Moving through the International Space Station might look effortless on camera, but for robots, it is one of the hardest navigation problems imaginable. There is no up or down, corridors are tight, equipment sticks out in every direction, and computing power is far more limited than what robots enjoy on Earth. Until now, that combination has kept autonomous robots largely on a short leash.
For the first time, researchers have shown that artificial intelligence can help a robot navigate the ISS on its own.
In a new experiment led by Stanford University, NASA’s cube-shaped Astrobee robot successfully planned and executed safe flight paths inside the space station using a machine-learning system. The work was presented at the 2025 International Conference on Space Robotics (iSpaRo) and marks the first known use of AI-assisted control for a robot operating aboard the ISS.
Also Read: AI Successfully Controls Satellite Attitude in Orbit for the First Time
“This is the first time AI has been used to help control a robot on the ISS,” said lead researcher Somrita Banerjee. “It shows that robots can move faster and more efficiently without sacrificing safety, which is essential for future missions where humans won’t always be able to guide them.”
Why space robots struggle with autonomy
On Earth, autonomous robots can rely on powerful processors and relatively predictable conditions. Spaceflight hardware is a different story. Computers aboard the ISS are radiation-hardened and intentionally conservative, designed for reliability over computational speed.
As senior author Marco Pavone explained, “The flight computers to run these algorithms are often more resource-constrained than those on terrestrial robots. Additionally, in a space environment, uncertainty, disturbances, and safety requirements are often more demanding than in terrestrial applications.”
To move safely under those constraints, Astrobee already relies on a traditionally used optimization method known as sequential convex programming. The method plans motion by breaking a complex navigation problem into a series of smaller optimization steps, each one carefully checking safety limits such as collision avoidance and allowable forces. The result is a trajectory that is feasible and safe, an essential requirement aboard the ISS.
The drawback is speed. In this method, each new step requires the system to solve those optimizations from scratch, which can strain limited onboard computing resources and slow planning, especially in tight, cluttered sections of the station.
Rather than replacing this safety-critical method, the Stanford team focused on speeding things up.
They trained a machine-learning model on thousands of previously solved paths, enabling the system to identify patterns such as where a corridor typically exists or where obstacles tend to be, inside ISS-like environments. When Astrobee needed to plan a new route, the AI generated an informed initial guess, known as a “warm start“, which the traditional optimizer could then refine while still enforcing every safety constraint.
Also Read: World’s Smallest Autonomous Robot That Can Sense, Think, and Act
“Using a warm start is like planning a road trip by starting with a route that real people have driven before, rather than drawing a straight line across the map,” said Banerjee. “You start with something informed by experience and then optimize from there.”
In this setup, AI never takes control away from the underlying planning system. It simply reduces the amount of work needed to reach a safe solution, allowing the robot to plan movements much faster without relaxing safety requirements.
Before sending the system into space, the researchers tested it at NASA’s Ames Research Center in Silicon Valley using a specialized testbed. There, a robot floated on a cushion of compressed air above a granite table, sliding “like a puck on an air-hockey table,” as Banerjee described it, to simulate aspects of microgravity.
When the experiment moved to the ISS, the test was run in what NASA calls a “crew-minimal” mode: astronauts handled only basic setup and cleanup, then stepped aside while Astrobee was commanded from the ground.
Over a four-hour session aboard the ISS, Astrobee was directed to fly 18 trajectories through station modules. Each trajectory was executed twice — first using the standard planning method, then again with the AI providing a warm start — allowing the researchers to make a direct, controlled comparison.
Also Read: This New AI Lets Robots “Imagine” How Objects Will Move Before Acting
The difference was clear. Paths generated with the AI-assisted approach were planned roughly 50 to 60% faster, with the largest gains appearing in tight, cluttered areas of the station and during maneuvers that required careful rotation rather than straight-line motion.
For Banerjee, watching Astrobee operate in orbit carried a more personal significance. “The coolest part was having astronauts float past during the experiment,” she said. “One of them was one of my childhood heroes, Sunita Williams. Seeing years of work actually perform in space and watching her there while the robot moved around was incredible.”
Why this matters beyond the ISS
While the demonstration involved short flights lasting just over a minute each, the implications extend far beyond the ISS. The system has now reached NASA’s Technology Readiness Level 5, meaning it has been validated in a real operational environment and is considered relatively low risk.
As missions push farther from Earth, continuous human control will become impractical. Communication delays to Mars alone can stretch into tens of minutes.
Also Read: Robot Learns 1,000 Tasks in a Single Day
“As robots travel farther from Earth and as missions become more frequent and lower cost, we won’t always be able to teleoperate them from the ground,” Banerjee said. “Autonomy with built-in guarantees isn’t just helpful; it’s essential for the future of space robotics.”
For now, Astrobee’s short flights through the ISS point to a practical change rather than a dramatic one. The robot is not acting independently in the human sense, but it no longer needs constant guidance to move safely. That small reduction in human oversight could matter a great deal as missions grow more complex and more distant.
Sources: arXiv, Stanford News