Controlling biomimetic underwater robots in unknown flow fields remains a challenge due to the strong nonlinearity of the fluid. This article investigates the attitude holding task of a robotic fish swimming in reality. Such a typical sensing-based control task requires the fish to keep a desired angle of attack in an unknown and even varied incoming flow. To this end, we propose a learning-based approach by using a deep neural network directly maps the raw data of sensors equipped on the robot to the continuous control signals in an end-to-end manner. First, based on experimental data of the physical robot, we construct a data-driven simulation environment including three modules of dynamic, sensor, and control. The dynamic and sensor modules are established to model the dynamics of the fish and to generate its sensors’ data, based on which a deep reinforcement learning (DRL) algorithm in the control module is trained to get a control policy. Then, we directly deploy the trained policy to a physical robotic fish for attitude holding task. Experimental results demonstrate the robustness and effectiveness of the DRL policy and, thus, verify the success of our approach to achieving sim-to-real transfer.