The Spacecraft That Will Ignore Its Own AI

The thing they built is called Solo. AstroForge describes it as a spacecraft intelligence model, software designed to coordinate the vehicle's onboard functions autonomously. TechCrunch reports that it is transformer based and built in-house, which is to say it comes from the same broad family of model as the chatbots, pointed at telemetry instead of text.

Here is the part that tends to fall out when a story like this gets retold. Solo is not the flight software. The company is explicit that it does not replace the physics based software and deterministic algorithms underneath it. It sits on top of them.

AstroForge's flight software lead, Armand Awad, describes three layers. At the bottom, conventional control algorithms: the well-understood, testable, do-exactly-this-when-that-happens code that has been flying for decades. Above that, smaller models trained on test data for specific subsystems, things like power generation and navigation. At the top, an intelligence layer trained on data from about 2,500 sensors across the spacecraft, whose job is to look at all of it at once.

Looking at all of it at once is the point. A power fault and a star tracker behaving oddly present as two separate problems to two separate subsystems. A layer that sees both at the same moment can notice they are one problem, and that turning the star tracker off and on again might be the entire fix.

So Solo is not being asked to fly the spacecraft in the stick-and-rudder sense. It is being asked to be the duty engineer. Read the room, work out what is off-nominal, decide what happens next, and hand the actual doing to code that has been doing it for years.

Recovery is a loop. You act, the world changes because you acted, you look at the changed world, you act again. A shadow system only ever gets the first half of the first step. Its decision is never executed, so the state it reads a minute later belongs to a world its own choice never touched. Every subsequent call it makes is a call about somebody else's situation.

Which means a clean shadow run earns you something real, and narrow: on live data, in conditions nobody could fake, the model's reading of the situation and its opening move both looked right to the people reviewing them. That is diagnosis, and diagnosis genuinely does test this way. What it does not earn you is the claim everyone will want to make, which is that the model can fly the vehicle out of a hole. Climbing out of a hole takes several moves in a row, each one made in a world the previous move just changed. The hole never opened. The model never had to climb.

There is a second, quieter problem. Off-nominal events are rare, which is the whole reason anyone wants autonomy in the first place. A week of shadow telemetry from a healthy spacecraft contains almost nothing to be right or wrong about. You can pass a shadow run by having an uneventful week, and an uneventful week is not evidence of anything.

None of this makes shadow mode a bad idea. It is the honest thing to do and it is enormously better than the alternative, which is finding out on the day. It simply has a ceiling, and the ceiling is this: you are grading the answers, not the outcome.

The firsts are the company's own, and the same announcement is fair enough to name the prior art. In May 1999 NASA handed Deep Space 1 to an experiment called Remote Agent: twenty-nine hours on the Monday, another six on the Friday. OSIRIS-REx touched an asteroid and collected a sample in 2020 with nobody on Earth in the loop, because at roughly eighteen minutes each way there was no possible way for anybody to be. What is new here is narrower than the first autonomous spacecraft. It is the no-uplink part, and the whole-mission part.

Remote Agent is worth stopping on, because of what happened in that six hour session on the Friday. The team simulated an electronics unit failing. The software noticed, and fixed it by switching the unit off and on. Then they simulated a sensor failing, and it correctly worked out that the sensor was the problem rather than the thing the sensor was watching. Then they stuck a thruster in the off position, and it switched to a different set of thrusters. That is the closed loop. That is the exact thing shadow mode cannot reach, and somebody demonstrated it, in flight, twenty-seven years ago.

The dates are soft. AstroForge's own announcement gives no launch date for either mission. Press reporting puts DeepSpace-2 alongside Intuitive Machines' third lunar mission, targeted for the end of this year, and Autonomy-1 in 2027. Space schedules move, and both of these depend on somebody else's rocket.

Solo has never flown. The two vehicles that came before it both hit anomalies.

And the industry has not settled this question. Most spacecraft autonomy still runs on traditional control algorithms, because people are not yet comfortable with how a neural network behaves the first time it meets something it has not seen. The first time a neural network controlled a satellite's own positioning in orbit was only last year.