SWAM-Posttrain / README.md
FreDERicia's picture
Use approved light UMI cover with original layout
38a0cca verified
|
Raw History Blame Contribute Delete
3.34 kB
metadata
pretty_name: SWAM Posttraining Data
license: cc-by-nc-sa-4.0
task_categories:
  - robotics
tags:
  - robotics
  - world-action-models
  - robotwin
  - lerobot
  - posttraining

SWAM-Posttrain

SWAM project cover

GitHub code Project website arXiv paper

This repository provides the RoboTwin Clean posttraining data used in SWAM: Scaling World Action Models with Cross-Modal Alignment.

The release contains 50 tasks and 2,500 episodes, distributed as 50 TAR archives totaling approximately 3.94 GB. Browse all data

Data format and directory structure

Each task follows a LeRobot-style layout with data/, videos/, and meta/ directories, containing actions and states, videos, language instructions, and statistics. Original episode IDs are preserved. Latents are not included.

The archives retain 16 episodes listed in the per-task meta/episode_exclusions_20261023.jsonl files. Apply these exclusions during training to obtain 2,484 eligible episodes. Task instructions are stored in meta/tasks_v1.jsonl.

Download and extract

Install the Hugging Face CLI.

pip install -U huggingface_hub

Download the dataset.

hf download FreDERicia/SWAM-Posttrain \
  --repo-type dataset --local-dir ./SWAM-Posttrain

Extract all task archives into the same directory on Linux or macOS.

mkdir -p ./lerobot_robotwin_eef_clean_50
for archive in ./SWAM-Posttrain/data/*.tar; do
  tar -xf "$archive" -C ./lerobot_robotwin_eef_clean_50
done

The extracted layout is lerobot_robotwin_eef_clean_50/<task-directory>/{data,videos,meta}/.

Source terms and attribution

The source data declares a CC BY-NC-SA 4.0 license. Please follow the upstream license and attribution requirements. The original documentation is retained in SOURCE_README.md.

Related resources

Model checkpoints · Pretraining data

Code and citation

Code is available in the SWAM code repository.

If you use this dataset, please cite the SWAM paper.

@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}
}