With the development of computational fluid dynamics and deep reinforcement learning, it has become possible to combine deep reinforcement learning with computational fluid dynamics to study the control problem of underwater robots. This article uses Computational Fluid Dynamics (CFD) and Deep Reinforcement Learning to study planar pose control of robotic fish. A two-dimensional model of the robotic fish is built in CFD, and the surrogate model environment is built based on CFD simulation results. Actor-Critic method is used to train the pose control in surrogate model environment. Sparse reward problem is solved by using course learning method during training. Finally, pose control experiment of robotic fish is carried out. The experimental results show that this pose control method is effective and feasible, control accuracy is higher.