This paper presents a reinforcement learning approach to improve the gait performance and training efficiency of the Unitree humanoid robot by introducing a novel motion smoothness reward. The proposed reward encourages more stable and natural locomotion during training. Experimental results demonstrate that, compared to the baseline policy provided by Unitree, our method achieves faster convergence and higher average rewards. Additionally, visual analysis of walking trajectories confirms that the learned gait is smoother, more symmetrical, and better suited for real-time deployment.