---
license: mit
pretty_name: PGPG — demonstrations, rollouts and pretrained policies
task_categories:
- robotics
tags:
- robotics
- robomimic
- offline-rl
- preference-based-rl
- diffusion-policy
---

# PGPG data

Data and pretrained reference policies for **[Progress-Guided Preference Generation for Aligning Generative Control Policies](https://pgpg2026.github.io/)** (CoRL 2026).

Code: [github.com/tunglm2203/pgpg](https://github.com/tunglm2203/pgpg) · Paper: [arXiv:2610.25692](https://arxiv.org/abs/2610.25692)

## Contents

Five robomimic low-dim tasks: `lift`, `can`, `square`, `transport`, `tool_hang`.

| folder | per task | what it is |
|---|---|---|
| `data/<task>.tar.gz` | HDF5 | expert demonstrations + 500 policy rollouts |
| `videos/<task>.tar` | MP4 | the rendered episodes for both of the above |
| `pretrain/<task>.tar` | 2 × `.ckpt` | reference policies π_ref: diffusion and rectified-flow |

`videos/` is **required**, not optional. Visual progress scoring encodes the demonstration
videos with R3M on every run, so training fails without them.

## Usage

The code ships a script that downloads, checksums and extracts in one step:

```sh
bash bash_script/others/download_data.sh                 # all five tasks
bash bash_script/others/download_data.sh square lift     # only these
```

Manually:

```sh
huggingface-cli download tungluu2203/pgpg-data --repo-type dataset \
  --local-dir /tmp/pgpg --include "data/square.tar.gz" "videos/square.tar" "pretrain/square.tar"

# every archive stores repo-root-relative paths, so extraction is always the same
for a in /tmp/pgpg/*/square.tar*; do tar -xf "$a" -C /path/to/pgpg; done
```

This yields `logs/collect_data/<task>_lowdim/...` and `logs/pretrain/<task>_lowdim/...`.

## How the demonstrations were produced

They are the [robomimic](https://robomimic.github.io/) multi-human (`mh`) datasets — proficient-human
(`ph`) for `tool_hang` — **replayed in simulation**. `actions`, `states`, `obs` and `dones` are
bit-identical to the originals, and all mask splits are preserved. The replay adds:

- `success`: per-step binary flag, equal to the original sparse robomimic reward.
- `rewards`: overwritten with dense shaped rewards on `lift`, `can` and `square`. `transport` and
  `tool_hang` stay sparse, because robosuite implements no shaped reward for those two environments.
- `videos/`: rendered during the same replay.

Rollouts are 500 episodes per task, collected by rolling out the pretrained policies in `pretrain/`.
They are not part of the robomimic release. You can regenerate them with `bash_script/data_collect/`
in the code repository, but sampling is stochastic, so the episodes will not be the same ones.

To regenerate the demonstrations instead of downloading them, see `bash_script/data_collect/` in the
code repository.

## License and attribution

MIT, following robomimic. Derived from the robomimic datasets (Mandlekar et al., CoRL 2021); please
cite robomimic alongside this work if you use these files.

```
@inproceedings{
    luu2026progress,
    title={Progress-Guided Preference Generation for Aligning Generative Control Policies},
    author={Tung M. Luu and Tri Ton and Duc M. Nguyen and Xuan Thanh Nguyen and Nguyen Gia Binh and Quang Tan Nguyen and Vu Duong and Daniel Sonntag and Duy M. H. Nguyen and Vien Anh Ngo and Chang D. Yoo},
    booktitle={Conference on Robot Learning},
    year={2026}
}
```