Hierarchical Trajectory Planning for Narrow-Space Automated Parking with Deep Reinforcement Learning: A Federated...

Area of Science:

• Robotics and Artificial Intelligence
• Autonomous Systems
• Path Planning

Background:

• Automated parking in narrow spaces presents significant trajectory planning challenges.
• Existing optimization methods struggle with complex constraints and time limitations.
• Neural network approaches offer speed but lack generalization and raise privacy concerns.

Purpose of the Study:

• To develop a collision-free trajectory planning method for narrow automated parking scenarios.
• To improve planning speed and generalization capabilities of automated parking systems.
• To address privacy concerns associated with centralized training in deep learning models.

Main Methods:

• Proposes HALOES (Hierarchical trajectory planning with deep reinforcement learning in federated learning scheme).
• Combines high-level deep reinforcement learning with low-level optimization-based approaches.
• Utilizes a decentralized federated learning scheme for privacy-preserving model parameter fusion.

Main Results:

• Achieves efficient automatic parking in multiple narrow spaces.
• Improves planning time by 12.15%–66.02% compared to state-of-the-art methods (e.g., hybrid A*, OBCA).
• Maintains trajectory accuracy and demonstrates strong model generalization.

Conclusions:

• HALOES effectively generates rapid, accurate, and collision-free parking trajectories in complex narrow environments.
• The federated learning approach enhances model generalization and ensures data privacy.
• The method offers a significant advancement in automated parking technology.