Policy Search and Model Predictive Control~(MPC) are two different paradigms\nfor robot control: policy search has the strength of automatically learning\ncomplex policies using experienced data, while MPC can offer optimal control\nperformance using models and trajectory optimization. An open research question\nis how to leverage and combine the advantages of both approaches. In this work,\nwe provide an answer by using policy search for automatically choosing\nhigh-level decision variables for MPC, which leads to a novel\npolicy-search-for-model-predictive-control framework. Specifically, we\nformulate the MPC as a parameterized controller, where the hard-to-optimize\ndecision variables are represented as high-level policies. Such a formulation\nallows optimizing policies in a self-supervised fashion. We validate this\nframework by focusing on a challenging problem in agile drone flight: flying a\nquadrotor through fast-moving gates. Experiments show that our controller\nachieves robust and real-time control performance in both simulation and the\nreal world. The proposed framework offers a new perspective for merging\nlearning and control.\n