This paper explores the dynamic torsion control strategy for bionic flapping-wing aircraft based on machine learning. Firstly, it outlines the importance of dynamic torsion control in bionic flapping-wing aircraft and the application of machine learning in this field. Subsequently, a comparative analysis of the energy efficiency of passive torsion and active torsion is conducted, and the challenges faced by traditional Deep Reinforcement Learning (DRL) in flapping-wing control are pointed out. To address these issues, this paper proposes an improved DRL algorithm incorporating an attention mechanism. The design of the new model, the establishment of the simulation environment, and the experimental setup are described in detail. Finally, through the analysis and discussion of the experimental results, the effectiveness of the improved algorithm in optimizing the dynamic torsion control of bionic flapping-wing aircraft is verified, providing insights for future work.