A Self-Organizing Federated Reinforcement Learning Framework for Cross-Scenario Collaborative AVP Trajectory Planning

Cross-scenario collaborative trajectory planning can improve Automated Valet Parking (AVP) by transferring policy knowledge learned from vehicle experiences in different environments. However, existing transfer approaches often rely on static architectures, coarse representations, and limited adaptability, restricting continual knowledge sharing across heterogeneous scenarios. We propose a Self-Organizing Federated Reinforcement Learning (SO-FRL) framework for cloud-edge-vehicle collaboration. The vehicle layer performs local perception and real-time planning, the edge layer aggregates vehicle-level knowledge into parking-lot repositories, and the cloud layer integrates multi-scenario knowledge for cross-scenario generalization. SO-FRL combines a self-organizing network with federated node-level knowledge transfer, enabling continual policy evolution without directly sharing raw trajectories or sensor streams. Experiments in controlled simulations constructed from multiple real parking-lot maps show that SO-FRL achieves the highest measured success rate and success-weighted path efficiency among the tested learning-based and classical baselines, reaching$79.0\% \pm 1.5\%$SR and$0.60 \pm 0.04$SPL. It also yields lower and more consistent navigation error while maintaining a compact knowledge structure and acceptable computational complexity. Note to Practitioners—This paper introduces an SO-FRL framework for cross-scenario trajectory planning and adaptive knowledge sharing in autonomous valet parking. By continually compressing and evolving a cloud-edge-vehicle node-based knowledge repository, SO-FRL is designed to reduce onboard model size and uplink traffic by exchanging compact knowledge nodes rather than raw trajectories or raw sensor streams. The current evaluation is conducted in a controlled simulator built from real parking-lot map files, including parking-space layouts, lane geometry, drivable boundaries, and entrance/exit markers. Therefore, the results should be interpreted as real-map-based algorithmic validation rather than field-deployment evidence. The present study does not yet include real-vehicle experiments, digital-twin deployment, or validation in a high-fidelity sensor/traffic simulator. Communication delay and packet loss, unstable edge connectivity, localization error, sensor noise, dynamic pedestrian behavior, unexpected obstacles, cybersecurity risks, and edge-server failures remain to be evaluated in future deployment-oriented validation.