Researchers trained an AI to navigate near unknown spacecraft

A recent research paper introduces a way for spacecraft to identify and track objects in orbit using only a monocular camera. This approach allows a robot to figure out where a target is without needing previous data about its shape or weight.

Typically, machines in space rely on expensive tools like lidar or multiple cameras to gauge distance. This new method skips those extras. It uses a transformer-based neural network to spot features on the target spacecraft and feed that information into a Multi-State Constraint Kalman Filter, or MSCKF.

Think of the Kalman filter as a constant guess-and-correct engine. It combines the visual data from the camera with simple physics models of how objects move in space to keep track of the target's position and orientation.

By performing specific maneuvers, the servicer spacecraft can confirm the exact distance to the object it is observing. The researchers tested this on synthetic images of over 100 different craft, including some the AI had never encountered before.

In these simulations, the system achieved a median error of about 3.7 degrees in attitude and 2.2 percent in range. It performs impressively even when it does not have a map of what it is looking at.

However, this model was only tested in simulated environments rather than actual space missions. It also relies on the servicer moving in specific ways to fully resolve the distance to the target, which might be difficult in tight timeframes. Have you ever wondered how we will manage the growing amount of space junk orbiting our planet?