Automatic Parking Assist System (APAS) is an important part of Advanced Driver Assistance System (ADAS). It frees drivers from the burden of maneuvering a vehicle into a narrow parking space. This paper deals with the motion planning, a key issue of APAS, for vehicles in automatic parking. Planning module should guarantee the robustness to various initial postures and ensure that the vehicle is parked symmetrically in the center of the parking slot. However, current planning methods can’t meet both requirements well. To meet the aforementioned requirements, a method combining neural network and Monte-Carlo Tree Search (MCTS) is adopted in this work. From a driver’s perspective, different initial postures imply different parking strategies. In order to achieve the robustness to diverse initial postures, a natural idea is to train a model that can learn various strategies. As artificial neural network has outstanding potential in representing and learning knowledge, a neural network is utilized to provide prior knowledge, which is trained through supervised learning by a novel method that imitates human learning style. However, the training accuracy and generalization ability of a neural network, together with the quality of training data, will interfere with the robustness and terminal parking performance of a planning module inevitably. To counter this, a particular variant of MCTS (P-MCTS), which can memorize the best sequence with the highest reward, is combined with the neural network to select the final action by taking Monte-Carlo samples among possible candidates produced by the neural network. Compared with the standard MCTS, P-MCTS performs better in one-player puzzle like automatic parking. Using this method, for the planning part, the robustness to various initial postures and good terminal parking performance can be guaranteed simultaneously, which is verified by the simulations of Matlab and hardware-in-the-loop (HIL) tests.