ParkingTransformer uses LLMs to boost autonomous parking success to 88.7%
LLM-powered trajectory planning eliminates black-box issues with 61.32 driving score in CARLA
ParkingTransformer is a novel autonomous parking framework that addresses the black-box nature of existing end-to-end methods by leveraging large language models (LLMs) for high-level semantic understanding and interpretability. The system combines multi-view perception with LLM implicit state features and trajectory queries to directly output planning trajectories from raw sensor data, eliminating the need for dense bird's-eye view (BEV) representations. To compensate for LLMs' weak spatial reasoning, the team introduces 3D positional encoding that explicitly injects spatial geometric awareness into the model.
A key innovation is the fixed-window streaming mechanism for historical information processing, which significantly improves long-term temporal efficiency and inference speed. Additionally, a coarse-to-fine decoding strategy progressively refines trajectory precision. In extensive closed-loop experiments on the CARLA simulator, ParkingTransformer achieved a driving score of 61.32, while real-world vehicle tests demonstrated an average success rate of 88.70%. These results validate the feasibility of using LLMs for end-to-end autonomous parking without relying on intermediate representations.
- Uses LLM implicit features and trajectory queries to skip dense BEV representations
- Introduces 3D positional encoding to overcome LLMs' weak spatial reasoning
- Achieves 61.32 driving score in CARLA and 88.70% success rate in real-world parking tests
Why It Matters
LLM-enhanced trajectory planning could make autonomous parking safer, more interpretable, and deployable without costly sensor arrays.