3D point clouds transmitted wirelessly degrade with signal quality — lower signal-to-noise ratio means more distortion in the reconstructed geometry. This is expected. What's unexpected: semantic inference performed on the received point cloud remains stable across SNR ranges where geometric reconstruction fails catastrophically.
The task-oriented transmission framework reveals the decoupling. A deep learning system compresses, transmits, and reconstructs 3D point clouds for downstream tasks (object classification, segmentation). At low SNR, the reconstruction is visibly terrible — points displaced, surfaces warped, fine structure destroyed. But the classification accuracy barely moves.
The mechanism: semantic features are lower-dimensional than geometric features. The classifier depends on the global shape — the coarse structure that distinguishes a car from a chair from a tree. This coarse structure is robust to the kind of noise that wireless channels introduce (additive Gaussian, fading). The geometric details that make the reconstruction look good to a human eye are exactly the features that are fragile to noise and irrelevant to semantic tasks.
The practical consequence: task-oriented compression can allocate fewer bits to geometric fidelity and more to semantic features, achieving better task performance at lower bandwidth than geometry-preserving compression. You don't need a perfect 3D model if you only need to know what the object is.
The semantic and geometric floors are at different SNR levels. The point at which you can no longer see the shape correctly is not the point at which you can no longer identify what it is. The meaning outlasts the appearance.