Abstract: This study presents the application of the deep reinforcement learning (DRL) method in water tunnel experiments, which establishes an automatic closed-loop optimization framework in the laboratory. The framework is used to optimize the propulsion efficiency of a NACA0012 airfoil model under pure pitching motion at a Reynolds number of Re = 1.3×10⁴. Existing related studies often limit the motion patterns to periodic functions, thanks to the DRL method, the optimization process can take place in a broader non-periodic action space. In the experiment, the airfoil automatically interacts with the water tunnel environment, and ultimately learns highly efficient non-periodic motions. By modifying the reward function, the efficiency optimization can be achieved above a given thrust threshold. The present findings demonstrate the DRL model can continuously enhance the propulsion efficiency by adjusting the amplitude and frequency of the flapping motion. Moreover, the optimal flapping motion obtained by DRL is close to the sinusoidal motion, and the achieved optimal propulsion efficiency lies on the upper bound of that of the sinusoidal motion with a similar amplitude. The present study demonstrates the feasibility of using the DRL method for complex flow control problems.