In peg-in-hole assembly using imitation learning, traditional reinforcement learning often struggles with low object positioning accuracy and inefficient exploration due to sparse rewards. To overcome these limitations, this study presents an imitation learning framework enhanced with an attention mechanism. A self-attention module builds 3D state representations and dynamically highlights object geometric features, boosting positioning accuracy. The model incorporates soft Q-learning to optimize the loss function, improving policy learning stability and efficiency. Experiments on RLbench tasks show the proposed method increases key state feature weights by 2.1–4.7 times, enhancing positioning accuracy. It also achieves a 5.3% faster convergence and a 20% higher success rate. In real-world scenarios, the model trains effective policies within minutes. This work offers a solution for peg-in-hole assembly by uniting precise positioning with efficient learning.