The invention discloses a compliant force control method based on fuzzy reinforced learning for a mechanical arm. A fuzzy reinforced learning algorithm is adopted, a real-time adjustment strategy for admittance parameters is trained in a manner of online learning, and a motor is controlled by the converged variable-admittance control strategy according to an external moment applied by an operator, and the current joint speed and acceleration to actively comply with a control intention of the operator, so as to complete an active following task of the mechanical arm, without the need of establishing corresponding task and environment model, so that a higher convergence speed and a stable actual effect are achieved. The method is capable of remarkably lowering the working intensity of the operator and improving the location accuracy, and conducive to reduce the structural dimension and the dead weight of the mechanical arm; a haptic human-machine interaction model is capable of greatly responding to the control intention of the operator, and high in self-adaptive capacity, so that haptic human-machine interaction experience is smoother and more natural, and more similar to haptic interaction experience during operation for an actual object in daily life.