MambaPan3D: Mamba-Transformer for 3D LiDAR Panoptic Segmentation with Adaptive Coordinate Fusion

With the advent of autonomous intelligent systems, such as humanoid robots, environmental perception should require rapid and accurate real-time 3D scene interpretation. LiDAR sensors are core and accurate distance measuring components, but it is complicated to process the unstructured, sparse, and unevenly distributed nature of LiDAR point cloud data while meeting the real-time object classification needs. To address the limitations of current 3D LiDAR-based panoptic segmentation methods, we propose a MambaPan3D design. It is a hybrid architecture that integrates Mamba and Transformer models for efficient and accurate 3D point cloud understanding. Our framework solves two key challenges: 1) geometric ambiguity caused by sparse and irregular LiDAR point cloud distributions, and 2) inefficient long-range dependency modeling in largescale scenes. Specifically, CartPolar-KAN embedding, a novel positional encoding strategy, is introduced to interpret between Cartesian and polar coordinates by adding a Kolmogorov-Arnold network (KAN) with learnable B-spline basis functions. The module dynamically fuses multi-coordinate features to overcome the limitations of fixed Bird's-Eye View (BEV) quantization. Additionally, our Mamba-Transformer Decoder combines the global attention capabilities of the Transformer and the linear computational efficiency of the Mamba state-space model to achieve real-time inference while maintaining the global receptive field. Extensive experiments on SemanticKITTI dataset demonstrated state-of-the-art performance. The panoptic quality (PQ) could reach 63.3 % in complex urban scenes, i.e., 1.3 % higher than the current optimal baseline method. The proposed framework provides a powerful solution for real-time situational awareness in autonomous driving systems, balancing accuracy, efficiency, and scalability. Our MambaPan3D model offers a robust solution for real-time situational awareness in autonomous systems through balancing accuracy, efficiency, and scalability.