Abstract: Active control of flow past bluff bodies is an effective approach for drag reduction, wake suppression, and the mitigation of flow-induced vibration and noise; however, under turbulent conditions, closed-loop flow control often faces major challenges caused by the strong nonlinearity of the system and the high dimensionality of the parameter space. Deep reinforcement learning (DRL) provides a new paradigm for solving these complex fluid dynamics problems through data-driven interaction and policy optimization, but its robustness verification, real-time applicability, and parameter sensitivity analysis in physical experimental environments are still insufficient. In this study, the flow past a circular cylinder at a Reynolds number of 3900 is investigated, and a closed-loop rotary control experimental platform based on a towing tank is established; particle image velocimetry (PIV) is adopted to provide real-time flow-field feedback, the proximal policy optimization (PPO) algorithm is used to realize closed-loop intelligent flow control, and the effects of key hyperparameters of the DRL algorithm on the experimental results are quantitatively studied by comparing different discount factors and hidden-layer widths. The results show that the agent can autonomously learn an effective low-energy control strategy from repeated interactions with the experimental wake; after control is applied, the mean recirculation length in the cylinder wake is reduced by approximately 50%, the symmetry of the flow field is significantly improved, and the periodic shedding of the Karman vortex street is effectively weakened, indicating that the learned rotary actuation can regulate the wake instability. In addition, the hyperparameter sensitivity analysis indicates that the parameter combination with a discount factor of γ = 0.97 and a hidden-layer width of 320 × 320 leads to the fastest policy convergence and achieves the best flow-control performance among the tested cases, resulting in the shortest recirculation region and local enhancement of turbulent kinetic energy in the near wake. This study verifies the feasibility of DRL under experimental conditions and provides important experimental evidence and reference for closed-loop active control of turbulent flows, especially for the deployment of intelligent control strategies in complex experimental and engineering flow systems.