How to sample point clouds 2.5× faster for robot vision

Farthest Point Sampling (FPS) downsamples point clouds while preserving geometry, but its cubic complexity kills real-time performance in robotics. RadiusFPS prunes redundant distance calculations using spherical voxels and a GPU kernel that fuses operations into memory-coalesced blocks. On SemanticKITTI and ScanNet, it matches or beats existing methods with 2.5× speedup and half the GPU memory of competing approaches.