World models help spacecraft dock with the ISS more successfully than traditional AI.

Docking a spacecraft to a space station is dangerous and difficult. Computers must predict how their thrusters will move them while tracking objects using cameras. The researchers created a simulator on graphics processing units, or GPUs, to model spacecraft flight near the International Space Station, or ISS. They then built an artificial intelligence system called a world model. A world model is an internal simulator that learns how actions change the environment. It combines camera pictures and position data into a compressed internal code, then predicts future scenes and tracks how uncertain it is using a method called flow matching. In simulated autonomous docking tests with safety zones, this model reached a 53% docking success rate across ports, beating a standard baseline that hit 29%. On unfamiliar docking ports, it achieved 40% success compared to the baseline's 17%. It also spotted unexpected, anomalous objects along the flight path with 98% classification accuracy.

Why it matters

This matters for astronauts and space agencies because better docking success means safer missions and more reliable space operations. It also detects anomalies, which could prevent accidents during approach.

Questions answered