A team at the Delft University of Technology has taught a palm-sized drone to find its way through pitch-black rooms the way a rodent does: by touch. Fitted with two artificial whiskers and running every calculation onboard, the aircraft steers along walls in complete darkness using navigation algorithms that occupy just 34 KB of memory. The study, “Whisker-based tactile flight for tiny drones,” appeared on 18 September 2026 in Nature Communications.
The authors — Chaoxiang Ye, Guido de Croon and Salua Hamaza of the Faculty of Aerospace Engineering — are reviving an idea as old as rats: when sight fails, touch takes over.
- Whiskers, not cameras: each sensor is a 200 mm nitinol thread, 0.4 mm thick, seated in a 3D-printed resin capsule that mimics a follicle.
- Ultra-light hardware: the sensing package weighs 3.2 g, while the Crazyflie Brushless test platform with both whiskers weighs 44.1 g.
- Blind navigation proven: the drone tracked a rigid wall at 40–100 mm for about seven seconds and found the 80 cm exit of a 2 x 2 m glass-walled room.
Pressure, not pixels: how the whiskers work
There is no camera in the loop. Each whisker is a 200 mm nitinol thread just 0.4 mm in diameter, mounted on a 3D-printed capsule filled with resin that acts as an artificial follicle. Inside sit three barometers per whisker, six channels in total. When a whisker bends against a surface, the pressure readings shift; a multilayer perceptron converts them into contact depth, and a Kalman filter smooths the estimate.
Two front whiskers are mounted 5 cm apart at a 45° angle. Other inclinations were tested, but only 45° held stable across every speed; the 15° setup accumulated the most collisions. The complete sensor assembly weighs 3.2 grams, and the Crazyflie Brushless testbed with both whiskers comes to 44.1 grams. Maximum forward speed is a cautious 20 cm/s — appropriate for something exploring by feel.
One widely repeated figure deserves context: the 34 KB refers to the memory consumed by the navigation algorithms, not the whole system. The microcontroller executing them carries 192 KB of RAM in total. That is still striking, but it is not the same as saying the drone runs on 34 KB.
Millimetre accuracy, transparent glass and a 2 x 2 metre room
Depth estimates land within a few millimetres of ground truth. On a whiteboard the mean absolute reconstruction error was 4.12 mm; on glass, 5.71 mm, per the preprint’s results table.
Glass is no small detail: transparent surfaces are a classic failure mode for cameras and lidar, and the authors acknowledge that their own reference laser sensor breaks down when the drone is not perpendicular to the panels. In a test with three transparent partitions, the whiskered drone reached its target on all five attempts per configuration.
In total darkness the team flew wall-following runs along a rigid surface and a cloth banner. On the rigid wall the drone held a 40–100 mm standoff for roughly seven seconds; the fabric caused oscillation, yet tracking still worked. The aircraft also explored a 2 x 2 metre room with glass walls and found its 80 cm exit — described in the preprint as a single real-world run. In simulation, four of five attempts found the way out.
The fine print: what remains unproven
This remains a laboratory demonstration, and the darkness flights lasted about seven seconds. The authors note that estimates drift as sweeps lengthen and that calibrating each whisker is laborious: barometer-to-barometer differences prevent a model trained on one sensor from transferring to another without recalibration.
Some coverage has overstated the scope, presenting the drone as ready for smoke, dust or battlefields. The preprint cites those scenarios only as motivation and does not test them. What is actually demonstrated is darkness, plus transparent or soft indoor surfaces.
“We wanted to demonstrate that touch doesn’t have to cost size or computing power,” Ye said.
Hamaza has framed the goal as giving drones a rich sense of touch for navigation, not for manipulating objects in mid-air. Cited potential applications include collapsed buildings after earthquakes, pipe and duct inspection, and even lunar craters and Martian lava tubes.
The real contribution is not the drone but the budget. Useful tactile perception from 3.2 grams of sensor and a few kilobytes of code means touch can be added to existing platforms rather than requiring a redesign. It is a cheap sense that complements expensive ones instead of replacing the camera.
What comes next is the difficult part: longer flights, environments packed with obstacles and, above all, trials in the smoke-and-dust conditions so often invoked as motivation. Until then, the drone that gropes through a collapsed building remains a promising result rather than a rescue tool.