Fly to an asteroid, land on it, and make no mistakes: it is not that simple.
NASA builds redundancy and automation into its spacecraft, yet much of the work relies on teams of flight controllers communicating from Earth. The OSIRIS-REx mission, which met an asteroid in 2018, required 100 operators per eight-hour shift.
Startups do not have such resources. AstroForge, a company developing technology to mine asteroids, is turning to artificial intelligence. It has built an autonomous control stack for its spacecraft called Solo. This is a transformer-based model developed in-house.
AstroForge plans to fly its first autonomous spacecraft in 2027 on the first rocket launched by Stoke Space. That mission will be backed by NASA and is expected to gather scientific data about the sun.
That would be an accomplishment. Most spacecraft autonomy depends on traditional control algorithms because of concerns about the unreliability of neural networks. The first use of a neural network to control a satellite’s positioning in orbit took place just last year.
AstroForge was founded in 2022 and has raised $56 million in venture funding. It has launched two prototype spacecraft, both of which suffered anomalies that prevented them from achieving most of their mission objectives. In 2025, its Odin spacecraft was launched into deep space, but the company had difficulty communicating with it. There are a limited number of antennas on Earth big enough to transmit to spacecraft hundreds of thousands of miles away, and the windows of time in which to do it are small.
Ultimately, AstroForge could not gain control of Odin. The experience spurred it to consider alternatives. Could it put sufficient intelligence onboard the spacecraft for it to solve its own problems?
“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,” Matthew Gialich, AstroForge’s co-founder and CEO, 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?”
Armand Awad, AstroForge’s head of flight software, said the company decided to use the advances in transformer models driven by frontier labs. It created a stack that includes traditional control algorithms, models trained on test data for specific subsystems like power generation or navigation, and an overall intelligence layer trained on about 2,500 sensors in the spacecraft.
“I’m not saying I’m going to make general spacecraft autonomy or general autonomy for the world,” Gialich said. “I’m making a constrained autonomy at a very low sensor input, following the basic training of a transformer model.”
In theory, the agent in control of the spacecraft will perform tasks like anomaly resolution. Awad imagines it realising that it has lost track of its position in space, correlating a power anomaly to issues in its star tracker, and fixing the whole thing. That might mean turning the system on and off.
The company’s third vehicle, DeepSpace-2, is currently set to launch alongside Intuitive Machines’ third moon mission, expected to head for space by the end of 2026. Solo will be onboard that vehicle, flying in shadow mode, so AstroForge’s engineers can put it through its paces before the Autonomy-1 mission.
Will this truly be an AI agent operating an independent spacecraft?
“I don’t plan on flying radios that can receive from Earth on Autonomy-1,” Gialich said. “We have to go all in right now. The team’s probably going to talk me into it by the time we fly it. But right now, I’m telling them no radios.”
What it means
For the engineers at AstroForge, this shift changes the daily work of monitoring a ship. Instead of waiting for a signal from Earth to diagnose a failure, the onboard model must act immediately. The team loses the ability to intervene remotely during critical moments, forcing them to rely entirely on the software’s ability to interpret sensor data and fix issues without human input.
