gpudad/so101_pick_cube_chunked

---
license: apache-2.0
task_categories:
- robotics
tags:
- lerobot
- so101
- robot-manipulation
- act
- episode-level-videos
configs:
- config_name: default
data_files:
- split: train
path: data/*/*.parquet
size_categories:
- 1M<n<10M
---

# SO101 Pick Cube Dataset (Chunked)


This is a restructured version of the [`gpudad/so101_pick_cube`](https://huggingface.co/datasets/gpudad/so101_pick_cube) dataset with **episode-level video files** for faster data loading during training.

## Why Chunked?

The original dataset has 3 monolithic video files (one per camera, 13+ hours each). Random access during training is slow because the decoder must seek through huge files.

This version splits videos into **1000 episodes per chunk**, making data loading **~50x faster**.

## Dataset Info

| Property | Value |
|----------|-------|
| Episodes | 10,990 |
| Frames | 1,456,443 |
| FPS | 30 |
| Cameras | front, overhead, wrist (512x512) |
| Action space | 6-DOF |

### Action Dimensions
- `elbow_flex`
- `gripper`
- `shoulder_lift`
- `shoulder_pan`
- `wrist_flex`
- `wrist_roll`

## Structure

```
videos/
├── observation.images.front/
│ ├── chunk-000/ # Episodes 0-999
│ │ ├── file-000.mp4
│ │ ├── file-001.mp4
│ │ └── ...
│ ├── chunk-001/ # Episodes 1000-1999
│ └── ...
├── observation.images.overhead/
└── observation.images.wrist/
```

## Usage

```python
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("gpudad/so101_pick_cube_chunked")
```

Or with roboport:

```bash
python -m roboport.train act gpudad/so101_pick_cube_chunked -o ./outputs/my_model
```

## Training Performance

With episode-level videos:
- **Data loading**: ~0.1s per batch (vs ~5.6s with monolithic videos)
- **Training iteration**: ~50x faster

## Original Dataset

This is derived from [gpudad/so101_pick_cube](https://huggingface.co/datasets/gpudad/so101_pick_cube).