cueng/so101_demo_bowl

---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
configs:
- config_name: default
data_files: data/*/*.parquet
size_categories:
- 100K<n<1M
---

# SO101 Put [object] into Bowl

## Dataset Summary

This dataset is formulated for use with the [LeRobot](https://github.com/huggingface/lerobot) framework. It contains robotic manipulation demonstrations for the **SO101** embodiment performing a standardized pick-and-place task: **"Put [object] into a bowl."** The dataset captures precise manipulation trajectories across 10 distinct household and office objects against a fixed background, recorded from multiple camera angles. It is designed to train and evaluate visuomotor policies for robotic manipulation.
Recorded and Validated by [Sorrawit Poomseetong](https://huggingface.co/chalkp)

---

## Dataset Details

### Statistics
- **Robot:** SO101 Follower
- **Total Rows/Frames:** 221,896
- **Control Frequency:** 30 Hz
- **Demonstrations:** 100 per object (1,000 total episodes)
- **Background Setup:** Fixed background
- **Action Horizon/Length:** ~200 steps per episode

### Sensor Configuration
The dataset includes synchronized multi-view RGB image observations:
* **Top Camera:** (`observation.images.top`) Provides a fixed, global overview of the workspace, the bowl, and the object's initial position.
* **Wrist Camera:** (`observation.images.wrist`) Provides an ego-centric view from the robot's end-effector, crucial for fine manipulation and grasping tasks.

---

## Task & Objects

**Task Instruction:** `"Put [object] into a bowl."`

The dataset covers variations of this task across the following 10 objects. Each object features exactly 100 successful demonstrations:

- A plastic spoon
- A metal spoon
- A scissors
- A roll of tissue paper
- A blue marker
- A black marker
- A nipper
- A glue
- A white tape
- A screw driver

---

## Data Structure

Data is formatted natively for `lerobot` and contains the following standard feature keys per step:

### Observations
- `observation.images.top` (`image`): RGB image from the overhead camera.
- `observation.images.wrist` (`image`): RGB image from the wrist camera.
- `observation.state` (`list` of `float32`): The state of the SO101 robot.

### Actions & Metadata
- `action` (`List[float32]`): The commanded actions at 30 Hz.
- `timestamp` (`float32`): Time elapsed in seconds since the start of the episode.
- `frame_index` (`int64`): Frame number within the current episode.
- `episode_index` (`int64`): Index of the demonstration episode (0 to 999).
- `index` (`int64`): Global index of the frame.
- `task_index` (`int64`): Corresponding ID for the specific object task.

---

## Usage with LeRobot

You can load this dataset directly using the LeRobot library:

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

# Load the dataset
dataset = LeRobotDataset("cueng/so101_demo_bowl")

# Load first frame
frame = dataset[0]
print(frame.keys())
```