The assembly process of robot is constrained by the characteristics of assembly object, assembly process and assembly control law. In order to solve the problem of pose uncertainty in the contact phase of assembly process, a self-learning method of assembly post adjustment skills is proposed in this paper. Firstly, the problem of robot assembly skill under multiple constraints is described, and the assembly system model based on force/torque, posture, joint angle and other multi-mode information is built. Then the robot decision network and strategy optimization network are constructed to learn assembly pose adjustment skills through continuous interaction with the environment. Finally, the test is carried out on the low-voltage electrical appliance plastic shell fasten assembly experimental platform. The results show that under the constraints of workpiece characteristics, assembly process and control law, the skill learning method that robots adopted has obtained the end pose adjustment strategy, and completed the fasten assembly. The success rate increases by 7.4 percent than the deep Q-learning network(DQN) based algorithm.