A new machine learning format supports the design of custom strain sensors to be integrated into soft machines. (Image: University of Maryland)

Soft machines — a subcategory of robotics that uses deformable materials instead of rigid links — are an emerging technology commonly used in wearable robotics and biomimetics (e.g., prosthetic limbs). Soft robots offer remarkable flexibility, outstanding adaptability, and evenly distributed force, providing safer human-machine interactions than conventional hard and stiff robots.

An essential component of soft machines is the high-precision strain sensor to monitor the strain changes of each soft body unit and achieve a high-precision control loop. First, the complex movements of soft machines require the strain sensors to monitor a wide strain range from <5% to >200%, which exceeds the capabilities of conventional strain sensors. Second, to monitor the coordinated motions of a soft machine, multiple strain sensors are required to satisfy different sensing tasks for separate robotic units, which demands tedious trial-and-error tests.

To address this problem, a University of Maryland (UMD) research team led by Po-Yen Chen, Professor of Chemical and Biomolecular Engineering, has created a machine learning (ML) framework to facilitate the construction of a prediction model, which can be utilized for two design tasks: (1) predict sensor performance based on a fabrication recipe and (2) recommend feasible fabrication recipes for adequate strain sensors.

“To use a food analogy, we gave a list of ingredients to a ‘chef,’ and that chef is able to design the perfect meal based on individual tastes of the customer,” said Chen.

This technology can be used in the fields of advanced manufacturing, underwater robot design, prosthesis design, and beyond.