---
license: mit
pretty_name: EPCC checkpoints and data
tags:
- robotics
- world-model
- model-predictive-control
- diffusion
- imitation-learning
---

# EPCC: checkpoints and data

Checkpoints, evaluation datasets and published run directories for
**Expert-Play Contouring Control: Faster-than-Demonstration Planning from Slow Expert and Fast Play**
([arXiv:2609.22798](https://arxiv.org/abs/2609.22798)).
Code: [github.com/epcc-robotics/epcc](https://github.com/epcc-robotics/epcc).

## Download

```bash
hf download epcc-robotics/epcc --repo-type dataset --local-dir epcc_data   # about 100 GB
export STABLEWM_HOME=$PWD/epcc_data/data              # read by every evaluation cell
export EPCC_PUBLISHED=$PWD/epcc_data/published_runs   # read by the Table I / II scorers
```

To rescore the paper's runs without running any policy, download only `published_runs/` (about 200 MB):

```bash
hf download epcc-robotics/epcc --repo-type dataset --local-dir epcc_data --include "published_runs/*"
```

## Contents

`data/` is the `STABLEWM_HOME` tree that the paper's evaluation cells
(`config/eval/cells/paper/` in the code repository) reference. `docs/DATA.md` in the code
repository lists every file with the cells that use it. File names are the ones the cells
were run with.

| Path | Contents |
|---|---|
| `data/checkpoints/agg_plainM2_object.ckpt`, `agg_slipM1_object.ckpt`, `agg_slipM2_object.ckpt`, `agg_cubeM2_object.ckpt` | EPCC world models (PushT, Slippery-PushT, Cube), Table I and II |
| `data/checkpoints/dtg/` | contour generators (latent diffusion, one per task), Table I and II, Fig. 4, data scaling |
| `data/checkpoints/pusht_wm/`, `slippery_pusht/lewm/`, `cube_wm/` | base world models referenced by the baseline cells (`pusht_wm/lewm_object.ckpt` is the public LeWM PushT checkpoint) |
| `data/checkpoints/dp/`, `skip/`, `sail/`, `data/cube/dp/`, `data/cube/sail/` | Diffusion Policy, SkiP and SAIL baselines |
| `data/{pusht,slippery_pusht,cube}/sup/` | SuP baseline (ADM models and IQL gate) |
| `data/pusht_invdyn/`, `data/slippery_pusht/invdyn_base/`, `data/cube/invdyn/` | inverse-dynamics models (Table II) |
| `data/checkpoints/fricd0*0_object.ckpt`, `data/checkpoints/friction/` | world models and contour generators per friction level, Fig. 3a |
| `data/checkpoints/f4cold/`, `wm_demoonly/`, `wm_playsub/` | world models of the data-composition study, Fig. 4 |
| `data/checkpoints/scaling/` | world models of the data-scaling study (appendix) |
| `data/datasets/` | evaluation datasets: slow expert demonstrations for PushT, Slippery-PushT and Cube (initial states and goals) |
| `data/slippery_pusht/data/`, `data/{pusht,cube}/latents/` | demonstration sets and their latent encodings used by the contour generators |
| `data/{pusht,slippery_pusht,cube}/sail/awe/` | SAIL waypoint labels |
| `data/stats/` | observation / action scaler statistics of each world model's training corpus (the corpora themselves are not included) |
| `published_runs/findings/260907_final_dtg_s1_warm/` | Table I (EPCC, Play+Goal) and Table II run directories and scores |
| `published_runs/findings/261023_sup_robot/` | Table I SuP run directories and scores |
| `published_runs/findings/260521_baseline_setup/` | baseline run directories read by the Table I scorer (Diffusion Policy, SkiP, SAIL at each k, SuP) |

Each run directory holds `rollout_traj.npz` (states, actions and success flags of every
episode), `meta.json` (the resolved configuration), the harness's `results.json` and, depending on
the policy, control-timing traces and candidate decisions. The paper's
success rates come from the scorers in the code repository (`paper/common/score_runs.py`), not
from `results.json`.

Absolute paths recorded in configurations, checkpoint metadata and HDF5 attributes were shortened to
`${STABLEWM_HOME}/…`, `${REPO}/…` and `${TMPDIR}/…`, and the experiment-logging account name was
replaced by `<wandb-entity>`; the paths document provenance and need not match this layout. The
SAIL label files (`data/*/sail/awe/*.h5`) keep HDF5 external links to the images of the
demonstration file they were labelled from; that file is not included and evaluation does not
read these links.

## License

MIT, as the code. The PushT world models build on the
[LeWorldModel](https://github.com/lucas-maes/le-wm) PushT checkpoint (MIT).

```
Copyright (c) 2026 Lucas Maes
Copyright (c) 2026 the EPCC authors

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
```

## Citation

```bibtex
@article{cho2026epcc,
  title={Expert-Play Contouring Control: Faster-than-Demonstration Planning from Slow Expert and Fast Play},
  author={Seunghoon Cho and Wonsuhk Jung and Sundhar Vinodh Sangeetha and Shreyas Kousik},
  journal={arXiv preprint arXiv:2609.22798},
  year={2026}
}
```