In a cluttered environment in which objects are lying very closely to each other, the arranging motion is required before the robot attempts to grasp the target object. Thus, a robot must determine which motion to perform based on a given situation. This study presents an approach to learning a decision-making ability for the robot to grasp the target object after rearranging the surrounding objects obstructing the target object. The learning is performed in the virtual environment, and the image, which is an input of the deep Q-network, is preprocessed to directly apply the results of the learning to the real environment. That is, the difference between the two environments is minimized by making the states obtained from the virtual and real environments similar to each other. In addition, image preprocessing can be used to generalize the results of learning so that the robot can determine the appropriate actions to take when objects that were not used for learning are given. A hierarchical structure, which consists of high-level and low-level motion selectors, is used for the learning: the former determines the grasping or pushing actions while the latter determines how to perform such selected actions. The results of various experiments show that the proposed scheme is effective in grasping the target object in a cluttered environment without the need for any additional learning in the real world.

在杂乱无章的环境中，对象彼此非常靠近，在机器人尝试抓住目标对象之前需要进行排列运动。因此，机器人必须根据给定情况确定要执行的运动。这项研究提出了一种学习方法，让机器人在重新布置阻碍目标物的周围物体后，能够抓住目标物。学习是在虚拟环境中执行的，并且对作为深度Q网络输入的图像进行了预处理，以将学习结果直接应用于实际环境。即，通过使从虚拟环境和真实环境获得的状态彼此相似，来最小化两个环境之间的差异。此外，图像预处理可用于概括学习结果，以便机器人可以确定在给出未用于学习的对象时要采取的适当动作。用于学习的层次结构由高层和低层的运动选择器组成：前者确定抓握或推动动作，而后者确定如何执行所选动作。各种实验的结果表明，所提出的方案可以有效地在杂乱的环境中抓住目标物体，而无需在现实世界中进行任何额外的学习。用于学习：前者确定抓握或推动动作，而后者确定如何执行选定的动作。各种实验的结果表明，所提出的方案可以有效地在杂乱的环境中抓住目标物体，而无需在现实世界中进行任何额外的学习。用于学习：前者确定抓握或推动动作，而后者确定如何执行选定的动作。各种实验的结果表明，所提出的方案可以有效地在杂乱的环境中抓住目标物体，而无需在现实世界中进行任何额外的学习。