AstroForge bets on AI to fly its next spacecraft without Earth in the loop

AstroForge is preparing to put artificial intelligence in charge of a spacecraft for the first time, a major test of whether a startup can rely on onboard autonomy instead of a large ground team to keep a mission alive. The company says its 2027 spacecraft will use an in-house transformer-based system called Solo, and it could become one of the clearest demonstrations yet of AI operating in deep space.

The move matters because spaceflight is unforgiving: once a probe is far from Earth, communication delays, narrow radio windows and limited staffing can turn a minor fault into a mission-ending failure. AstroForge’s answer is to move more decision-making into the vehicle itself, where software can diagnose and correct problems without waiting for instructions from the ground.

Why AstroForge is betting on onboard AI

AstroForge is trying to solve a problem that has long shaped spacecraft design: humans and Earth-based control rooms are reliable, but they are expensive, slow and limited by distance. The company, which is building technology to mine asteroids, argues that the next generation of missions will need more independence than traditional systems can provide.

Large agencies like NASA still lean on layers of redundancy and extensive teams of flight controllers, even on missions with significant automation. The contrast is stark. On the OSIRIS-REx asteroid mission, which rendezvoused with the asteroid Bennu in 2018, the operation reportedly required 100 operators on each eight-hour shift. That model works for a public space agency with deep budgets. It is much harder for a venture-backed startup trying to reduce cost while increasing mission complexity.

AstroForge’s founders say the company arrived at AI after experience showed how fragile deep-space operations can be when a spacecraft cannot easily talk back to Earth. With a limited number of large antennas capable of reaching a vehicle hundreds of thousands of miles away, every communications window becomes precious. If a spacecraft goes quiet at the wrong time, even a technically recoverable issue can become impossible to fix.

Matthew Gialich, AstroForge’s co-founder and chief executive, said the company began asking whether a spacecraft could diagnose itself after launch if all the data were already onboard. He framed the choice as one between building an expensive global communications network or pushing more intelligence into the vehicle itself.

How Solo works

Solo is AstroForge’s answer to spacecraft autonomy, and the company describes it as an in-house control stack built around transformer technology. Rather than replacing every traditional control system, the software is designed to sit above a conventional foundation of flight algorithms and subsystem-specific models.

That layered approach matters. In high-risk aerospace systems, engineers are generally cautious about handing critical decisions to neural networks because those models can be difficult to predict and verify. AstroForge is not claiming to have built a universal robot pilot. Instead, the company says it is developing a constrained system that ingests a relatively small set of sensor inputs and uses that information to make tightly defined decisions.

According to the company, Solo combines three elements:

• traditional control algorithms that handle core flight behavior;
• models trained on test data for particular subsystems such as navigation and power; and
• an overarching intelligence layer trained on data from roughly 2,500 spacecraft sensors.

The goal is not humanlike reasoning in the broad sense. It is practical recovery. If the spacecraft loses track of its attitude, power generation or another vital function, the onboard system should be able to identify the failure, connect it to the right subsystem and attempt a fix.

What kind of problems could AI solve in flight?

AstroForge says the most valuable use case is anomaly resolution, especially when a vehicle is too far away to receive immediate help from mission control. One example the company has discussed is a scenario in which the spacecraft loses its position in space and the software traces the issue to a star tracker or related power problem, then restores the system by restarting the affected hardware.

That sounds simple, but in the context of deep-space flight it could be decisive. Many missions have been lost not because they lacked technical sophistication, but because they could not respond quickly enough when something unexpected happened after launch.

Why this mission is a bigger test than it sounds

The move to AI comes after AstroForge has already experienced the pain of limited autonomy. The company, founded in 2022 and backed by $56 million in venture capital, has launched two prototype spacecraft, both of which ran into anomalies that kept them from completing most of their intended work.

In 2025, AstroForge sent its Odin spacecraft into deep space, but the company struggled to maintain communication with it. The problem was not simply technical elegance; it was operational reality. Deep-space communications depend on a small pool of powerful antennas on Earth, and those dishes can only point at a spacecraft during short time windows. If the link is lost or a craft behaves unpredictably, the recovery options are limited.

That experience helped convince AstroForge that future missions need to carry more intelligence onboard. If the spacecraft itself can notice a failure and react before ground controllers intervene, the chances of salvaging the mission rise considerably.

What makes AI spacecraft control different from traditional flight software?

AI-based control differs from standard flight software because it is designed to infer context, not just follow prewritten rules. Traditional systems can be excellent at planned maneuvers and known contingencies, but they are less flexible when a spacecraft encounters a combination of failures that engineers did not predict in advance.

That flexibility is exactly why AI is attractive and why engineers remain cautious. Neural networks can be powerful pattern recognizers, but the aerospace industry has long preferred deterministic systems that are easier to test and certify. AstroForge is trying to bridge that gap by keeping its AI within narrow boundaries and surrounding it with conventional controls.

Armand Awad, AstroForge’s head of flight software, said the company’s approach was influenced by progress in frontier transformer models. In practical terms, the company is borrowing techniques that have made modern AI systems better at pattern recognition and multi-step decision-making, then constraining those techniques for spacecraft use.

Awad said the company is not trying to create general-purpose autonomy for every spacecraft or a universal autonomous system for all machines. Instead, he described a limited model designed to work with a narrow sensor feed and a tightly controlled mission profile.

That restraint is important. In spaceflight, autonomy is only useful if it is trustworthy. A model that makes clever but unstable decisions would be worse than a less capable system that behaves predictably every time.

When will Solo fly?

AstroForge expects its first meaningful test of Solo to arrive before the fully autonomous mission. The company’s third vehicle, DeepSpace-2, is scheduled to launch alongside Intuitive Machines’ third lunar mission, with liftoff expected by the end of 2026. On that spacecraft, Solo will operate in what the company calls shadow mode, meaning it will observe and process conditions without taking control of the vehicle.

That step is designed to let engineers compare Solo’s judgments against the actual spacecraft behavior before trusting it with a mission of its own. Shadow mode is often how high-risk autonomy systems are validated: the software runs in parallel, records what it would have done, and gives operators a chance to evaluate its performance without allowing it to steer the mission.

The more consequential test is slated for 2027. AstroForge plans to fly what it calls the first autonomous spacecraft on the first rocket launched by Stock Space. That mission is expected to receive support from NASA and is intended to collect scientific data about the Sun.

If it succeeds, the flight would mark a significant milestone not just for AstroForge but for the broader space industry, because it would demonstrate that a small company can field spacecraft with much less dependence on a ground-based operations army.

Why the company may still keep a radio link

AstroForge says it is not yet certain whether its fully autonomous mission will completely sever contact with Earth. Gialich has suggested that the team is debating whether to include radios capable of receiving commands from the ground. For now, however, he says the company is leaning toward removing that safety net.

That would be a bold choice. A radio link offers a fallback if the onboard system behaves unexpectedly. But it also blunts the central experiment: whether the AI can truly handle a mission without human intervention. The company appears to be weighing the value of redundancy against the value of proving that autonomy can stand on its own.

Gialich has said that the team is still arguing over the design, but that the current direction is to avoid building a system that can simply be rescued from Earth if problems arise.

The tension is familiar in aerospace. Every safety layer adds resilience, but every extra layer can also increase cost, complexity and mission mass. AstroForge’s choice will say as much about the company’s appetite for risk as it does about the maturity of current AI.

How AstroForge’s strategy reflects a broader shift in space tech

AstroForge’s effort is part of a larger movement toward more autonomous spacecraft, especially as the economics of launch and deep-space exploration change. Private companies are increasingly expected to do more with less, and that means designing systems that can survive failures without a large standing operations team.

There is also a technological backdrop. Transformer models have transformed software development in many other fields, and some space engineers are now asking whether the same underlying advances can support navigation, diagnostics and mission planning. The appeal is obvious: if AI can recognize patterns in language, images and complex data streams, perhaps it can also recognize faults in a spacecraft.

But space is not a chat app or a productivity tool. A bad answer in orbit can destroy a mission, waste years of work and strand hardware far from Earth. That is why even supporters of autonomy often talk in cautious terms. The question is not whether AI can be useful. It is whether it can be trusted enough to act when there is no human there to take over.

Why startups are pushing autonomy faster than big agencies

Startups have a stronger incentive to automate because they cannot afford the staffing model that agencies have historically used. NASA can mobilize many specialists for a complex mission. A startup with limited venture funding must think differently. Automation is not just a performance upgrade; it is a cost strategy and, in some cases, a survival strategy.

For AstroForge, that logic is especially relevant because its long-term mission involves asteroid mining, a field where distances are large, assets are expensive and decisions need to be made far from Earth. If autonomy can work for science missions, it could eventually help commercial operations that depend on repeatable, low-touch spacecraft behavior.

What happens next?

The next 12 to 18 months will likely determine whether AstroForge’s approach is a breakthrough or an instructive experiment. DeepSpace-2 should provide the first opportunity to see whether Solo can match human expectations without being allowed to control the mission directly. If that test goes well, the company will have a stronger case for Autonomy-1.

Even if the fully autonomous mission runs with a backchannel to Earth, the test would still be important. Spaceflight autonomy tends to advance incrementally. Each successful layer reduces the need for human intervention and makes future systems more ambitious.

For now, AstroForge is trying to answer a question that sits at the frontier of both AI and aerospace: can a spacecraft think enough like an operator to save itself when there is no operator available? The company is betting that the answer, at least in a narrow but meaningful sense, is yes.

Key dates and milestones

• 2022: AstroForge is founded and begins developing asteroid-mining technology.
• 2025: Odin launches into deep space, but communications problems prevent full mission success.
• End of 2026: DeepSpace-2 is expected to launch with Solo running in shadow mode.
• 2027: AstroForge aims to fly its first autonomous spacecraft, potentially without Earth radio command capability.

For the space industry, the significance goes beyond one startup. If AstroForge can show that AI can detect, diagnose and recover from spacecraft faults on its own, the result could influence how future probes, landers and commercial missions are built. If it fails, it will still offer a valuable lesson about the limits of machine autonomy where distance, delay and danger all collide.