---
license: mit
---

Paper: https://arxiv.org/abs/2506.23725

## Download

Grab the full dataset as a single tarball and expand it locally:

```bash
wget https://huggingface.co/datasets/lens-lab/pacbench/resolve/main/pacbench-20250218.tar.gz
tar -xzf pacbench-20250218.tar.gz
```

After extraction you will get the `pacbench/` directory with the same structure used on the Hub (constraint images, humanoid captures, open images, RoboCasa objects, metadata, etc.).

If you prefer the classic Hub workflow, the original files remain accessible via `hf download lens-lab/pacbench --repo-type dataset`, although non-pro account requests may be rate limited.

## Citation

Please cite **PAC Bench** if you use this benchmark or framework in your research:

```bibtex
@inproceedings{pacbench2025,
  title     = {PAC Bench: Do Foundation Models Understand Prerequisites for Executing Manipulation Policies?},
  author    = {Atharva Gundawar and Som Sagar and Ransalu Senanayake},
  booktitle = {NeurIPS 2025 Datasets and Benchmarks Track},
  year      = {2025}
}
```