On March 19, MIT News reported on Wave-Former, a system from a Media Lab research team that uses millimeter-wave wireless signals — the same band Wi-Fi occupies — to penetrate cardboard, wood, and drywall, then hands the measurements to a generative AI model that completes the shape. The result: full 3D reconstructions of objects that are completely hidden from view. The project is led by Fadel Adib, associate professor in MIT’s EECS department and director of the Signal Kinetics group at the Media Lab, and the paper, “Wave-Former: Through-Occlusion 3D Reconstruction via Wireless Shape Completion,” has been accepted to CVPR 2026. Across roughly 70 everyday objects, reconstruction accuracy improved by nearly 20 percent over state-of-the-art baselines.
For robotics and warehouse automation, this matters: the old constraint — if the sensor can’t see it, the arm can’t grasp it — now has a systematic workaround, and one that runs on radio waves ordinary hardware can already emit.
The Problem Wave-Former Solves
Millimeter waves pass through drywall, plastic, and cardboard and reflect off hidden objects back to a sensor. That is the physical basis of “wireless vision.” But the approach has always hit a wall of its own: mmWave reflections are specular, meaning energy bounces away in a single direction. As research assistant Laura Dodds puts it, the sensor effectively “sees” only the top surface of an object, while the sides and bottom stay blank. A system that tried to reconstruct, say, a mug from its lid alone would fail no matter how clean the measurement was. Earlier methods from Adib’s group interpreted reflections through physics alone, and that physics capped how accurate any reconstruction could get.
Specular Reflections and Generative Shape Completion
Wave-Former’s move is to wire physics and generative modeling into a single pipeline. The system first proposes a set of potential object surfaces from the reflected signal, feeds those to a generative AI model that completes the overall shape, then iteratively refines the surfaces until a full 3D reconstruction converges. Sensing supplies the skeleton; the generative model fills in what the physics cannot. Adib frames the contribution as “using AI to finally unlock wireless vision.” Generative shape completion is not new in computer vision — the novelty is attaching it to a wireless signal and keeping physical constraints in the loop across the entire pipeline.
No Dataset? Simulate One
Generative models need training data, and no large-scale dataset of millimeter-wave scans of occluded objects exists. The team’s workaround: take established computer vision datasets and embed simulated specular reflections and noise into them, effectively baking mmWave reflection physics into synthetic training data. Research assistant Maisy Lam notes that collecting real data instead would have taken years. The technique deserves attention on its own: when sensor data is scarce, using a physics model as a data amplifier is a legitimate way around the cold-start problem.
Results, the Companion System, and Applications
In testing, Wave-Former reconstructed around 70 everyday objects — cans, boxes, utensils, fruit — hidden behind or under cardboard, wood, drywall, plastic, and fabric, improving accuracy by nearly 20 percent over the best existing baselines. The team also published a companion system called RISE, which uses a single static radar plus the reflections produced by human motion to extend the same idea to whole-room reconstruction, roughly doubling precision versus existing techniques.
The most direct applications are robotic: reliably grasping objects outside the robot’s line of sight, and warehouse robots that verify the contents of sealed boxes against an order before shipment to cut return waste. Both cases share the same economics — a camera can only confirm what a human packer already saw, while a radio probe can audit a sealed box at arbitrary points in the logistics chain, and the verification step costs seconds rather than an unpack-repack cycle. The research was partially funded by the National Science Foundation, the MIT Media Lab, and Amazon. The other side of the ledger deserves saying out loud: 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 engage with that question, but no real-world deployment will be able to avoid it.
Sources
- Generative AI improves a wireless vision system that sees through obstructions — MIT News
- Wave-Former: Through-Occlusion 3D Reconstruction via Wireless Shape Completion — arXiv
- Generative AI improves vision system for robots to see through walls — Tomorrow’s World Today
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.