A Self-Interactive Action-Search Method for Autonomous Vehicle Parking with Gated Recurrent Temporal Modeling

Autonomous parking tasks in enclosed environments (e.g., ports and mines) often face limited accuracy due to data collection constraints and insufficient temporal modeling. To overcome these challenges, a Self-Interactive Action-Search Method for Autonomous Vehicle Parking with Gated Recurrent Temporal Modeling (SIS-GTM) is proposed. SIS-GTM performs optimal action search through Monte Carlo Tree Search (MCTS) without prior data, enabling self-optimization by extracting prioritized experiences from historical self-interactions with the environment. It also uses Gated Recurrent Unit (GRU) to model temporal features of environmental observations thereby enhancing state representation. We evaluate SIS-GTM in a reinforcement learning (RL) environment designed for autonomous parking tasks. In this environment, the experimental results show that SIS-GTM achieves an average parking success rate of 81.33 %, an average position error of 0.41 m, and an average angular error of 7.43°, all of which outperform commonly used reinforcement learning methods such as Deep Deterministic Policy Gradient (DDPG) and Proximal Policy Optimization (PPO). The ablation study results show that SIS-GTM with GRU as the representation network achieves superior accuracy in parking tasks compared to baseline models using Multilayer Perceptron (MLP) and Convolution Neural Network (CNN). Overall, SIS-GTM outperforms baseline RL methods in both success rate and parking accuracy, validating the effectiveness of combining tree-search planning and temporal modeling.