Automated robotic assembly of shaft sleeve based on reinforcement learning

Shaft sleeve assembly is a common task in industrial manufacturing. The fitting approach for shaft sleeve assembly is usually interference fit, which requires significant contact forces. Conventional assembly methods, though focused on safety, often struggle to achieve high efficiency. Reinforcement learning can effectively select appropriate assembly actions through interaction with the environment, making it well-suited for shaft sleeve assembly tasks. Firstly, a comprehensive workflow for shaft sleeve assembly is formulated, including system initialization, insertion, push, and completion. Our research focuses mainly on the insertion process. Secondly, the core control algorithm adopts a deep reinforcement learning method based on the Actor-Critic architecture. The reward function includes safety reward, step length reward, and step reward. Safety reward ensures assembly security, while step length and step reward enhance assembly efficiency from different perspectives. Finally, real-world experiments on shaft sleeve assembly are conducted, including ablation experiments, parameter tuning experiments on reward function, and comparative experiments with conventional methods. The results of the ablation experiments and parameter tuning experiments indicate that after combining safety reward, step length reward, and step reward, the assembly effect achieves the best, verifying the effectiveness of the proposed reward function. Comparative experimental results demonstrate that our approach not only enhances safety compared to conventional methods but also significantly improves assembly efficiency, indicating the feasibility of this method.