AstroForge, the startup building technology to mine asteroids, will let a transformer model it built in-house, called Solo, make real-time flight decisions aboard its next spacecraft. That is a job NASA has typically assigned to rooms full of human flight controllers working in shifts. The company’s co-founder and CEO, Matthew Gialich, and its head of flight software, Armand Awad, described the system to TechCrunch.
The stakes are concrete. AstroForge has already flown two prototype spacecraft, and both suffered anomalies that kept them from completing most of their objectives. Its Odin probe launched into deep space in 2025, but the company struggled to stay in contact with it: Earth has only a limited number of antennas large enough to reach a craft hundreds of thousands of miles out, and the transmission windows are short. AstroForge never regained full control of Odin.
That failure is what pushed the company toward putting more intelligence onboard rather than relying on Earth-based operators. “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,” Gialich told TechCrunch.
Gialich framed the choice as a straightforward cost tradeoff against building AstroForge’s own communications infrastructure. “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?” he said.
Awad said the flight software team built a layered stack rather than a single model, combining traditional control algorithms with narrower models trained for specific subsystems such as power generation and navigation. Sitting above those, a broader model was trained on roughly 2,500 onboard sensors to tie the whole system together. Gialich was careful to scope the ambition down. “I’m not saying I’m going to make general spacecraft autonomy or general autonomy for the world,” he said. “I’m making a constrained autonomy at a very low sensor input, following the basic training of a transformer model.”
Solo will not fly unsupervised right away. It is set to ride along on AstroForge’s third vehicle, DeepSpace-2. That spacecraft is scheduled to lift off by the end of 2026, sharing its launch with Intuitive Machines’ third trip to the moon. On that flight Solo runs in “shadow mode,” observing and making recommendations without controlling the spacecraft, so engineers can evaluate it before the company’s Autonomy-1 mission. Autonomy-1 itself is planned for 2027, launching on the first rocket flown by Stoke Space in a NASA-backed mission to gather data on the sun. Gialich said he currently intends to fly that mission without a way to intervene from the ground at all: “I don’t plan on flying radios that can receive from Earth on Autonomy-1. We have to go all in right now.”
The approach is unusual for the industry: most spacecraft autonomy still runs on traditional control algorithms rather than neural networks, and a neural network was used to manage a satellite’s orbital positioning for the first time only last year. AstroForge’s bet is that a company running on just $56 million in venture backing cannot afford the ground infrastructure NASA-scale missions depend on, making it cheaper to train a model that makes the call onboard.
For any space or robotics startup watching Autonomy-1’s shadow-mode data next year, the real signal will be whether Solo’s recommendations match what a human controller would have chosen with the same sensor feed, not whether the mission reaches its target.
TechCrunch reported the AstroForge autonomy plans on September 22, 2026.
