Diffusion policy has recently advanced visuomotor imitation learning by effectively capturing multimodal behaviors from expert demonstrations. However, diffusion policy relies on multi-step denoising to generate actions. This process leads to high inference latency, limiting its applicability to real-time robotic control. To address this issue, we introduce the 2D Consistency Flow Matching Policy (CFMP), a novel method that enables 1-step action generation by enforcing velocity field consistency within a U-Net architecture conditioned on images and the robot state. By integrating visual observation and robot state information, CFMP learns a deterministic mapping from noise distribution to action distribution, enabling accurate and efficient action generation. We evaluate CFMP on 23 robotic manipulation tasks. The experimental results show that CFMP achieves 12–40 \(\times \) faster inference speed compared to the diffusion policy with 100-step DDPM, and 3–4 \(\times \) faster compared to the diffusion policy with 10-step DDIM, while achieving comparable or higher success rates.