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

A new autonomous-parking framework called ParkingTransformer replaces dense bird’s-eye-view maps with large-language-model scene understanding, according to research submitted June 12, 2026 [1]. The authors report a 61.32 driving score in the CARLA simulator and an 88.70% average success rate in real-world tests [1]. The work targets a persistent weakness in end-to-end parking systems: their black-box nature, which limits high-level semantic understanding and interpretability [1]. By combining trajectory queries with implicit state features from large language models, ParkingTransformer interacts directly with raw sensor data and historical information to produce planning trajectories [1]. The architecture eliminates the need for dense bird’s-eye-view representations, a departure from many current approaches [1]. Large language models are built on transformer architectures that use attention mechanisms to model long-range dependencies in data [4]. The researchers compensate for the limited spatial reasoning of such models by introducing 3D positional encoding, which explicitly injects spatial geometric awareness [1]. A fixed-window streaming mechanism processes historical information, improving long-term temporal efficiency and inference speed [1]. A coarse-to-fine decoding strategy then progressively refines trajectory precision [1]. Closed-loop experiments were conducted on the CARLA simulator and on real-world vehicle platforms [1]. The driving score of 61.32 in CARLA and the 88.70% average success rate in physical tests validate the feasibility of the proposed algorithms, the authors state [1]. The paper does not disclose the vehicle platforms used, but the broader autonomous-driving industry relies heavily on high-performance computing hardware. Nvidia, for example, controlled more than 80% of the market for GPUs used in training and deploying AI models as of 2025 [3]. The submission appears on arXiv, a preprint server, and has not yet been peer-reviewed [1]. The authors frame long-distance autonomous parking—from the road to a target spot—as a critical task within autonomous driving, and they argue that interpretability gains from language-model integration could help bridge the gap between research prototypes and deployment [1].