Robotic Assembly of Shaft Sleeves in Different Sizes Based on Deep Reinforcement Learning

Shaft sleeve is a kind of usual component in industrial manufacturing, and its assembly is also a common task. The sizes of shaft sleeves are usually diverse due to the various application scenarios. Traditional methods often fail to balance efficiency and safety. A deep reinforcement learning method is proposed to obtain the most effective assembly strategy in the interaction with environment. Firstly, a four-stage method for assembling different sizes of shaft sleeves is developed. After the shaft sleeve to be assembled is identified via the monocular vision with a global camera, its position is computed. Then a manipulator is employed to grasp the shaft sleeve and align it to the corresponding hole with hole search algorithm. The shaft sleeve is ultimately inserted into the hole. Secondly, the insertion process adopts the deep reinforcement learning method based on Actor-Critic architecture, which feeds the size of shaft sleeve as an additional feature into the reinforcement learning network for training. Finally, ablation experiments and comparative experiments are conducted. The experimental results show that the proposed method has good performance in terms of safety and efficiency during the assembly for different sizes of shaft sleeves.