is sound. Human flight controller teams are expensive, antenna time is scarce, and deep space does not wait for scheduling conflicts. NASA ran 100 operators per shift on its OSIRIS-REx asteroid mission. AstroForge has $56 million in total venture funding. The math was always going to land here.
What happens next
Solo launches in 2027. If the mission succeeds, the template for replacing hundred-person operations teams with a transformer model trained on sensor data will exist, documented, replicable, and priced within reach of any well-funded startup with an asteroid it would like to visit.
The spacecraft will be completely alone, hundreds of thousands of miles from the nearest human opinion, making decisions in real time. The humans describe this as the goal.