Traveling to an asteroid and landing on it sounds straightforward, but the reality is quite complex.
When NASA dispatches spacecraft for solar system exploration, they incorporate extensive redundancy and automation. However, much relies on flight control teams on Earth communicating with the spacecraft. The OSIRIS-REx mission, which reached an asteroid in 2018, employed around 100 operators during each eight-hour shift.
In contrast, startups may not have access to such resources. AstroForge, a company focused on asteroid mining technology, is turning to artificial intelligence to fill the gap. They have created an autonomous control system for their spacecraft known as “Solo,” which utilizes a transformer-based model developed in-house.
AstroForge aims to launch its first autonomous spacecraft in 2027 aboard the inaugural rocket from Stoke Space. This mission, supported by NASA, is anticipated to gather valuable scientific information about the sun.
Achieving this would be significant since most autonomous spacecraft utilize traditional control algorithms due to concerns surrounding the reliability of neural networks. Just last year marked the beginning of neural networks being employed for satellite orbit control.
Founded in 2022 and with $56 million in venture capital backing, AstroForge has launched two prototype spacecraft, but both faced technical problems that hindered their operational success. In 2025, its Odin spacecraft ventured into deep space, but the company encountered communication challenges. There are limited ground antennas capable of reaching spacecraft positioned hundreds of thousands of miles away, and the available time slots for contact are brief.
Ultimately, AstroForge lost control of Odin. However, this experience inspired the company to explore whether sufficient intelligence could be integrated into the spacecraft to permit it to self-resolve issues.
“If I had data on the spacecraft, could it have been salvaged? I can’t say for sure, but I wish we had tried something onboard,” stated Matthew Gialich, co-founder and CEO of AstroForge.
Gialich posed a crucial decision: creating an extensive ground network costing approximately $200 million to facilitate global operations, or enhancing autonomy through a new model.
Armand Awad, head of flight software at AstroForge, explained that the business opted to utilize advancements in transformer models coming from leading research labs. They developed a stack combining traditional control algorithms with models trained on specific subsystems such as navigation and power generation, along with an overarching intelligence layer trained on data from around 2,500 sensors in the spacecraft.
“I’m not claiming to create a fully autonomous spacecraft or general autonomy,” Gialich remarked. “What I’m aiming for is limited autonomy within a constrained environment, utilizing minimal sensor inputs and following transformer training basics.”
Theoretically, the controlling agent could carry out functions such as anomaly resolution. Awad envisages the system recognizing a loss of spatial orientation, linking a power problem to issues with its star tracker, and adjusting everything—likely by cycling power on and off.
AstroForge’s third vehicle, DeepSpace-2, is slated for launch alongside Intuitive Machines’ upcoming lunar mission, expected to launch by late 2026. Solo will be aboard this vehicle, operating in “shadow mode” to allow engineers to evaluate its performance before the major Autonomy-1 mission.
Will this be a genuinely independent AI-driven spacecraft?
“I’m currently not planning to include receiver radios for Earth communication in Autoonomy-1,” Gialich stated. “We have to fully commit at this point. The team might change my mind before the launch, but as of now, no radios.”
