Cyborg cockroaches recognize terrain in a research system

A cockroach reaches a wall, and the machine guiding it has a choice: steer around or let it climb. Researchers from the University of Osaka and Universitas Diponegoro have built a navigation system that recognizes the terrain in real time and adapts control so cyborg cockroaches can use their natural climbing ability. The study appears in Device.

The problem was not simply getting an insect to move. Conventional systems tend to avoid obstacles, even when climbing would be possible. A reactive-climbing controller combined goal-seeking, obstacle avoidance, wall-following and innate climbing behavior, but it could not tell what surface the insect was on. During a climb, it kept sending steering commands, producing hesitation and inefficient movement.

The new layer uses a multilayer perceptron, or MLP — a machine-learning model — fed by onboard sensor data. It classifies flat ground, uphill, downhill and holes in real time. The controller then adjusts stimulation to the recognized terrain, reducing unnecessary steering and supporting forward movement across challenging surfaces.

The classifier reached 92% accuracy in offline evaluation. Professor Keisuke Morishima of the University of Osaka said the central challenge was recognizing terrain without compromising the insect's natural locomotion. The researchers describe the approach as “biohybrid physical AI,” combining a living body's movement with miniature electronics.

So what changes in practice? A future cyborg insect could take a more direct route through a search-and-rescue site or an area needing infrastructure inspection, rather than automatically going around every obstacle. The work is still at the research stage: the report does not describe field deployment, and the reported accuracy comes from offline evaluation rather than a field deployment.