Astrobee is NASA’s free-flying robotic system. Using Astrobee, Stanford researchers became the first to test AI-based robotic control aboard the International Space Station. Credit: NASA

Stanford researchers have become the first to demonstrate that machine-learning control can safely guide a robot aboard the ISS, laying the groundwork for more autonomous space missions.

Imagine a robot about the size of a toaster floating through the tight corridors of the International Space Station, quietly moving supplies or checking for leaks—all without an astronaut at the controls. Such technology could free up valuable time for astronauts and open new opportunities for robotics-based exploration.

That sci-fi vision is coming closer to reality now that Stanford researchers have become the first to show that machine-learning-based control can operate aboard the ISS.

New research presented at the 2025 International Conference on Space Robotics (iSpaRo), introduces a system designed to help Astrobee, a cube-shaped, fan-powered robot, autonomously navigate the International Space Station. The paper is also published on the arXiv preprint server.

The ISS is a complex environment made up of interconnected modules filled with computers, storage, wiring, and experimental hardware. This makes planning safe motion for Astrobee far from trivial, said Somrita Banerjee, lead researcher who conducted this work as part of her Stanford Ph.D.

The traditional autonomous planning approaches that have gained traction on Earth are largely impractical for space-rated hardware.

"The flight computers to run these algorithms are often more resource-constrained than ones on terrestrial robots. Additionally, in a space environment, uncertainty, disturbances, and safety requirements are often more demanding than in terrestrial applications," said senior author Marco Pavone, associate professor of aeronautics and astronautics in the School of Engineering and director of Stanford’s Autonomous Systems Laboratory.

Despite these challenges, the team pushed the field forward with a noteworthy space research achievement. "This is the first time AI has been used to help control a robot on the ISS," said 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."

Training AI for space

Banerjee compares the challenge of optimizing Astrobee’s routes through the ISS to planning a road trip from San Francisco to Los Angeles: You want the fastest path, the most energy-efficient one, and, above all, a safe one.

To tackle that task in the ISS’s compact environment, the team’s route planning system relies on a traditionally used optimization method called sequential convex programming, which breaks a difficult planning problem into a series of smaller, simpler steps.

This process is designed to produce a final trajectory that is safe and feasible. However, solving each step from scratch can be demanding for Astrobee’s onboard computer and can slow the process—one of the key limitations of conventional techniques.

With the aim of speeding things up, the team enhanced their system with a machine-learning-based model that they trained on thousands of past path solutions. The model can reveal patterns such as where a corridor always exists and where obstacles tend to be.

Providing the robot with foundational knowledge before further refinements is known as a "warm start." The optimization technique still enforces all the safety constraints; the machine learning model just helps it reach the answer much faster.

"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," Banerjee said. "You start with something informed by experience and then optimize from there."

A milestone for AI in space

Before sending their AI to space, the team applied the system to a special testbed at NASA Ames Research Center. There, they had the AI model operate a robot similar to Astrobee, as it floated just above the surface of a granite table, buoyed by compressed air that mimics partial microgravity. "It’s like a puck on an air-hockey table," Banerjee said.

When the real test day arrived, the Stanford team joined by video call while astronauts on the ISS completed what NASA calls a "crew-minimal" setup. The astronauts handled only preparation and cleanup, then stepped aside.

For the next four hours, Banerjee sent instructions to ground operators at NASA’s Johnson Space Center in Houston. Then, the NASA team relayed the commands to Astrobee, specifying its starting point and destination, simulating obstacles to avoid, and trying both warm and cold starts.

Multiple safety measures kept the experiment secure, including replacing physical obstacles with virtual ones to eliminate collision risk, maintaining a backup robot, and allowing operators to abort a run if necessary.

The team tested 18 trajectories, each lasting more than a minute. Each was run twice: first with a cold start using the standard planning method, and then with a warm start, where the AI provided a first draft of the path that the system could quickly adjust.

The tests showed that giving Astrobee a warm start significantly sped up motion planning. "We showed that it’s 50 to 60% faster, especially in more challenging situations," Banerjee said. Those harder cases included cluttered areas, tight corridors, and maneuvers requiring rotation instead of a straight path.

Astronaut Sunita Williams helped with setup and supervision of the experiments. Video shows how the research was conducted with Astrobee aboard the ISS and describes the findings. Credit: NASA

Watching Astrobee in orbit was a deeply personal experience for Banerjee. "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."

The future of robots in orbit

After their experiment on the ISS, the team’s warm start system reached Technology Readiness Level 5, a NASA designation indicating successful testing in a real operational environment. The upgrade indicates that this technology is low risk, which is important for proposing new experiments or future missions.

Looking ahead, Banerjee said this type of mathematically grounded, safety-focused AI will be crucial as robots take on more tasks independently, and as NASA sends crewed missions to the moon and Mars.

"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," she said.

Such technologies will allow astronauts to focus on higher-priority work and use their time more effectively. "Autonomy with built-in guarantees isn’t just helpful; it’s essential for the future of space robotics," she said.

Pavone highlighted that his lab will continue to research and advance warm starting techniques.

"As part of the Center for Aerospace Autonomy Research (CAESAR), we are collaborating with the Stanford Space Rendezvous Lab to explore more powerful AI models—the same kinds used in modern language tools and self-driving systems. With stronger generalization capabilities, these models would enable robots to navigate even more challenging situations in future space missions."

More information: Somrita Banerjee et al, Deep Learning Warm Starts for Trajectory Optimization on the International Space Station, arXiv (2025). DOI: 10.48550/arxiv.2505.05588

Journal information: arXiv

Citation: Astrobee: AI-guided robot navigates space station corridors with improved speed and safety (2025, December 5) retrieved 5 December 2025 from https://phys.org/news/2025-12-ai-advances-robot-international-space.html

This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.