TU Delft fits whisker-based drone navigation into 34 kilobytes of onboard memory

This work goes after an ordinary failure. A micro-drone flies into a dark, dusty or smoke-filled space, its cameras stop giving it anything useful, and it crashes ^[6]. Carrying a LiDAR instead is not an option under 100 grams, and neither is the processor to run one ^[5].

So the drone learns about the world at the moment it touches it. Each whisker on the nose sits on three miniature pressure sensors that read the bend and turn it into contact depth and location in three dimensions ^[2]. There are two whiskers, so six pressure sensors in all ^[3]. Earlier touch sensing on flying robots used robotic arms or bumpers ^[4].

According to the write-up, the hard part was the drone's own air. Propeller airflow throws noise and distortion into the readings and makes them drift, and the Delft pipeline runs onboard in 34 kilobytes of memory, continuously subtracting the drone's own turbulence to leave depth data at millimeter precision ^[7][8]. "We wanted to show that touch does not have to come at the cost of size or computational power," said Chaoxiang Ye of TU Delft ^[9].

In the tests the drone flew pitch-black enclosures, mapped unknown room layouts, followed surface contours and found exits with no visual cue ^[10]. The write-up does not report the drone's mass or airspeed, and it carries no measured comparison against a camera in the same space ^[17]. The study is in Nature Communications ^[13].

Salua Hamaza, associate professor of aerial physical interaction and embodied intelligence in aerial robots at TU Delft, describes the goal as "a novel concept of tactile navigation: using touch to explore and fly through the unknown" ^[12], and distinguishes it from manipulation in the air. "Inspired by nature, we found the answer in whiskers," Hamaza said ^[11].

Whether any of this belongs in a plan depends on the space. Can a camera see in it, once you account for dust, smoke and light. Can the vehicle touch what it is inspecting without hurting it or itself. Whiskers only help where the first answer is no and the second is yes, because the drone learns about a surface by reaching it and the map fills in one contact at a time. In a lit shaft with room to hover, I would still fly a camera.

The write-up names uses that sit in exactly that corner: narrow pipelines, ventilation shafts, sewers and underground mines ^[15], collapsed smoke-choked buildings where visual and thermal cameras become useless ^[14], and dark lunar craters and Martian lava tubes ^[16]. Those industrial inspection uses are stated as future possibilities ^[15].

Anyone sizing the memory and the nose clearance on the next sub-100 gram airframe can act on this. For a buyer comparing inspection drones this quarter, what exists is a Nature Communications paper and a set of lab flights ^[13][10].