Aiming at the existing peg-in-hole assembly method problems of dependence on accurate contact state models, difficulties in data acquisition, low sampling efficiency, and poor security, a simulation research method for robot peg-in-hole assembly strategy based on DRL is proposed. A simulation environment of robot peg-in-hole assembly based on ROS-Gazebo is built, and a method of gravity compensation for force/torque sensor based on a least square method is proposed. The reinforcement learning paradigm is employed to model the robot peg-in-hole assembly, and a method based on soft actor-critic(SAC) algorithm is proposed. The communication mechanism between the simulation environment and the deep reinforcement learning algorithm is established through ROS. Simulation experiments show that the proposed SAC algorithm enables robots to accomplish the peg-in-hole assembly task autonomously and compliantly with good generalization ability.