کنفرانس مهندسی برق ایران

صفحه اصلی / سی و دومین کنفرانس بین المللی مهندسی برق

Mapping Human Grasping to 3-Finger Grippers: A Deep Learning Perspective

نویسندگان :

Fatemeh Naeinian¹ Elnaz Balazadeh² Mehdi Tale Masouleh³

1- University of Tehran 2- University of Tehran 3- University of Tehran

کلمات کلیدی :

Grasping،Grasp Detection،MediaPipe،Deep Learning،Object Detection

چکیده :

Grasping is a fundamental aspect of human interaction with the environment, and replicating this ability in robotic systems is challenging due to the complexity of hand-object interactions. This paper explores the mapping of human grasping behavior to a 3-finger robotic gripper using a deep learning approach. The study focuses on the development of a model that predicts grasping points based on RGB images of objects and finger coordinates during the approaching state. The research contributes a novel dataset of human-like grasping with 3 fingers, capturing various object orientations and scenarios. Finger coordinates are extracted in both the approaching and grasping states by using Mediapipe. The proposed model employs a pre-trained Faster R-CNN for object detection, coupled with fully connected networks for regression of grasping points. Evaluation metrics include Intersection over Union (IoU), grasp angle, and accuracy of predicted points for each finger. The trained network achieves compelling results, demonstrating 72% accuracy, 96.5% in grasp angle, and a 30% IoU in grasp rectangle representation. Results indicate the efficacy of the approach, offering a promising avenue for advancing robotic grasping capabilities.