RandLA-Net is an efficient and lightweight neural architecture designed for semantic segmentation of large-scale 3D point clouds. The key innovation is the use of random point sampling instead of complex point selection methods, which significantly reduces computational and memory requirements. To address the potential loss of key features due to random sampling, RandLA-Net introduces a novel local feature aggregation module that progressively increases the receptive field for each 3D point, preserving geometric details. Extensive experiments demonstrate that RandLA-Net can process 1 million points in a single pass, up to 200 times faster than existing approaches, and outperforms state-of-the-art methods on the Semantic3D and SemanticKITTI benchmarks. The paper also includes an analysis of different sampling approaches and an ablation study to validate the effectiveness of the proposed local feature aggregation module.RandLA-Net is an efficient and lightweight neural architecture designed for semantic segmentation of large-scale 3D point clouds. The key innovation is the use of random point sampling instead of complex point selection methods, which significantly reduces computational and memory requirements. To address the potential loss of key features due to random sampling, RandLA-Net introduces a novel local feature aggregation module that progressively increases the receptive field for each 3D point, preserving geometric details. Extensive experiments demonstrate that RandLA-Net can process 1 million points in a single pass, up to 200 times faster than existing approaches, and outperforms state-of-the-art methods on the Semantic3D and SemanticKITTI benchmarks. The paper also includes an analysis of different sampling approaches and an ablation study to validate the effectiveness of the proposed local feature aggregation module.