An AI controller leans on cockroach instinct to cross rough ground

A research team at Osaka University has shown that a small machine-learning controller strapped to a cockroach can turn the insect into a self-navigating scout over rocky, unknown terrain, with the insect's own climbing and wall-following behavior doing the actual locomotion work.

The paper, published in Soft Robotics and announced by the university in February 2025, comes from a collaboration between Osaka University and Diponegoro University in Indonesia. Mochammad Ariyanto led the work; Keisuke Morishima is the senior author. The full text sits in the Osaka University institutional repository.

The setup is part of a field called biohybrid robotics, which pairs evolution-built bodies, in this case Madagascar-hissing cockroaches, with small electronic backpacks. The electronics carry a sensor stack and a controller that nudges the insect left or right; the cockroach's own nervous system handles balance, climbing, and recovery. As Ariyanto puts it: "By simply attaching electronic devices to insects, we can avoid the finer details of robotics engineering and focus on achieving our goals."

Prior cyborg-insect demonstrations split the navigation work the other way: insects walked flat or gently structured ground, while external cameras, beacons, or pre-mapped routes told them where to go. The Osaka team reports tests on sandy ground strewn with stones and pieces of wood, with the controller selecting left and right turn commands on the fly from onboard sensor readings. The cyborgs reached the target without external navigation aids and recovered when they stumbled.

"In our work, the insect's own autonomous behavior, including climbing, walking, and wall-following, manages recovery from falls," Morishima said. "Our autonomous biohybrid navigation system overcomes problems that have traditionally challenged robots, such as recovering from falling. This is what is needed for stepping outside the laboratory and into real-life scenarios like wilderness."

The paper is roughly 18 months old, and Mirage News's recent pickup restates the original university release rather than adding a new result. Follow-on work in 2024 and 2025, including insect-scale soft-robot and terrain-classification studies, leans on the same control architecture. The next test is whether that lineage moves from a single lab demo toward a deployable scout.

The press materials point to post-disaster site inspection, search-and-rescue support, pipe and collapsed-building exploration, and possible use in low-oxygen, deep-sea, or space settings. None of those have been demonstrated in the paper; they are proposed scenarios. A realistic near-term use case is the kind of small, rough, partially mapped interior a wheeled robot cannot roll through and a legged micro-robot cannot yet balance on for long, such as a partially collapsed building or a debris-choked crawlspace.

On the other side, insect-scale electronics are constrained by payload, battery, and onboard compute, and a cockroach cannot carry a manipulator or a heavier sensor such as a thermal camera or a gas sniffer. The paper is a navigation result, not a platform claim, and any move from sandy test plots to rubble piles, smoke, or flooding will need its own field validation.