Peng Gang, Liao Jinhu, Guan Shangbin, Yang Jin, Li Xinde
School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, 430074, China.
Key Laboratory of Image Processing and Intelligent Control, Ministry of Education, Huazhong University of Science and Technology, Wuhan, 430074, China.
Sci Rep. 2022 Mar 10;12(1):3927. doi: 10.1038/s41598-022-07900-2.
In the field of intelligent manufacturing, robot grasping and sorting is important content. However, there are some disadvantages in the traditional single-view-based manipulator grasping methods by using a 2D camera, where the efficiency and the accuracy of grasping are both low when facing the scene of stacking and occlusion for the reason that there is information missing by single-view 2D camera-based methods while acquiring scene information, and the methods of grasping only can't change the difficult-to-grasp scene which is stack and occluded. Regarding the issue above, a pushing-grasping collaborative method based on the deep Q-network in dual viewpoints is proposed in this paper. This method in this paper adopts an improved deep Q-network algorithm, with an RGB-D camera to obtain the information of objects' RGB images and point clouds from two viewpoints, which solved the problem of lack of information missing. What's more, it combines the pushing and grasping actions with the deep Q-network, which make it have the ability of active exploration, so that the trained manipulator can make the scenes less stacking and occlusion, and with the help of that, it can perform well in more complicated grasping scenes. In addition, we improved the reward function of the deep Q-network and propose the piecewise reward function to speed up the convergence of the deep Q-network. We trained different models and tried different methods in the V-REP simulation environment, and it drew a conclusion that the method proposed in this paper converges quickly and the success rate of grasping objects in unstructured scenes raises up to 83.5%. Besides, it shows the generalization ability and well performance when novel objects appear in the scenes that the manipulator has never grasped before.
在智能制造领域，机器人抓取与分拣是重要内容。然而，传统基于单视角的二维相机机械手抓取方法存在一些缺点，即面对堆叠和遮挡场景时，抓取效率和准确性都很低，原因是基于单视角二维相机的方法在获取场景信息时存在信息缺失，且仅有的抓取方法无法改变堆叠和遮挡这种难以抓取的场景。针对上述问题，本文提出了一种基于双视角深度Q网络的推抓协同方法。本文的方法采用改进的深度Q网络算法，利用RGB-D相机从两个视角获取物体的RGB图像和点云信息，解决了信息缺失问题。此外，它将推和抓的动作与深度Q网络相结合，使其具有主动探索能力，从而使训练后的机械手能够减少场景中的堆叠和遮挡情况，借助于此，它能在更复杂的抓取场景中表现良好。另外，我们改进了深度Q网络的奖励函数，提出了分段奖励函数以加速深度Q网络的收敛。我们在V-REP仿真环境中训练了不同模型并尝试了不同方法，得出本文提出的方法收敛速度快，在非结构化场景中抓取物体的成功率提高到83.5%的结论。此外，当新物体出现在机械手之前从未抓取过的场景中时，它展示了泛化能力和良好性能。
