NASA Put Claude to Work on Mars — Autonomous Spacecraft Are Next

Last December, NASA's Jet Propulsion Laboratory used Anthropic's Claude models to help plan two Mars drives for the Perseverance rover — the first time a large language model contributed to route planning for a mission on another planet. Human planners checked and adjusted the routes before upload, but the milestone marks a turning point in how space agencies think about machine intelligence.

The Perseverance drives were not an isolated experiment. In May, NASA and IBM placed a compressed AI model on the International Space Station and on a satellite to identify things like floods and clouds from orbit — the first model of its kind demonstrated in space. In July, astronauts on the ISS tested a large language model to see whether it could help answer questions about maintenance procedures.

These experiments point to a broader shift in space engineering. For decades, engineers on Earth decided what a machine in space would do, and the machine executed those instructions exactly. Now researchers are testing whether non-deterministic systems — including generative AI — can give spacecraft the flexibility to interpret their surroundings, plan tasks, and one day make decisions on their own. The technology remains far from trustworthy enough to hand over control of a spacecraft. But as missions become more complex, more distant, and more numerous, engineers are beginning to ask a different question: whether they can afford not to.

Why engineers resisted autonomy — and why that is changing

Spacecraft have operated autonomously for decades. Autonomy, however, has never been the dominant design philosophy, and the reasons run to the core of how space engineering works: engineers prize systems whose behavior they can predict.

"Autonomy often does not have a deterministic outcome, which means that how you got into a certain situation changes the behavior," says Robert Ambrose, former chief of NASA's Software, Robotics and Simulation Division. "So if you come into the same situation but from different paths, the outcome could be different. And so engineers hate that."

Ambrose spent much of his career working on autonomous systems at NASA, including autonomy for the Orion spacecraft — NASA's deep space and lunar orbiter — and Robonaut 2, a humanoid robot designed to work alongside astronauts that flew to space in 2011. With Orion, he saw firsthand how quickly the testing burden multiplies. Engineers had to evaluate not only what the spacecraft might do, but all the different paths that could have led it to a given decision.

That combinatorial problem did not kill the program. Ambrose says engineers found ways to manage the complexity, including automating the testing itself. "We fought the challenges of autonomy using autonomy," he says. "That actually works."

The case for autonomy strengthens the farther a mission travels from Earth. Ambrose points to a possible mission to Europa, Jupiter's ice-crusted moon, where a spacecraft might need to dive through a water plume erupting from beneath the surface. Such a plume could appear too quickly for engineers on Earth to direct the spacecraft into it.

"It's up to the spacecraft to make a decision, and we'll be watching what happened an hour ago," Ambrose says. A mission like that, he adds, is "totally impossible" without giving the machine real autonomy.

Robots built for a different physics

As AI and robotics advance on Earth, a commercial space boom is creating new openings to deploy those technologies in orbit. The catch: what works on the ground does not necessarily work in space.

Icarus Robotics is developing what it calls a robotic labor force for space, including Joy, a free-flying robotic system. Joy recently completed zero-gravity testing in Canada ahead of a planned deployment to the ISS, where one of its first tasks would be moving cargo bags between modules. The company intends to begin with teleoperation and use the resulting data to eventually train the robots to work independently.

"What the rollout will probably look like is something much closer to beginning with partial autonomy, so you still have human supervision in the loop at all times," says Jamie Palmer, Icarus's co-founder and CTO.

The core problem is physics. A robot trained on Earth learns from the environment around it — and in orbit, that environment behaves in fundamentally different ways. "If you take the newest Gemini robotics model, or you take the newest physical intelligence model, and you put it in zero-G there, it's just going to fail immediately," Palmer says. On Earth, a robot learns that pushing an object off a table makes it fall. In orbit, the object keeps moving.

That leaves Icarus facing a data problem the terrestrial robotics industry knows well, but in a more extreme form: there is very little real-world data from the environment where its robots will operate. The company is combining demonstrations from its robots in microgravity with simulations and ground tests to build its own dataset. "I wish there was" an available dataset Icarus could download and use, says Ethan Barajas, the company's co-founder and CEO. "But there's not today, not in a meaningful way."

The real question is risk management, not perfection

The challenge is not simply teaching spacecraft to act on their own. It is figuring out how to manage the risks of giving them more freedom.

"It has been mind-boggling to me how little autonomy we have in space applications," says Ufuk Topcu, an engineering professor at The University of Texas at Austin and director of the Center for Autonomy. "Because it's exactly the place where human involvement is extremely hard, the stakes are high, and you need to act fast."

Topcu argues the goal cannot be to guarantee that autonomous systems never make mistakes. These systems are most useful precisely in situations humans cannot anticipate, he says. The practical path forward is to start with restricted applications, learn how the systems behave, and gradually expand where and how they are used. In his view, the real question is how well the risk of deployment is managed — not whether risk can be eliminated entirely.

The economics of the space industry may force that reckoning sooner rather than later. For most of the space age, a small number of government agencies designed missions that took decades to develop and operate. Commercial companies are now putting more spacecraft into orbit, and new missions can go from concept to launch far faster than before.

"Space used to have very slow innovation cycles," Topcu says. "They would think of a mission concept and spend 10 or 15 years on it. It's not like that anymore. Everything is evolving faster now."

The near-term picture is one of graduated trust: humans in the loop, as with the Perseverance drives, while the systems accumulate flight experience. But the Europa example defines the boundary. Missions to the outer solar system — and a commercial orbital economy operating at increasing speed and scale — will not wait for instructions from Earth. The spacecraft that fly them will have to decide for themselves, and the experiments now running on the ISS and Mars are the first steps toward making that decision-making trustworthy enough to fly.