The primary objective of generative adversarial imitation learning (GAIL) is to learn expert behavior from trajectories without relying on a predefined reward function. However, conventional GAIL algorithms often depend on expert experience to guide decision-making and lack active exploration of the environment, resulting in sub-optimal solutions and instability. To address these issues, we propose ICM-GAIL: Generative Adversarial Imitation Learning Algorithm based on an Improved Curiosity Module. ICM-GAIL introduces an internal reward function to promote active exploration of the environment, coupled with an uncertainty metric to rectify inaccuracies in state prediction. Additionally, to ensure efficiency and stability, the algorithm utilizes Wasserstein divergence to quantify the discrepancy between generated policy and the expert’s policy. Experimental results demonstrate that ICM-GAIL surpasses other imitation learning algorithms in Mujoco robot control tasks.