ln2697/transfuser-carla-123d

TransFuser CARLA 123D

Pre-trained TransFuser-style end-to-end driving checkpoints, trained with the kesai-labs/lead pipeline on the LEAD 123D dataset. The model consumes camera and lidar inputs and predicts waypoints for closed-loop driving on CARLA Leaderboard 2.0 routes.

Website · Blog · Paper · Code

Contents

Path	Description
resnet34_v1.5.0/seed0/	config.yaml + model_0030.pth
resnet34_v1.5.0/seed1/	same, independent training seed
resnet34_v1.5.0/seed2/	same, independent training seed

Checkpoints are named after their backbone and the lead release that produced them. Each directory is self-contained: the config.yaml next to the weights is the exact training configuration and is all the code needs to rebuild the model.

These checkpoints belong to the main branch of the LEAD repository and are not compatible with the cvpr2026 branch. The checkpoints for the CVPR paper are hosted on that branch's release instead.

Results

Closed-loop Driving Score, three training seeds:

Benchmark	seed0	seed1	seed2	overall
Bench2Drive	92.9	94.7	93.2	93.6 ± 1.0
Longest6 v2	57.1 ± 4.1	50.3 ± 3.8	55.6 ± 2.8	54.3 ± 4.4

Bench2Drive is one evaluation run per seed; Longest6 is three evaluation runs per seed (nine total), so the Longest6 mean also absorbs evaluation variance.

Training

Trained on the ~1 TB LEAD 123D dataset (normal + perturbated view) for 60 epochs, roughly 40 hours on 4×H100.

Usage

Follow the setup in kesai-labs/lead, then point the evaluation scripts at a checkpoint directory to drive the model closed-loop in CARLA 0.9.16.

Citation

If these checkpoints are useful to you, 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},
}

@article{Dauner2026ARXIV,
  author  = {Dauner, Daniel and Charraut, Valentin and Berle, Bastian and Li, Tianyu and Nguyen, Long and Wang, Jiabao and Jing, Changhui and Igl, Maximilian and Caesar, Holger and Ivanovic, Boris and Geiger, Andreas and Chitta, Kashyap},
  title   = {123D: Unifying Multi-Modal Autonomous Driving Data at Scale},
  journal = {arXiv preprint arXiv:2605.08084},
  year    = {2026},
}