A simulation model is often used to guide the decision-making for a real complex stochastic system. To faithfully assess the mean performance of the real system, it is necessary to efficiently calibrate the simulation model. Existing calibration approaches are typically built on the summary statistics of simulation outputs and ignore the dynamic information carried by the detailed sample paths. Given a tight simulation budget, in this paper, we develop a new calibration approach that collects the detailed output sample paths in a sequential manner. Both theoretical and empirical studies demonstrate that we can efficiently use the simulation resources and achieve better calibration accuracy by exploring the system dynamic behaviors.