The motion planning task of the manipulator in a dynamic environment is relatively complex. This paper uses the improved Soft Actor Critic Algorithm (SAC) with the maximum entropy advantage as the benchmark algorithm to implement the motion planning of the manipulator. In order to solve the problem of insufficient robustness in dynamic environments and difficulty in adapting to environmental changes, it is proposed to combine Euclidean distance and distance difference to improve the accuracy of approaching the target. In addition, in order to solve the problem of non-stability and uncertainty of the input state in the dynamic environment, which leads to the inability to fully express the state information, we propose an attention network fused with Long Short-Term Memory (LSTM) to improve the SAC algorithm. We conducted simulation experiments and present the experimental results. The results prove that the use of fused neural network functions improved the success rate of approaching the target and improved the SAC algorithm at the same time, which improved the convergence speed, success rate, and avoidance capabilities of the algorithm.