Asteroid mining startup AstroForge is betting its next deep space mission on an AI that can fly the spacecraft by itself. According to TechCrunch AI, the company built an in-house transformer-based control stack called “Solo.” It plans to fly its first fully autonomous spacecraft, Autonomy-1, in 2027 on the first rocket launched by Stoke Space. NASA is backing the mission, which will gather scientific data about the sun.
The riskiest part is how little contact Earth will have with it. CEO Matthew Gialich says he doesn’t plan to put radios on Autonomy-1 that can receive commands from Earth. If something goes wrong, the spacecraft has to fix it alone.
🚀 Why a Startup Is Doing This
Most deep space missions depend on large human teams. NASA’s OSIRIS-REx mission, which reached an asteroid in 2018, had 100 operators on each eight-hour shift. A startup can’t pay for that.
AstroForge learned this the hard way. Founded in 2022 with $56 million in venture funding, it has launched two prototype spacecraft, and anomalies kept both from reaching most of their goals. In 2025 it lost contact with its Odin spacecraft in deep space. Only a few antennas on Earth are big enough to reach spacecraft hundreds of thousands of miles away, and the time windows for using them are short.
Gialich describes the choice plainly: “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?”
He also said that when Odin went dark, “nothing onboard tried” to recover it. That’s the gap Solo is meant to close.
🧠 How Solo Works
Solo doesn’t hand every decision to one big neural network. AstroForge’s head of flight software, Armand Awad, describes three layers:
- Traditional control algorithms: the proven, predictable base that most spacecraft already use.
- Subsystem models: trained on test data for specific jobs like power generation and navigation.
- An overall intelligence layer: trained on data from about 2,500 sensors across the spacecraft.
The main job is fixing problems. Awad gives an example: the spacecraft notices it has lost track of its position, links a power anomaly to a problem with its star tracker, and fixes it, “probably turning it on and off in this case.”
Gialich is careful about the scope. “I’m making a constrained autonomy at a very low sensor input, following the basic training of a transformer model,” he said. He isn’t trying to build general autonomy. It’s one narrow agent for one vehicle.
📡 Why This Is a Big Deal
The space industry has avoided neural networks for good reasons. They can behave in ways that are hard to predict, and you can’t send a technician when something breaks 500,000 miles away. That’s why spacecraft autonomy still runs mostly on traditional control algorithms. The first time a neural network controlled a satellite’s position in orbit was just last year.
AstroForge wants to jump from that first step to a spacecraft whose recovery depends on AI. It’s a big leap, and the company’s track record of two troubled missions makes it bolder still.
AstroForge will test first. Solo will fly in “shadow mode” on the company’s third vehicle, DeepSpace-2, which is set to launch alongside Intuitive Machines’ third moon mission by the end of 2026. In shadow mode, the AI watches and makes decisions without actually controlling the spacecraft, so engineers can compare its choices with what really happened.
🔭 What to Watch Over the Next 1-3 Years
Late 2026: the shadow mode results. DeepSpace-2 will be the first real test of whether Solo’s decisions hold up in space. If the data looks good, expect other small space companies to take agentic control more seriously.
2027: Autonomy-1 flies. If a spacecraft with no uplink can detect and fix its own faults, the economics of deep space change. Startups won’t need huge ground teams or their own antenna networks to run missions.
Beyond 2027: the pattern spreads. Look past space, too. “Constrained autonomy” is where agentic AI is heading everywhere: narrow agents with a limited job, built on top of deterministic controls instead of replacing them. If you’re building agents for industrial systems, robotics, or infrastructure, AstroForge’s layered design is worth studying.
One question is still open. Gialich admits his team “is probably going to talk me into” adding radios before launch. Whether Autonomy-1 flies fully cut off or keeps a backup link will show how much trust the industry is ready to give AI when there’s no human to step in. The full story is available at TechCrunch AI.