ParkingTransformer: LLM-Enhanced End-to-End Trajectory Planning for Autonomous Parking

ParkingTransformer is a novel framework for end-to-end autonomous parking that integrates Large Language Models (LLMs) with multi-view perception to enhance trajectory planning. It introduces 3D positional encoding to improve spatial reasoning and employs a fixed-window streaming mechanism for efficient historical data processing, achieving a driving score of 61.32 in the CARLA simulator and an 88.70% success rate in real-world tests. This approach addresses the limitations of traditional methods by providing high-level semantic understanding and improved interpretability, making it significant for practitioners focused on seamless autonomous parking solutions.