AI Now Controls Squishy Robots With Pinpoint Accuracy

Most robots are hard, heavy, and dangerous around people. Soft pneumatic actuators are different: they're squishy, flexible tubes that move when filled with air. They're naturally safe to touch, but they've been very hard to control because their behavior is bendy, stretchy, and unpredictable. This new research tackles that problem with a clever machine learning approach.

The key idea is to split the robot's movement into two parts: a simple "resting shape" model and a "learning" model that captures the wobbly, dynamic motion. The learning model, built using a data-driven method called EDMDc, lets the robot predict and correct its own movements in real time. The result: it can track a desired path with error of just one millimeter at low speeds, and less than ten millimeters even when moving ten times faster.

The researchers also showed something even more practical: the robot could follow wild, unpredictable hand-drawn paths, including sharp twists and turns with accelerations over 25 meters per second squared, all while dodging obstacles in its way. That's like a robot arm that can catch a tossed object or smoothly navigate around a person without stopping.

Finally, they built a way to analyze whether the control system will stay stable or wobble out of control. They proved it can accurately predict stable, borderline, and unstable behaviors. This matters because you can't deploy a robot in a hospital or workshop if you don't know when it might suddenly jerk or fail. Together, these advances move soft robots from lab curiosity to a real option for safe, precise automation.