AI Plus Radar Lets Robots See Through Smoke and Fog

Cameras are the eyes of most robots and cars, but they go blind in exactly the moments that matter most: a burning building, a foggy highway, a dark tunnel. Radar — the same technology your car uses for parking sensors — keeps working through smoke and darkness because it uses radio waves, not light. The catch is that radar alone gives a blurry, low-detail picture. So the research team paired the two.

Their system, called GRADE, works in two steps. First it reads radar signals and produces a rough map of how far away everything is. Then a generative AI model — the same kind of technology behind image generators like DALL-E — fills in the missing structure, like an artist sketching details onto a faint outline. When a working camera is available, extra visual hints are blended in. As smoke thickens, the system quietly leans more on radar.

The team tested GRADE on roughly 95,000 frames collected inside 12 real buildings, including scenes filled with genuine smoke rather than computer-generated haze. It estimated distance to within about 30 centimetres — roughly one foot — in clear conditions, and stayed almost as accurate at about 31 centimetres under smoke. That consistency is the headline: performance barely drops when visibility disappears. The work will appear at ACM MobiCom 2026, and the code and dataset are publicly available.

Why care now? Drones inspecting burning buildings, robots in mines, warehouse machines in dusty aisles, and eventually self-driving cars could all navigate where today's sensors fail. Search-and-rescue teams could get 3D maps of smoke-filled rooms before sending people inside. There are limits: this is a research prototype, not a product you can buy, radar hardware adds cost, and the tests covered single frames indoors — not fast-moving outdoor traffic. Expect this in specialised robots long before it reaches your car.