Check out our Large Driving Model Series!

Doe-1: Closed-Loop Autonomous Driving with Large World Model

Wenzhao Zheng* $\dagger$, Zetian Xia*, Yuanhui Huang, Sicheng Zuo, Jie Zhou, Jiwen Lu

* Equal contribution $\dagger$ Project leader

Doe-1 is the first closed-loop autonomous driving model for unified perception, prediction, and planning.

• [2024/12/13] Evaluation code released.
• [2024/12/13] Paper released on arXiv.
• [2024/12/13] Demo released.

Doe-1 is a unified model to accomplish visual-question answering, future prediction, and motion planning.

We formulate autonomous driving as a unified next-token generation problem and use observation, description, and action tokens to represent each scene. Without additional fine-tuning, Doe-1 accomplishes various tasks by using different input prompts, including visual question-answering, controlled image generation, and end-to-end motion planning.

We explore a new closed-loop autonomous driving paradigm which combines end-to-end model and world model to construct a closed loop.

1. Download nuScenes V1.0 full dataset data HERE.

2. Download the annotations data_nusc from OmniDrive and unzip it.

3. Download the VQVAE weights from HERE and put them to the following directory as HERE:

Doe/
- model/
    - lumina_mgpt/
        - ckpts/
            - chameleon/
                - tokenizer/
                    - text_tokenizer.json
                    - vqgan.yaml
                    - vqgan.ckpt
    - xllmx/
- ...

1. Generate the conversation data for inference and set the max :

# max length: 1 for qa, 5 for planning
python dataset/gen_data.py \
--info_path path/to/infos_var.pkl \
--qa_path path/to/OmniDriveDataset \
--nusc_path path/to/nuscenes \
--save_path path/to/save/outputs \
--max_length 1

2. Inference with a model ckpt:

# set split and id for multi gpus
CUDA_VISIBLE_DIVICES=0 python inference/eval.py \
--anno_path path/to/val_infos.pkl \
--nusc_path path/to/nuscenes \
--save_path path/to/save/output \
--model_path path/to/model/ckpt \
--data_path path/to/generated/data.json \
--task qa

Our code is based on the excellent work Lumina-mGPT.

If you find this project helpful, please consider citing the following paper:

@article{doe,
    title={Doe-1: Closed-Loop Autonomous Driving with Large World Model},
    author={Zheng, Wenzhao and Xia, Zetian and Huang, Yuanhui and Zuo, Sicheng and Zhou, Jie and Lu, Jiwen},
    journal={arXiv preprint arXiv: 2412.09627},
    year={2024}
}