The aviation manufacturing industry requires extremely high assembly accuracy of parts, and there are many shortcomings in traditional assembly methods. The development of artificial intelligence (AI) technology provides a new way to solve this problem. In the aspect of path planning, aiming at the limitations of traditional algorithms in dynamic environment and complex scenes, a hybrid architecture combining Deep Reinforcement Learning (DRL) and RRT* is proposed. This architecture achieves dynamic obstacle avoidance and path optimization under multiple constraints by extracting obstacle features through the 3D convolutional network in the perception layer, making decisions using the Proximal Policy Optimization (PPO) algorithm in the decision layer, and further optimizing the path in the optimization layer. Experimental results show that compared with the traditional RRT* algorithm, the proposed method has significantly improved in path length, planning time, success rate in dynamic obstacle environment and joint torque fluctuation. In the aspect of error compensation, a multi-source error model is established, including thermal deformation error, part processing error, visual positioning error and dynamic error caused by assembly stress. The spatio-temporal feature fusion model is combined with LSTM network and graph convolution network (GCN) to predict the error, and then a real-time error compensation strategy based on reinforcement learning is designed. Experiments verify the superior performance of this compensation strategy in reducing all kinds of error sources. Compared with non-compensation and traditional PID compensation methods, it significantly reduces the errors caused by thermal deformation, visual positioning and dynamic stress, and the comprehensive error is far below the requirements of ISO standard, showing good real-time and high-precision characteristics.