COMPUTER VISION-BASED GRASP DETECTION FOR A METAMATERIAL SOFT GRIPPER IN ROBOTIC VEGETABLES HARVESTING
DETECTAREA PRINDERII UNUI GRIPPER SOFT DIN METAMATERIALE UTILIZAT LA RECOLTAREA ROBOTIZATĂ A LEGUMELOR, FOLOSIND VEDERE ARTIFICIALĂ
DOI : https://doi.org/10.35633/inmateh-77-71
Authors
Abstract
This paper presents a computer vision-based methodology for evaluating the grasping performance of a soft robotic gripper fabricated from mechanical metamaterials, designed specifically for fruit and vegetables harvesting applications. Due to the fragile nature of fruits such as tomatoes or strawberries, the ability to assess and control the deformation of the gripper during interaction is critical to avoid damage while ensuring a secure grasp. A deep learning approach is proposed, leveraging convolutional neural networks (CNNs) to classify grasp outcomes from visual input. The model is trained on a custom dataset of images captured during robotic harvesting trials and optimized to detect subtle variations in gripper shape and fruit contact. The integration of soft metamaterial-based grippers with computer vision algorithms enables a robust, non-invasive grasp assessment pipeline, contributing toward fully autonomous and adaptive fruit-picking robots. The proposed method achieved an accuracy of 94.0% for correct grasps, 91.5% for failed grasps, and 95.9% for no-object cases, with an average inference time of 87 ms (ranging from 75 to 98 ms).
Abstract in Romanian



