Researchers from Northwestern University presented Phantom Twist at RSS 2026 in Sydney — a drone designed to be more than ten times harder to spot with the naked eye than a standard quadrotor. The trick is spinning the entire platform at 15–25 Hz, turning a solid structure into a persistence-of-vision blur that fades into the background.
Key takeaways
- Phantom Twist spins at 15–25 Hz, exploiting persistence of vision to blend into the background
- LPIPS visibility score: 0.0104 vs ~0.2 for a human-designed layout vs ~0.1+ for a same-size quadrotor
- Shape optimized by an iterative algorithm from a pool of ~20,000 feasible configurations
- Control achieved by pulsing a single motor's speed — no control surfaces
- Paper presented at RSS 2026, published as arXiv:2605.11296
How it works — some physics
Human vision needs roughly 100 ms to integrate an image before sending it to the brain. If an object moves fast enough within that window, the eye averages the motion — instead of a distinct shape, the observer sees a blurred smear.
Phantom Twist exploits exactly this. The platform spins at 15–25 Hz, and with the right structural geometry, the result is a drone that nearly disappears into its surroundings.
Spinning drones are not new in robotics — earlier designs like Picolissimo and the Samsara series inspired by maple seeds showed that a stable, controllable platform with a single motor is entirely feasible. What makes Phantom Twist different is that the structural layout was designed specifically for minimal visual detectability, not just for stability and maneuverability.
Computational optimization of invisibility
The team led by Michael Rubenstein at Northwestern University built a two-stage design pipeline. Input: a set of functional components (batteries, control PCB, motor-propeller assembly, counterweights) connected by 0.8 mm carbon fiber rods. Output: a component placement that minimizes LPIPS?LPIPS: A perceptual metric measuring how different two images appear to human vision. (Learned Perceptual Image Patch Similarity) — a metric based on how the human visual system perceives the difference between two images.
The lower the LPIPS score between the background and the background-plus-spinning-drone, the harder the drone is to spot. The optimized design scored 0.0104 — a human-designed layout without algorithmic support reached about 0.2, and a quadrotor?quadrotor: A four-rotor drone — the most common multirotor design. of the same size would be more than ten times more visible.
The algorithm started from a pool of roughly 20,000 feasible configurations (satisfying inertial and aerodynamic constraints) and iteratively converged toward minimum perceptibility.
| Design | LPIPS |
|---|---|
| Phantom Twist | 0.0104 |
| Human-designed layout | ~0.2 |
| Same-size quadrotor | ~0.1+ |
The core insight: components must not visually overlap each other during rotation from any observation angle. The more background visible through the spinning structure, the smaller the visible trace in the observer's field of view.
Control without control surfaces
Phantom Twist has one motor and no traditional control surfaces. How do you fly precisely with that? By pulsing motor speed at precisely timed moments during each rotation. Speeding up or slowing the motor at the right phase of rotation generates an asymmetric thrust vector that translates the drone horizontally. Altitude is controlled by adjusting overall thrust. The spinning motion itself provides passive stability — no active leveling needed as with a quadrotor.
Current test prototypes operate with an external optical tracking system for navigation data, confining Phantom Twist to controlled lab environments. Rubenstein points to earlier work by his team on similar single-motor platforms that flew successfully outdoors, and is optimistic about escaping the lab.
Why this matters?
Phantom Twist represents a rare approach to robotic hardware design: optimizing not just mechanical performance but how a human observer perceives the machine. Until now, drone engineers focused on performance, durability, and energy efficiency — treating visibility as an incidental side effect of size. Here, visibility becomes a measurable engineering objective to minimize, and the human visual system becomes part of the design specification.
Practical applications include wildlife observation (less disruptive to animals), environmental monitoring, and inspection in hard-to-reach areas — anywhere a drone needs to be present but as unobtrusive as possible. The same property naturally opens discussion about surveillance and military use cases, though the authors emphasize civilian and scientific applications.
There is also a broader methodological contribution: automatic shape design optimized for human perception. The same technique — minimizing LPIPS or similar perceptual metrics — could be applied to other moving platforms, packaging, or visual elements of robotic systems.
What's next?
- Rubenstein points to his team's earlier outdoor-capable single-motor drones as the direct path to getting Phantom Twist out of the lab
- Mounting a camera on the spinning platform is a noted next step — a 15–25 Hz spinning drone could capture a 360° panoramic view for onboard navigation
- Acoustic signature reduction was not a goal of this paper, but the authors reference existing approaches as potentially compatible directions
Sources
IEEE Spectrum — How to Make an Invisible Drone
arXiv — Computational Design of a Low-Visibility UAV Using a Human-Aligned Perceptual Metric
RSS 2026 — Robotics: Science and Systems
