Advanced Grasping Sensor Technologies for Autonomous Robotic Apple Harvesting Using Tactile Data and CNNs

This research investigates how to achieve an optimal grasp of an apple using a four-finger soft robotic grasper equipped with force-resistive sensors. Specifically, we sought to determine whether a convolutional neural network (CNN) could accurately classify the grasper's state and recommend adjustments ("in," "out," or "good" grasp) based on tactile data from the sensors. Spatiotemporal tactile images were developed from the sensors and fed into our CNN, achieving near 100% accuracy on unseen test data. This work suggests that CNN-based processing of tactile images can be a powerful tool for real-time control of soft robotic grippers.