Generative AI Lets Spacecraft Make Their Own Decisions: A Shift Engineers Once Feared

Space was once seen as the ultimate frontier for human exploration. Today it is becoming the ultimate testing ground for artificial intelligence. NASA's Jet Propulsion Laboratory is using Anthropic's Claude model to plan driving routes for the Perseverance Mars rover. NASA and IBM are sending compressed AI models to the International Space Station and satellites to identify floods and clouds. Astronauts aboard the ISS are testing whether large language models can help with maintenance procedure queries. These experiments point to a deeper engineering shift: engineers are beginning to seriously consider whether spacecraft should have genuine autonomous decision-making capabilities.

Why Spacecraft Need True Autonomy

Spacecraft have operated autonomously for decades, but autonomy has never been the dominant mode. Space engineers have long favored systems with predictable behavior. Robert Ambrose, former head of NASA's Software, Robotics, and Simulation Division, explains that autonomous systems often lack deterministic outcomes—different paths into the same situation produce different behaviors, and engineers detest that.

During his NASA career, Ambrose worked on autonomous systems for the Orion spacecraft and the Robonaut 2 humanoid robot. On the Orion project, he saw firsthand how testing problems could balloon rapidly—engineers had to consider not only what the spacecraft might do, but also the many different paths it might take to reach a given decision point. But Ambrose says engineers found ways to manage that complexity, including automating the testing itself. "We used autonomy to fight the challenge of autonomy," he says. "And it actually works."

The farther a mission is from Earth, the more obvious the need for autonomy becomes. Ambrose gives a hypothetical Europa mission as an example: a spacecraft would need to dive through water plumes erupting from the surface of Jupiter's icy moon. The plumes appear too quickly for engineers on Earth to command in time. "Only the spacecraft can make the decision itself, and we can only watch afterward what happened an hour ago," he says. Without true autonomy, such missions are "completely impossible."

Deterministic vs. non-deterministic systems is a core distinction in space software engineering. A deterministic system always produces the same output for the same input, and its behavior can be fully predicted and verified—which is why it has long been the mainstream choice for space software, since engineers can formally prove every state transition. A non-deterministic system (such as a neural-network-based model or a reinforcement learning policy) may take different decision paths in the same situation due to internal state, sampling randomness, or tiny environmental differences, making verification and certification extremely difficult. In the space domain, this non-determinism is compounded by unique factors such as communication delays, radiation-induced bit flips, and sensor noise, making traditional software test coverage metrics hard to apply directly. What Ambrose means by "using autonomy to fight the challenge of autonomy" is precisely leveraging automated testing, simulation, and AI-assisted verification tools to manage the complexity that AI itself introduces.

How Robots Adapt to Microgravity Physics

As AI and robotics advance on Earth, the commercial space boom is creating new opportunities to apply these technologies in orbit. But what works on Earth may not work in space.

Icarus Robotics is developing what it calls a space robotic workforce, including a free-flying robotic system called Joy. Joy recently completed zero-gravity testing in Canada and is slated for deployment to the International Space Station, where one of its first tasks will be moving cargo bags between modules. The company plans to start with teleoperation and use the data it collects to train the robot to eventually work autonomously.

"The deployment process will likely be closer to starting with partial autonomy, always with a human in the loop," says Jamie Palmer, co-founder and CTO of Icarus. Palmer points out that robots trained on Earth learn the physics of their surroundings, but in orbit the physics are completely different. "If you take the latest Gemini robotics model or the latest physical intelligence model and put it in a zero-gravity environment, it will fail immediately," he says. On Earth, a robot learns that pushing an object off a table makes it fall; in orbit, the object keeps moving.

This presents Icarus with a more extreme version of a problem the terrestrial robotics industry also faces: very little real-world data from the environment where the robot operates. The company is combining microgravity demonstrations with ground-based simulation and testing to build its own dataset. "I wish there were an off-the-shelf dataset I could just download and use," says co-founder and CEO Ethan Barajas. "But there isn't today—there's no meaningful dataset."

The differences in physics under microgravity are a key challenge in training space robots. On Earth, robotic models learn implicit rules of gravity, friction, and contact forces through large-scale interaction data—objects fall due to gravity after being pushed and slow down due to friction. Once in orbit, gravitational acceleration drops to roughly 10⁻⁶g, and a pushed object continues moving in a near-straight line until it hits a module wall or is altered by other forces (such as air drag or static electricity). This means vision-motion policies, grasping models, and physics prediction networks trained on the ground quickly fail in orbit, exhibiting the "push it and it flies away" phenomenon. Solutions include collecting real microgravity data in parabolic flights, drop towers, or neutral buoyancy pools; using domain randomization to broaden ground simulation distributions; and adopting online adaptive fine-tuning strategies so robots can quickly calibrate their internal dynamics models in orbit with a small amount of real interaction.

How to Manage the Risks of Autonomous Spacecraft

The challenge is not only teaching spacecraft to act autonomously, but also managing the risks of giving them more freedom.

"The lack of autonomy in space applications is incredible to me," says Ufuk Topcu, an engineering professor at the University of Texas at Austin and director of its Autonomy Center. "Because it is precisely in space that human intervention is extremely difficult, the risks are extremely high, and you need to act quickly."

Topcu argues that the goal cannot be to guarantee autonomous systems never fail. They are most useful in situations humans cannot foresee, so researchers need to start with constrained applications, understand system behavior, and then gradually expand the scope of use. In his view, the real question is how to manage deployment risk, not whether risk can be completely eliminated.

This may become increasingly important as the space industry changes. For most of the Space Age, missions designed by a handful of government agencies could take decades to develop and operate. Today commercial companies are sending more spacecraft into orbit, and new missions can be developed and launched faster. "Space used to have a very slow innovation cycle," Topcu says. "They would spend 10 to 15 years studying a mission concept. Now it's different—everything is accelerating."

From Claude planning Mars routes to LLM maintenance assistants on the ISS, from zero-gravity robots to autonomous decision architectures, space engineering is undergoing a paradigm shift. Engineers once feared non-deterministic systems. Now they are beginning to ask: in the face of more complex, more distant, and more numerous missions, can we still afford the cost of not giving spacecraft autonomy?

Constrained applications and incremental deployment are common paths in safety engineering for autonomous systems. What Topcu means by "starting with constrained applications" in the space context usually means first limiting the operational domain—for example, enabling autonomous decisions only in specific modules, under specific lighting, and during specific mission phases—while retaining human-in-the-loop oversight and emergency takeover channels. As the system accumulates operational data and exposes edge cases in real environments, constraints are gradually relaxed and autonomous authority expanded. This approach borrows from the tiered thinking of aviation autopilot and ground autonomous driving, turning the question of "can risk be completely eliminated" into an engineering management question of "how to expand capability in an orderly way under acceptable risk." For deep-space missions, because communication delays are measured in minutes or even hours, human-in-the-loop can only degrade to after-the-fact oversight, making early constrained verification and formal guarantees especially critical.

One-line Summary

From Mars rovers to deep-space probes, space engineering is shifting from a preference for deterministic systems to seriously considering giving spacecraft true autonomous decision-making capabilities.

Paragraph Summary

The article traces an engineering shift in which autonomy is moving from the margins to the core of the space field. NASA has used Claude to plan driving routes for Perseverance and tested large language models aboard the ISS to assist with maintenance; Icarus Robotics is developing a free-flying robot called Joy in an effort to build a space robotic workforce. But AI and robotics solutions that work on Earth often fail in microgravity, and real on-orbit data is extremely scarce. Former NASA division head Ambrose notes that deep-space missions such as diving through Europa's plumes can only rely on onboard autonomous decisions. UT Austin professor Topcu argues the key is not eliminating risk but starting with constrained applications and gradually expanding autonomous authority. Commercial space is accelerating the innovation cycle, making autonomy shift from optional to essential.