DONUT: A Decoder-Only Model for Trajectory Prediction

ICCV 2025

arXiv | Project Page | YouTube | BibTeX

Markus Knoche¹, Daan de Geus^1,2, Bastian Leibe¹

¹ RWTH Aachen University ² Eindhoven University of Technology

Installation

Clone repository:

git clone https://github.com/MKnoche/DONUT.git
cd DONUT

Install dependencies either using uv:

pip install uv
uv sync
source .venv/bin/activate

or using pip:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Training

Adjust the root paths in train_donut.py according to your setup. Data will be downloaded and preprocessed automatically when running for the first time. This will take some time.

Distributed training is supported via the devices and nodes parameters. Gradient accumulation makes sure that the effective batch size is always acc_batch_size, as long as batch_size * devices * nodes <= acc_batch_size.

Training for 60 epochs on 4 NVIDIA H100 GPUs with a batch size of 8 per GPU takes about 4.5 days.

python train_donut.py

Evaluation

python eval_donut.py

Model Checkpoint

Checkpoint	b-minFDE6	minFDE6	minADE6	MR6
DONUT	1.814	1.181	0.726	0.144

Store the checkpoint as {args.ckpt_root}/donut/donut.ckpt.

Citation

If you use our work in your research, please use the following BibTeX entry.

@inproceedings{knoche2025donut,
  title     = {{DONUT: A Decoder-Only Model for Trajectory Prediction}},
  author    = {Knoche, Markus and de Geus, Daan and Leibe, Bastian},
  booktitle = {ICCV},
  year      = {2025}
}

Acknowledgements

This project builds upon code from QCNet (Apache-2.0 License).