基于Deep Q Networks的机械臂推动和抓握协同控制

Collaborative control of mechanical arm pushing and grasping based on Deep Q Networks

针对目前机械臂在复杂场景应用不足以及推动和抓握自主协同控制研究不多的现状,发挥深度Q网络(Deep Q Networks)无规则、自主学习优势,提出了一种基于Deep Q Networks的机械臂推动和抓握协同控制方法。通过2个完全卷积网络将场景信息映射至推动或抓握动作,经过马尔可夫过程,采取目光长远奖励机制,选取最佳行为函数,实现对复杂场景机械臂推动和抓握动作的自主协同控制。在仿真和真实场景实验中,该方法在复杂场景中能够通过推动和抓握自主协同操控实现对物块的快速抓取,并获得更高的动作效率和抓取成功率。

In view of the insufficient application of mechanical arms in complex scenes and the lack of research on autonomous collaborative control of pushing and grasping,the advantages of Deep Q Networks′irregular autonomous learning have been brought into full play and a collaborative control method of pushing and grasping based on Deep Q Networks was proposed.Two fully convolutional networks were used to map the scene information to the pushing or grasping actions.After Markov process,the long-term vision reward mechanism was adopted to select the best behavior function to realize the autonomous collaborative control of the pushing and grasping actions of the mechanical arm in complex scenes.In the experiments of simulated and real scenes,the method presented can realize the rapid grasping of blocks in complex scenes through the autonomous collaborative control of pushing and grasping,and achieve higher movement efficiency and grasping success rate.