Facing the complex three-dimensional environment, the computational complexity of the traditional path planning algorithm is extremely increased, and the original effect is lost.Reinforcement learning can be independent of the accurate environment model, and its overall efficiency is much higher than that of the traditional algorithms.Aiming at the path planning problem of AUV in 3D environment, based on the establishment of collision avoidance detection model and gym simulation environment, the optimized PPO algorithm design and model training of improved network structure are carried out.Through simulation experiments, the accuracy and effectiveness of the algorithm are verified.