README.md

---
pretty_name: ESRP-PD
task_categories:
- robotics
- reinforcement-learning
language:
- en
tags:
- embodied-ai
- rearrangement
- indoor-scenes
- omnigibson
- 3d-front
- imitation-learning
- benchmark
size_categories:
- 1K<n<10K
viewer: false
---

# ESRP-PD: Paired Scene Data for Embodied Scene Rearrangement Planning

Scene pairs and expert demonstrations for **Embodied Scene Rearrangement Planning (ESRP)**, published at
IEEE Robotics and Automation Letters (RA-L), 2026.

[Project page](https://bit-pie.github.io/ESRP/) · [Code](https://github.com/BIT-PIE/ESRP) · Built on
[OmniGibson](https://behavior.stanford.edu/omnigibson/) with scenes from
[3D-FRONT](https://tianchi.aliyun.com/specials/promotion/alibaba-3d-scene-dataset)

In ESRP, an embodied agent rearranges furniture in a fully 3D indoor scene to match a target layout given as a
single top-down image, using only egocentric RGB observations. This dataset provides the initial-target scene
pairs the agent is evaluated on, the train/validation/test splits used in the paper, and an example set of expert
demonstrations for imitation learning.

**5,549 scene pairs** across bedrooms, living rooms, dining rooms, libraries, and other room types, holding more
than 8,100 movable furniture items.

## Files

| File | Size | Contents | Needed for |
| --- | --- | --- | --- |
| `data.zip` | 27.5 GB | `data/3d_front/scenes/` (the scene pairs) and `data/3d_front/usd_objects/` (converted furniture USD models), plus `data/assets/` | Running the simulator: every baseline |
| `train_data.txt`, `valid_data.txt`, `test_data.txt` | 313 KB | The split lists, one scene name per line (also bundled as `data_splits.zip`) | Every baseline |
| `imitation_example.zip` | 83 MB | `imitation_example/*.npz`, one expert demonstration per scene | ESRP-BC (imitation learning) |
| `3d_front.zip` | 3.7 GB | `3D-FRONT/*.json`, the raw 3D-FRONT scene descriptions | Only if you regenerate scene pairs from scratch |

Download everything, or just the parts you need:

```bash
pip install -U huggingface_hub

# Everything (~31 GB)
hf download serendipity800/ESRP-PD --repo-type dataset --local-dir ESRP-PD

# Just what you need to run the benchmark
hf download serendipity800/ESRP-PD data.zip train_data.txt valid_data.txt test_data.txt \
    --repo-type dataset --local-dir ESRP-PD
```

On older versions of the client, `hf download` is called `huggingface-cli download`.

## Setup

Install ESRP-Bench first by following [INSTALL.md](https://github.com/BIT-PIE/ESRP/blob/main/INSTALL.md), then:

1. Unzip `data.zip` and move the resulting `data/` folder into the `omnigibson/` package directory, so that you
   end up with `omnigibson/data/3d_front/` and `omnigibson/data/assets/`. If you keep the data somewhere else,
   point `ThreeD_FRONT_DATASET_PATH` in `omnigibson/macros.py` at it instead.
2. Place `train_data.txt`, `valid_data.txt`, and `test_data.txt` into `omnigibson/data/3d_front/`. Some scripts
   read the lists relative to the repository root, so keep a copy in the directory above `omnigibson/` as well.
3. Optional, for imitation learning: unzip `imitation_example.zip` into a directory named `imitation_data/`
   alongside `omnigibson/`, then split it with `python -m omnigibson.baseline.IL.split_scenes`.
4. Build the mesh materials once, before running anything else:

   ```bash
   python -m omnigibson.create_mesh
   ```

## Scene pair format

Every scene lives in its own directory under `data/3d_front/scenes/`, named as the 3D-FRONT scene UID followed by
the room, for example `00110bde-f580-40be-b8bb-88715b338a2a_LivingDiningRoom-44785`:

| File | Role |
| --- | --- |
| `<scene>_target.json` | Goal layout: the object poses the agent has to reproduce |
| `<scene>_initial.json` | Start layout: disarranged poses, plus a `target_bbox` for every rearrangeable object |
| `<scene>_target.png` | Top-down render of the goal layout, resized to 128×128 and given to the agent |
| `robot_pos_ori.txt` | Robot spawn position and orientation |

Both JSON files follow OmniGibson's scene-state format. Objects carry an `is_to_rearrange` flag marking the ones
the agent has to move; placement is judged against the `target_bbox` stored in the initial file. The task loads
the initial layout and derives the goal by swapping `initial` for `target` in the filename, so the two files must
stay side by side with matching names.

## Splits

| Split | Scenes | File |
| --- | --- | --- |
| Train | 4,995 | `train_data.txt` |
| Validation | 200 | `valid_data.txt` |
| Test | 354 | `test_data.txt` |

Each file is one scene directory name per line. Difficulty is defined by how many objects have to be rearranged:
easy (1 object), medium (2 to 3), and hard (4 to 6).

## About the demonstrations

`imitation_example.zip` is a **proof-of-concept subset**, not the full demonstration set used to train the
baseline in the paper. It is enough to get a training run working end to end. To train a policy that reproduces
the reported numbers, generate demonstrations yourself with the data generation pipeline described in
[DATASET.md](https://github.com/BIT-PIE/ESRP/blob/main/DATASET.md).

## Citation

```bibtex
@article{chen2026esrp,
  title     = {Embodied Scene Rearrangement Planning},
  author    = {Chen, Canzhi and Wang, Zan and Zhu, Siqi and Wu, Qi and Li, Yixuan and Liang, Wei},
  journal   = {IEEE Robotics and Automation Letters},
  year      = {2026},
  publisher = {IEEE}
}
```

## Contact

Please open an issue in the [code repository](https://github.com/BIT-PIE/ESRP/issues) for questions about the
dataset or the benchmark.