Fly to the asteroid, land and don’t fail. I wish it were that easy.
When NASA sends spacecraft to explore the solar system, redundancy and automation are built in. But much of the work will be done by a team of flight controllers who will communicate with the vehicle from Earth. The OSIRIS-REx mission, which rendezvoused with the asteroid in 2018, had 100 operators on each eight-hour shift.
But such resources are not available to startups, so AstroForge, which is developing technology to mine asteroids, is turning to AI. The company has developed an autonomous control stack for the spacecraft called “Solo,” an in-house transformer-based model.
AstroForge plans to fly its first autonomous spacecraft on the first rocket launched by Stoke Space in 2027. The mission is supported by NASA and is expected to collect scientific data about the Sun.
That would be an accomplishment. Due to concerns about the unreliability of neural networks, most spacecraft autonomy relies on traditional control algorithms. Only last year, neural networks were first used to control the position of satellites in orbit.
Founded in 2022 and backed by $56 million in venture funding, AstroForge has launched two prototype spacecraft, both of which encountered anomalies and missed most of their mission goals. In 2025, the company’s Odin spacecraft was launched into deep space, but communication was difficult. There are only a limited number of antennas on Earth large enough to transmit to spacecraft hundreds of thousands of miles away, and the time window in which they can transmit is narrow.
In the end, AstroForge was unable to control Odin, but the experience led him to consider alternatives. Can we put enough intelligence onboard a spacecraft to solve its own problems?
AstroForge co-founder and CEO Matthew Gialich said, “With all the data on the spacecraft, could it have been recovered? I don’t know, but I can tell you that no one has tried anything on board. If the spacecraft was unrecoverable at launch, we’d like to try something on board.”
“The trade for me is whether we build our own terrestrial network, which costs (about) $200 million to install five dishes around the world and operate on them. Or do we try to use a model to remove the network?”
Armand Awad, AstroForge’s director of flight software, said the company decided to take advantage of advances in transformer models being advanced by Frontier Laboratories. We created a stack that includes traditional control algorithms, models trained on test data for specific subsystems such as power generation and navigation, and an overall intelligence layer trained on approximately 2,500 sensors within the spacecraft.
“I’m not saying we’re going to achieve general spacecraft autonomy or global autonomy,” Gearich said. “Following the basic training of the transformer model, we achieve constrained autonomy with very low sensor inputs.”
In theory, an agent controlling the spacecraft would perform tasks such as anomaly resolution. Realizing that it has lost its position in space, Awad associates the power glitch with a problem with the star tracker and fixes the whole thing — “in this case, I imagine, probably by toggling it on and off.”
The company’s third spacecraft, DeepSpace-2, is currently scheduled to launch at the same time as Intuitive Machines’ third moon mission, which is expected to head into space by the end of 2026. Solo will be aboard the spacecraft and fly in “shadow mode,” allowing AstroForge engineers to put it through its paces before the Autonomy-1 mission.
Is this really an AI agent piloting an independent spaceship?
“We’re not going to fly a radio on Autonomy 1 that can receive signals from Earth,” Gearich said. “Right now I have to go all in. The team will probably convince me about it by the time we fly. But for now, I’m telling them not to use the radio.”
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
