AI-Powered Terrain Recognition Helps Cyborg Cockroaches Navigate Faster
On 21 August 2026 a University of Osaka and Universitas Diponegoro team published a navigation system in Device that lets a living cockroach read the ground it is walking on. A multilayer perceptron running on the insect’s own 2.3 gram backpack sorts terrain into flat, ascent, descent and hole at 92% offline accuracy, and the controller uses that call to decide whether to steer or to stay out of the way. The speed gain comes from suppressing commands during climbs rather than issuing better ones. This piece covers the hardware budget, the four terrain classes, the 20-insect swarm result, the companion internal-state paper, the welfare questions, and what a four-class model on a fractional-gram power budget teaches ordinary AI projects.
