Multi-Objects Robotic Grasping Optimization Employing a 2D camera

Abstract

Artificial intelligence algorithms can be exploited to enhance identification, localization, and grasping performance in robotics applications, employing low-cost vision systems (such as 2D cameras). The aim of this paper is, indeed, to optimize the camera pose to improve object detection tasks considering a mul- tiple objects scenario. Therefore, transfer learning capabilities are required to minimize the experimental effort in subsequent grasps. Bayesian optimization (BO) with transfer learning (TL) capabilities has been proposed to address the mentioned scenario. A grasping task of multiple parts has been considered, being executed by an ABB Yumi single-arm manipulator IRB 14050 with a 2D Cognex AE3 In-Sight camera mounted at its end- effector. The proposed BO+ TL methodology has been compared with BO (without TL). The achieved results show that BO+TL is more efficient than BO exploiting the already available data.