Rosettasim

RosettaSim

Official model weights for:

Long-term Traffic Simulation via Structured Autoregressive Modeling
Lingyu Xiao, Zexin Feng, Xintao Yan
ECCV 2026

Paper: https://arxiv.org/abs/2606.31209

Model Description

RosettaSim is a structured autoregressive framework for interactive and long-term traffic simulation.

The model represents scene topology, agent states, motion, and agent spawning as a structured autoregressive sequence with variable token cardinality, enabling both short-term traffic simulation and long-horizon rollouts with agents entering and leaving the scene.

The released model is based on Qwen2.5-0.5B-Instruct and is trained using the Waymo Open Motion Dataset (WOMD).

Intended Use

This model is intended for non-commercial research, teaching, scientific publication, benchmarking, and personal experimentation in areas including:

• traffic simulation
• autonomous-driving research
• multi-agent motion generation
• long-horizon traffic modeling

This model is not licensed for commercial or production use.

In particular, users must comply with the restrictions of the Waymo Dataset License Agreement for Non-Commercial Use.

License

Important: this model was trained using the Waymo Open Dataset.

This model was made using the Waymo Open Dataset, provided by Waymo LLC under the Waymo Dataset License Agreement for Non-Commercial Use, available at:

https://waymo.com/open/terms/

Access and use of this model are governed by the terms and conditions of that agreement.

This model, including its released weights, is distributed under the Waymo Dataset License Agreement for Non-Commercial Use. Any downstream use, redistribution, modification, or derivative model must remain subject to the applicable terms of that agreement, including its non-commercial-use restrictions.

Users should review the complete Waymo Dataset License Agreement before using or redistributing this model.

The underlying Qwen2.5-0.5B-Instruct model is separately subject to its applicable license and terms.

Restrictions

Under the Waymo Dataset License Agreement, this model and its weights must not be used:

• in the operation of a vehicle or to assist in the operation of a vehicle;
• in production systems;
• for primarily commercial purposes.

Please refer to the complete Waymo Dataset License Agreement for the authoritative terms and definitions.

Training Data

RosettaSim is trained and evaluated using the Waymo Open Motion Dataset (WOMD), including the Waymo Open Sim Agents Challenge (WOSAC) setting.

The Waymo Open Dataset is not redistributed in this model repository.

Users who wish to reproduce training or evaluation should obtain the dataset independently from:

https://waymo.com/open/

and accept the applicable Waymo Dataset License Agreement.

Base Model

RosettaSim uses:

Qwen2.5-0.5B-Instruct

as its autoregressive sequence-model backbone.

Citation

If you find this model useful, please cite:

@article{xiao2026longterm,
  title={Long-term Traffic Simulation via Structured Autoregressive Modeling},
  author={Xiao, Lingyu and Feng, Zexin and Yan, Xintao},
  journal={arXiv preprint arXiv:2606.31209},
  year={2026}
}