This Drone Is Practically Invisible to the Human Eye. AI Designed It That Way.
Roboticists at Northwestern University used a computer optimiser to build a spinning drone that blurs itself into the background. The result is ten times harder to spot than a normal quadrotor.
Key points
- Northwestern University researchers presented the Phantom Twist drone at the RSS 2026 conference in Sydney this week.
- The drone spins at 15 to 25 times per second, fast enough to exploit a quirk of human vision and become nearly transparent in flight.
- An AI optimiser tested roughly 20,000 possible layouts before settling on a design with a visibility score of 0.0104, compared to over 0.1 for a conventional drone of the same size.
- The lead researcher, Michael Rubenstein, says the most exciting near-term use is watching wildlife without disturbing animals.
- The drone currently needs an indoor tracking system to fly, but the team is working on taking it outside.
Most drones announce themselves loudly. They buzz, they hover in unnaturally straight lines, and your eye goes straight to them. Phantom Twist wants to be the drone you never notice.
Built by roboticists at Northwestern University in Evanston, Illinois, Phantom Twist spins its entire body at between 15 and 25 times per second during flight. That speed exploits something called persistence of vision: the roughly 100 milliseconds your eye needs to process a scene before sending it to your brain. Move something fast enough inside that window and your eye averages the motion into a transparent smear against whatever is behind it. The drone effectively disappears.
The spinning trick is not new on its own. What makes Phantom Twist different is that its physical shape was designed by a computer optimiser with one explicit goal: be as hard to see as possible.
As first reported by IEEE Spectrum AI, the team started with around 20,000 physically workable arrangements of the drone's parts: a small motor, a propeller, a pair of batteries, a microcontroller, carbon-fibre rods, and some counterweights. A human engineer could not hold all those trade-offs in their head at once. The software could.
The optimiser judged each design using a metric called LPIPS (Learned Perceptual Image Patch Similarity), a measurement that scores how different a background image looks once the spinning drone is layered on top. A score close to zero means the drone is nearly invisible. The winning layout scored 0.0104. A hand-designed version of the same drone scored around 0.2. A standard quadrotor of the same size scored above 0.1, making it more than ten times more visible.
The key insight the algorithm found: keep parts spread apart so they never overlap each other as the drone rotates, and keep as much material as possible away from the central axis where overlap is hardest to avoid.
Could this drone be used to spy on people?
Honestly, yes, covert surveillance is the obvious concern. A near-silent, nearly invisible microdrone would be a serious privacy problem. The sound issue is not yet solved, and the drone currently requires an indoor optical tracking system (a network of cameras that follow reflective tags) to navigate, so it is not ready for uncontrolled outdoor use. The researchers are working on both problems.
Lead researcher Michael Rubenstein is open about the possibilities but says his personal interest is in wildlife observation, where a less intrusive drone could watch animals without altering their behaviour. He also notes that mounting a small camera on the spinning body could give the drone a full 360-degree view of its surroundings, useful for onboard navigation.
The full paper, "Computational Design of a Low-Visibility UAV Using a Human-Aligned Perceptual Metric", is available from the RSS 2026 proceedings for anyone who wants the technical detail.
