AeroMamba: An Efficient Mamba-Based Approach for Large-Scale Point Cloud Registration in Aircraft Assembly

In aircraft assembly, precise component docking is crucial, yet traditional manual alignment often results in extended assembly times and inconsistent outcomes. Recent advances in virtual assembly, relying on LiDAR-based 3D scanning, reduce both physical assembly attempts and potential damage to large aircraft parts. However, the extensive size and smooth geometry of these components require multiple LiDAR devices that produce large-scale and featureless point clouds, making accurate point cloud registration challenging. Conventional methods, including ICP-based variants, rely heavily on favorable initial poses and are prone to local minima, while Transformer-based approaches often require high computational costs. To address these issues, we propose AeroMamba, a learning-based framework for efficient and high-precision registration of large-scale aircraft point clouds. It integrates a state-space model for local feature extraction, combing a Hilbert curve strategy to maintain spatial coherence by reordering raw point clouds in a one-dimensional sequence and a global feature module for robust representation. Finally, we use Singular Value Decomposition to compute transformations. Validations on a self collected fuselage dataset demonstrate AeroMamba’s superior accuracy, scalability, and computational efficiency, highlighting its potential for real-world aircraft assembly applications.