Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Xubo Luo, Zhaojin Li, Xue Wan, Wei Zhang, Leizheng Shu

TL;DR

KANLoc addresses the critical need for accurate, real-time $6$-DoF pose estimation in autonomous lunar landings by tightly fusing high-frequency visual odometry with sparse, learned absolute pose anchors produced by a Kolmogorov-Arnold Network. The method integrates a $KAN$-based absolute pose regressor, a CPU-based ORB VO module, and a constrained bundle adjustment backend to achieve globally consistent trajectories with real-time performance. Key contributions include the novel $KAN$ architecture for efficient pose regression, a tightly-coupled VO-AVL fusion strategy, and a mask-augmented training regime that boosts robustness under occlusion and sim-to-real transfer. Results on synthetic Unreal/AirSim data and real Chang'e-3 imagery show substantial reductions in translation and rotation error (e.g., up to $32\%$ and $45\%$ respectively) and real-time operation (>$15$ FPS), highlighting its practical impact for resource-constrained lunar landers and demanding illumination conditions.

Abstract

Accurate and real-time 6-DoF localization is mission-critical for autonomous lunar landing, yet existing approaches remain limited: visual odometry (VO) drifts unboundedly, while map-based absolute localization fails in texture-sparse or low-light terrain. We introduce KANLoc, a monocular localization framework that tightly couples VO with a lightweight but robust absolute pose regressor. At its core is a Kolmogorov-Arnold Network (KAN) that learns the complex mapping from image features to map coordinates, producing sparse but highly reliable global pose anchors. These anchors are fused into a bundle adjustment framework, effectively canceling drift while retaining local motion precision. KANLoc delivers three key advances: (i) a KAN-based pose regressor that achieves high accuracy with remarkable parameter efficiency, (ii) a hybrid VO-absolute localization scheme that yields globally consistent real-time trajectories (>=15 FPS), and (iii) a tailored data augmentation strategy that improves robustness to sensor occlusion. On both realistic synthetic and real lunar landing datasets, KANLoc reduces average translation and rotation error by 32% and 45%, respectively, with per-trajectory gains of up to 45%/48%, outperforming strong baselines.

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

TL;DR

KANLoc addresses the critical need for accurate, real-time -DoF pose estimation in autonomous lunar landings by tightly fusing high-frequency visual odometry with sparse, learned absolute pose anchors produced by a Kolmogorov-Arnold Network. The method integrates a -based absolute pose regressor, a CPU-based ORB VO module, and a constrained bundle adjustment backend to achieve globally consistent trajectories with real-time performance. Key contributions include the novel architecture for efficient pose regression, a tightly-coupled VO-AVL fusion strategy, and a mask-augmented training regime that boosts robustness under occlusion and sim-to-real transfer. Results on synthetic Unreal/AirSim data and real Chang'e-3 imagery show substantial reductions in translation and rotation error (e.g., up to and respectively) and real-time operation (> FPS), highlighting its practical impact for resource-constrained lunar landers and demanding illumination conditions.

Abstract

Accurate and real-time 6-DoF localization is mission-critical for autonomous lunar landing, yet existing approaches remain limited: visual odometry (VO) drifts unboundedly, while map-based absolute localization fails in texture-sparse or low-light terrain. We introduce KANLoc, a monocular localization framework that tightly couples VO with a lightweight but robust absolute pose regressor. At its core is a Kolmogorov-Arnold Network (KAN) that learns the complex mapping from image features to map coordinates, producing sparse but highly reliable global pose anchors. These anchors are fused into a bundle adjustment framework, effectively canceling drift while retaining local motion precision. KANLoc delivers three key advances: (i) a KAN-based pose regressor that achieves high accuracy with remarkable parameter efficiency, (ii) a hybrid VO-absolute localization scheme that yields globally consistent real-time trajectories (>=15 FPS), and (iii) a tailored data augmentation strategy that improves robustness to sensor occlusion. On both realistic synthetic and real lunar landing datasets, KANLoc reduces average translation and rotation error by 32% and 45%, respectively, with per-trajectory gains of up to 45%/48%, outperforming strong baselines.
Paper Structure (29 sections, 10 equations, 4 figures, 7 tables, 1 algorithm)

This paper contains 29 sections, 10 equations, 4 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Monocular Localization
  4. Hybrid Localization
  5. Kolmogorov-Arnold Networks (KAN)
  6. Methodology
  7. KAN-based Absolute Pose Regressor
  8. Relative Pose Estimation
  9. Globally-Consistent Trajectory Estimation via Constrained BA
  10. Experiments
  11. Experiments on Synthetic Datasets
  12. Synthetic Dataset Generation
  13. Implementation Details
  14. Results and Analysis
  15. Experiments on Real Flight Dataset
  16. ...and 14 more sections

Figures (4)

  • Figure 1: Illustration of lunar landing phases, such as those in the Chang'e-3 mission, which require continuous and accurate localization across different altitudes and sensor conditions.
  • Figure 2: The proposed KANLoc framework integrates: (a) AVL, combining coarse retrieval with KAN-based 6-DoF refinement; (b) RVL, utilizing standard visual odometry for high-frequency tracking; and (c) a bundle adjustment back-end ensuring global consistency via AVL constraints.
  • Figure 3: Sample images from (a) real Chang'e-3 imagery and (b) high-fidelity simulation used for training or evaluation.
  • Figure 4: (a) Box plot of the localization errors on the Chang'e-3 landing sequence. (b) Error vs. Efficiency of different models (lower error is better). Bubble size indicates model parameters.