LEAD: Minimizing Learner–Expert Asymmetry in End-to-End Driving

Project Page | Documentation | Checkpoints | Dataset | Supplementary Material | Paper

video.mp4

Driving stress test: Closed-loop execution of TransFuser v6, the latest iteration of the TransFuser family, demonstrating stable control in a complex urban scenario under degraded perception and adversarial traffic.

Overview

We release the complete pipeline (covering routes description, expert driver, data preprocessing, training, and evaluation) required to achieve state-of-the-art closed-loop performance on the Bench2Drive benchmark. Built around the CARLA simulator, the stack features a data-centric design with:

• Extensive visualization suite and runtime type validation for easier debugging.
• Optimized storage format, packs 72 hours of driving in ~200GB.
• Native support for real-world datasets, including NAVSIM and Waymo Vision-based E2E.

Table of Contents

• Roadmap
• Updates
• Quick Start (Get Driving in 20 Minutes)
• Further Documentation
• External Resources
• Acknowledgements
• Citation
• License

Roadmap

• ✅ Checkpoints and inference code (stable)
• 🚧 Documentation, training pipeline and expert code (partial release)
• Full dataset release on HuggingFace
• Cross-dataset training tools and documentation

Status: Active development. Core code and checkpoints are released; remaining components coming soon.

Updates

• 2025/12/24 Arxiv paper and code release

Quick Start (Get Driving in 20 Minutes)

1. Environment initialization

Clone the repository and map the project root to your environment

git clone https://github.com/autonomousvision/lead.git
cd lead

# Set the project root directory and configure paths for CARLA, datasets, and dependencies.
{
  echo -e "export LEAD_PROJECT_ROOT=$(pwd)"  # Set project root variable
  echo "source $(pwd)/scripts/main.sh"       # Persist more environment variables
} >> ~/.bashrc  # Append to bash config to persist across sessions

source ~/.bashrc  # Reload config to apply changes immediately

Note

Please verify that ~/.bashrc reflects these paths correctly.

2. Setup experiment infrastructure

We utilize Miniconda and conda-lock for a more deterministic build:

# Install conda-lock and create conda environment
pip install conda-lock && conda-lock install -n lead conda-lock.yml
# Activate conda environment
conda activate lead
# Install dependencies and setup git hooks
pip install uv && uv pip install -r requirements.txt && uv pip install -e .
# Install other tools needed for development
conda install conda-forge::ffmpeg conda-forge::parallel conda-forge::tree conda-forge::gcc
# Optional: Activate git hooks
pre-commit install

While waiting for dependencies installation, we recommend CARLA setup on parallel:

bash scripts/setup_carla.sh # Download and setup CARLA at 3rd_party/CARLA_0915

3. Model zoo

Pre-trained driving policies are hosted on HuggingFace for reproducibility. These checkpoints follow the TFv6 architecture (detailed in Fig. 1), but differ in their sensor configurations, vision backbones or dataset composition.

Figure 1: TFv6 architecture.

Tab. 1 shows available checkpoints with their performance on three major CARLA benchmarks. As first step, we recommend tfv6_resnet34 as it provides a good balance between performance and resource usage.

Checkpoint	Description	Bench2Drive	Longest6 v2	Town13
tfv6_regnety032	TFv6	95.2	62	5.01
tfv6_resnet34	ResNet34 Backbone	94.7	57	3.31
4cameras_resnet34	Additional rear camera	95.1	53	-
noradar_resnet34	No radar sensor	94.7	52	-
visiononly_resnet34	Vision-only driving model	91.6	43	-
town13heldout_resnet34	Generalization evaluation	93.1	52	2.65

Table 1: Performance of pre-trained checkpoints. We report Driving Score, for which higher is better.

To download one checkpoint:

bash scripts/download_one_checkpoint.sh

Or download all checkpoints at once with git lfs

git clone https://huggingface.co/ln2697/TFv6 outputs/checkpoints
cd outputs/checkpoints
git lfs pull

4. Verify driving stack

To initiate closed-loop evaluation and verify the integration of the driving stack, execute the following:

# Start driving environment
bash scripts/start_carla.sh
# Start policy on one route
bash scripts/eval_bench2drive.sh

Driving logs will be saved to outputs/local_evaluation with the following structure:

outputs/local_evaluation
├── 23687
│   ├── checkpoint_endpoint.json
│   ├── debug_images
│   ├── demo_images
│   └── metric_info.json
├── 23687_debug.mp4
└── 23687_demo.mp4

Tip

1. Disable video recording in config_closed_loop by turning off produce_demo_video and produce_debug_video.
2. If memory is limited, modify the file prefixes to load only the first checkpoint seed. By default, the pipeline loads all three seeds as an ensemble.

5. Verify autopilot

Verify the expert policy and data acquisition pipeline by executing a test run on a sample route:

# Start CARLA if not done already
bash scripts/start_carla.sh
# Run expert on one route
bash scripts/run_expert.sh

Data collected will be stored at data/expert_debug and should have following structure:

data/expert_debug
├── data
│   └── BlockedIntersection
│       └── 999_Rep-1_Town06_13_route0_12_22_22_34_45
│           ├── bboxes
│           ├── depth
│           ├── depth_perturbated
│           ├── hdmap
│           ├── hdmap_perturbated
│           ├── lidar
│           ├── metas
│           ├── radar
│           ├── radar_perturbated
│           ├── results.json
│           ├── rgb
│           ├── rgb_perturbated
│           ├── semantics
│           └── semantics_perturbated
└── results
    └── Town06_13_result.json

Further Documentation

For more detailed instructions, see the full documentation. In particular:

• Data collection
• Training
• Evaluation

Acknowledgements

Special thanks to carla_garage for the foundational codebase. We also thank the creators of the numerous open-source projects we use:

• PDM-Lite, leaderboard, scenario_runner, NAVSIM, Waymo Open Dataset

Other helpful repositories:

• SimLingo, PlanT2, Bench2Drive Leaderboard, Bench2Drive, CaRL

Long Nguyen led development of the project. Kashyap Chitta, Bernhard Jaeger, and Andreas Geiger contributed through technical discussion and advisory feedback. Daniel Dauner provided guidance with NAVSIM.

Citation

If you find this work useful, please consider giving this repository a star ⭐ and citing our work in your research:

@article{Nguyen2025ARXIV,
  title={LEAD: Minimizing Learner-Expert Asymmetry in End-to-End Driving},
  author={Nguyen, Long and Fauth, Micha and Jaeger, Bernhard and Dauner, Daniel and Igl, Maximilian and Geiger, Andreas and Chitta, Kashyap},
  journal={arXiv preprint arXiv:2512.20563},
  year={2025}
}

License

This project is released under the MIT License. See LICENSE for details.