README.md

metadata

license: cc-by-nc-sa-4.0
language:
  - en
tags:
  - Robotics

The missing infrastructure for Physical AI post-training in AD. Open-source. Production-validated.

Joint effort by OpenDriveLab at The University of Hong Kong, Huawei Inc. and Shanghai Innovation Institute (SII).

Highlights

• A post-training framework for Physical AI: Systematically addresses the long-tail safety-critical data scarcity problem in autonomous driving.
• Data-driven long-tail discovery: Failure-prone scenarios are automatically identified from real-world driving logs by the pre-trained agent itself — no manual design, no synthetic perturbations.
• Photorealistic interactive simulation via 3D Gaussian Splatting (3DGS): Each discovered scenario is reconstructed into a fully controllable, real-time-renderable simulation environment.
• Behavior-driven scenario generation: Leverages Behavior World Model (BWM) to generalize and synthesize diverse traffic variations from long-tail scenarios, expanding sparse safety-critical events into a dense, learnable distribution.
• RL-based post-training on safety-critical rollouts substantially outperforms scaling pre-training data alone — competitive with a ~10x increase in pre-training data.
• Production-scale validation: Deployed on a mass-produced ADAS platform trained on 80,000+ hours of driving logs, reducing collision rate by up to 45.5% and achieving zero disengagements in a 200 km on-road test.

News

• [2026/06/19] Paper released on arXiv. See World Engine: Towards the Era of Post-Training for Autonomous Driving.
• [2026/04/09] Official data release.

---

Table of Contents

• Highlights
• News
• Dataset Overview
• Directory Structure
• Environment Setup
• Usage
• Citation
• License
• Related Links
• Contact

📦 Dataset Overview

This dataset uses a modular data structure where each subsystem (AlgEngine, SimEngine) has its own data requirements while sharing common formats.

Module	Function	Data Types
Raw Data	nuPlan & OpenScene base datasets	Sensor data, maps, annotations
AlgEngine	End-to-end model training & evaluation	Preprocessed annotations, ckpts, caches
SimEngine	Closed-loop simulation environments	Scene assets, config files

WorldEngine/
└── data/                          # Main data directory
   ├── raw/                       # Raw datasets (nuPlan, OpenScene)
   ├── alg_engine/                # AlgEngine-specific data
   └── sim_engine/                # SimEngine-specific data

---

📂 Directory Structure

1️⃣ Raw Data (data/raw/)

Click to expand full directory structure

After downloading the nuPlan and OpenScene raw datasets, set up the following structure via symlinks (ln -s):

data/raw/
├── nuplan/                        # nuPlan raw dataset
│   └── dataset/
│      ├── maps/                  # HD maps (required for all modules)
│      │   ├── nuplan-maps-v1.0.json
│      │   ├── us-nv-las-vegas-strip/
│      │   ├── us-ma-boston/
│      │   ├── us-pa-pittsburgh-hazelwood/
│      │   └── sg-one-north/
│      └── nuplan-v1.1/
│          ├── sensor_blobs/      # Camera images and LiDAR
│          └── splits/            # Train/val/test splits
│   
│
└── openscene-v1.1/                # OpenScene dataset (nuPlan-based)
    ├── sensor_blobs/
    │   ├── trainval/              # Training sensor data
    │   └── test/                  # Test sensor data
    └── meta_datas/
        ├── trainval/              # Training metadata
        └── test/                  # Test metadata

2️⃣ AlgEngine Data (data/alg_engine/)

Click to expand full directory structure

Data for end-to-end model training and evaluation:

data/alg_engine/
├── openscene-synthetic/           # Synthetic data generated by SimEngine (need to generate)
│   ├── sensor_blobs/
│   ├── meta_datas/
│   └── pdms_pkl/
│
├── ckpts/                         # Pre-trained model checkpoints
│   ├── bevformerv2-r50-t1-base_epoch_48.pth
│   ├── e2e_vadv2_50pct_ep8.pth
│   ├── track_map_nuplan_r50_navtrain_100pct_bs1x8.pth
│   └── track_map_nuplan_r50_navtrain_50pct_bs1x8.pth
│
├── pdms_cache/                    # Pre-computed PDM metric caches
│   ├── pdm_8192_gt_cache_navtest.pkl
│   └── pdm_8192_gt_cache_navtrain.pkl
│
├── merged_infos_navformer/        # Preprocessed annotations
│   ├── nuplan_openscene_navtest.pkl
│   └── nuplan_openscene_navtrain.pkl
│
└── test_8192_kmeans.npy          # K-means clustering for PDM

3️⃣ SimEngine Data (data/sim_engine/)

Click to expand full directory structure

Data for closed-loop simulation:

data/sim_engine/
├── assets/                        # Scene assets for simulation
│   ├── navtest
│   │   ├── assets
│   │   └── configs
│   ├── navtrain/
│   └── navtest_failures/
│
└── scenarios/                     # Scenario configurations
    ├── original/                  # Original logged scenarios
    │   ├── navtest_failures/
    │   ├── navtrain_50pct_collision/
    │   ├── navtrain_ep_per1/
    │   ├── navtrain_failures_per1/
    │   └── navtrain_hydramdp_failures/
    │
    └── augmented/                 # Augmented scenarios (from BWM)
        ├── navtrain_50pct_collision/
        ├── navtrain_50pct_ep_1pct/
        └── navtrain_50pct_offroad/

---

⚙️ Environment Setup

Configure the following environment variables for proper data access:

Quick Configuration

# Add to ~/.bashrc or ~/.zshrc
export WORLDENGINE_ROOT="/path/to/WorldEngine"
export NUPLAN_MAPS_ROOT="${WORLDENGINE_ROOT}/data/raw/nuplan/maps"
export PYTHONPATH=$WORLDENGINE_ROOT:$PYTHONPATH

Apply Changes

source ~/.bashrc  # or source ~/.zshrc

💡 Tip: After adding the above to your shell config file, these environment variables will be automatically loaded every time you open a new terminal.

---

📖 Usage

Quick Start

Follow these steps to set up the dataset:

Step	Action	Description
1	Download dataset	Use Hugging Face Hub or Git Clone
2	Extract scene assets	Extract split archives in data/sim_engine/assets/
3	Set environment variables	Configure WORLDENGINE_ROOT and related paths
4	Create symlinks	Link raw datasets (if needed)
5	Verify installation	Run the quick test script

Detailed Setup

4. Create Symlinks (Optional)

If you have already downloaded nuPlan and OpenScene data, use symlinks to avoid data duplication:

cd WorldEngine/data/raw
ln -s /path/to/nuplan nuplan
ln -s /path/to/openscene-v1.1 openscene-v1.1
cd openscene-v1.1
ln -s ../nuplan/maps maps

Next Steps

After dataset setup, refer to the main project documentation:

• 📘 Installation Guide
• 🚀 Quick Start
• 🎮 SimEngine Usage Guide
• 🧠 AlgEngine Usage Guide

---

📝 Citation

If this project is helpful to your research, please consider citing:

@misc{li2026worldengineeraposttraining,
      title={World Engine: Towards the Era of Post-Training for Autonomous Driving}, 
      author={Tianyu Li and Li Chen and Caojun Wang and Haochen Liu and Kashyap Chitta and Zhenjie Yang and Yuhang Lu and Naisheng Ye and Yihang Qiu and Yufei Wang and Luoxi Zou and Jiaxin Peng and Jin Pan and Zhaoyu Su and Andrei Bursuc and Shengbo Eben Li and Andreas Geiger and Peng Su and Hongyang Li},
      year={2026},
      eprint={2606.19836},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2606.19836}, 

---

📄 License

This dataset is released under the CC-BY-NC-SA-4.0 license.

Terms of Use

• ✅ Allowed: Modification, distribution, private use
• 📝 Required: Attribution, share alike
• ⚠️ Restricted: No commercial use; copyright and license notices must be retained

---

🔗 Related Links

Resource	Link
🏠 Project Home	WorldEngine GitHub
🤗 Hugging Face	Dataset Page
📦 ModelScope	Dataset Page
💬 Discussions	Hugging Face Discussions
📖 Full Documentation	Documentation
🎨 Scene Reconstruction	MTGS Repository

---

📧 Contact

For questions or suggestions, feel free to reach out:

• 🐛 Bug Reports: GitHub Issues
• 💬 Discussions: Hugging Face Discussions

---

⭐ If you find WorldEngine useful, please consider giving us a Star! ⭐

Thank you for your support of the WorldEngine project!