BevAD

Official code release for What Matters for Scalable and Robust Learning in End-to-End Driving Planners? — accepted at CVPR Findings 2026.

David Holtz, Niklas Hanselmann, Simon Doll, Marius Cordts and Bernt Schiele
Mercedes-Benz AG & Max-Plack-Institute for Informatics, SIC

---

📰 News

• 2026-06 — Initial code and checkpoint release.
• 2026-06 — Publish Fail2Drive results.
• 2026-03 — Paper release on arXiv.

🌐 Overview

bevad-workspace/
├── bevad/                  # Main codebase
├── bevad-sim/              # Connector between bevad and CARLA
├── checkpoints/            # Model checkpoints (download separately)
├── data/b2d-xml/           # Bench2Drive route definitions (XML)
├── external/               # Third-party dependencies (CARLA, mmcv)
├── recordings/             # Simulation episode stored here (created by inference.py)
└── inference.py            # Example script for closed-loop inference

⚙️ Installation

Requirements:

• Ubuntu 22.04 (recommended)
• Python 3.10
• CUDA 12 with a GPU of compute capability ≥ 7.0
• CARLA 0.9.15 with additional large maps installed
• uv package manager (recommended)

Setup:

# Set the number of parallel compilation jobs (adjust to your CPU core count)
export MAX_JOBS=24
# Specify target CUDA architectures
export TORCH_CUDA_ARCH_LIST="7.0+PTX 7.5 8.0+PTX"
# Install all dependencies
uv sync -p 3.10

Note: Installation takes 5–15 minutes depending on hardware. The majority of time is spent compiling mmcv from source — setting MAX_JOBS to match your available CPU cores speeds this up significantly.

🚗 Closed-Loop Simulation

1. Download the BevAD checkpoint from Hugging Face and place it at checkpoints/bevad-m.ckpt.
2. Run closed-loop simulation on a Bench2Drive route:

python inference.py \
    --checkpoint checkpoints/bevad-m.ckpt \
    --config bevad/bevad/configs/cvpr/scaling_diffusion.py \
    --route data/b2d-xml/b2d-24224.xml \
    --output-dir recordings

All arguments are optional and default to the values shown above. To run a different route, replace the --route path with any XML file from data/b2d-xml/.

🐘 Fail2Drive

Fail2Drive evaluates both in-distribution performance and generalization under distribution shift. These results were conducted by Karol Fedurko during his research internship in our lab.

Method	In-Distribution DS ↑	In-Distribution SR ↑	In-Distribution HM ↑	Generalization DS ↑	Generalization SR ↑	Generalization HM ↑
TCP	24.7	39.1	30.3	24.5	31.4	27.5
UniAD	47.5	36.3	41.2	44.0	27.6	33.9
Orion	53.0	52.0	52.5	51.2	46.0	48.5
HiP-AD	74.1	70.7	72.4	67.1	56.7	61.5
SimLingo	82.6	79.3	80.9	71.7	55.0	62.2
TF++	83.3	78.5	80.8	75.4	61.1	67.5
PlanT 2.0	87.8	85.0	86.4	73.3	58.0	64.8
BevAD (ours)	87.4	83.3	85.3	82.3	68.7	74.9

📚 Citation

@InProceedings{Holtz_2026_CVPRF,
    author    = {Holtz, David and Hanselmann, Niklas and Doll, Simon and Cordts, Marius and Schiele, Bernt},
    title     = {What Matters for Scalable and Robust Learning in End-to-End Driving Planners?},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings},
    month     = {June},
    year      = {2026},
    pages     = {931-941}
}

📄 License

This project is licensed under the MIT License. Note that it includes third-party components in the external/ directory that are subject to their own license terms. See the LICENSE file for details.