AI-Powered Robot Navigates ISS 60% Faster | Astrobee Update 2025

In a significant step forward for autonomous robotics, scientists successfully piloted a small robot, Astrobee, inside the International Space Station (ISS) using artificial intelligence (AI) in February 2025. The development of the flying robot was led by Dr. Somrita Banerjee under the supervision of Professor Marco Pavone.

According to research published in November, Astrobee can now independently plan routes through the ISS 50 to 60 percent faster than previously possible. This represents a major advancement, particularly as similar robots are envisioned for future missions to the Moon and Mars.

Astrobee robots were first deployed to the ISS in 2019, with three units – named Honey, Queen, and Bumble – arriving at the station. At that time, however, the robots were not equipped with AI capabilities.

The Astrobee robots are cube-shaped, measuring 30 centimeters on each side and weighing nine kilograms. They are propelled by electric fans and equipped with cameras, obstacle sensors, RFID readers, and even a small robotic arm. They can also autonomously recharge at a docking station. Their tasks include routine monitoring, inventory management, and assistance with experiments.

Consider planning a route from Prague to Brno: you might prioritize the shortest distance, fastest time, or most fuel-efficient path. Traditional navigation systems use a method called sequential convex programming (SCP) to solve this, relying on a pre-existing map. SCP breaks down complex planning problems into a series of smaller, simpler steps. The system first identifies all possible routes and then calculates the length, duration, and fuel consumption for each, ultimately selecting the optimal option.

Navigating the ISS, however, is far more complex. Unlike driving, which operates primarily in two dimensions, Astrobee navigates a continuous three-dimensional space. Furthermore, the robot may initially have no prior knowledge of its surroundings.

The robot must avoid various obstacles – cables, equipment, boxes, and modules – and uses fans instead of wheels. Traditionally, this requires dozens of refinement cycles, even with an accurate and up-to-date map of the station. This process is slow, even with AI assistance, given the limited processing power onboard Astrobee.

Switching to AI algorithms that use incremental neural network learning can sometimes be even slower, requiring the robot to learn its way around the ISS gradually.

This is where the “hot start” approach comes into play, according to the scientific portal Phys.org. Instead of starting with a blank slate, the robot begins with thousands of virtual flights completed in a simulated ISS environment, effectively pre-training its neural network.

To illustrate the difference, consider a cold start as finding a route from point A to point B in an unknown, complex environment, starting from scratch at every intersection. A hot start, on the other hand, is like having a local resident provide a rough map that already accounts for major roads, detours, and one-way streets.

The network first receives input data for its calculations: the robot’s starting point, destination, orientation, speed, and nearby obstacles. It then immediately proposes an initial route – an approximate path and control sequence.

Somrita Banerjee compares this to real-world trip planning: instead of starting with a blank page, you begin with established routes and advice from experienced drivers.

Before the live test in space, the team tested the approach on another robot at NASA’s Ames Research Center. They used a small robotic puck, similar to Astrobee, that levitated above a granite slab, training its software neural network for autonomous control.

For Astrobee, the final test took place in orbit: a four-hour trial on the ISS with minimal crew intervention. Astronaut Sunita Williams set up the scenario, which took 35 minutes, and then stepped back.

Astrobee then completed 18 flights lasting over a minute, successfully avoiding virtual obstacles, with a backup robot and emergency stop button available. The researchers first tested the robot with a “cold start,” having no prior experience. They then ran a new set of 18 flights using the “hot start” method, leveraging the previously trained procedures from the simulator.

Within the ISS’s confined spaces, requiring robot rotation, the AI-powered “hot start” variant saved up to 60 percent of the time.

Banerjee carefully monitored the robot’s safe flight, even with astronauts nearby, which she described as the best moment of her life.

Astrobee is now, thanks to AI, more than just a test platform. The successful testing in the real environment of the ISS prepares it for fully autonomous operation on larger missions. For example, the communication delay between Earth and Mars can exceed 30 minutes, making remote control from Earth impractical.

It will also be valuable for robots to independently build, repair, or monitor systems during crewed missions to more distant destinations.

Stanford plans to further improve the results, leveraging AI experience from autonomous vehicles and models like ChatGPT for even more complex and chaotic environments (e.g., unexpected obstacles).

“Robot autonomy in space with safety guarantees is the future,” stated Marco Pavone. According to ENGtechnica, this means cheaper and more frequent space missions, saving human lives and time. This breakthrough demonstrates that today’s AI can function even in the harshest conditions of space.