AINEWS
A 30 gram Northwestern prototype spins at 25 rev/sec so the human eye merges it with the background when its colour matches the scene. The geometry came from a three stage AI design loop, not a human engineer.
A 30-gram drone the size of a hand hides not by absorbing radar but by spinning at 25 revolutions a second until the human visual system merges every moving part with the background, but only when its colour matches the scene. The mechanism is the point: Phantom Twist, built by Emma Alexander's group at Northwestern University, is a demonstration that stealth can be a perception problem solved by computation, not an exotic-materials problem solved by chemistry.
Every component, propellers, frame, body, whirls in circles. Nothing is fixed and recognizable to a stationary observer; motion blur blends the parts into a smeared disk roughly the colour of whatever sits behind them. David Whitaker at Cardiff University, who studies visual perception, said the human visual system merges fast-moving objects with their background above roughly 60 hertz, but the drone runs fast enough to be missed when its colour matches the scene. The motion-blur effect is established vision science; the novelty here is the geometry that exploits it.
What selects the geometry is the more interesting story. Alexander's team used a three-stage automated design pipeline rather than hand-tuning. The first stage generated millions of candidate geometries. A second model filtered that set down to roughly 20,000 designs that could theoretically fly, then adjusted component placement to minimize average visibility from every viewing angle. A third model, trained to mimic human sight, scored the shortlist against varied backgrounds. The researchers, not the algorithm alone, selected and built the final design. The AI is a search tool for a design space too large for a human engineer, not an autonomous inventor.
The result weighs 30 grams and fits in a palm. The source is New Scientist's write-up of the work; the underlying primary paper, venue, and authors beyond Alexander are not yet pinned in the current source basis, so any "breakthrough" framing should be hedged until the paper itself surfaces. What the source does support is a concrete instance of a perception-first stealth pipeline producing a flying object.
Conventional stealth approaches rely on radar-absorbent materials, camouflage netting, or simply not being where the observer is looking. Phantom Twist collapses those approaches into one loop: design, simulate against many backgrounds, pick the geometry that disappears best. The same pipeline could in principle optimise other perception-first objects, from sensor mounts to ground robots, against any imaging model the researchers can simulate.
Peter Lee at the University of Portsmouth, who was not involved in the research, flagged the obvious military relevance of perception-first stealth. The source excerpt cuts off on that point and does not document flight endurance, payload, battery life, or control scheme, so claims about operational use would not be supported on the current source basis. What is supported is the design idea: a cheap, additive-manufacturable object whose invisibility is a property of geometry and motion, not of materials.
A 30-gram prototype is not a weapon system, but it is a proof that an AI loop can search a design space and pick a shape that exploits the human visual system rather than physics. Whitaker's caveat is the right read: missed, not invisible, and only when colour matches. The work shows how "missed" becomes a design target instead of an accident.