RandLA-Net is an efficient neural architecture for semantic segmentation of large-scale 3D point clouds. It uses random sampling to downsample point clouds while preserving key features through a local feature aggregation module that increases the receptive field for each point. This approach is significantly faster than existing methods, achieving up to 200× speedup on large point clouds. RandLA-Net outperforms state-of-the-art methods on two large-scale benchmarks, Semantic3D and SemanticKITTI. The key innovations include random sampling for efficient processing and a local feature aggregation module that progressively enhances the receptive field to preserve geometric details. The architecture is memory and computationally efficient, using shared MLPs to achieve high performance without requiring pre/post-processing steps. Experiments show that RandLA-Net can process 1 million points in a single pass, demonstrating its effectiveness on real-world large-scale point clouds. The method is efficient, scalable, and suitable for real-time applications.RandLA-Net is an efficient neural architecture for semantic segmentation of large-scale 3D point clouds. It uses random sampling to downsample point clouds while preserving key features through a local feature aggregation module that increases the receptive field for each point. This approach is significantly faster than existing methods, achieving up to 200× speedup on large point clouds. RandLA-Net outperforms state-of-the-art methods on two large-scale benchmarks, Semantic3D and SemanticKITTI. The key innovations include random sampling for efficient processing and a local feature aggregation module that progressively enhances the receptive field to preserve geometric details. The architecture is memory and computationally efficient, using shared MLPs to achieve high performance without requiring pre/post-processing steps. Experiments show that RandLA-Net can process 1 million points in a single pass, demonstrating its effectiveness on real-world large-scale point clouds. The method is efficient, scalable, and suitable for real-time applications.