LEAD: Minimizing Learner–Expert Asymmetry in End-to-End Driving
Project Page | Paper | Code
Official model weights for LEAD and TransFuser v6 (TFv6), an expert-student policy pair for autonomous driving research in the CARLA simulator.
LEAD addresses the misalignment between privileged expert demonstrations and sensor-based student observations in imitation learning. By narrowing these asymmetries, the TFv6 student policy achieves state-of-the-art performance on major CARLA closed-loop benchmarks.
Main Features
- Lean pipeline: Pure PyTorch implementation with minimal dependencies.
- Cross-dataset training: Support for NAVSIM and Waymo datasets, with optional co-training on synthetic CARLA data.
- Data-centric infrastructure: Enforced tensor typing with BearType/JaxTyping and extensive visualizations for debugging.
- State-of-the-Art Performance: TFv6 reaches 95 DS on Bench2Drive and significantly outperforms prior models on Longest6 v2 and Town13.
Evaluation Results (Bench2Drive)
| Method | Driving Score (DS) | Success Rate (SR) |
|---|---|---|
| TF++ (TFv5) | 84.21 | 67.27 |
| TFv6 (Ours) | 95.28 | 86.80 |
Usage
For setup instructions, data collection, and evaluation scripts, please refer to the official GitHub repository and the full documentation.
Example evaluation command:
bash scripts/start_carla.sh # Start CARLA server
bash scripts/eval_bench2drive.sh # Evaluate one Bench2Drive route
Citation
If you find this work useful, please cite:
@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.