Generative AI improves a wireless vision system that sees through obstructions

MIT researchers developed two systems that use generative AI to improve wireless vision through obstructions. The first, Wave-Former, uses millimeter wave (mmWave) signals to partially reconstruct hidden 3D objects, then employs a specially trained generative AI model to fill in missing geometry caused by specular reflections — achieving ~20% accuracy improvement over baselines. The second system, RISE, reconstructs entire indoor scenes using multipath mmWave reflections off moving humans from a single stationary radar, producing reconstructions twice as precise as existing methods. A key innovation was adapting large computer vision datasets to simulate mmWave reflection physics, bypassing the lack of large mmWave training datasets. Applications include warehouse robots verifying packed items and smart home robots tracking human location while preserving privacy.