Abstract: To address the limitations of traditional magnetic levitation control algorithms, which often rely on precise mathematical models and exhibit poor adaptability, this paper proposes an improved deep deterministic policy gradient (IDDPG) controller based on reinforcement learning. A mathematical model of the electromagnetic levitation system is first established, and its dynamic characteristics are analyzed. To enhance the conventional DDPG algorithm’s applicability in maglev control, a segmented inverse proportional reward function is designed to improve steady-state accuracy and response speed. The control framework is further optimized to meet real-time deployment requirements. Comprehensive simulations and experiments are conducted to evaluate the impact of current loop tracking, reward function design, training step size, and model variations on control performance. Results demonstrate that the proposed IDDPG controller significantly enhances system response speed while reducing steady-state error and overshoot，and the optimized control flow is suitable for real system deployment. Moreover, it maintains consistent control performance across varying system models, confirming its robustness and adaptability without reliance on exact model knowledge.