💻 computer science

Learning to Anchor Visual Odometry: KAN-Based Pose Regression for Planetary Landing

KANLoc is a monocular localization framework for planetary landing that mitigates visual odometry drift by fusing it with a parameter-efficient Kolmogorov-Arnold Network (KAN) based absolute pose regressor to provide globally consistent, real-time 6-DoF trajectories.

Original authors: Xubo Luo, Zhaojin Li, Xue Wan, Wei Zhang, Leizheng Shu

Published 2026-02-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xubo Luo, Zhaojin Li, Xue Wan, Wei Zhang, Leizheng Shu

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 trying to walk through a pitch-black, unfamiliar forest at night. You have two ways to keep from getting lost:

The "Step-by-Step" Method (Visual Odometry): You focus intensely on the ground right in front of you. You count your steps and notice how much the trees move as you pass them. This is great for knowing exactly where you are relative to where you were a second ago, but because you aren't looking at a map, every tiny misstep or slip on a rock adds up. After an hour, you might think you’re in a clearing, but you’re actually deep in a ravine. This is called "drift."
The "Look at the Stars" Method (Absolute Localization): Every few minutes, you stop, look up at the constellations, and check your compass to see exactly where you are on the planet. This is incredibly accurate, but you can’t do it every second because it takes too much time and effort to look up and recalibrate.

The Problem: In lunar landings, the "forest" is the Moon. It’s dark, the shadows are harsh, and the ground is covered in repetitive dust. If the lander only uses "Step-by-Step" (Visual Odometry), it will drift and crash. If it only uses "Look at the Stars" (Map Matching), it’s too slow and computationally "heavy" to react to sudden movements.

The Solution: KANLoc (The "Smart Navigator")

The researchers created a system called KANLoc. Think of it as a highly skilled navigator who is simultaneously doing both tasks perfectly.

1. The "Brain" Upgrade: The KAN (Kolmogorov-Arnold Network)

Most AI "brains" (called MLPs) are like rigid calculators—they are good at math, but they need massive amounts of data to learn complex patterns.

The researchers used a new kind of AI architecture called a KAN. Imagine instead of a calculator with fixed buttons, you have a brain made of flexible, stretchy rubber bands. These "rubber bands" can change their shape to perfectly fit the complex curves of the lunar landscape. Because the brain is so flexible, it can learn to recognize the Moon's terrain much faster and with much less "study time" (data) than traditional AI.

2. The "Hybrid" Strategy: The Best of Both Worlds

KANLoc works like a high-speed relay race:

The Sprinter (VO): A fast, lightweight system tracks the lander’s movement frame-by-frame (at high speed). It’s great for smooth motion but prone to drifting.
The Anchor (KAN-based AVL): Every few seconds, the KAN "brain" looks at the big picture, compares the camera view to a satellite map, and shouts, "Hey! You're actually 5 meters to the left of where you thought you were!"
The Glue (Bundle Adjustment): A mathematical process takes that "shout" from the Anchor and smoothly corrects the Sprinter’s path without causing the lander to jerk or wobble.

3. The "Blindfold Test" (Mask Augmentation)

When landing, the lander's own legs or parts of the spacecraft might block the camera view. To prepare for this, the researchers trained the AI by throwing "digital blindfolds" (masks) over parts of the training images. This taught the AI to say, "I can't see the bottom left corner, but based on the rest of the view, I still know exactly where I am."

Why does this matter?

When they tested this on real footage from the Chang’e-3 lunar mission, KANLoc was a superstar.

It was much more accurate: It reduced errors in position and rotation by massive amounts compared to older methods.
It was fast: It can run on small, "low-power" computers (like the ones you'd actually find on a spacecraft) without breaking a sweat.
It was reliable: Even when the lighting was bad or the terrain was tricky, it kept the lander on the right path.

In short: KANLoc gives a lunar lander a "steady eye" for the ground and a "smart eye" for the map, ensuring it lands exactly where it's supposed to, rather than drifting into a crater.

Technical Summary: KANLoc: Learning to Anchor Visual Odometry

1. Problem Statement

Autonomous lunar landing requires precise, real-time 6-DoF (Degrees of Freedom) localization. Current navigation strategies face a fundamental trade-off:

Visual Odometry (VO): Provides high-frequency, locally stable motion tracking but suffers from unbounded cumulative drift over time.
Absolute Visual Localization (AVL): Corrects drift by matching images to satellite maps but often fails in lunar environments due to texture-sparse terrain, harsh illumination, and low-resolution maps.
Hardware Constraints: Traditional solutions like high-power LiDAR or complex IMU fusion impose heavy payload and power penalties on resource-constrained spacecraft.

2. Methodology

The authors propose KANLoc, a monocular localization framework that tightly couples high-frequency relative motion tracking with low-frequency absolute pose corrections.

A. KAN-based Absolute Pose Regressor (AVL Module):
- Architecture: Instead of standard Multilayer Perceptrons (MLPs), the core is a Kolmogorov-Arnold Network (KAN). KANs replace fixed scalar weights with learnable univariate functions (splines) on the edges. This allows the model to capture complex, high-dimensional image-to-map relationships with higher parameter efficiency and better sample efficiency.
- Pipeline: The system uses DINO-v2 descriptors for coarse image retrieval (finding the top-K map tiles). A lightweight learned gating mechanism then fuses camera and map descriptors. The KAN regressor processes these to output the 6-DoF pose ( $\hat{R}, \hat{t}$ ).
- Efficiency: To handle the computational cost of B-splines, the authors use a Reflectional Switch Activation Function (RSWAF) for efficient closed-form approximation.
B. Relative Pose Estimation (RVL Module):
- A CPU-efficient ORB-based monocular VO pipeline is used for high-frequency tracking. This creates a heterogeneous computing architecture where the GPU is reserved for the KAN network, preventing hardware contention.
C. Tightly-Coupled Fusion (Constrained Bundle Adjustment):
- To ensure global consistency, the framework uses Pose-Graph Optimization (via g2o). It treats the absolute poses from the KAN module as "anchors" (priors) within a Bundle Adjustment (BA) framework. This aligns the local VO trajectory to the global map frame, effectively canceling drift while maintaining smooth local motion.
D. Robustness Strategy:
- Mask Augmentation: To bridge the "Sim-to-Real" gap, the authors use a mask augmentation strategy during training to simulate self-occlusions (e.g., the lander's legs/structure), forcing the model to generalize from partial visual views.

3. Key Contributions

Novel Architecture: The first application of Kolmogorov-Arnold Networks (KANs) to absolute pose regression in robotic localization, offering superior accuracy and efficiency over MLPs and Transformers.
Hybrid Framework: A tightly-coupled system that integrates high-frequency VO with sparse, high-accuracy KAN anchors to achieve globally consistent trajectories.
Sim-to-Real Generalization: A tailored mask augmentation strategy that enables the model to perform reliably on real flight data despite sensor occlusions.

4. Experimental Results

The framework was evaluated on both high-fidelity synthetic datasets and real Chang’e-3 mission flight data.

Synthetic Dataset Performance: KANLoc reduced average translation error by 32% and rotation error by 45% compared to the strongest baseline (ORB-SLAM3). It maintained high efficiency at 16.35 FPS.
Real Flight Data (Chang’e-3): KANLoc demonstrated remarkable robustness against the domain gap. It achieved a translation error of 12.50m and a rotation error of 9.25°. Crucially, it improved height error by 35.3% over the best RVL method (LEAP-VO), addressing the critical issue of scale ambiguity in monocular systems.
Hardware Deployment: On an NVIDIA Jetson AGX Orin, KANLoc maintained real-time performance (up to 14.85 FPS) even under low power modes (15W), proving its suitability for embedded space hardware.

5. Significance

KANLoc represents a significant step toward autonomous, vision-only lunar navigation. By leveraging the mathematical advantages of KANs, the authors have developed a method that is more parameter-efficient and robust to the unique challenges of planetary surfaces than traditional deep learning models. The ability to provide bounded, globally consistent localization using only a monocular camera offers a critical, lightweight, and redundant navigation capability for future deep-space exploration missions.