Reinforcement learning parking path planning method and system for non-structured scene

Abstract

This invention proposes a reinforcement learning-based parking path planning method and system for unstructured scenarios. It encodes an unstructured obstacle map into latent feature vectors using a variational autoencoder to extract hidden features from the obstacle map. Then, it represents the vehicle's own state information and parking space information as vectors, concatenating these three to form a state vector. A hybrid A* algorithm is used as an expert to search for an initial parking path. The state vector and path information are input into an imitation learning network, enabling the autonomous vehicle to learn to plan a basic parking path. Reinforcement learning is then used to optimize the imitation learning planning strategy, allowing it to not only optimize the basic path but also adapt to parking scenarios not covered in the imitation learning. This transforms reinforcement learning from image processing to latent vector processing, reducing model training time and making the strategy easier to generalize. Furthermore, it allows continued interaction with the environment after a parking strategy has been established, addressing the problems of insufficient datasets and low learning efficiency in the imitation learning approach.

Images