The purpose of this paper is to explore how to improve the motion accuracy and adaptability of quadruped robots in complex environments through deep reinforcement learning techniques. A proximal policy optimization (PPO) algorithm and a gated recurrent unit (GRU) network model are adopted, and the system is simulated and trained on NVIDIA Isaac Gym simulation platform. The experimental results show that the deep reinforcement learning strategy combining the PPO algorithm and the GRU model significantly improves the accuracy and adaptability of the quadruped robot in complex environments, which provides a new idea for the application of quadruped robots in extreme scenarios.