Vascular interventional robots reduce radiation exposure and alleviate operator fatigue, offering an improvement for interventional physicians’ working conditions. Furthermore, guidewire autonomous operation can greatly assists doctors in performing operations. This paper presents a novel approach for guidewire autonomy. This work generate a large number of realistic simulation tasks by deriving them directly from an open-source coronary angiography dataset. This approach eliminates the need for additional preprocessing steps like image segmentation or reliance on synthetic 3D models, providing richer and more challenging samples than existing methods. This paper further proposes a logarithmic reward function combined with the A* algorithm, which can effectively guide an agent to make optimal decisions in different guidewire tasks; it also proposes an Actor-Critic network with ResNet and Vision Transformers, which can improve the performance of reinforcement learning agents compared to other baseline networks. A multi-process reinforcement learning framework has been proposed to accelerate the data interaction process of this method. This paper finally deploys the proposed agent on a robotic system, demonstrating its effectiveness in real-world guidewire autonomous tasks and its promising potential for practical clinical applications.