LiLoDriver is a lifelong learning framework for closed-loop motion planning in long-tail autonomous driving scenarios. It leverages Large Language Models (LLMs) for adaptive planner selection and memory-based knowledge retrieval, enabling continual scene understanding and strategy evolution during inference without the need for retraining.

Our method integrates four core modules:

• Environment and Perception: Fuses map and historical trajectory data.
• Scene Encoder: Extracts scene embeddings via a multi-modal MLP-LSTM-Transformer stack.
• Memory Bank: Maintains long-tail clusters and associated few-shot planner experiences.
• LLM Reasoning and Planner Selection: Generates task-specific prompts and selects behavior planners in real-time.

LiLoDriver is the first framework to introduce lifelong learning and retrieval-augmented planning into closed-loop autonomous driving at scale.

• ✅ Closed-loop motion planning with long-tail generalization
• ✅ Online planner adaptation without model retraining
• ✅ Memory-augmented scenario clustering and few-shot planner retrieval
• ✅ LLM-based reasoning for scene-specific decision making
• ✅ Evaluated on the nuPlan real-world benchmark

git clone https://github.com/your-org/LiLoDriver.git
cd LiLoDriver
conda create -n lilodriver python=3.10
conda activate lilodriver
pip install -r requirements.txt

We recommend using CUDA 11.7+ and PyTorch 2.0+ for GPU-accelerated training and inference.

LiLoDriver/
├── assets/                     # Diagrams and figures
├── configs/                   # Config files for model and experiments
├── data/                      # Dataset loaders and preprocessing scripts
├── memory/                    # Memory bank and planner retrieval logic
├── models/                    # Scene encoder and planner network modules
├── planners/                  # Rule-based and learned behavior planners
├── prompts/                   # Prompt engineering for LLM calls
├── utils/                     # Common utilities
├── main.py                    # Entry point for training and evaluation
└── README.md

We evaluate LiLoDriver on the nuPlan benchmark, which provides large-scale real-world driving data with support for closed-loop simulation.

To use nuPlan:

# Install nuPlan devkit
pip install nuplan-devkit
# Set up simulation environment
python setup_nuplan.py --download_data --prepare_maps

To run closed-loop simulation:

python main.py --config configs/lilodriver_nuplan.yaml --eval_only

Key evaluation metrics:

• Success Rate (SR)
• Collision Rate (CR)
• Comfort Index (CI)
• Long-tail Scenario Completion (LTSC)

LiLoDriver outperforms existing rule-based and learning-based planners across rare and complex driving scenarios:

If you find this work helpful, please cite us:

@article{yao2025lilodriver,
  title={LiloDriver: A Lifelong Learning Framework for Closed-loop Motion Planning in Long-tail Autonomous Driving Scenarios},
  author={Yao, Huaiyuan and Li, Pengfei and Jin, Bu and Zheng, Yupeng and Liu, An and Mu, Lisen and Su, Qing and Zhang, Qian and Chen, Yilun and Li, Peng},
  journal={arXiv preprint arXiv:2505.17209},
  year={2025}
}

This work builds upon nuPlan, LLaMA, and LangChain. We thank the open-source community for their contributions.

For questions or collaboration inquiries, please contact:
📧 [email protected]