• bounding box
• grasp
• human-like
• neural network
• object pose
• primitives
• soft hand
• vision system

This paper is based on the implementation of a neural network using the Darknet/YOLO framework, with a vision system consisting of an RGBD Intel RealSense D435 camera that recognizes a generic object observed by the camera itself and associates it with a human-like grasp achievable by a robotic manipulator, to whose end-effector a soft robotic hand is attached. The chosen primitives were selected with reference to the Grasp Taxonomy by Feix et al., based on experiments conducted in the laboratory using the soft robotic hands (specifically IIT/Pisa Soft-Hand 2). The neural network, in addition to the most probable grasp primitive to be performed, also outputs a 3D bounding box of the object (calculated based on the depth information from the camera, the position of the object's centroid and the data provided by the initial calculation of the 2D bounding box) and reconstructs the 2D object pose using a PCA algorithm. CUDA is the system architecture used, while for the software implementation, libraries from OpenCV were utilized among others. The codes were written in C++ and Python. In the project, ROS nodes were employed to easily interact with the hardware used and with various network inputs and outputs. The goal of this work is the identification of human-like primitives for grasping objects of various sizes, shapes, and weights with soft robotic hands. The obtained results make this project applicable in various fields of automation, particularly useful for the grasping and manipulation of objects in industrial settings, and it can also be useful for the development of human-robot interaction for assisted handover grasp in collaborative environments.