Asteroid-mining startup AstroForge is building a transformer-based control system intended to let a spacecraft diagnose and fix problems without waiting for Earth. The system, called Solo, is expected to fly in test mode by the end of 2026 before taking operational control of a mission planned for 2027.
The company is pursuing onboard intelligence after two prototype spacecraft suffered anomalies that prevented most mission objectives. One of them, Odin, traveled into deep space in 2025, but AstroForge could not maintain communication or regain control.
Communication Failure Drove the Autonomy Plan
Deep-space communication depends on a limited number of sufficiently large antennas and narrow contact windows. Large public missions can support extensive ground teams: NASA’s OSIRIS-REx asteroid mission used 100 operators during each eight-hour shift after its 2018 rendezvous.
AstroForge does not have those resources. Cofounder and chief executive Matthew Gialich estimated that building five ground dishes around the world and operating them would cost about $200 million. His alternative is to place enough decision-making capacity aboard the vehicle that it can attempt a recovery independently.
The company was unable to determine whether Odin could have saved itself using its onboard data. Gialich’s conclusion was that the spacecraft contained nothing able to try. That experience became the case for an agent that can act when communication has failed and the normal option is losing the mission.
AstroForge was founded in 2022 and has raised $56 million. Both launched prototypes encountered problems, making the autonomy program a response to demonstrated operational limits rather than a purely theoretical experiment.
Solo Combines Models With Conventional Controls
Solo is not presented as unrestricted general intelligence. The stack includes traditional control algorithms, models trained for specific subsystems such as navigation and power generation, and a higher-level intelligence layer trained on input from roughly 2,500 spacecraft sensors.
Head of flight software Armand Awad described a possible recovery sequence in which the system notices that the spacecraft has lost its position, connects a power anomaly to the star tracker and restarts the affected equipment. The example illustrates anomaly resolution; TechCrunch does not report that Solo has performed that sequence in space.
Gialich calls the goal constrained autonomy with relatively low sensor input. That boundary matters because neural networks have generally been viewed as too unreliable for spacecraft control. TechCrunch reports that the first neural-network control of a satellite’s orbital positioning occurred only last year.
The architecture therefore keeps established algorithms while adding learned models where AstroForge believes they can interpret complex telemetry and coordinate a response.
Shadow Mode Comes Before the Radio-Light Mission
AstroForge plans to place Solo aboard its third vehicle, DeepSpace-2, alongside Intuitive Machines’ third lunar mission, which is expected to launch by the end of 2026. Solo will operate in shadow mode, allowing engineers to compare its proposed actions with the spacecraft’s actual controls without giving it command.
The operational test is planned for Autonomy-1 in 2027, scheduled to fly on Stoke Space’s first rocket. NASA is backing that mission, which is expected to collect scientific information about the sun.
Gialich said he currently does not intend to include radios capable of receiving instructions from Earth, although he acknowledged the team may persuade him before launch. The no-radio plan is therefore an expressed intention, not a finalized spacecraft specification.
Solo’s value remains unproven until it survives flight testing. Shadow mode can reveal whether its decisions match real conditions, while the later mission would test whether constrained AI can replace a costly ground connection without creating a new failure point. AstroForge is betting that, in deep space, an imperfect attempt at recovery may be more useful than a spacecraft waiting for instructions that never arrive.
