clap-ad/clap-ad-ltfv6-navsim

CLAP-AD with LTFv6 for NAVSIM

This repository contains the NAVSIM LTFv6 checkpoint for CLAP-AD (Closed-Loop Aware Planning via Imitation-Free Speed Scoring for End-to-End Autonomous Driving), accepted at WACV 2027. See the paper, project page, and code repository.

Model

The model uses LEAD's LTFv6 backbone and the CLAP-AD planning head, adapted to NAVSIM's four-second planning horizon. config.json contains the inference architecture and training settings; model_0060.pth is the released checkpoint.

Evaluation

Results below are from the paper's NAVSIM v2.2 navhard two-stage protocol. Detailed evaluation output is included in results/navhard_two_stage.csv.

Stage	NC ↑	DAC ↑	DDC ↑	TLC ↑	EP ↑	TTC ↑	LK ↑	HC ↑	EC ↑
1	98.8	87.6	99.0	99.6	78.2	98.0	93.1	98.0	37.8
2	90.3	79.1	89.6	99.2	78.8	87.7	53.6	97.2	51.6

Overall EPDMS ↑: 38.0. All metric values are reported on a 0–100 scale.

NC: No at-fault Collisions; DAC: Drivable Area Compliance; DDC: Driving Direction Compliance; TLC: Traffic Light Compliance; EP: Ego Progress; TTC: Time to Collision; LK: Lane Keeping; HC: History Comfort; EC: Extended Comfort.

Use

Download the model files with huggingface_hub:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="clap-ad/clap-ad-ltfv6-navsim",
    local_dir="outputs/checkpoints/clap-ad-ltfv6-navsim",
)

Set agent.checkpoint_path to outputs/checkpoints/clap-ad-ltfv6-navsim/model_0060.pth in the NAVSIM agent configuration. The NAVSIM supplement documents the repository's NAVSIM setup and evaluation workflow.

Intended use and limitations

This checkpoint is a research artifact for NAVSIM benchmark evaluation. It is not validated for deployment in real vehicles or safety-critical settings. Reported scores depend on the NAVSIM release, split, and evaluation protocol.

License

The CLAP-AD checkpoint is released under the MIT License (see LICENSE). The image encoder uses timm/regnety_032.ra_in1k; its pretrained weights are licensed under Apache-2.0. See the upstream model card and the included LICENSE-APACHE-2.0.

Citation

If you find this work useful, please cite:

@inproceedings{aghajanzadeh2027clapad,
  title={Closed-Loop Aware Planning via Imitation-Free Speed Scoring for End-to-End Autonomous Driving},
  author={Aghajanzadeh, Emad and Schwarzfischer, Thomas and Kaufeld, Marc and Thoma, Moritz and P{\'e}goud, Ryan and Vemparala, Manoj Rohit and Mori, Pierpaolo and Fasfous, Nael and Betz, Johannes and Mueller-Gritschneder, Daniel and Schlichtmann, Ulf},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  year={2027}
}