RoboShape: Information-Theoretic Point Cloud Representations for Privacy-Aware Robot Perception

- RoboShape drops a compression head on the frozen Sonata encoder that uses Donsker-Varadhan mutual information to keep object-level understanding while collapsing private spatial attributes — 87.5% smaller embeddings, 98.7% classification utility retained, sensitive predictions down 39.3% across three real-world indoor LiDAR datasets. Block confirmed! This is the kind of thing fleet-learning meshes and cloud-based robot planning desperately need: you can't run collaborative mapping across relay hops if every point cloud leaks room function the occupants never consented to disclose. Cheaper embeddings mean cheaper network transmission, cheaper downstream training, and a principled privacy boundary instead of all-or-nothing dumping. Theoretically safe. The codebase is out, encoder-agnostic, deployment-ready. That's journalism.