SWAM-Pretrain

SWAM pretraining data is organized into General, Ego, UMI, and Robotics, covering general videos, egocentric videos, UMI manipulation demonstrations, and robot demonstrations. We standardize the video format, preserve source-specific episode records and annotations, and add detailed annotations for manipulation data. This repository releases the publicly distributable portion in stages.

Data is being uploaded in batches. See the repository file list for currently available downloads.

Data scope

Category	Data sources	Release scope
General	OpenVid, Panda70M	OpenVid included; Panda70M excluded from this release
Ego	EgoVerse, Ego4D, EgoSuite	EgoVerse included; Ego4D and EgoSuite excluded from this release
UMI	UMI manipulation demonstrations	Awaiting a future release by Noematrix
Robotics	AgiBot World, RoboCoin, RoboMIND, InternData A1, Galaxea, ABC-130K	Released in batches according to the release manifest

The planned release covers 26 source directories, approximately 2.94 TB (decimal units). Directory names preserve the original organization; resolution labels describe the source data specifications.

Why some data is excluded: Panda70M, Ego4D, and EgoSuite are subject to upstream licensing and redistribution conditions; please obtain them through their official channels. UMI data awaits a future release by Noematrix, with release arrangements to be announced through its official channels.

Research data statistics

The figures below describe the paper's full pretraining data pool of approximately 47,000 hours, including Panda70M, Ego4D, EgoSuite, and UMI, which are excluded from this release. These figures do not represent the amount of data publicly released or uploaded to this repository.

Left: data sources and video hours. Right: shares of total video hours and episodes, and video-duration distributions within each category.

Data format and directory structure

• Video: 320 pixels wide by 256 pixels high, AV1 encoded in MP4 containers. Source frame rates and temporal ordering are preserved.
• Episode records: Parquet files retain the actions, states, and language annotations provided by each source. Available fields vary across datasets.
• Metadata: Source-specific metadata is retained, with video-related dimensions and codec information updated for the processed files.
• Layout: LeRobot-style data/, meta/, and videos/ directories are packaged as TAR archives. This is an archive distribution, not a single uniform tabular dataset.

Each source directory contains metadata.tar and one or more chunk-*.tar archives. Extract all archives for a source into the same source directory. For example:

archives/data/galaxea/
  metadata.tar
  chunk-000.tar
  ...
  chunk-020.tar

# After extraction:
data/galaxea/
  meta/
  data/chunk-000/...
  videos/chunk-000/...
  ...

Download and extract

Install the Hugging Face CLI:

pip install -U huggingface_hub

Download the complete Galaxea subset:

hf download FreDERicia/SWAM-Pretrain \
  --repo-type dataset \
  --include "data/galaxea/*.tar" \
  --local-dir ./archives

Extract it on Linux or macOS:

mkdir -p ./data/galaxea
for archive in ./archives/data/galaxea/*.tar; do
  tar -xf "$archive" -C ./data/galaxea
done

For a smaller download, fetch metadata.tar and one chunk:

hf download FreDERicia/SWAM-Pretrain \
  data/galaxea/metadata.tar data/galaxea/chunk-019.tar \
  --repo-type dataset --local-dir ./sample-archives

Extract archives from the same source into the same directory. For partial downloads, load only the corresponding episodes. Video readers need AV1 decoding support; consult each source's meta/info.json for field definitions.

Source terms and attribution

This collection includes data from multiple upstream projects. Their respective licenses, attribution requirements, and usage restrictions continue to apply, including non-commercial restrictions where specified. Processing and repackaging do not grant additional rights or place the entire collection under a single permissive license.

Code

The accompanying SWAM code repository integrates RLinf, DreamZero, DMuon, and RoboTwin, with entry points for training, checkpoint conversion, and RoboTwin evaluation. The code repository is currently private and requires authorized access.