Stanford University researchers have demonstrated an AI-based control system for NASA’s Astrobee robot aboard the International Space Station, marking the first use of machine-learning-assisted robotic control on the station. The work, presented at the 2025 International Conference on Space Robotics, focuses on improving Astrobee’s ability to plan safe and efficient trajectories through the station’s confined and obstacle-filled interior.
Astrobee operates with space-rated flight computers that offer less processing capacity than systems typically used for terrestrial robotics. According to senior author Marco Pavone, this constraint, combined with environmental uncertainties and stringent safety requirements, limits the practicality of widely used autonomous planning methods. The Stanford team addressed these constraints by combining sequential convex programming, a conventional trajectory-optimization method, with a machine-learning model trained on historical path data. The model provides an initial “warm start” trajectory that reduces the computational load required to generate a final, constraint-compliant path.
Lead researcher Somrita Banerjee said the approach allows the optimization process to converge more quickly while retaining the safety guarantees of the underlying method. Prior to testing in orbit, the system was validated on a microgravity-analog testbed at NASA Ames Research Center.
The ISS experiment was conducted through NASA’s “crew-minimal” setup, with astronauts overseeing preparation and cleanup while ground operators executed the tests. Eighteen trajectories were evaluated, each run once with a standard cold start and once with a warm start. The team reported that warm starts reduced planning time by about 50 to 60 percent, particularly in areas with greater clutter or where rotational maneuvers were required. Safety protocols included the use of virtual obstacles, a backup robot, and operator-controlled abort options.
Following the experiment, the system reached NASA Technology Readiness Level 5, indicating successful operation in a relevant environment. Banerjee said such approaches will be important as robots increasingly support astronauts and operate at greater distances from Earth. Pavone noted that future work through the Center for Aerospace Autonomy Research will examine more advanced AI models to improve generalization and support more complex autonomous navigation tasks in space.
Photo: NSA
