In industrial IoT, the design of control algorithms is pivotal to industrial servo systems. Unlike existing work, we study the industrial servo system control through edge computing with semi-closed-loop feedbacks, high-order nonlinear disturbances and lightweight implementation requirements. Particularly, in this paper, we take permanent magnet synchronous motor (PMSM) driven servo system as a typical industrial IoT device, propose a novel deep reinforcement learning based semi-closed-loop control algorithm, i.e., DRL-SCLC, and successfully deploy it on an edge server to provide real-time edge intelligent decision making. The control problem is formulated as a Markov decision process and then solved by the designed DRL-based algorithm, for minimizing the absolute error between reference signal and corresponding system response. To guarantee robustness, we further integrate “three-loop control structure” in traditions to DRL-SCLC for restricting outputs within a desired limit. Experiments on a real-world aerospace servo testbed show that the proposed solution is not only effective but also superior over counterparts.