Quadruped robots are increasingly required to operate in unstructured and unpredictable terrains, where stable autonomous locomotion remains a significant challenge due to uneven ground geometry, changing slopes, and complex interaction dynamics. To address these difficulties, this paper proposes an autonomous motion control framework that integrates a Regulated Central Pattern Generator (RCPG) with deep reinforcement learning, enabling quadruped robots to adapt their gait to diverse environments in real time. Firstly, a terrain estimation algorithm is developed to obtain real-time ground inclination and contact characteristics using foot-end kinematics and IMU feedback. This module provides essential environmental perception for subsequent gait adaptation. Secondly, an RCPG network based on an elliptical nonlinear oscillator is constructed, and a terrain-feedback regulation mechanism is introduced to dynamically adjust gait parameters such as step length, step height, and oscillation frequency. To ensure smooth switching between swing and stance phases, a gating function is incorporated to generate continuous and stable foot-end trajectories. Finally, a Proximal Policy Optimization (PPO) based reinforcement learning strategy is designed to optimize RCPG parameters through embodied interaction, supported by a multi-term reward function emphasizing trajectory tracking accuracy, attitude stability, terrain adaptability, and anti-fall robustness. Simulation results across multi-slope terrains demonstrate that the proposed framework enables stable locomotion, accurate posture adjustment, and strong adaptability to complex, unseen environments, validating its effectiveness and robustness.