Human-robot collaboration is an important area with great potential in intelligent manufacturing. Due to the diversity of collaboration tasks, robot collaboration skills should have the ability to adapt to different skills. However, problems such as skill expression and generalization are challenging. Meanwhile, the differences in the skills of various operators bring difficulties to collaborative robots. This work develops a variable impedance learning method for human-robot collab oration assembly. Unlike most previous work that mainly dis cussed a special human collaborator with the fixed impedance parameters, this work learns a robot impedance by reinforce ment learning. We aim to make the inertia, damping, and stiffness parameters adaptive by Proximal Policy Optimization (PPO) algorithm. Hence, we can let the robot collaborate with various human collaborators to accomplish a high-precision assembly task. Two experiment results illustrate the validity of the proposed method. The detailed experimental videos are available at https://youtu.be/AJyjW2NwA74.