Digital Twin Enhanced Assembly Based on Deep Reinforcement Learning

Discrete manufacturing is becoming a popular modality, which places a higher demand on the flexibility of the production line. Traditional assembly lines require extensive manual design and cannot meet the need for flexibility. Due to the rise of reinforcement learning, we suspect that modern algorithms are crucial to further improve the flexibility of assembly. In this paper, we propose a digital twin enhanced assembly method with deep reinforcement learning. A digital twin model of the assembly line is first built. Then, the deep deterministic policy gradient based reinforcement learning agent is trained on the digital twin model. The simulation of the reinforcement learning environment is based on a mixture of simulation engine and real signals. Thus, we can balance the training efficiency and the simulation accuracy. Finally, to validate our proposed method, peg-in-hole assembly experiments were conducted and good results were observed.