A Regression Network for Bend Angle Prediction of a Soft Gripper via Ultrasonic and Bend Sensor Data

Soft grippers are used in a variety of applications that involve delicate objects. Depending on the shape of the object, each segment of an individual finger of a soft gripper undergoes different bending motions. In this article, we propose a data-driven approach to predict the appropriate bending angle for a soft gripper to effectively grasp objects of different shapes and orientations from their ultrasonic reflections. A classification-guided regression network is designed for the prediction task. On an experimental setup, prediction accuracies [for bend angle (BA)] of approximately 80% are obtained. The ultrasonic sensor-based setup with a neural network for BA prediction obviates the need for a sophisticated feedback mechanism for grasping.