Check out our Large Driving Model Series!

GPD-1: Generative Pre-training for Driving

Zixun Xie*, Sicheng Zuo*, Wenzhao Zheng* $\dagger$, Yunpeng Zhang, Dalong Du, Jie Zhou, Jiwen Lu, Shanghang Zhang$\ddagger$

* Equal contribution $\dagger$ Project leader $\ddagger$ Corresponding author

GPD-1 proposes a unified approach that seamlessly accomplishes multiple aspects of scene evolution, including scene simulation, traffic simulation, closed-loop simulation, map prediction, and motion planning, all without additional fine-tuning.

• [2024/12/12] Code released.
• [2024/12/12] Paper released on arXiv.

The pre-trained GPD-1 can accomplish various tasks without finetuning using different prompts.

Our model adapts the GPT-like architecture for autonomous driving scenarios with two key innovations: 1) a 2D map scene tokenizer based on VQ-VAE that generates discrete, high-level representations of the 2D BEV map, and 2) a hierarchical quantization agent tokenizer to encode agent information. Using a scene-level mask, the autoregressive transformer predicts future scenes by conditioning on both ground-truth and previously predicted scene tokens during training and inference, respectively.

• 1. Installation and download
• 2. Feature cache
• 3. Train map tokenizer
• 4. Train GPD

Our work is inspired by these excellent open-sourced repos: PlanTF Sledge Navsim

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

  @article{gpd-1,
    title={GPD-1: Generative Pre-training for Driving},
    author={Xie, Zixun and Zuo, Sicheng and Zheng, Wenzhao and Zhang, Yunpeng and Du, Dalong and Zhou, Jie and Lu, Jiwen and Zhang, Shanghang},
    journal={arXiv preprint arXiv:2412.08643},
    year={2024}
}