Summary: An MIT Media Lab team introduced Wave-Former, which transmits millimeter-wave signals through common barriers and uses generative AI to complete hidden objects into 3D reconstructions. MIT News profiled the CVPR 2026 paper on March 19, reporting tests on roughly 70 everyday objects with accuracy nearly 20 percent above prior baselines. The work could aid robots and sealed-box checks, while through-wall sensing also raises privacy questions the paper does not resolve.
Wave-Former is the work of a Media Lab research team, and MIT News profiled the system on March 19. The setup transmits millimeter-wave wireless signals — the same band Wi-Fi occupies — through cardboard, wood, and drywall; whatever echoes back from a fully hidden object is passed to a generative AI model, which completes the shape into a full 3D reconstruction. Fadel Adib leads the project, as an associate professor in MIT’s EECS department and director of the Signal Kinetics group at the Media Lab. The paper, “Wave-Former: Through-Occlusion 3D Reconstruction via Wireless Shape Completion,” has been accepted to CVPR 2026. Tested on roughly 70 everyday objects, the system’s reconstruction accuracy landed nearly 20 percent above state-of-the-art baselines.
For robotics and warehouse automation, this is a weighty step: the long-standing rule that the arm cannot reach what the sensor never saw finally has a systematic answer — and one that rides on radio waves ordinary hardware can already emit.
What Wave-Former Sets Out to Fix
“Wireless vision” rests on a physical fact: millimeter waves travel through drywall, plastic, and cardboard, strike a hidden object, and carry its reflection back to a sensor. The technique has nonetheless always dragged a stubborn flaw of its own — mmWave reflections are specular, so the energy departs in a single direction. Research assistant Laura Dodds offers an image for it: the sensor effectively “sees” only an object’s top surface, while the sides and bottom sit nearly blank. Try to rebuild a mug from nothing but its lid and the attempt fails, however clean the measurement happens to be. Earlier systems from Adib’s group read the reflections through physics alone, and that physics is where reconstruction accuracy hit its ceiling.
Pairing Specular Physics With Generative Completion
Wave-Former’s answer is to make physical sensing and a generative model work as one pipeline. The reflected signal is first turned into a set of candidate object surfaces; a generative AI model takes over and completes the overall shape; the surfaces are then revised pass after pass until the 3D reconstruction converges on a finished form. The outline comes from physical sensing; the generative model closes whatever gaps remain. Adib frames the contribution as “using AI to finally unlock wireless vision.” Generative shape completion has a history in computer vision — what is genuinely new is binding it to a wireless signal and keeping physical constraints in the loop across the entire pipeline.
When the Data Doesn’t Exist, Simulate It
Generative models cannot train without large volumes of data, yet no large-scale dataset of millimeter-wave scans of occluded objects exists anywhere. The team’s detour: start from established computer vision datasets and embed simulated specular reflections and noise, so the synthetic training data itself carries mmWave reflection physics baked in. Collecting real measurements instead would have taken years, as research assistant Maisy Lam points out. The maneuver merits attention in its own right: wherever sensor data runs scarce, pressing a physics model into service as a data amplifier is a fair route around the cold-start problem.
Test Results, RISE, and Where It Applies
In testing, cans, boxes, utensils, and fruit were among the roughly 70 everyday objects Wave-Former rebuilt, hidden behind or under cardboard, wood, drywall, plastic, and fabric, with accuracy nearly 20 percent above the strongest existing baselines. The team also released a companion system, RISE: a single static radar, joined by the reflections that human motion generates, carries the same idea to whole-room reconstruction at roughly twice the precision of current techniques.
The applications that arrive soonest are robotic: dependable grasping of objects outside the robot’s line of sight, and warehouse robots that check a sealed box against an order before shipment, trimming the waste that returns create. Both cases share the same economics — a camera can vouch only for what a human packer already saw, while a radio probe can audit a sealed box at any point along the logistics chain, and the verification step costs seconds instead of an unpack-repack cycle. Partial funding came from the National Science Foundation, the MIT Media Lab, and Amazon. The other side of the ledger deserves to be said plainly: a sensing capability that sees through walls carries built-in privacy questions, and the same properties that let a warehouse robot audit a sealed carton let an unattended device profile a room it was never invited into. The paper does not take up that question, but every real-world deployment will have to.