---
pretty_name: SAGE Manipulation Datasets
task_categories:
  - robotics
tags:
  - robotics
  - world-model
  - planning
  - libero
  - robotwin
---

# SAGE Datasets

Manipulation datasets for **[SAGE: Subgoal-Conditioned Action Generation for Latent World Model Planning](https://arxiv.org/abs/2607.17973)**.

[Paper](https://arxiv.org/abs/2607.17973) | [Code](https://github.com/PKU-ML/SAGE) | [Models](https://huggingface.co/CLTRAY/SAGE)

## Data

| Environment | File prefix | Train | Validation | Evaluation |
|:---|:---|---:|---:|---:|
| LIBERO Scene2 | `libero_scene2` | 78.92 GB | 9.88 GB | 19.34 GB |
| LIBERO Caddy | `libero_caddy` | 68.43 GB | 8.56 GB | 8.64 GB |
| RoboTwin A2B | `robotwin_a2b` | 70.58 GB | 8.95 GB | 1.36 GB |

Sizes are archive sizes in decimal GB. Training and validation archives are under `training/`; evaluation archives are under `evaluation/`. Evaluation archives contain the inputs required by the released query manifests, not a complete training dataset. Simulator assets and model weights are installed separately.

## Download

Install the download client:

```bash
pip install -U huggingface_hub
```

Download only the evaluation data you need, for example LIBERO Scene2:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="CLTRAY/SAGE-data",
    repo_type="dataset",
    allow_patterns=["evaluation/libero_scene2_evaluation.tar"],
    local_dir="sage-data",
)
```

To download Scene2 training and validation instead, use `allow_patterns=["training/libero_scene2_*"]`. Replace `libero_scene2` with `libero_caddy` or `robotwin_a2b` for the other environments. Omitting `allow_patterns` downloads all data (approximately 275 GB).

## Extract

Evaluation and validation archives are ordinary tar files:

```bash
mkdir -p datasets/libero_scene2
tar -xf sage-data/evaluation/libero_scene2_evaluation.tar -C datasets/libero_scene2
```

Training archives are split into numbered binary parts. **Merge them before extraction; individual parts are not separate tar archives.** After downloading all parts and the corresponding `.parts.json`, run:

```python
import hashlib
import json
from pathlib import Path

root = Path("sage-data")
manifest = json.loads((root / "training/libero_scene2_train.tar.parts.json").read_text())
archive = root / "training" / manifest["archive"]
full_hash = hashlib.sha256()
with archive.open("wb") as output:
    for part in manifest["parts"]:
        source = root / part["path"]
        assert source.stat().st_size == part["bytes"], source
        part_hash = hashlib.sha256()
        with source.open("rb") as stream:
            for block in iter(lambda: stream.read(8 * 1024 * 1024), b""):
                part_hash.update(block)
                full_hash.update(block)
                output.write(block)
        assert part_hash.hexdigest() == part["sha256"], source
assert archive.stat().st_size == manifest["bytes"]
assert full_hash.hexdigest() == manifest["sha256"]
print(f"Verified: {archive}")
```

Then extract the merged archive:

```bash
mkdir -p datasets/libero_scene2_train
tar -xf sage-data/training/libero_scene2_train.tar -C datasets/libero_scene2_train
tar -xf sage-data/training/libero_scene2_validation.tar -C datasets/libero_scene2_train
```

Allow disk space for the downloaded parts, merged archive, and extracted files. For environment setup, query manifests, and training/evaluation commands, see the [native benchmark guide](https://github.com/PKU-ML/SAGE/blob/main/NATIVE.md).

## Citation

```bibtex
@article{cheng2026sage,
  title={SAGE: Subgoal-Conditioned Action Generation for Latent World Model Planning},
  author={Cheng, Letian and Zhang, Qi and Wang, Yisen},
  journal={arXiv preprint arXiv:2607.17973},
  year={2026}
}
```