README.md

---
license: mit
tags:
  - robotics
  - imitation-learning
  - diffusion-policy
  - robomimic
  - manipulation
---

# ReGuide — Checkpoints

[![arXiv](https://img.shields.io/badge/arXiv-2606.28939-b31b1b.svg)](https://arxiv.org/abs/2606.28939)
[![Project Page](https://img.shields.io/badge/Project-Page-blue.svg)](https://reguide-project.github.io)
[![Code](https://img.shields.io/badge/GitHub-Code-181717.svg?logo=github)](https://github.com/tzuhsiangl/reguide)

Pretrained **diffusion policy** and **visual dynamics model** checkpoints for
**ReGuide: From Test-Time Guidance to Self-Improving Diffusion Policies**.

ReGuide is a self-improving framework that recycles *guided rollouts* as on-policy
recovery data to fix covariate shift in behavior-cloned diffusion policies, improving
base-policy success by **1.3–7.7×** on Robomimic. See the
[paper](https://arxiv.org/abs/2606.28939) and [project page](https://reguide-project.github.io)
for method details.

Checkpoints are organized by benchmark:

- [`robomimic/`](robomimic) — Can, Square, Transport, Tool Hang (available now)
- `libero/` — coming soon

## Robomimic tasks

| Task        | Demos | Notes |
|-------------|:-----:|-------|
| `can`       | 15    | policy + dynamics model |
| `square`    | 30    | policy + dynamics model |
| `transport` | 10    | policy + dynamics model |
| `tool_hang` | 80    | policy + dynamics model |

## Repository layout

```
robomimic/diffusion_policy/<task>/<variant>/
├── checkpoints/<epoch>.ckpt     # policy weights (~4.3–5.3 GB)
├── normalizer.pth               # observation/action normalizer
└── .hydra/                       # full training config

robomimic/dyn_model/<task>/
├── checkpoints/model_<epoch>.pth # visual dynamics model (~0.5–1.5 GB)
├── normalizer.pth
└── hydra.yaml                    # training config
```

The number in each checkpoint filename (e.g. `880.ckpt`) is the training epoch.

### Diffusion policy variants

| Variant                 | What it is |
|-------------------------|------------|
| `base_policy`           | Diffusion policy trained on the original demonstrations only. |
| `ReGuide-FS`            | Retrained from scratch on demos + guided rollouts. |
| `ReGuide-FT-iteration1` | Base checkpoint fine-tuned on demos + guided rollouts (first iteration). |
| `ReGuide-FT-iteration2` | Second fine-tuning iteration on freshly collected guided rollouts. |
| `ReGuide-FS-FT`         | ReGuide-FT applied on top of a ReGuide-FS policy (best on Can/Square/Transport). |

Each checkpoint ships with its full Hydra training config (`.hydra/` for policies,
`hydra.yaml` for dynamics models), which records the exact hyperparameters used. Paths
inside the configs are relative.

## Code

Training and evaluation code: https://github.com/tzuhsiangl/reguide

## Citation

```bibtex
@article{lin2026reguide,
  title         = {ReGuide: From Test-Time Guidance to Self-Improving Diffusion Policies},
  author        = {Lin, Tzu-Hsiang and Shakkottai, Srinivas and Kalathil, Dileep and Kumar, P. R.},
  journal       = {arXiv preprint arXiv:2606.28939},
  year          = {2026},
  eprint        = {2606.28939},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2606.28939}
}
```

## License

Released under the [MIT License](https://opensource.org/license/mit).