AstroForge puts an AI in command of its next spacecraft
AstroForge spent two missions learning how hard it is to talk to a spacecraft. Its answer is to need the conversation less.
AstroForge has built an in-house, transformer-based autonomy stack called Solo, and it plans to fly its first fully autonomous spacecraft in 2027 on the maiden launch of a Stoke Space rocket — a NASA-backed mission expected to gather scientific data about the Sun. The company is framing Solo as the fix for the constraint that killed its last mission: in 2025 it launched the Odin spacecraft into deep space and then could not gain control of it. There are only a handful of antennas on Earth big enough to reach a vehicle hundreds of thousands of miles out, and the windows to use them are short. AstroForge's co-founder and chief executive, Matthew Gialich, put the choice bluntly — spend around $200 million putting five dishes around the world and staffing operations, or try to remove the ground network with a model.
What Solo actually is matters more than the ambition. It does not replace the physics. The stack sits as an intelligence layer above AstroForge's existing deterministic flight software, combining conventional control algorithms, subsystem models trained on test data for power generation and navigation, and an overall layer trained on roughly 2,500 spacecraft sensors. Armand Awad, the company's head of flight software, describes the target behaviour as noticing it has lost track of its position in space, correlating a power anomaly to a problem in its star tracker, and correcting it — likely powering the unit off and back on. Gialich is careful about the scope: "constrained autonomy at a very low sensor input," not general autonomy for the world.
The economics are the argument. Ground operations — time on a dish, telemetry analysis, commanding — account for nearly a third of AstroForge's mission cost, according to the company's own breakdown, and that bill scales badly as more spacecraft fly to deep space. An onboard agent also sees far more than could ever be downlinked: NASA's Osiris-Rex ran about 100 operators per eight-hour shift, and a startup with $56 million in venture funding since 2022 cannot staff anywhere near that.
Our read: autonomy is becoming the answer to a comms problem, not a capability showcase, and that reframes what "AI in command" means — Solo earns its place by making one dish optional, not by being cleverer than a flight controller. The caution is that the same company's two prototype spacecraft, Odin and its predecessor, each suffered anomalies that prevented most of their objectives. Autonomy is exactly the fix you cannot A/B test on a vehicle you cannot reach. AstroForge is hedging accordingly: its third vehicle, DeepSpace-2, heads to space alongside Intuitive Machines' third moon mission near the end of 2026, with Solo aboard in "shadow mode" so engineers can watch it make decisions the real software is still making.
What to watch: whether Solo's shadow-mode telemetry gets published in a form outsiders can check, and whether the 2027 autonomous flight rides Stoke's first rocket as planned.
Would you trust an onboard model to power-cycle a subsystem when no one on Earth is listening? Tell us in the comments.
