Natural object manipulation using anthropomorphic robotic hand through deep reinforcement learning and deep grasping pro
- PDF / 2,433,650 Bytes
- 15 Pages / 595.276 x 790.866 pts Page_size
- 70 Downloads / 245 Views
Natural object manipulation using anthropomorphic robotic hand through deep reinforcement learning and deep grasping probability network Edwin Valarezo Añazco 1 & Patricio Rivera Lopez 1 & Nahyeon Park 1 & Jiheon Oh 1 & Gahyeon Ryu 1 & Mugahed A. Al-antari 1 & Tae-Seong Kim 1
# Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Human hands can perform complex manipulation of various objects. It is beneficial if anthropomorphic robotic hands can manipulate objects like human hands. However, it is still a challenge due to the high dimensionality and a lack of machine intelligence. In this work, we propose a novel framework based on Deep Reinforcement Learning (DRL) with Deep Grasping Probability Network (DGPN) to grasp and relocate various objects with an anthropomorphic robotic hand much like a human hand. DGPN is used to predict the probability of successful human-like natural grasping based on the priors of human grasping hand poses and object touch areas. Thus, our DRL with DGPN rewards natural grasping hand poses according to object geometry for successful human-like manipulation of objects. The proposed DRL with DGPN is evaluated by grasping and relocating five objects including apple, light bulb, cup, bottle, and can. The performance of our DRL with DGPN is compared with the standard DRL without DGPN. The results show that the standard DRL only achieves an average success rate of 22.60%, whereas our DRL with DGPN achieves 89.40% for the grasping and relocation tasks of the objects. Keywords Anthropomorphic robotic hand . Natural object grasping and relocation . Deep reinforcement learning . Human grasping hand poses . Deep grasping probability network . Natural policy gradient
1 Introduction Human hands are dexterous in manipulating variously shaped objects [1, 2]. Anthropomorphic robotic hands are designed to perform object manipulation like human hands [3–5]. However, the autonomous human-like manipulation of objects is still a challenging problem for robotic hands. This is because the subspace of natural human-like object grasping hand poses is considerably smaller than the entire space of potential object grasping hand poses produced with the high Degrees of Freedom (DoF) [6–8]. Since the natural object grasping hand poses accommodate the object geometry, research efforts must focus on training strategies that perform an efficient exploration of the object grasping space considering the object geometry to learn object manipulation like human hands. * Tae-Seong Kim [email protected] 1
Department of Biomedical Engineering, College of Electronics and Information, Kyung Hee University, Yongin-si, Republic of Korea
Recently, Deep Reinforcement Learning (DRL) is actively adapted to teach a robotic hand for natural object manipulation like human hands. For instance, DRL has been used to relocate objects using two-fingered and three-fingered robotic hands [9–12]. However, only standard DRL might not be sufficient to teach an anthropomorphic robotic hand humanlike natural grasping of obj
Data Loading...
