We're building a sensor that tells rock from rubble, in the dark, without a lens.
Umbra reads the material and shape of objects using radio-frequency echoes instead of light — for the moments when cameras and lidar go blind: total darkness, smoke, dust, and buried ground.
Talk to usThe thesis
A bat flies through a cluttered cave, in total darkness, and still tells a moth from a leaf. It does this with sound, not light — sending out a call and reading the echo for shape, texture, and motion, even with a wall of clutter bouncing signal back at it. No camera has ever come close to that in the dark.
Umbra is built on the same idea, using radio waves instead of sound. The bet is that object recognition doesn't have to mean vision. It can mean reading an echo well enough to know not just that something is there, but what it's made of.
The problem
A camera needs light. Lidar needs clean air. Neither survives a collapsed building, a dust-choked mine shaft, or a smoke-filled room — exactly where finding the right object, or the right person, matters most.
Existing radio-frequency sensors can tell you that something is there. They're much worse at telling you what it is. We're building the piece that closes that gap.
How it works
Emit
A low-power RF pulse goes out into the space — through dust, smoke, or darkness — and bounces off everything it touches.
Reject clutter
A neural net trained on real cluttered environments separates the return you care about from the noise of rubble, walls, and debris.
Classify
The cleaned signal is matched against material and texture signatures — metal, flesh, stone, void — to identify what's actually there.
The neural network
Every RF pulse comes back as a mess: the object you care about, buried in echoes off walls, dust, and debris. A hand-built formula can't separate the two — there are too many environments and too many kinds of clutter.
So instead of writing rules, we train a network on real cluttered returns until it learns the difference on its own: first to pull a clean signal out of the noise, then to match that signal's shape and texture against known materials. It's the same principle as image recognition, just applied to echoes instead of pixels — the model never sees a picture, only the pattern of a reflection.
Built for
Search & rescue
Locate survivors under rubble by distinguishing a person from concrete, wood, and debris — before anyone has to dig blind.
Mining
Tell ore from rock and detect unstable voids in shafts where dust and darkness make cameras useless.
Defense
Sense what's behind a wall or under disturbed ground, in conditions where visual and thermal sensors are already blind.