---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- robotics
- lerobot
- so101
- mujoco
- sim
- pick-cube
pretty_name: Pick-Cube SO-101 (MuJoCo sim twin)
configs:
- config_name: default
  data_files: data/*/*.parquet
---

# Pick-Cube SO-101 — MuJoCo sim twin

A **1-to-1 MuJoCo reproduction** of the real teleoperated
[**dobri420/pick-cube-so101**](https://huggingface.co/datasets/dobri420/pick-cube-so101)
dataset. Every real episode is replayed through a MuJoCo sim of the SO-101 cell
and re-rendered from the same three camera views — see the source card for the
full breakdown of tranches, grasp distribution, and recording setup.

Generator code: [**dyordan1/so101-mujoco**](https://github.com/dyordan1/so101-mujoco).

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

## How it was made

- **Proprioception is verbatim.** `action` and `observation.state` (6-DoF joint
  positions) are copied from the real dataset unchanged — only the pixels are
  synthetic.
- **Pixels are MuJoCo.** The three views (`camera1`/`camera2`/`camera3`, 480×640)
  are re-rendered in sim, with the cube welded at the recorded grasp frame, the
  tote at the release frame, and each episode's distractors placed in the fan.
- **Sim-fidelity filter.** Each episode is first rolled physics-only to check the
  cube actually lands in the tote; episodes that don't land in sim are dropped,
  and the survivors renumbered. So this twin has **535 episodes** (vs the source's
  540) — the ones the sim faithfully reproduces.

## Train and test your own policy — no robot needed

Because both the data **and** the environment are in the loop, you can train a
policy on this dataset and evaluate it in the same MuJoCo sim it came from — with
**no physical arm and no need to reproduce the real scene lighting/geometry**. The
generator repo [**dyordan1/so101-mujoco**](https://github.com/dyordan1/so101-mujoco)
ships the whole loop:

```bash
git clone https://github.com/dyordan1/so101-mujoco && cd so101-mujoco
pip install -r requirements.txt

python download.py                       # this dataset -> datasets/
scripts/train                            # finetune SmolVLA -> checkpoints/  (needs a GPU)
python mujoco_policy.py \
  checkpoints/pick-cube-so101-sim-smolvla/checkpoints/last/pretrained_model --grid
```

`scripts/train` wraps `lerobot-train` with the SmolVLA recipe; `mujoco_policy.py`
places the cube in the sim and lets the policy drive (`--reach`/`--azim` to move
the cube, `--view` for the 3D viewer). Same loop works on the real
[dobri420/pick-cube-so101](https://huggingface.co/datasets/dobri420/pick-cube-so101).

## Usage

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

ds = LeRobotDataset("dobri420/pick-cube-so101-sim")
print(ds.num_episodes, ds.num_frames)  # 535 205329
```

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