AstroForge plans an AI-controlled spacecraft with no radio

The economics of autonomy

Deep-space missions normally carry redundancy and automation, but they still depend heavily on flight-control teams on Earth. During the OSIRIS-REx mission’s approach to an asteroid in 2018, 100 operators worked on each eight-hour shift.

That infrastructure is difficult for a startup to build. AstroForge was founded in 2022, has raised $56 million in venture funding and has already launched two spacecraft prototypes. Both encountered anomalies that prevented them from completing most of their mission tasks.

The more consequential failure came in 2025, when Odin traveled into deep space and AstroForge ran into communications problems. Only a limited number of antennas on Earth are large enough to transmit data to spacecraft hundreds of thousands of miles away, and the available communication windows are short. AstroForge ultimately could not restore control of Odin.

The company considered two alternatives:

AstroForge chose the second path. Co-founder and CEO Matthew Gialich said he does not know whether Odin could have been recovered if all the necessary data had been onboard. During the failure, however, none of the spacecraft’s systems tried to solve the problem on their own. He wants future spacecraft to have that option if they lose control immediately after launch.

What Solo actually contains

Solo is not a general-purpose pilot for spacecraft. Armand Awad, AstroForge’s head of spacecraft software, said the company built its stack by applying recent advances in transformer models to a narrow set of spacecraft functions.

The system combines:

AstroForge says it is not trying to create universal spacecraft autonomy, or intelligence capable of handling every possible situation. Solo will work with a limited number of sensor inputs and use basic transformer-model training principles.

In theory, the agent could trace an anomaly through several related systems. If it lost information about its position, for example, it might connect that failure to a power problem and a possible star-sensor issue, then repair the chain by powering the relevant system off and on again.

That is a much narrower proposition than replacing a flight-control team. It is closer to giving a spacecraft a bounded troubleshooting routine for situations in which the ground team may not be reachable.

The test before the commitment

AstroForge plans to launch its third spacecraft, DeepSpace-2, with Intuitive Machines’ third lunar mission by the end of 2026. Solo will run on DeepSpace-2 in “shadow mode”: active on the spacecraft, but initially observed by AstroForge engineers while they assess its capabilities before the Autonomy-1 mission.

The announcement is quiet about how that assessment will establish that Solo is safe to trust beyond the scenarios it has already seen. The company has described the model’s inputs and its intended troubleshooting logic, but not the boundary between an anomaly it can repair and one it might misread.

My guess is that this boundary matters more than the transformer architecture. AstroForge is not choosing AI because a model is inherently better at spacecraft control; it is choosing a constrained model because a global antenna network would cost about $200 million. The business case is clear, but it also makes every unhandled failure look like a cost the company decided not to pay on the ground.

A mission without radio commands

Gialich currently does not plan to install radio systems on Autonomy-1 capable of receiving commands from Earth. He says the company needs to rely fully on autonomy now, although the team will probably try to persuade him to add radio communications before launch. For now, his decision is to fly without them.

That turns Autonomy-1 into more than a demonstration of onboard software. It will test whether a startup can substitute a limited AI control layer for the communications infrastructure that established missions use to recover from failure. If Solo works, AstroForge will have shown that autonomy can be an economic necessity as much as a technical ambition; if it fails, the missing ground network will be difficult to ignore.