README.md

---
license: other
task_categories:
  - other
tags:
  - cloth
  - deformable-objects
  - state-estimation
  - point-cloud
  - simulation
  - dedo
pretty_name: DEDO GSE Demos
size_categories:
  - 100K<n<1M
---

# DEDO GSE Demos (`gse_demos_v1.h5`)

Cloth-manipulation demonstrations collected in the [DEDO](https://github.com/contactrika/dedo)
(Dynamic Environments with Deformable Objects) simulator, used as an
out-of-distribution evaluation set for graph-based cloth **state estimation (GSE)**.

Each demo is a single trajectory of a deforming cloth, recorded with synchronized
ground-truth mesh state, a single-camera depth point cloud, and an RGB image per
timestep.

## Contents

| Split        | Cloths | Trajectories | Steps per traj. | Vertices |
|--------------|--------|--------------|-----------------|----------|
| `training`   | 181    | 1 per cloth  | 44–121          | 209–221  |
| `validation` | 21     | 1 per cloth  | 58–121          | 209–221  |

## File structure

HDF5 layout (`gse_demos_v1.h5`, ~1.5 GB):

```
<split>/                              # "training" or "validation"
  cloth_XXX/
    rest_positions   (V, 3)  float32  # canonical / rest mesh, constant per cloth
    faces            (F, 3)  int64    # triangle topology (F = 368 for V = 219)
    trajectory_0/
      step_YYYY/
        positions              (V, 3)        float32  # deformed mesh state at t
        pointclouds/cam_0      (512, 3)       float32  # single-camera depth back-projection
        rgb                    (128, 128, 3) uint8     # RGB render at t
```

Vertex count `V` varies per cloth (209–221); topology (`faces`) is fixed within a cloth.

## Coordinate & scale conventions

> ⚠️ This dataset does **not** share the frame or scale of the `fold_unfold_lift`
> cloth training data. Pre-process before feeding it to state-estimation models.

- **Axis convention** — GSE cloths lie flat in the **XZ plane** with gravity along
  **+Y**. The `fold_unfold_lift` training cloths lie flat in **XY** with gravity
  along **+Z**. Map GSE → training with a **+90° rotation about X**:
  `(x, y, z) -> (x, -z, y)`.
- **Scale** — GSE cloths are ~3–6 world units wide; training cloths are ~12 cm.
  Scale each GSE cloth so its rest-mesh max extent ≈ `0.1225`.
- **Mesh density** — GSE meshes have 209–221 vertices vs. training's 44–151, and
  point clouds are single-camera (`cam_0`, ~512 points). Both are out-of-distribution
  for models trained on the multi-camera fixed-template data.

Without the rotation fix, Chamfer error on the GPS state-est model was ~982 mm;
with it, ~300–780 mm (a real residual domain gap).

## Usage

```python
import h5py

with h5py.File("gse_demos_v1.h5", "r") as f:
    grp = f["training"]["cloth_000"]
    rest = grp["rest_positions"][:]          # (V, 3)
    faces = grp["faces"][:]                  # (F, 3)
    traj = grp["trajectory_0"]
    step = traj["step_0000"]
    positions = step["positions"][:]         # (V, 3) deformed mesh
    pcd = step["pointclouds"]["cam_0"][:]    # (512, 3) point cloud
    rgb = step["rgb"][:]                     # (128, 128, 3)
```

An interactive viewer (`visualize_gse_demos.py`) accompanies the dataset; it shows
the RGB image, deformed mesh, point cloud, and canonical mesh side-by-side with a
timestep slider and per-demo selector.