ln2697/ltfv6-navsim

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

⚠️ Coordinate System Warning

This model was trained in the left-handed coordinate system of CARLA (x-forward, y-right, z-up), not the ISO 8855 convention used by NAVSIM / nuPlan / most AD stacks (x-forward, y-left, z-up).

If you use ltfv6.py directly, the predicted waypoints and headings are in CARLA's left-handed frame. You must convert the planning output back to ISO 8855 before feeding it to any downstream planner, simulator, or evaluation tool that expects the right-handed convention.

✅ Recommended: use the prepared NAVSIM workspaces

For correct, reproducible evaluation, Use one of the prepared workspaces below — they already wire up the model with the correct coordinate conversion, input preprocessing, and metric computation:

• NAVSIM v1.1: 3rd_party/navsim_workspace/navsimv1.1
• NAVSIM v2.2: 3rd_party/navsim_workspace/navsimv2.2

These are the only configurations we have validated end-to-end against the reported numbers. If you evaluate outside of them, results may silently disagree with the paper.

Manual conversion (only if you must integrate the model yourself)

waypoints_iso[..., 0] =  waypoints_carla[..., 0]   # x unchanged
waypoints_iso[..., 1] = -waypoints_carla[..., 1]   # flip y
headings_iso          = -headings_carla            # flip yaw sign

Project Page | Paper | Code

Official model weights for Latent TransFuser v6 (LTFv6), a NAVSIM checkpoint accompanies our paper LEAD: Minimizing Learner–Expert Asymmetry in End-to-End Driving.

We release the complete pipeline 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 NAVSIM and Waymo Vision-based E2E and extending those benchmarks through closed-loop simulation and synthetic data for additional supervision during training.

Find more information on https://github.com/autonomousvision/lead.

Usage

Install dependencies

pip install torch timm numpy opencv-python jaxtyping beartype omegaconf huggingface_hub

See example.ipynb to inspect data format and example inference.

Data Format

We also provide example NAVSIM cache here.

Input:

• RGB: (256, 1920, 3), range [0, 255]
• Command: [left, straight, right, unknown], e.g. [0, 1, 0, 0] for straight
• Speed: m/s
• Acceleration: m/s²

Output:

• waypoints: (N, 2) predicted positions
• headings: (N,) predicted angles

Citation

If you find this work useful, please cite:

@inproceedings{Nguyen2026CVPR,
    author = {Long Nguyen and Micha Fauth and Bernhard Jaeger and Daniel Dauner and Maximilian Igl and Andreas Geiger and Kashyap Chitta},
    title = {LEAD: Minimizing Learner-Expert Asymmetry in End-to-End Driving},
    booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)},
    year = {2026},
}