Feiyang Jia*, Lin Liu*, Ziying Song, Caiyan Jia†, Hangjun Ye, Xiaoshuai Hao† and Long Chen⊥ [📄 Paper (arXiv)]

We present DriveWorld-VLA, a tightly coupled framework where a world model serves as the reasoning engine bridging action and prospective imagination.

• Feb. 01th, 2026: We released our paper on Arxiv. NavSim Code/Models are released!

• Release Paper
• Release NavSim Models and Training/Evaluation Framework
• Release NuScenes Models and Training/Evaluation Framework

Training conducted on 8 NVIDIA H20 GPUs.

Legend • NC: No Collision • DAC: Drivable Area Compliance • EP: Ego Progress • TTC: Time to Collision • Comfort: Comfort • PDMS: Predictive Driver Model Score

root/
├── ckpts/
│   └── resnet34.pth
├── internvl_chat/
│   └── Internvlm checkpoint
├── dataset/
│   ├── maps/
│   ├── navsim_logs/
│   │   ├── test/
│   │   └── trainval/
│   ├── sensor_blobs/
│   │   ├── test/
│   │   └── trainval/
└── exp/
    └── metric_cache/

To obtain the navsim dataset:

bash download/download_maps.sh
bash download/download_navtrain.sh
bash download/download_test.sh

refer to https://github.com/xiaomi-research/recogdrive to download checkpoint

bash scripts/evaluation/run_metric_caching.sh

Create the conda environment:

conda env create -f environment.yml
conda activate Driveworld-vla

Install dependencies:

pip install -r requirements.txt
pip install git+https://github.com/motional/nuplan-devkit.git@nuplan-devkit-v1.2#egg=nuplan-devkit

Add environment variables to ~/.bashrc (modify paths as needed):

export NUPLAN_MAP_VERSION="nuplan-maps-v1.0"
export NUPLAN_MAPS_ROOT="$HOME/navsim_workspace/dataset/maps"
export NAVSIM_EXP_ROOT="$HOME/navsim_workspace/exp"
export NAVSIM_DEVKIT_ROOT="$HOME/navsim_workspace/"
export OPENSCENE_DATA_ROOT="$HOME/navsim_workspace/dataset"

Update paths in:

——navsim/agents/WoTE/configs/default_stage1.py
——navsim/agents/WoTE/configs/default_stage2.py
——navsim/agents/WoTE/configs/default_stage3.py

Then launch training stage 1:

bash scripts/training/run_ImagineWorld_stage1.sh # stage1_training

Then launch training stage 2:

bash scripts/training/run_ImagineWorld_stage2.sh # stage2_training

Then launch training stage 3:

bash scripts/training/run_ImagineWorld_stage3.sh # stage3_training

Evaluation (stage 3):

bash scripts/evaluation/eval_driveworld_vla.sh

Visualization examples of navsim dataset. Top label: source of trajectory.

Visualization examples of nuScenes validation dataset. Top label: source of trajectory.

DriveWorld-VLA is greatly inspired by the following outstanding contributions to the open-source community: NAVSIM, DPPO, LightningDiT, DiffusionDrive, WOTE.

If you find DriveWorld-VLA is useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.

@article{
update soon
}