Integrating Vision Localization and Deep Reinforcement Learning for High-Precision, Low-Cost Peg-in-Hole Assembly

Combining vision and force assembly strategies constitutes a crucial component of industrial automation production. However, there are two challenges. First, the positioning error of existing vision systems often fails to meet the demands of precision assembly, necessitating the design of assembly strategies based on force sensors tailored to the characteristics of the assembly process. Second, the extensive data required to train vision positioning systems represents a significant burden. Therefore, we propose an assembly method based on object location with a relatively low deployment cost and fine-tuning with reinforcement learning. This method reduces the training cost with few training samples through data augmentation and the incorporation of attention modules. Furthermore, by integrating a deep reinforcement learning module, it performs high-precision hole searching and insertion operations learned by itself, mitigating the impact of visual positioning errors. The experimental results indicate that by improving the vision model, the required labeled samples for training can be reduced to 10. Furthermore, the method proposed in this paper achieves a faster average assembly time by 7.8 seconds and reduces the average number of assembly steps by 28 compared to traditional methods.