# 星海图 GalaxeaManipSim 安装、任务环境与数据采集实操

> 原文链接：https://robotics.hellobike.com/tutorials/galaxea-manipsim-quickstart
> 频道：tutorials | 更新日期：2026-10-22 | 标签：GalaxeaManipSim, Galaxea R1, LeRobot, Diffusion Policy, 数据采集, 机器人仿真
> 核心摘要：本文按 Galaxea Manipulation Simulator 文档整理安装、资产下载、示例任务、演示数据采集、回放过滤、LeRobot 转换以及训练评估流程。

# 星海图 GalaxeaManipSim 安装、任务环境与数据采集实操

> 本文依据以下官方文档整理：[raw.githubusercontent.com/OpenGalaxea/GalaxeaManipSim/main/R](https://raw.githubusercontent.com/OpenGalaxea/GalaxeaManipSim/main/README.md)。命令与版本以文档当前版本为准。

## 前置条件

- Linux + CUDA GPU
- Python 3.10+
- CUDA 11.8
- conda
- LeRobot
- 可选：使用 --use_video 时需要 ffmpeg installed

## 1. 创建 GalaxeaManipSim 环境并安装

在 Galaxea Manipulation Simulator 代码目录中创建 conda 环境，文档指定 Python 版本为 3.10，并以 editable 方式安装当前项目。

```bash
## Create conda environment
conda create -n galaxea-sim python=3.10 -y
conda activate galaxea-sim
pip install -e .
```

## 2. 安装 LeRobot 并设置 PYTHONPATH

退出当前目录到上一级，克隆 LeRobot，切换到文档指定提交，再进行 editable 安装。最后把 LeRobot 代码路径加入 PYTHONPATH。

```bash
## Install lerobot
cd ..
git clone https://github.com/huggingface/lerobot.git
cd lerobot
git checkout a5e0aae13a3efd0080ac7ab6b461980d644014ab
pip install -e .
export PYTHONPATH="your_lerobot_codebase_path:$PYTHONPATH"
```

## 3. 下载并放置仿真资产

按文档下载并解压资产包，然后将 `robotwin_models` 移动到 `galaxea_sim/assets/`。这是注册和运行仿真任务所需的模型资产。

```bash
gdown https://drive.google.com/file/d/1ZvtCv1H4FLrse_ElUWzsVDt8xRK4CyaC/
unzip robotwin_models.zip
mv robotwin_models galaxea_sim/assets/
```

## 4. 选择机器人、任务和控制器

文档示例支持 R1、R1 Pro、R1 Lite。R1 示例任务为 `R1DualBottlesPickEasy`，R1 Pro 为 `R1ProBlocksStackEasy`，R1 Lite 为 `R1LiteBlocksStackEasy`。控制器可选 `bimanual_joint_position` / `bimanual_relaxed_ik`，但 R1 Lite 示例仅列出 `bimanual_joint_position`。

## 5. 用 mplib 生成原始演示数据

使用 `collect_demos` 生成 raw demos。默认 `--obs_mode` 是 `state`，速度更快，推荐给 replay 时再获取图像观测的 eef policies。默认数据保存到 `datasets/<env-name>/<data-time>`，格式为 h5。

```bash
## Example of R1 Picking up Bottles
python -m galaxea_sim.scripts.collect_demos --env-name R1DualBottlesPickEasy --num-demos 100

## Example of R1 Pro Stacking Blocks
python -m galaxea_sim.scripts.collect_demos --env-name R1ProBlocksStackEasy --num-demos 100

## Example of R1 Lite Stacking Blocks:
python -m galaxea_sim.scripts.collect_demos --env-name R1LiteBlocksStackEasy --num-demos 100
```

## 6. 指定控制器回放并过滤演示

使用 `replay_demos` 将记录的末端执行器位姿轨迹传给 relaxed_ik 控制器，并检查任务是否完成。带有 inverse kinematics solutions 的 demos 会被过滤，image/depth observations 会保存到 `datasets/<env-name>/final`。

```bash
## Example of R1 Picking up Bottles by Joints Control
python -m galaxea_sim.scripts.replay_demos --env-name R1DualBottlesPickEasy --target_controller_type bimanual_joint_position --num-demos 100

## Example of R1 Pro Stacking Blocks by End Effector Control
python -m galaxea_sim.scripts.replay_demos --env-name R1ProBlocksStackEasy --target_controller_type bimanual_relaxed_ik --num-demos 100

## Example of R1 Lite Stacking Blocks by Joint Control
python -m galaxea_sim.scripts.replay_demos --env-name R1LiteBlocksStackEasy  --target_controller_type bimanual_joint_position --num-demos 100
```

## 7. 转换为 LeRobot Dataset

若使用 Lerobot Diffusion Policy implementation，使用 `convert_single_galaxea_sim_to_lerobot`。`--tag` 决定转换哪些 demos；使用 `--use_eef` 时，arm observation 和 action 会替换为末端位姿。

```bash
## If you want to use Lerobot Diffusion Policy implementation, please use script convert_single_galaxea_sim_to_lerobot
## For Joint Control
## Example of R1 Picking up Bottles 
python -m galaxea_sim.scripts.convert_single_galaxea_sim_to_lerobot --task R1DualBottlesPickEasy --tag final --robot r1

## Example of R1 Pro Stacking Blocks
python -m galaxea_sim.scripts.convert_single_galaxea_sim_to_lerobot --task R1ProBlocksStackEasy --tag final --robot r1_pro

## Example of R1 Lite Stacking Blocks
python -m galaxea_sim.scripts.convert_single_galaxea_sim_to_lerobot --task R1LiteBlocksStackEasy --tag final --robot r1_lite

## For End Effector Control
## Example of R1 Pro Picking up Bottles 
python -m galaxea_sim.scripts.convert_single_galaxea_sim_to_lerobot --robot r1_pro --task R1ProDualBottlesPickEasy --tag final --use_eef

## Example of R1 Pro Stacking Blocks 
python -m galaxea_sim.scripts.convert_single_galaxea_sim_to_lerobot --robot r1_pro --task R1ProBlocksStackEasy --tag final --use_eef
```

## 8. 训练并评估 LeRobot Diffusion Policy

转换后可训练 LeRobot Diffusion Policy。默认策略保存到 `outputs/train/<env-name>/diffusion/<date-time>`。评估结果保存到指定 checkpoint dir，可用 `--save-video` 保存评估视频。

```bash
## Example of R1 Picking up Bottles 
python -m galaxea_sim.scripts.train_lerobot_dp_policy --task R1DualBottlesPickEasy

## Example of R1 Pro Stacking Blocks 
python -m galaxea_sim.scripts.train_lerobot_dp_policy --task R1ProBlocksStackEasy

## Example of R1 Lite Stacking Blocks 
python -m galaxea_sim.scripts.train_lerobot_dp_policy --task R1LiteBlocksStackEasy
```

```bash
## Example of R1 Picking up Bottles 
python -m galaxea_sim.scripts.eval_lerobot_dp_policy --task R1DualBottlesPickEasy --pretrained-policy-path outputs/train/R1DualBottlesPickEasy/diffusion/.../checkpoint --target_controller_type bimanual_joint_position 

## Example of R1 Pro Stacking Blocks
python -m galaxea_sim.scripts.eval_lerobot_dp_policy --task R1ProBlocksStackEasy --pretrained-policy-path outputs/train/R1ProBlocksStackEasy/diffusion/.../checkpoint  --target_controller_type bimanual_joint_position

## Example of R1 Lite Stacking Blocks 
python -m galaxea_sim.scripts.eval_lerobot_dp_policy --task R1LiteBlocksStackEasy --pretrained-policy-path outputs/train/R1LiteBlocksStackEasy/diffusion/.../checkpoint --target_controller_type bimanual_joint_position
```

## 验证是否跑通

安装后能完成资产放置；采集后在 `datasets/<env-name>/<data-time>` 看到 h5 files；回放后在 `datasets/<env-name>/final` 看到输出；转换后数据生成到 HF_LEROBOT_HOME，未指定时为 `~/.cache/huggingface/lerobot`；训练后在 `outputs/train/<env-name>/diffusion/<date-time>` 生成策略，评估结果保存到指定 checkpoint dir。

## 文档里点名的注意事项

- If installing together with Galaxea-DP, please refer to this [link]().
- If you find `dataset` package version conflict. Please clear the package cache in `~/.cache/huggingface/datasets`.
- The default `--obs_mode` is `state`, which is faster so it is recommended for eef policies that will get image observations during replay.
- Replay demonstrations will pass the recorded end effector's pose trajectory to the relaxed_ik controller and check if task is completed.
- The demos with inverse kinematics solutions will be filtered out, and image/depth observations will be saved.
- `--use_video` requires ffmpeg installed, and you may want to change vcodec to "libh264" in `.../site-packages/lerobot/common/datasets/video_utils.py` if default "libsvtav1" is not available.
- `--tag` determines which demos will be convert, by defult it will convert all demos in `datasets/<env-name>`.
- If `--use_eef` is used, the arm's observation and action will be replace by end effector's pose(x,y,z,qw,qx,qy,qz).
- Restoring from checkpoints is currently not supported.

## 原始参考来源 (Sources)

- [raw.githubusercontent.com/OpenGalaxea/GalaxeaManipSim/main/README.md](https://raw.githubusercontent.com/OpenGalaxea/GalaxeaManipSim/main/README.md)