Abstract: In response to the issues of insufficient interaction compliance, high collision risk, and poor tracking performance in human-robot interaction tasks, a reinforcement learning strategy based on admittance control is proposed. Firstly, a smooth and differentiable reference trajectory is generated using a soft saturation function and an expected admittance model. Secondly, a reinforcement learning method based on the AC (actor-critic) structure is introduced to deal with the system dynamic uncertainties. Unlike existing studies, a controller incorporating a time-varying asymmetric barrier Lyapunov function is constructed. This controller ensures accurate trajectory tracking of the end-effector while strictly satisfying position constraints, thus guaranteeing the system safety and reliability in complex interaction scenarios. Finally, the semi-global uniform ultimate boundedness of the closed-loop system is demonstrated through Lyapunov stability theory. A series of experiments are conducted on the Baxter robot experimental platform, and the proposed control method is compared with adaptive impedance control, fuzzy adaptive impedance control and traditional impedance control. The comparison results show that the proposed control method outperforms the other methods in terms of tracking accuracy, compliance and collision avoidance, thereby fully validating the feasibility of the proposed method.