The traditional robust multi-objective optimization (RMO) is computationally expensive, making it difficult to apply in engineering. To this end, a selecting RMO (SRMO) framework that considers multi-source uncertainties and the adaptive update of the surrogate model is proposed. The SRMO framework is more efficient and has been applied to optimize a micro-electro mechanical system (MEMS) electro-thermal actuator (ETA). The application of sensitivity analysis and surrogate modeling techniques improves efficiency in terms of reducing the problem dimensionality and the computational cost of fitness values, respectively. A multi-objective particle swarm optimization algorithm is adopted to find the better-performing points where time-consuming finite element analysis (FEA) will be performed, thus ensuring the prediction accuracy of the global surrogate (GS) model around them. A greedy algorithm is employed to select design points from the candidate solutions set to train local surrogate (LS) models that perform the robust evaluation taking into account multi-source uncertainties in dimensions, material parameters, and surface topography features. In the optimization of ETA, only 2.9% and 0.5% of the particles require performing FEA to update the GS model and train LS models, respectively. The final solutions show good robustness, and the effectiveness of the SRMO framework is verified by experiments and simulations.