Generative AI improves a wireless vision system that sees through obstructions

MIT researchers have developed generative artificial intelligence models that improve the accuracy of wireless vision systems used to see through obstructions. The techniques build on more than a decade of work with surface-penetrating millimeter wave signals that reflect off hidden objects. The new approach creates a partial reconstruction from reflected signals and employs a specially trained generative model to fill in missing shape details. This addresses limitations caused by specular reflections that previously rendered portions of surfaces invisible to sensors. One resulting system reconstructs individual objects with nearly 20 percent greater accuracy than prior methods. A second system reconstructs entire indoor scenes from signals reflected off moving humans using only one stationary radar unit. The systems preserve privacy by avoiding camera-based methods and could support applications such as warehouse inventory verification and safer human-robot interactions. Both papers describing the work will be presented at the IEEE Conference on Computer Vision and Pattern Recognition. Fadel Adib, associate professor in the Department of Electrical Engineering and Computer Science and director of the Signal Kinetics group in the MIT Media Lab, is the senior author. The research received support from the National Science Foundation, the MIT Media Lab, and Amazon.