AstroForge is handing control of its next spacecraft to a transformer model. The asteroid-mining startup has built an in-house autonomous control stack called Solo, trained on roughly 2,500 onboard sensors, and plans to fly it on its 2027 Autonomy-1 mission — the first rocket launched by Stoke Space, with NASA backing to gather scientific data about the sun. Founded in 2022 and backed with $56 million in venture funding, AstroForge is betting that a neural network can do the job that NASA typically staffs with a hundred flight controllers per shift.
The scale gap is the whole point. NASA's OSIRIS-REx mission, which rendezvoused with an asteroid in 2018, ran with 100 operators on each eight-hour shift. AstroForge has no such bench, and building the ground infrastructure to talk to a deep-space vehicle is prohibitively expensive for a venture-funded startup. CEO Matthew Gialich frames the decision as a straight cost trade between hardware and software.
The pivot to onboard intelligence came out of failure. AstroForge has launched two prototype spacecraft, both of which suffered anomalies that killed most of their mission objectives. In 2025, the company's Odin spacecraft made it to deep space but AstroForge could not maintain communications, and there are only a handful of antennas on Earth big enough to reach a vehicle hundreds of thousands of miles away — and only narrow windows in which to use them. Odin was lost.
Key facts
- 01AstroForge built Solo, an in-house transformer model trained on 2,500 onboard sensors, to run its spacecraft without ground operators.
- 02The company's Autonomy-1 mission is scheduled to fly in 2027 on the first launch by Stoke Space, backed by NASA to gather solar data.
- 03AstroForge has raised $56 million in venture funding since its 2022 founding and has lost both prototype spacecraft to anomalies.
- 04CEO Matthew Gialich pegs the alternative — building a private ground network of five dishes — at roughly $200 million.
- 05DeepSpace-2, launching alongside Intuitive Machines' third moon mission by end of 2026, will fly Solo in shadow mode ahead of Autonomy-1.
Gialich says the Odin failure exposed the core problem: nothing on the spacecraft was trying to save itself. "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 said.
“The trade for me is: Do I go build my own ground network, which is going to cost [around] $200 million to put up five dishes around the world and then do operations on it, or do I try to remove it with a model?”— Matthew Gialich, AstroForge co-founder and CEO
Head of flight software Armand Awad said the team decided to piggyback on transformer advances from the frontier AI labs rather than reinvent them. Solo combines traditional control algorithms, smaller models trained on test data for specific subsystems like power generation and navigation, and an overall intelligence layer trained across the full 2,500-sensor telemetry stream. The pitch is not general intelligence — it is a narrow agent for a spacecraft that already knows what it is supposed to do.
Awad is explicit about the scope. The agent's job, in practice, is anomaly resolution: recognizing that the vehicle has lost its position fix, correlating a power irregularity to a star-tracker fault, and cycling the affected component — "probably turning it on and off in this case," Awad said. That framing matters because most spacecraft autonomy today still leans on deterministic control code; the first use of a neural network to control a satellite's positioning in orbit happened only last year.
The path to Autonomy-1 runs through a rehearsal. AstroForge's third vehicle, DeepSpace-2, is set to fly alongside Intuitive Machines' third moon mission by the end of 2026 and will carry Solo in shadow mode — the model runs, its decisions are logged, but a human loop stays in control. That gives AstroForge's engineers roughly a year of in-flight data before Solo is trusted with a live mission.
The Autonomy-1 configuration Gialich is currently pushing for is uncompromising. He wants the 2027 vehicle to fly without a receive radio at all, cutting off any possibility of an Earth-side override. It is the kind of statement that gets softened between now and launch, and Gialich admits as much, but it captures the strategic bet: if a startup cannot afford $200 million in dishes, the model has to be the mission.
The obvious counterweight is that transformer models are famously bad at the failure modes that matter in space — rare, out-of-distribution events for which there is no training data. Traditional aerospace control code is verifiable; a neural network's decision boundary is not. Both AstroForge prototypes launched to date have failed for reasons unrelated to autonomy, and the company will need Solo to prove itself in shadow mode on DeepSpace-2 before anyone regulates or insures a fully autonomous deep-space vehicle. The startup has no track record of a successful mission end-to-end.
Still, AstroForge is doing something the big primes have been unwilling to try: putting a modern learned model in the control loop of a real spacecraft, in deep space, with real capital behind it. If Solo works even in shadow mode on DeepSpace-2, it is a template other cash-constrained space startups will copy — because the alternative, building your own Deep Space Network, is not a business plan. If it fails, the industry gets a very expensive data point on where transformer autonomy still cannot go. Either outcome moves the field.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
