README.md

---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- so101
- pick-place
- manipulation
pretty_name: Pick-Cube SO-101
configs:
- config_name: default
  data_files: data/*/*.parquet
---

# Pick-Cube SO-101

A **deliberately structured** teleoperated pick-and-place dataset on the
[SO-101](https://github.com/TheRobotStudio/SO-ARM100) arm — **540 episodes** of
*"grab the cube and put it in the box"*, structured to systematically vary grasp
**coverage**, **reach depth**, **wrist angle**, and **scene clutter**.

<a href="https://huggingface.co/spaces/lerobot/visualize_dataset?path=dobri420/pick-cube-so101">
<img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/></a>

> **MuJoCo sim twin:** [**dobri420/pick-cube-so101-sim**](https://huggingface.co/datasets/dobri420/pick-cube-so101-sim)
> — a 1-to-1 re-render of these episodes in simulation (same proprioception, synthetic pixels).

## Trained model

A **SmolVLA** policy finetuned on this dataset (first 340 episodes, through the
twist-CCW tranche) reaches **89% grab rate (72/81)** on the real arm — balanced
across reach depth: [**dobri420/pick-cube-smolvla-so101**
(model)](https://huggingface.co/dobri420/pick-cube-smolvla-so101). A further
checkpoint trained on the full 540 episodes (2.56M samples, batch 96) is
pending real-arm evaluation.

## Composition

Seven tranches, recorded in order:

| Episodes | Tranche | What |
|---|---|---|
| `0–49` | **discrete @ 25 cm** | 5 target angles (0°, ±45°, ±90°) × 10 demos |
| `50–69` | **dense jitter** | 20 demos clustered at 45° right, ±2 cm placement jitter |
| `70–119` | **discrete @ 15 cm** | the same 5 angles at the near depth × 10 demos |
| `120–239` | **blanket jitter** | 120 demos spread across the workspace |
| `240–339` | **twist-CCW jitter** | 100 demos with a counter-clockwise–twisted wrist for grasp-angle variability |
| `340–439` | **twist-CW jitter** | 100 demos with a clockwise–twisted wrist — the mirror of the CCW tranche |
| `440–539` | **distractor jitter** | 100 blanket-style demos with 1–3 distractor objects on the workspace |

## Grasp distribution

Each arrow is one episode's grasp — its **position** on the workspace half-disk
(rings = reach in cm from the shoulder-pan axis, radials = target angle) and its
**wrist yaw** (arrowhead points toward the upper jaw), recovered by SO-101 forward
kinematics on the jaw center. Red arrows highlight the named tranche within the
full dataset; a stamped digit is that episode's distractor count. The grid is
calibrated from this dataset's own registration clusters.

![all 540 grasps](https://huggingface.co/datasets/dobri420/pick-cube-so101/resolve/main/media/grasp-all.png)

<table>
<tr>
<td><img src="https://huggingface.co/datasets/dobri420/pick-cube-so101/resolve/main/media/grasp-discrete-25.png" alt="discrete @ 25 cm"/></td>
<td><img src="https://huggingface.co/datasets/dobri420/pick-cube-so101/resolve/main/media/grasp-dense.png" alt="dense jitter"/></td>
</tr>
<tr>
<td><img src="https://huggingface.co/datasets/dobri420/pick-cube-so101/resolve/main/media/grasp-discrete-15.png" alt="discrete @ 15 cm"/></td>
<td><img src="https://huggingface.co/datasets/dobri420/pick-cube-so101/resolve/main/media/grasp-blanket.png" alt="blanket jitter"/></td>
</tr>
<tr>
<td><img src="https://huggingface.co/datasets/dobri420/pick-cube-so101/resolve/main/media/grasp-twist-ccw.png" alt="twist-CCW jitter"/></td>
<td><img src="https://huggingface.co/datasets/dobri420/pick-cube-so101/resolve/main/media/grasp-twist-cw.png" alt="twist-CW jitter"/></td>
</tr>
<tr>
<td colspan="2" align="center"><img width="50%" src="https://huggingface.co/datasets/dobri420/pick-cube-so101/resolve/main/media/grasp-distractor.png" alt="distractor jitter"/></td>
</tr>
</table>

The **discrete** tranches sit as tight clusters on the 15/25 cm rings; **blanket
jitter** fills the gaps; the **twist** tranches visibly rotate the arrows off-radial
(counter-clockwise, then clockwise) — the wrist variability the earlier tranches lack;
**distractor jitter** repeats the blanket spread with cluttered scenes (the
stamped counts), so the policy must find *the cube*, not *the object*.

## Recording

- **Robot:** SO-101 leader→follower pair, teleoperated at 30 fps.
- **Cameras:** three views — `camera1` top, `camera2` wrist, `camera3` side (480×640).
- **State / action:** 6-DoF joint positions (`shoulder_pan … gripper`), in degrees.
- **Task:** *"Grab the cube and put it in the box."*

## Usage

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

ds = LeRobotDataset("dobri420/pick-cube-so101")
print(ds.num_episodes, ds.num_frames)  # 540 ...
```

Built with [LeRobot](https://github.com/huggingface/lerobot) (`codebase_version: v3.0`).