🤖 machine learning

Deep Learning-Based Lunar Crater Terrain Relative Navigation

This paper proposes a deep learning-based terrain relative navigation system that integrates a specialized crater detector with an Extended Kalman Filter to accurately estimate spacecraft position and reduce navigation errors to a few hundred meters, even when starting with initial location uncertainties of up to 5 km.

Original authors: Batu Candan, Simone Servadio

Published 2026-06-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Batu Candan, Simone Servadio

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are flying a spaceship toward the Moon, but your GPS is broken, and your internal compass is slowly drifting off course. You are flying blind over a landscape that looks like a giant, gray cookie covered in chocolate chips (craters). If you don't know exactly where you are, you might crash into a dangerous spot or miss your landing zone entirely.

This paper presents a new way for a spaceship to find its way using a combination of super-smart computer vision and mathematical guessing, specifically designed for the Moon.

Here is how the system works, broken down into simple parts:

1. The "Eyes": A Deep Learning Detective

First, the spaceship needs to "see" the craters. The authors built a special computer program (a deep learning network based on a model called U-Net++) that acts like a detective.

Training: This detective was trained on thousands of pictures of the Moon, learning to spot the rims and centers of craters even when the lighting is weird or shadows are long. It's like teaching a child to recognize a cookie by its chocolate chips, no matter how the light hits the plate.
The Job: When the spaceship flies over the Moon, this detective scans the images and draws little circles around every crater it sees.

2. The "Map": Matching Chips to a Catalog

Once the detective spots the craters, the spaceship needs to know which craters they are.

The Global Database: The Moon has a giant, pre-made map of every known crater (a "catalog").
The Matching Game: The system takes the craters the detective found and tries to match them to the ones on the map.
The "Quadrant" Trick: To make sure the match is perfect, the system doesn't just look at the nearest crater. It picks four craters spread out across the image (like the four corners of a room). This ensures the spaceship gets a wide, stable view of its surroundings, preventing it from getting confused if one crater looks slightly different than expected.

3. The "Brain": The Extended Kalman Filter (EKF)

Now that the spaceship knows which craters it is looking at, it needs to calculate exactly where it is. This is done by a mathematical engine called an Extended Kalman Filter (EKF).

The Guess-and-Check: The EKF is like a very smart guesser. It starts with a "best guess" of where the ship is.
The Correction: When the detective finds a match (e.g., "That crater on the left is actually Crater A, not Crater B"), the EKF says, "Oh, I was wrong! I need to move my position estimate."
The Result: It constantly corrects the ship's path, shrinking the error from being potentially 5 kilometers off down to just a few hundred meters (or even less).

4. The Safety Net: Altitude Aiding

The paper mentions a small but important trick: the system also uses a simple measurement of how high the ship is (altitude). Think of this as a "height gauge" that stops the ship from drifting up or down too much, keeping the horizontal position (left/right/forward/back) accurate.

What Did They Prove?

The authors didn't just build this; they tested it in a computer simulation that mimics the Moon's surface.

The Test: They started the simulation with the spaceship completely lost (up to 5 km away from where it thought it was).
The Outcome: The system quickly realized its mistake, corrected its course, and stabilized the navigation error to within roughly 50 to 100 meters.
The Key Insight: They found that for this to work, the "training" of the detective must match the "reality" of the flight. If the pictures the computer learns from look different than the pictures it sees in space (different sizes, different lighting), the system gets confused.

Summary

In short, this paper describes a navigation system that lets a spaceship "read" the Moon's surface like a map. By using a highly trained AI to spot craters and a mathematical filter to correct its position, the spaceship can find its way home even if it starts off completely lost, without needing expensive radar equipment. It turns the Moon's chaotic, cratered surface into a reliable guidepost.