Robotics
Safetensors
world-action-models
robotwin
SWAM / README.md
FreDERicia's picture
Use approved light UMI cover with original layout
dd79a3f verified
|
Raw History Blame Contribute Delete
2.75 kB
---
tags:
- robotics
- world-action-models
- robotwin
- safetensors
datasets:
- FreDERicia/SWAM-Pretrain
- FreDERicia/SWAM-Posttrain
---
# SWAM
[![SWAM project cover](https://huggingface.co/datasets/FreDERicia/SWAM-Pretrain/resolve/main/visual-previews/swam-cover-umi-light-original-layout.png)](https://char-lotte-xia.github.io/swam-official/)
<p align="center">
<a href="https://github.com/shaaaaron/SWAM"><img src="https://img.shields.io/badge/GitHub-Code-black.svg?logo=github&amp;logoColor=white" alt="GitHub code"></a>
<a href="https://char-lotte-xia.github.io/swam-official/"><img src="https://img.shields.io/badge/Project-Page-0A7F5F.svg" alt="Project website"></a>
<a href="https://arxiv.org/abs/2611.12870"><img src="https://img.shields.io/badge/arXiv-Paper-b31b1b.svg" alt="arXiv paper"></a>
</p>
This repository provides the pretrained and RoboTwin posttrained checkpoints for SWAM: Scaling World Action Models with Cross-Modal Alignment.
## Model checkpoints
| Checkpoint | Download directory |
| --- | --- |
| Pretrained model | [pretrain](https://huggingface.co/FreDERicia/SWAM/tree/main/pretrain) |
| RoboTwin posttrained model | [posttrain](https://huggingface.co/FreDERicia/SWAM/tree/main/posttrain) |
The posttrained model is initialized from the pretrained checkpoint and trained on RoboTwin Clean data. Each checkpoint is approximately 27.98 GB and stored as sharded Safetensors.
## Download and use
Install the Hugging Face CLI.
```bash
pip install -U huggingface_hub
```
Download the pretrained checkpoint.
```bash
hf download FreDERicia/SWAM --include "pretrain/*" --local-dir ./SWAM
```
Download the RoboTwin posttrained checkpoint.
```bash
hf download FreDERicia/SWAM --include "posttrain/*" --local-dir ./SWAM
```
See the [SWAM code repository](https://github.com/shaaaaron/SWAM) for training and evaluation entry points. Set the resource paths in your loading configuration to the local files and prepare the required Wan and UMT5 components separately.
## Related data
[Pretraining data](https://huggingface.co/datasets/FreDERicia/SWAM-Pretrain) · [Posttraining data](https://huggingface.co/datasets/FreDERicia/SWAM-Posttrain)
## Code and citation
Code is available in the [SWAM code repository](https://github.com/shaaaaron/SWAM).
If you use these models, please cite the SWAM paper.
```bibtex
@misc{zhou2026swam,
title = {SWAM: Scaling World Action Models with Cross-Modal Alignment},
author = {Yu Zhou and Xinyuan Xia and Xuankun Yang and Yinghao Yang and Wendi Chen and Bowen Liu and Jianhan Ma and Weicheng Duan and Jun Lv and Cewu Lu and Yanwei Li},
year = {2026},
eprint = {2611.12870},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.12870}
}
```