With the development of science and technology, quadruped robots have shown excellent athletic ability on various terrains through reinforcement learning (RL). Compared with the traditional model-based algorithm, RL is not limited by accurate modeling and has better environmental adaptability. Therefore, this paper introduces a path planning method for quadruped robots traversing risky terrains which have non-continuous footholds based on reinforcement learning. Firstly, the kinematics model and the static stability of the robot are analyzed. Meanwhile, non-continuous foothold terrains are concretized as plum piles, whose simulation model satisfying the robot’s motion constraints is established. Then, the feasible footholds of the quadruped robot walking on the plum pile are selected through Markov decision process. Finally, the global path planning of the body centroid is obtained by the Deep Deterministic Policy Gradient (DDPG) algorithm. Through simulation, the effectiveness of this method is verified, achieving path planning for quadruped robots under the constraints of non-continuous foothold terrains.