Toward Anthropomorphic Grasping in Food Industries: A Dual-Arm Mobile Robot With Human-Like Reaching Function for Adaptive Grasping

Abstract

Performing unstructured grasping tasks in cluttered or obstacle-rich food processing environments is a key challenge in robotic systems. This work presents a task-adaptive grasping approach for a dual-arm anthropomorphic robot, named Herdsman, developed for the food industry. With an articulated torso, Herdsman is able to perform human-like reaching motions for more flexible grasping operations. To recognize the target object and extract the features for grasping, a vision pipeline, including a lightweight network GDC-YOLO for real-time object detection and a U-ReSENet network for grasping detection enhancement, is designed based on convolutional neural networks. After the detection comes the grasp execution, where a task-adaptive grasping strategy that works with the articulated torso is put forward to carry out grasping tasks in unstructured environments. Comparative experiments are designed to evaluate the detection performance between the proposed network and other popular networks for object detection and grasping detection. In addition, the task-adaptive grasp strategy for the Herdsman robot is experimentally validated by grasping the objects at different heights. The results have shown that the task-adaptive grasping solution exhibits robustness against variations in the target object position, which could be a promising approach for its application in unstructured environments requiring autonomous grasping.