Goal-Conditioned Reinforcement Learning Approach for Autonomous Parking in Complex Environments

In complex urban environments, autonomous vehicle parking is challenged by irregularly parked cars, static obstacles, and narrow spaces. Traditional parking methods are often limited to well-structured parking lots and fail to adapt to unstructured scenarios like alleyways with obstacles and randomly positioned vehicles. This paper addresses this limitation by applying goal-conditioned reinforcement learning to enable autonomous vehicles to park in congested environments based on specified target positions and orientations. A custom simulation environment is developed using Pygame to simulate parking scenarios across four progressive levels of difficulty, with obstacles to test adaptability. Three RL algorithms, SAC, HER, and an improved HER variant, are implemented to compare the performance in the simulation environment. Experimental results demonstrate that the proposed approach significantly improves parking success rates and trajectory efficiency in complex scenarios, contributing to robust, adaptable parking solutions for autonomous vehicles in real-world applications.