---
library_name: lerobot
tags:
  - mujoco
  - lerobot
  - robotics
  - imitation-learning
  - act
---

# mujoco_lerobot_train

![Pipeline](./assets/pipeline.svg)

![Scene](./assets/scene.svg)

Minimal MuJoCo + LeRobot pipeline for:

1. collecting a standard LeRobot dataset
2. visualizing the dataset with `lerobot-dataset-viz`
3. training an ACT policy with `lerobot-train`
4. running closed-loop MuJoCo evaluation with the trained policy

This directory is intentionally small. All parameters are read from one file:

- `config.json`

这是一个最小化的 MuJoCo + LeRobot 工程，包含 4 步：

1. 用 MuJoCo 采集标准 LeRobot 数据集
2. 用 `lerobot-dataset-viz` 可视化数据
3. 用 `lerobot-train` 训练 ACT 策略
4. 在 MuJoCo 中闭环评估训练好的策略

整个目录尽量保持小而清晰，所有参数都只从一个文件读取：

- `config.json`

## Dependencies

- `mujoco`
- `lerobot`

See:

- `requirements.txt`

## 依赖

- `mujoco`
- `lerobot`

依赖文件见：

- `requirements.txt`

## Files

- `collect_dataset.py`: collect a MuJoCo pick-place dataset in LeRobot format
- `viz_dataset.py`: open `lerobot-dataset-viz` for the configured dataset
- `train_policy.py`: Python entry that reads config and launches training
- `eval_policy.py`: closed-loop MuJoCo evaluation using the trained policy
- `common.py`: shared minimal implementation
- `config.json`: all parameters

## 文件说明

- `collect_dataset.py`：采集 MuJoCo 抓取放置数据，并写成 LeRobot 标准格式
- `viz_dataset.py`：调用 `lerobot-dataset-viz` 可视化当前数据集
- `train_policy.py`：读取配置后启动训练
- `eval_policy.py`：在 MuJoCo 中闭环评估训练好的策略
- `common.py`：公共最小实现
- `config.json`：全部参数

## Run

Activate your environment first:

```bash
conda activate lerobot
cd mujoco_lerobot_train
```

## 运行

先激活环境并进入目录：

```bash
conda activate lerobot
cd mujoco_lerobot_train
```

Collect dataset:

```bash
python collect_dataset.py
```

Visualize dataset:

```bash
python viz_dataset.py
```

Train with the Python entry:

```bash
python train_policy.py
```

Closed-loop MuJoCo evaluation:

```bash
python eval_policy.py
```

闭环 MuJoCo 评估：

```bash
python eval_policy.py
```

## Config

`config.json` controls:

- dataset repo id and local root
- image size and fps
- number of episodes
- ACT training hyperparameters
- evaluation episodes and playback speed

## 配置

`config.json` 统一控制：

- 数据集 repo id 和本地路径
- 图像分辨率和 fps
- episode 数量
- ACT 训练参数
- 评估轮数和播放速度

## Upload To Hugging Face

Login first:

```bash
huggingface-cli login
```

## 上传到 Hugging Face

先登录：

```bash
huggingface-cli login
```

Then upload this folder:

```bash
bash ./upload_to_hf.sh <user_or_org>/<repo_name>
```

Private repo:

```bash
HF_PRIVATE=1 bash ./upload_to_hf.sh <user_or_org>/<repo_name>
```

Dataset repo instead of model repo:

```bash
HF_REPO_TYPE=dataset bash ./upload_to_hf.sh <user_or_org>/<repo_name>
```

Ignored during upload:

- `outputs/`
- `__pycache__/`
- `*.pyc`

上传时会自动忽略：

- `outputs/`
- `__pycache__/`
- `*.pyc`