For simultaneous localization and mapping (SLAM) technology has the disadvantages of high demand for computing resources, limited environmental adaptability, cumulative error problem, high system complexity, high cost, limited large scene processing capacity and lack of effective loop detection mechanism, so a method combining artificial potential field method and deep reinforcement learning is proposed. The graph theory is used to simulate the interaction between robots and the potential force between robots and the destination, and the twin delayed deep deterministic policy gradient algorithm is used to optimize the robot's perception and processing of obstacle information. The simulation results show that the method can make the robot locate and move quickly and accurately in the unknown environment, while maintaining the stability and consistency of the formation.