---
configs:
  - config_name: pose_data
    data_files: "pose_data/metadata.jsonl"
  - config_name: force_data
    data_files: "force_data/metadata.jsonl"
  - config_name: tacniq_gsmini
    data_files: "tacniq_gsmini/metadata.jsonl"
  - config_name: xela_9dtact
    data_files: "xela_9dtact/metadata.jsonl"
---

# PairTouch 13M Dataset

Multi-modal tactile dataset with pose, force, and tactile sensor data.

## Configs

| Config | Description | Sensors |
|--------|-------------|---------|
| `pose_data` | Pose estimation data | tac02/xela + camera |
| `force_data` | Force measurement data | tac02/xela + gelsight |
| `tacniq_gsmini` | TacNIQ + GSMini data | tacniq + gsmini |
| `xela_9dtact` | XELA + 9DTact data | xela + 9dtact |

## Usage

```python
from datasets import load_dataset

# Load specific config
ds = load_dataset("BorisGuo/pair_touch_13m", "pose_data")
```

## Tactile Sensors

### tac02 / tacniq

- **Taxel layout**: 11 x 6 = 66 taxels
- **Force type**: Z-axis normal force only
- **Dead taxels**: Index 0 (top-left) and 5 (top-right) always read 0, should be ignored

```
Taxel indices:
[ 0][ 1][ 2][ 3][ 4][ 5]   <- Row 0 (index 0, 5 are dead)
[ 6][ 7][ 8][ 9][10][11]   <- Row 1
...
[60][61][62][63][64][65]   <- Row 10
```

### xela

- **Taxel layout**: 4 x 6 = 24 taxels
- **Force type**: 3D force (X, Y, Z axes), total 72 values (24 x 3)
- **Data format**: Flattened as [x0, y0, z0, x1, y1, z1, ..., x23, y23, z23]

## Sampling Frequency

### force_data (1 tactile sample per frame)

| Dataset | Frame Rate | Tactile Sensor |
|---------|------------|----------------|
| 9dtact_force_h5 | ~10 Hz | - |
| gelsight_force_h5 | ~25 Hz | - |
| tac02_force_h5 | ~200 Hz | tac02 |
| xela_force_h5 | ~100 Hz | xela |

### pose_data (multiple tactile samples per frame)

| Dataset | Frame Rate | Samples/Frame | Effective Tactile Rate |
|---------|------------|---------------|------------------------|
| tac02_pose_h5 | ~7.5 Hz | 20 | ~150 Hz |
| xela_pose_h5 | ~7.5 Hz | 10 | ~75 Hz |
| gelsight_pose_h5 | ~7.5 Hz | - | - |
| 9dtact_pose_h5 | ~7.5 Hz | - | - |

### Paired sensor data

| Dataset | Frame Rate | Sensors |
|---------|------------|---------|
| tacniq_gsmini | 10 Hz | tacniq + gsmini |
| xela_9dtact | 10 Hz | xela + 9dtact |

## Data Fields

- `file_name`: Path to image file
- `episode_id`: Episode identifier
- `frame_idx`: Frame index within episode
- `timestamp`: Timestamp
- `rotation`: 3D rotation
- `translation`: 3D translation
- `tactile`: tac02 sensor matrix (11x6 taxels, Z-axis only)
- `xela`: XELA sensor matrix (4x6x3 taxels, 3D force)
- `tacniq`: TacNIQ sensor matrix (same as tac02)
- `video`: Episode video path

## Data Dimensions

- `images`: (N, S, 224, 224, 3) - N frames, S samples per frame
- `tactile` (pose_data): (N, 20, 66) - N frames, 20Hz sampling, 66 taxels (11x6)
- `xela` (pose_data): (N, 10, 72) - N frames, 10Hz sampling, 72 values (4x6x3)
- `tactile/xela` (force_data): (N, 66) or (N, 72) - 1 sample per frame

## Preprocessing

Use `preprocess.py` for data processing:

```bash
# Parse H5 files to JSON/images
python preprocess.py extract

# Check H5 structure (without parsing)
python preprocess.py extract --check

# Update metadata (add heatmap/video paths)
python preprocess.py extract --update

# Generate heatmaps
python preprocess.py heatmap
python preprocess.py heatmap --test              # Test mode
python preprocess.py heatmap --type tac02_pose   # Specific type

# Generate xela marker flow (point displacement visualization)
python preprocess.py marker_flow
python preprocess.py marker_flow --test
python preprocess.py marker_flow --type xela_pose

# Generate videos
python preprocess.py video
python preprocess.py video --test

# Pack images into tar files (reduce file count for HF upload)
python preprocess.py pack              # Pack only
python preprocess.py pack --delete     # Pack and delete original PNGs

# Unpack images from tar files
python preprocess.py unpack            # Unpack only
python preprocess.py unpack --delete   # Unpack and delete tar files

# Clean up (delete all PNGs, keep only videos)
python preprocess.py clean

# Upload to Hugging Face
python preprocess.py upload            # 只上传/更新，不删除远端文件
python preprocess.py upload --sync     # 同步模式：删除远端存在但本地不存在的文件

# Full pipeline (extract -> heatmap -> video -> update)
python preprocess.py all
```

### Heatmap Types

| Type | Description |
|------|-------------|
| `tac02_pose` | pose_data/tac02_pose_h5 |
| `xela_pose` | pose_data/xela_pose_h5 |
| `pose` | All pose_data |
| `tac02_force` | force_data/tac02_force_h5 |
| `xela_force` | force_data/xela_force_h5 |
| `force` | All force_data |
| `tacniq_gsmini` | tacniq_gsmini |
| `xela_9dtact` | xela_9dtact |
| `all` | Everything |

### Video Types

| Type | Description |
|------|-------------|
| `9dtact_force` | force_data/9dtact_force_h5 |
| `xela_force` | force_data/xela_force_h5 |
| `gelsight_force` | force_data/gelsight_force_h5 |
| `tac02_force` | force_data/tac02_force_h5 |
| `gelsight_pose` | pose_data/gelsight_pose_h5 |
| `9dtact_pose` | pose_data/9dtact_pose_h5 |
| `tac02_pose` | pose_data/tac02_pose_h5 |
| `xela_pose` | pose_data/xela_pose_h5 |
| `xela_pose_flow` | pose_data/xela_pose_h5 marker_flow (video_flow.mp4) |
| `xela_force_flow` | force_data/xela_force_h5 marker_flow (video_flow.mp4) |
| `xela_9dtact_flow` | xela_9dtact marker_flow (video_flow.mp4) |
| `tacniq_gsmini` | tacniq_gsmini (video_gsmini.mp4 + video_tacniq.mp4) |
| `xela_9dtact` | xela_9dtact (video_9dtact.mp4 + video_xela.mp4) |
| `all` | Everything |