Tohoku University Recreates Knee-Jerk Reflex Using Mechanics Alone, Achieves Bipedal Walking

A research group at Tohoku University developed a reflex module for musculoskeletal robots that automatically contracts an artificial muscle when it is stretched beyond a threshold. It generates movement through mechanical action alone, without relying on electronic sensors or circuits. Using the module, the group reproduced the knee-jerk reflex and achieved bipedal walking. The paper was published in Advanced Intelligent Systems on Oct. 22, 2026 (Japan time), and TechXplore reported the findings in an article dated Nov. 4, 2026.

According to Tohoku University, the design uses mechanical structures to handle the sensing and processing required for a reflex, minimizing the need for electronic sensors and processing. By demonstrating both a knee-jerk reflex and bipedal walking, the work provides a concrete example of incorporating mechanical reflexes into movement generation in musculoskeletal robots. However, no reduction in power consumption was reported.

The technical focus is not replacing all electronic control for walking, but assigning part of the reflex process—detecting muscle stretch and triggering contraction—to a mechanical structure. Tohoku University said the mechanism contracts an artificial muscle when it is stretched beyond a threshold, reproducing the knee-jerk reflex and enabling bipedal walking. The work can therefore be seen as an example of a design that incorporates the body’s own mechanics into movement generation, as an alternative to control that depends on complex electronic sensors and circuits. However, the findings do not show that electronic control in general is unnecessary. Our report, Tohoku University and partners develop six-legged robot that learns stick insect walking with AI, covered a six-legged robot trained by AI on stick insect walking cycles. That was a separate study, and the difference in approaches to generating walking is notable. The earlier work used AI to learn biological walking cycles, while this study assigns reflex sensing and processing to a mechanical structure. When incorporating biological movement into robots, learning-based control and mechanical responses built into the body are distinct design options. With this approach, a mechanical structure replaces some of the functions involved in acquiring and processing sensor data with electronic circuits. This raises questions about how it compares with designs that depend on sensor placement and control computations. Tohoku University said it hopes the approach will offer design guidelines for reducing computational load and power consumption, but its announcement provided no quantitative reduction figures. How mechanical implementation affects not only individual reflexes but also walking stability and responses to disturbances will also be important to assess. The next points to verify are the duration and stability of bipedal walking, walking conditions, the reflex module’s durability, and which functions remain under electronic control. The announcement provided no measured figures for power consumption or computational load, nor evidence of reproducibility across different robots or environments. Experiments clarifying how functions are divided between mechanical structures and electronic control will help determine the scope of the design approach.

The module mechanically feeds back the artificial muscle’s length and produces a contraction when the muscle is stretched beyond a threshold. Tohoku University said that assigning the sensing and processing required for this reflex to a mechanical structure minimizes the need for electronic sensors and processing. The research group reproduced the knee-jerk reflex in a musculoskeletal robot—the response in which the leg kicks when the area below the kneecap is tapped. The paper is titled “Purely Mechanical Muscle Length Feedback Enables Embodied Reflexive Control in Musculoskeletal Robots.” The authors are Shunsuke Koseki, Mitsuhiro Hayashibe and Dai Owaki.