A compliant floating force-controlled end-effector was designed, in order to resolve the problem of poor adaptability of industrial robots for the compliant grinding of workpieces. A robotic grinding force control method with the active adaptive variable impedance was proposed, using the reinforcement-learning based on the ensemble Bayesian neural networks model. According to the contact environment information of the robotic grinding, the multiple sampling samples from the small amount of data were obtained by the Bootstrapping method, and the ensemble Bayesian neural network model was trained to characterize the interactions between the robotic grinding system and the grinding condition environment. The optimal impedance parameters were solved by the covariance matrix adaptation evolution strategy (CMA-ES). A virtual prototype platform of the robotic grinding system was constructed. A robotic grinding simulation experiment of a blade workpiece was conducted, and the effectiveness of the proposed method was verified. Experimental results show that the proposed method reduces the absolute tracking error of the grinding force to a small value after a dozen training, realizes the active adaptive variable impedance for the grinding force control of the robotic grinding system, and improves the flexibility and the robustness of the robotic grinding force control.