Abstract
Motion planning is one of the key technologies of automatic parking, which receives environment information and vehicle states to guide the vehicle to the parking space. The planning module should consider the multi-objective optimization of the generated path including safety, comfort, parking efficiency, final parking performance, and the high efficiency of the algorithm. However, current planning methods can only meet part of above requirements. In order to satisfy the demands mentioned above, a method with neural network and MCTS is proposed in this work. First, the neural network is designed to classify the drivers’ actions by supervised learning from the experienced drivers parking data. The parking environment information and vehicle states are network inputs, and the corresponding drivers’ different actions are the network target output. The network actual output is the probability distribution over possible actions corresponding to the vehicle state. The neural network trained by drivers’ parking data implies the demands of safety and comfort. Second, the final action is selected among possible candidates of different possibility output by the neural network to ensure the final parking performance and the parking efficiency to be optimal through large numbers of simulations of MCTS. In addition, the parking experience learned by the network can provide prior probability guidance, so it will improve search efficiency of MCTS compared with random simulation of general MCTS. Finally, simulations of Matlab/Simulink and HIL (hardware in the loop) tests verify the effectiveness of the proposed method.
Keywords: Automatic Parking, Motion Planning, Neural Network, Monte Carlo Tree Search (MCTS), Supervised Learning