DEVELOPMENT AND DEPLOYMENT OF A SOFT ROBOTIC GRIPPER: HARNESSING PRESSURE FOR OBJECT RECOGNITION THROUGH DEEP LEARNING

Abstract

Soft robotic manipulators are uniquely suited to handle objects with diverse textures, shapes, and stiffnesses, thanks to their highly deformable structure; however, this same characteristic complicates the collection and interpretation of data for exteroceptive sensing. This challenge is often overcome by using machine learning on a wide array of strain sensors, but this work introduces a methodology that leverages the time-series pressure response of a novel, buckling-enabled soft robotic gripper to enable passive sensing capabilities. The behaviors of this pneumatically controlled, thin-shelled gripper are investigated with numerical modeling and validated experimentally. We detail a robust manufacturing strategy designed to minimize defects and demonstrate several representative use cases of the gripper. To develop built-in sensing capabilities, a dataset of the time series pressure responses during a variety of grasping events is constructed and analyzed. The dataset is initially explored for qualitative insights into exteroceptive sensing. It is then used to train two machine learning classifiers to distinguish between the geometry and size of a given object at over 85% accuracy across multiple classes. Finally, the ability to recalibrate the model on a new gripper using transfer learning is demonstrated via two grippers with intentionally produced dimensional modifications. By eliminating the need for embedded electronics or structural changes, this strategy enables rapid, low-cost haptic feedback from a single interaction, with broad implications for constrained environments such as surgery and industrial automation.