README.md

# Cloth-Sphere Push Datasets

Four splits of a cloth-being-pushed-over-a-sphere dataset, generated with
`examples/generate_cloth_sphere_push.py` (see `gendata.sh` at repo root).

The train/test split is defined by **cloth size** (particles per side):

| Split | Directory | Episodes | Cloth size range | Seed |
|-------|-----------|----------|-----------------|------|
| Train | `cloth_push_sphere/` | 2000 | 100–150 × 100–150 | 42 |
| ID test | `cloth_push_sphere_id_test/` | 100 | 100–150 × 100–150 | 88 |
| OOD test (smaller) | `cloth_push_sphere_ood_test_smaller/` | 50 | 50–100 × 50–100 | 42 |
| OOD test (larger) | `cloth_push_sphere_ood_test_larger/` | 50 | 150–200 × 150–200 | 42 |

All splits share the same scene setup, horizon, resolution, and action parameters.

---

## Directory Layout (per split)

```
<split>/
├── dataset.h5        ← actions + picker positions for all episodes
└── episodes/
    ├── episode_00000.mp4
    └── ...
```

No `observations` array is stored in the HDF5 file — the MP4s are the
canonical source of visual observations (`hdf5+mp4` format).

---

## Common Parameters

| Parameter | Value |
|-----------|-------|
| Horizon (steps/episode) | 64 |
| Image resolution | 512 × 512 |
| Video FPS | 16 |
| Grab steps | 3 |
| Action repeat | 16 |
| Sphere radius | 0.25 m |
| Sphere centre | (0.0, 0.25, 0.0) m |
| Num variations | 20 |

---

## HDF5 Layout

```
dataset.h5
├── episode_00000/
│   ├── actions          (64, 8)     float32
│   └── picker_positions (65, 2, 3)  float32
│   attrs: cloth_dimx, cloth_dimy
├── episode_00001/
│   └── ...
attrs: num_episodes, horizon, img_size, grab_steps, action_repeat,
       cloth_dimx_range, cloth_dimy_range, sphere_radius, sphere_center,
       seed, num_variations, notes
```

### `actions` — `(T, 8)` float32

```
actions[t] = [dx1, dy1, dz1, grab1,  dx2, dy2, dz2, grab2]
              ←—  picker 0  —→        ←—  picker 1  —→
```

| Field | Units | Notes |
|-------|-------|-------|
| `dx`, `dy`, `dz` | metres / sub-step | Actual executed delta: `(pos_after − pos_before) / action_repeat` |
| `grab` | — | Always 1.0; cloth is held throughout |

**Faithful recording**: the stored delta is the *actual* motion observed in the
simulator, not the intended command. Any clamping by the physics engine is
already reflected, so `actions[t]` exactly explains the motion in the MP4.

First `grab_steps = 3` actions have `dx = dy = dz = 0` (grab-only, no motion).
Net picker displacement per step = `delta × action_repeat` (up to `16 × 0.01 = 0.16 m`).

### `picker_positions` — `(T+1, 2, 3)` float32

World-space XYZ in metres (Y-up). Index `t` is the state *before* `actions[t]`.
Index `T` is the terminal state after the last action.

### Episode attributes

| Attribute | Description |
|-----------|-------------|
| `cloth_dimx` | Cloth width in particles (sampled from `cloth_dimx_range`) |
| `cloth_dimy` | Cloth height in particles (sampled from `cloth_dimy_range`) |

---

## Temporal Alignment

```
t=0   picker_positions[0]  →  actions[0]  →  picker_positions[1]
      video frame 0                           video frame 1
...
t=63  picker_positions[63] →  actions[63] →  picker_positions[64]
      video frame 63                          video frame 64
```

`picker_positions[t]` and video frame `t` are the state *before* `actions[t]`.
Each MP4 has `T + 1 = 65` frames (including the terminal frame).

---

## Reading the Dataset

```python
import h5py
import numpy as np

with h5py.File('data/cloth_push_sphere/dataset.h5', 'r') as f:
    horizon       = int(f.attrs['horizon'])        # 64
    action_repeat = int(f.attrs['action_repeat'])  # 16

    for ep_key in sorted(f.keys()):
        grp  = f[ep_key]
        acts = grp['actions'][:]          # (64, 8)    float32
        ppos = grp['picker_positions'][:] # (65, 2, 3) float32
        dimx = int(grp.attrs['cloth_dimx'])
        dimy = int(grp.attrs['cloth_dimy'])

        # (state_t, action_t, state_{t+1}) triples
        for t in range(horizon):
            pos_before = ppos[t]    # (2, 3)
            action     = acts[t]    # (8,)
            pos_after  = ppos[t+1]  # (2, 3)

# Load the corresponding video (frame t matches picker_positions[t])
import imageio
frames = imageio.v2.mimread('data/cloth_push_sphere/episodes/episode_00000.mp4')
# frames[t]: (512, 512, 3) uint8, pre-action frame for actions[t]
```

---

## Storage

| Split | HDF5 | Videos | Total |
|-------|------|--------|-------|
| `cloth_push_sphere` (train) | ~11 MB | ~172 MB | ~183 MB |
| `cloth_push_sphere_id_test` | ~0.6 MB | ~8.7 MB | ~9.3 MB |
| `cloth_push_sphere_ood_test_smaller` | ~0.3 MB | ~2.4 MB | ~2.7 MB |
| `cloth_push_sphere_ood_test_larger` | ~0.3 MB | ~6.6 MB | ~6.9 MB |
| **Total** | | | **~202 MB** |