Complete Tutorial on Deploying Isaac Lab + GR00T on Alibaba Cloud

Guide to deploying an embodied AI development environment for engineers and researchers with a Linux/GPU foundation

Author’s tested environment: Alibaba Cloud A10 GPU preemptive instance

⚠️ Version Declaration: This tutorial is written based on the January 2026 software version, which strongly relies on the following version combinations:

Isaac Sim 4.2.0 + Isaac Lab v1.4.1
GR00T N1.6 + transformers==4.51.3
PyTorch 2.5.1 + CUDA 12.1

Only guaranteed to be reproduction under the above version. If using a newer version, you may need to adjust dependencies and configurations manually.

⚠️ Safety Notice: This tutorial is designed for quick experimentation. It uses simplified configurations such as the root user, --network host, and xhost +.It is not recommended for production environments or long-term running servers.

Environment Overview
Creating Alibaba Cloud Instances
Basic Environment Configuration
VNC Remote Desktop Configuration
Deployment of Isaac Sim + Isaac Lab
GR00T Environment Configuration
GR00T + MuJoCo Evaluation
Common Issues and Troubleshooting Records

1. Environment Overview

Final Architecture

宿主机 (Ubuntu 22.04 + A10 GPU)
├── Docker: isaac-sim-gui
│   ├── Isaac Sim 4.2.0
│   ├── Isaac Lab v1.4.1
│   └── VNC显示 (端口6080)
└── Conda: groot环境
    ├── Isaac-GR00T代码
    ├── GR00T-N1.6-3B模型 (通用)
    └── GR00T-N1.6-G1-PnPAppleToPlate模型 (G1专用)

Why is it designed this way?

Isaac Lab within Docker: Isaac Sim has complex dependencies; the official Docker image is more convenient.
GR00T on host Conda: GR00T requires PyTorch 2.5+, which conflicts with Isaac Sim’s PyTorch 2.4.
VNC remote desktop: WebRTC Livestream has compatibility issues on cloud servers; VNC is more stable.

Key Design Decisions

Decision	Why do it this way	What happens if not done
GR00T installed via Conda instead of Docker	Conflict with PyTorch 2.5+ and Isaac Sim	Dependency hell, environment crash
Skip flash-attn	Compilation is time-consuming and not necessary	OOM/freeze, server restart
Lock transformers==4.51.3	New API changes	Model loading errors
Install GR00T with --no-deps	Avoid duplicate compilation of flash-attn	pip gets stuck
Manually clone submodules	curl download lacks .git	setup script errors

Deployment order (must be executed in sequence)

基础配置 (3) → VNC (4) → Isaac Docker (5) → GR00T Conda (6) → MuJoCo 评估 (7)
     │              │              │                │
     └── NVIDIA 驱动 → Docker → Container Toolkit ──┘

⚠️ Do not skip steps or disrupt the order, as each step depends on the previous configuration.

Hardware Requirements

GPU: NVIDIA A10 24GB or higher (RTX 3090/4090 also available)
Memory: 32GB+
Hard drive: 150GB+ SSD (actual usage is about 80-100GB)
- Ubuntu system: ~5GB
- Isaac Sim Docker image: ~25GB
- Docker runtime cache: ~10GB
- Conda environment (PyTorch + CUDA): ~10GB
- GR00T model (3B + G1): ~12GB
- Logs and temporary files: ~10GB
- Reserved space: ~30GB

2. Creating Alibaba Cloud Instances

2.0 Preparations

Register a NGC account (required to pull the Isaac Sim image)

Access https://ngc.nvidia.com
Click “Sign Up” to register with an email address (or log in using Google/GitHub)
After logging in, click the avatar in the top-right corner → “Setup” → “Generate API Key”
Click “Generate API Key”, then copy and save it

⚠️ The API key is displayed only once; make sure to save it. It will be needed in subsequent docker login nvcr.ios.

2.1 Cost Description

Actual cost: Completing the entire tutorial costs approximately 30 RMB (preemptible instance + data usage fee)
Account requirements: An Alibaba Cloud account must have a balance of 100 RMB or more to purchase a preemptible instance
The price of preemptible instances fluctuates, so the actual cost may vary

2.2 Select Instance Specifications

Log in to the Alibaba Cloud console → Cloud Virtual Machine ECS → Create instance.
Select the region: Southwest 1 (Chengdu) is recommended (lowest price).
Instance specification: Search for ecs.gn7i and choose the A10 24GB specification.
- ecs.gn7i-c16g1.4xlarge (recommended): 16 vCPU + 60GB memory
- ecs.gn7i-c8g1.2xlarge: 8 vCPU + 32GB memory

💡 Cost-saving tips: Prices vary greatly across regions!

Chengdu c16g1.4xlarge:约 2.4 yuan/hour (high configuration, low price)

Beijing/Hangzhou c8g1.2xlarge: about 3.1 yuan/hour

It is recommended to choose Chengdu as it offers higher configurations at lower prices.

Pay-as-you-go mode: Preemptible instances (80-90% cheaper than pay-per-use)
- Set the maximum price to 50-70% of the pay-per-use rate
- Check “Instance release protection”

2.2 Image Selection

Operating system: Ubuntu 22.04 64-bit
Or choose the NVIDIA GPU cloud acceleration image (with pre-installed drivers)

2.3 Network Configuration

Assign a public IP address (billed by traffic)
Bandwidth: 5-10 Mbps is sufficient

2.4 Security Group Configuration

Create or modify a security group to open the following ports:

Port	Purpose
22	SSH

⚠️ Security advice: Do not directly open the VNC port (5901/6080). Using an SSH tunnel is safer. See below for details.

3. Basic Environment Configuration

3.1 SSH Connection

bashssh root@<你的公网IP>

⚠️ Important: When connecting using Alibaba Cloud Workbench, do not select “Password-free connection”!

Password-free connection uses a containerized Web terminal with restricted permissions, which prevents Docker from starting properly.

Please choose “Password” or “Key Pair” for connection, or directly use the local SSH client.

3.2 Installing NVIDIA drivers (if not pre-installed in the image)

bash# 检查驱动
nvidia-smi

# 如果没有，安装驱动
apt update
apt install -y nvidia-driver-535
reboot

3.3 Installing Docker

bash# 检查是否已安装
docker --version

# 如果没有，安装 Docker
apt update
apt install -y docker.io
systemctl enable docker
systemctl start docker

# 验证安装
docker --version

3.4 Configuring Docker Domestic Images

Docker Hub is unstable to access in China. Configure image acceleration:

bashcat > /etc/docker/daemon.json << 'EOF'
{
  "registry-mirrors": [
    "https://docker.1ms.run",
    "https://docker.xuanyuan.me"
  ]
}
EOF

systemctl restart docker

3.5 Installing NVIDIA Container Toolkit

Download the deb package directly for installation (stable reproduction in China):

bashcd /tmp

# 下载 4 个必需的包
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/libnvidia-container1_1.17.4-1_amd64.deb
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/libnvidia-container-tools_1.17.4-1_amd64.deb
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/nvidia-container-toolkit-base_1.17.4-1_amd64.deb
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/nvidia-container-toolkit_1.17.4-1_amd64.deb

# 按顺序安装
dpkg -i libnvidia-container1_1.17.4-1_amd64.deb
dpkg -i libnvidia-container-tools_1.17.4-1_amd64.deb
dpkg -i nvidia-container-toolkit-base_1.17.4-1_amd64.deb
dpkg -i nvidia-container-toolkit_1.17.4-1_amd64.deb

# 配置 Docker
nvidia-ctk runtime configure --runtime=docker
systemctl restart docker

# 验证（应该能看到 nvidia-smi 输出）
docker run --rm --gpus all nvidia/cuda:12.2.0-base-ubuntu22.04 nvidia-smi

💡 Tip: The download link will automatically redirect to the domestic mirror (nvidia.cn), which is faster.

4. VNC Remote Desktop Configuration

4.1 Installing the Desktop Environment and VNC

bashapt update
apt install -y xfce4 xfce4-goodies tigervnc-standalone-server tigervnc-common novnc websockify

💡 Installation instructions:

If the “Daemons using outdated libraries” dialog box appears, press Tab to select <Ok> and then enter to continue.

This is a normal phenomenon when restarting services after an Ubuntu system update.

4.2 Configure VNC

bash# 设置 VNC 密码
vncpasswd
# 输入密码（至少6位），view-only 选 n

# 创建 VNC 配置
mkdir -p ~/.vnc
cat > ~/.vnc/xstartup << 'EOF'
#!/bin/sh
unset SESSION_MANAGER
unset DBUS_SESSION_BUS_ADDRESS
exec startxfce4
EOF
chmod +x ~/.vnc/xstartup

# 启动 VNC 服务
vncserver :1 -geometry 1920x1080 -depth 24 -localhost no

# 验证 VNC 是否正常启动
netstat -tlnp | grep 5901
# 应该看到 Xtigervnc 在监听 5901 端口，类似：
# tcp        0      0 0.0.0.0:5901            0.0.0.0:*               LISTEN      26811/Xtigervnc

# 启动 noVNC（Web 访问）
nohup websockify --web=/usr/share/novnc/ 6080 localhost:5901 > /dev/null 2>&1 &

4.3 Accessing VNC

Method 1: SSH Tunnel (Recommended, safer)

There is no need to open the VNC port in the security group. Execute on the local computer:

bash# Windows PowerShell / Mac Terminal / Linux
ssh -L 6080:localhost:6080 root@<你的公网IP>

Maintain the SSH connection, and then open the browser: http://localhost:6080/vnc.html

Method 2: Direct access (port must be opened)

If you choose direct access, you need to open port 6080 in the security group (only for your IP):

Browser open: http://<你的公网IP>:6080/vnc.html

Enter the VNC password to enter the desktop.

⚠️ Security Warning: The VNC protocol has weak security, and directly exposing ports is vulnerable to attacks. It is strongly recommended to use an SSH tunnel.

4.4 VNC Self-starting (Optional)

bashcat > /etc/systemd/system/vncserver.service << 'EOF'
[Unit]
Description=VNC Server
After=network.target

[Service]
Type=forking
User=root
ExecStart=/usr/bin/vncserver :1 -geometry 1920x1080 -depth 24
ExecStop=/usr/bin/vncserver -kill :1

[Install]
WantedBy=multi-user.target
EOF

systemctl enable vncserver

5. Deployment of Isaac Sim + Isaac Lab

5.1 Log in to NGC and pull images

Isaac Sim image is hosted on NVIDIA NGC (GPU Cloud) and requires login first:

bash# 登录 NGC
docker login nvcr.io

The system will prompt you to enter your username and password:

Username: Enter $oauthtoken (a fixed value, copy directly)
Password: Enter the NGC API Key you obtained in Section 2.0

When Login Succeeded is seen, it indicates successful login.

bash# 拉取 Isaac Sim 4.2.0 镜像（约 15GB，需要一些时间）
docker pull nvcr.io/nvidia/isaac-sim:4.2.0

After the download is successful, the following will be displayed:

Status: Downloaded newer image for nvcr.io/nvidia/isaac-sim:4.2.0

Verified that the image has been downloaded:

bashdocker images | grep isaac-sim
# 应该看到类似输出：
# nvcr.io/nvidia/isaac-sim   4.2.0   <IMAGE_ID>   <SIZE>

5.2 Starting Containers

bash# 在宿主机设置 X11 权限
export DISPLAY=:1
xhost +local:docker

# 启动容器
# 注意：--network host 和 xhost + 是为了简化配置，仅建议用于短期实验
docker run -it --name isaac-sim-gui --gpus all --network host \
  -e DISPLAY=:1 \
  -e ACCEPT_EULA=Y \
  -v /tmp/.X11-unix:/tmp/.X11-unix \
  --entrypoint bash \
  nvcr.io/nvidia/isaac-sim:4.2.0

5.3 Preparation: Download dependencies on the host machine

Due to the network environment limitations within the container, it is recommended to download Isaac Lab and robomimic on the host machine:

bash# 退出容器，回到宿主机
exit

# 在宿主机执行（不要进入容器）
cd ~

# 安装下载工具（如果没有）
apt update
apt install -y git wget unzip

# 下载 Isaac Lab（使用国内镜像）
git clone --depth 1 --branch v1.4.1 https://gitclone.com/github.com/isaac-sim/IsaacLab.git

# 下载 robomimic（用于模仿学习任务）
wget https://github.com/ARISE-Initiative/robomimic/archive/refs/heads/master.zip
unzip master.zip
mv robomimic-master robomimic

# 重新启动容器
docker start isaac-sim-gui

# 复制进容器
docker cp IsaacLab isaac-sim-gui:/isaac-sim/
docker cp robomimic isaac-sim-gui:/tmp/

5.4 Installing dependencies within the container

bash# 进入容器
docker exec -it isaac-sim-gui bash

# 安装编译工具和 EGL 库（robomimic 需要）
apt update
apt install -y build-essential cmake pkg-config git
apt install -y libegl1-mesa-dev libgl1-mesa-dev libgles2-mesa-dev

# 安装 robomimic
cd /tmp/robomimic
/isaac-sim/python.sh -m pip install -e .

5.5 Fixing Isaac Lab configuration files

There are two issues in the dependency configuration of Isaac Lab that need to be fixed:

bashcd /isaac-sim/IsaacLab

# 备份原文件
cp source/extensions/omni.isaac.lab_tasks/setup.py source/extensions/omni.isaac.lab_tasks/setup.py.bak

# 修复 1：rsl-rl 名称不匹配
# 原因：setup.py 中引用的是 GitHub 仓库名 "rsl-rl"，但 PyPI 上的包名是 "rsl-rl-lib"
# 这是上游的命名不一致问题，不影响功能
sed -i '46s/"rsl-rl@git+https:\/\/github.com\/leggedrobotics\/rsl_rl.git"/"rsl-rl-lib==2.3.0"/' source/extensions/omni.isaac.lab_tasks/setup.py

# 修复 2：robomimic 已手动安装，删除 git 克隆配置
# 原因：setup.py 试图从 GitHub 克隆 robomimic，但容器内网络不稳定
# 我们已在宿主机下载并复制进来，所以删除这个配置
sed -i '/robomimic@git/d' source/extensions/omni.isaac.lab_tasks/setup.py

# 验证修复
echo "=== 检查第 46 行（rsl-rl）==="
sed -n '46p' source/extensions/omni.isaac.lab_tasks/setup.py
echo "=== 检查第 50-56 行（robomimic 应该已删除）==="
sed -n '50,56p' source/extensions/omni.isaac.lab_tasks/setup.py

Expected output:

=== 检查第 46 行（rsl-rl）===
"rsl-rl": ["rsl-rl-lib==2.3.0"],
=== 检查第 52-56 行（robomimic）===
# Cumulation of all extra-requires
EXTRAS_REQUIRE["all"] = list(itertools.chain.from_iterable(EXTRAS_REQUIRE.values()))
# Remove duplicates in the all list to avoid double installations
EXTRAS_REQUIRE["all"] = list(set(EXTRAS_REQUIRE["all"]))

If line 46 is still rsl-rl@git+https://..., it means the sed command did not take effect. Check whether the line number is correct.

5.6 Installing Isaac Lab

bashcd /isaac-sim/IsaacLab

# 创建符号链接
ln -s /isaac-sim _isaac_sim

# 修复 pip（Isaac Sim 容器内 pip 可能损坏）
/isaac-sim/python.sh -m ensurepip --upgrade
/isaac-sim/python.sh -m pip install --upgrade pip setuptools

# 安装 Isaac Lab（约 5-10 分钟）
./isaaclab.sh --install

During the installation process, you will see the download and installation of many dependency packages. Finally, it should display:

Successfully installed omni-isaac-lab_tasks-0.10.18 ...

5.7 Testing Isaac Lab

bash# 在容器内运行 demo
export DISPLAY=:1
cd /isaac-sim/IsaacLab

# 机械臂 demo
./isaaclab.sh -p source/standalone/demos/arms.py

# 双足机器人 demo
./isaaclab.sh -p source/standalone/demos/bipeds.py

# 人形机器人强化学习训练
./isaaclab.sh -p source/standalone/workflows/rsl_rl/train.py \
  --task Isaac-Humanoid-Direct-v0 \
  --num_envs 64 \
  --max_iterations 100

The simulation image should be visible on the VNC desktop. Press Ctrl+C to stop the program.

6. GR00T Environment Configuration

6.1 Installing Miniconda

bash# 退出 Docker 容器，回到宿主机
exit

# 安装 Miniconda
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh -b -p ~/miniconda3
~/miniconda3/bin/conda init
source ~/.bashrc

# 接受 Conda 服务条款（新版 Conda 要求）
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r

6.2 Creating the GR00T Environment

bash# 创建 Python 3.10 环境
conda create -n groot python=3.10 -y
conda activate groot

# 安装 PyTorch（CUDA 12.1）
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

Install flash-attn (optional, recommended to skip)

⚠️ Important note: The compilation of flash-attn is very time-consuming (15-60 minutes), and it consumes a lot of system resources, which may cause SSH/VNC disconnection. GR00T can operate normally without flash-attn, but the inference speed is slightly slower. It is recommended to skip this step and directly proceed to Section 6.3.

If you really need flash-attn (for ultimate performance):

bash# 安装编译依赖
pip install numpy psutil packaging ninja

# 设置 CUDA 环境变量（必须）
export CUDA_HOME=/usr/local/cuda
export PATH=$CUDA_HOME/bin:$PATH

# 限制并行编译数，防止内存不足
# 32GB 内存用 MAX_JOBS=4
# 58GB+ 内存用 MAX_JOBS=6
MAX_JOBS=6 pip install flash-attn --no-build-isolation -v

⏳ Compilation process description:

The running bdist_wheel stage will hang for 1–3 minutes, which is normal (due to cmake configuration)

Later, [1/73] ... will indicate the start of actual compilation (73 files in total)

The entire compilation process takes 15–60 minutes, depending on the number of CPU cores and MAX_JOBS settings

⚠️ Abnormal phenomena during compilation (normal!):

New SSH connections may fail: Compilation consumes a lot of resources, causing slow sshd responses

VNC connection may drop: Same reason as above

Existing terminals may become unresponsive: High system load

As long as the CPU usage remains above 50%, it means compilation is in progress. Do not force a restart!

The compilation is complete when CPU usage drops below 5%

💡 You can monitor CPU usage on the “Instance Monitoring” page of the Alibaba Cloud console without needing an SSH connection

⚠️ 可能需要 restart after compilation:

SSH may still not connect after compilation

Restart the instance in the Alibaba Cloud console

After restart, run MAX_JOBS=6 pip install flash-attn --no-build-isolation -v again

This time, the cached wheel will be used directly, and installation will complete in just a few seconds

Verify installation:

bashpython -c "import flash_attn; print(flash_attn.__version__)"
# 应该输出 2.8.3 或类似版本

6.3 Installing Isaac-GR00T

bashcd ~

# 下载 Isaac-GR00T（注意：仓库在 NVIDIA 组织下，不是 NVIDIA-Omniverse）
# 方法1：git clone（国内可能较慢）
git clone --depth 1 https://github.com/NVIDIA/Isaac-GR00T.git

# 方法2：使用 gitclone 镜像
git clone --depth 1 https://gitclone.com/github.com/NVIDIA/Isaac-GR00T.git

# 方法3（推荐）：用 ghproxy 加速下载
curl -L -o isaac-groot.tar.gz https://ghproxy.cn/https://github.com/NVIDIA/Isaac-GR00T/archive/refs/heads/main.tar.gz
tar -xzf isaac-groot.tar.gz
mv Isaac-GR00T-main Isaac-GR00T

cd Isaac-GR00T

Install GR00T

⚠️ Important: Do not use pip install -e . directly, as it causes the flash-attn to be recompiled repeatedly, leading to freezing.

Reason: The setup.py of GR00T lists flash-attn as a dependency. Even if it is already installed, pip will attempt to rebuild it. Using --no-deps can skip the automatic dependency resolution, and manual installation of the core package can avoid this issue.

bash# 使用 --no-deps 跳过依赖自动安装
pip install -e . --no-build-isolation --no-deps

# 手动安装核心依赖
# transformers 必须锁定 4.51.3，更高版本会导致模型加载报错
pip install transformers==4.51.3 safetensors einops peft diffusers tyro omegaconf pandas dm-tree termcolor av albumentations huggingface_hub deepspeed accelerate
pip install click datasets gymnasium lmdb matplotlib msgpack-numpy pyzmq wandb torchcodec

# 验证安装
python -c 'from gr00t.policy.gr00t_policy import Gr00tPolicy; print("GR00T loaded!")'

💡 Regarding dependency version warnings: After installation, you may see a bunch of pip's dependency resolver warnings indicating incompatible versions. This is because GR00T’s setup.py locks very strict version numbers, but actually slightly newer versions can also work properly. As long as the verification command output GR00T loaded! indicates success, these warnings can be ignored.

💡 This approach avoids the issue of pip recompiling flash-attn. If flash-attn has been installed before, GR00T will automatically use it; if not, it will fall back to the standard attention.

6.4 Download GR00T Model

Note: For use in China, HuggingFace requires the use of an image.

bashconda activate groot
cd ~/Isaac-GR00T

# 下载通用模型 GR00T-N1.6-3B
HF_ENDPOINT=https://hf-mirror.com python -c "from huggingface_hub import snapshot_download; snapshot_download('nvidia/GR00T-N1.6-3B', local_dir='/root/groot_n16_model')"

# 下载 G1 机器人专用模型（用于评估）
HF_ENDPOINT=https://hf-mirror.com python -c "from huggingface_hub import snapshot_download; snapshot_download('nvidia/GR00T-N1.6-G1-PnPAppleToPlate', local_dir='/root/groot_g1_model')"

6.5 Testing GR00T Inference

bashcd ~/Isaac-GR00T

# 创建测试脚本
cat > test_groot.py << 'EOF'
import numpy as np
from gr00t.policy.gr00t_policy import Gr00tPolicy
from gr00t.data.embodiment_tags import EmbodimentTag

print('Loading GR00T N1.6...')
policy = Gr00tPolicy(
    model_path='/root/groot_n16_model',
    embodiment_tag=EmbodimentTag('gr1'),
    device='cuda',
)

obs = {
    'video': {
        'ego_view_bg_crop_pad_res256_freq20': np.random.randint(0, 255, (1, 1, 256, 256, 3), dtype=np.uint8),
    },
    'state': {
        'left_arm': np.random.rand(1, 1, 7).astype(np.float32),
        'right_arm': np.random.rand(1, 1, 7).astype(np.float32),
        'left_hand': np.random.rand(1, 1, 6).astype(np.float32),
        'right_hand': np.random.rand(1, 1, 6).astype(np.float32),
        'waist': np.random.rand(1, 1, 3).astype(np.float32),
    },
    'language': {
        'task': [['pick up the red apple']],
    },
}

action = policy.get_action(obs)
print('Action output:')
for k, v in action[0].items():
    print(f'  {k}: shape={v.shape}')
print('GR00T Inference Success!')
EOF

# 运行测试
python test_groot.py

7. GR00T + MuJoCo Evaluation

7.1 Preparation Work

bashconda activate groot
cd ~/Isaac-GR00T

# 安装依赖
pip install uv
apt-get update && apt-get install -y libegl1-mesa-dev libglu1-mesa git-lfs
git lfs install

# 配置 git 使用 ghproxy 加速（关键！）
git config --global http.version HTTP/1.1
git config --global url."https://ghproxy.cn/https://github.com/".insteadOf "https://github.com/"

# 克隆 GR00T-WholeBodyControl 子模块（必须用 git clone，不能用 tar.gz）
mkdir -p external_dependencies
cd external_dependencies
rm -rf GR00T-WholeBodyControl
git clone https://github.com/NVlabs/GR00T-WholeBodyControl.git

cd ~/Isaac-GR00T

# 验证 LFS 文件已下载（应该是 100KB+ 而不是 131 字节）
ls -la ~/Isaac-GR00T/external_dependencies/GR00T-WholeBodyControl/gr00t_wbc/control/robot_model/model_data/g1/meshes/left_hip_pitch_link.STL

⚠️ Important:

You must use git clone instead of curl to download the tar.gz file, as the repository contains large Git LFS files (robot STL models)

The tar.gz download only provides a LFS pointer file (131 bytes), not the actual model file

If the STL file is only 131 bytes, MuJoCo simulation will report an error Failed to determine STL storage representation

7.2 Modify the setup script and run it

bashcd ~/Isaac-GR00T

# 注释掉 git submodule 和 git lfs pull 命令（我们已经手动处理了）
sed -i 's/^git submodule update/#git submodule update/' gr00t/eval/sim/GR00T-WholeBodyControl/setup_GR00T_WholeBodyControl.sh
sed -i 's/^git -C/#git -C/' gr00t/eval/sim/GR00T-WholeBodyControl/setup_GR00T_WholeBodyControl.sh

# 注释掉 robosuite 的删除和 git clone（我们手动下载）
sed -i '26s/^rm/#rm/' gr00t/eval/sim/GR00T-WholeBodyControl/setup_GR00T_WholeBodyControl.sh
sed -i '27s/^git clone/#git clone/' gr00t/eval/sim/GR00T-WholeBodyControl/setup_GR00T_WholeBodyControl.sh

# 手动下载 robosuite（setup 脚本里的 git clone 经常失败）
cd ~/Isaac-GR00T/external_dependencies/GR00T-WholeBodyControl/gr00t_wbc/dexmg
curl -L -o robosuite.tar.gz https://ghproxy.cn/https://github.com/xieleo5/robosuite/archive/refs/heads/leo/support_g1_locomanip.tar.gz
tar -xzf robosuite.tar.gz
mv robosuite-leo-support_g1_locomanip gr00trobosuite
rm robosuite.tar.gz

# 运行 setup 脚本
cd ~/Isaac-GR00T
bash gr00t/eval/sim/GR00T-WholeBodyControl/setup_GR00T_WholeBodyControl.sh

The script will:

Create a uv virtual environment
Install dependencies such as robosuite, robocasa, and lerobot (downloading large packages like torch takes several minutes)
Verify the environment (Imports OK and Env OK should be visible)

💡 Network issue troubleshooting:

If the error occurs with HTTP2 framing layer: Ensure git config --global http.version HTTP/1.1 has been executed.

If the error occurs with Connection timed out: Verify that ghproxy acceleration is configured.

If SSH connection fails: Reconnect and run the setup script again; downloaded packages will be cached.

7.3 Modify the code to enable real-time visualization

bash# 备份原文件
cp ~/Isaac-GR00T/gr00t/eval/rollout_policy.py ~/Isaac-GR00T/gr00t/eval/rollout_policy.py.bak

# 修改渲染模式
sed -i 's/os.environ\["MUJOCO_GL"\] = "egl"/os.environ["MUJOCO_GL"] = "glx"/' ~/Isaac-GR00T/gr00t/eval/rollout_policy.py
sed -i 's/onscreen=False/onscreen=True/' ~/Isaac-GR00T/gr00t/eval/rollout_policy.py

7.4 Run Evaluation

Two terminals are required:

Terminal 1 - Start GR00T Server:

bashcd ~/Isaac-GR00T
conda activate groot
python gr00t/eval/run_gr00t_server.py \
  --model-path /root/groot_g1_model \
  --embodiment-tag UNITREE_G1 \
  --use-sim-policy-wrapper

Waiting to display Server is ready and listening on tcp://127.0.0.1:5555

Terminal 2 - Start the evaluation Client (executed in VNC):

💡 VNC Copy-Paste Tips:

Click the expand menu on the left side of the noVNC web page to find “Clipboard”.

Paste the following command into the clipboard text box.

Open the terminal in the VNC desktop (right-click the desktop → Terminal).

Use Ctrl+Shift+V to paste the command in the terminal.

Press Enter to execute it.

⚠️ Note: The command has been combined into a single line to avoid hidden characters during multi-line copying that could cause errors (such as invalid int value: '5555~').

bashcd ~/Isaac-GR00T && export DISPLAY=:1 && gr00t/eval/sim/GR00T-WholeBodyControl/GR00T-WholeBodyControl_uv/.venv/bin/python gr00t/eval/rollout_policy.py --n_episodes 3 --max_episode_steps 500 --env_name gr00tlocomanip_g1_sim/LMPnPAppleToPlateDC_G1_gear_wbc --n_action_steps 20 --n_envs 1 --policy_client_host 127.0.0.1 --policy_client_port 5555

A MuJoCo window will appear on the VNC desktop, displaying in real time how the G1 robot performs the “pick apple and place on plate” task.

8. Common Issues and Pitfalls Records

Q1: How to install NVIDIA Container Toolkit?

Recommended method: Directly download the deb package. See Section 3.4. This method is the most stable and reproducible.

Alternative method (if using the apt source):

bashdistribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
curl -fsSL https://nvidia.github.io/libnvidia-container/$distribution/libnvidia-container.list | \
  sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | \
  sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
apt-get update
apt-get install -y nvidia-container-toolkit

Note: nvidia.github.io may be blocked in China.

Q2: WebRTC Livestream Black Screen

Solution: Switch to VNC remote desktop, which is more stable.

Q3: Isaac Lab cloning failed

Symptoms: fatal: unable to access 'https://github.com/isaac-sim/IsaacLab.git/'

Solution: Copy the downloaded file from the domestic mirror into the container after using it on the host machine.

bash# 在宿主机
cd ~
git clone --depth 1 --branch v1.4.1 https://gitclone.com/github.com/isaac-sim/IsaacLab.git
docker cp IsaacLab isaac-sim-gui:/isaac-sim/

Q4: Robomimic installation failed

Symptoms: Failed to build 'robomimic' or egl_probe compilation errors

Solution:

bash# 在宿主机下载
cd ~
wget https://github.com/ARISE-Initiative/robomimic/archive/refs/heads/master.zip
unzip master.zip
mv robomimic-master robomimic
docker cp robomimic isaac-sim-gui:/tmp/

# 在容器内安装 EGL 库和编译工具
docker exec -it isaac-sim-gui bash
apt update
apt install -y build-essential cmake pkg-config
apt install -y libegl1-mesa-dev libgl1-mesa-dev libgles2-mesa-dev

# 安装 robomimic
cd /tmp/robomimic
/isaac-sim/python.sh -m pip install -e .

Q5: Setup.py configuration error

Symptoms: IndentationError or Could not find a version that satisfies the requirement rsl-rl

Solution: Fix the configuration file according to Section 5.5 of the tutorial

bashcd /isaac-sim/IsaacLab
sed -i '46s/"rsl-rl@git+https:\/\/github.com\/leggedrobotics\/rsl_rl.git"/"rsl-rl-lib==2.3.0"/' source/extensions/omni.isaac.lab_tasks/setup.py
sed -i '53,54d' source/extensions/omni.isaac.lab_tasks/setup.py

Q6: Error in the latest version of Isaac Lab

Reason: The latest version requires Isaac Sim 4.5+. Solution: Use git clone --depth 1 --branch v1.4.1 to clone the specified version directly.

Q7: rsl-rl version conflict

Solution: Fixed in Section 5.5 of the tutorial, use rsl-rl-lib==2.3.0

Q8: HuggingFace cannot be accessed

Solution: Use the mirror HF_ENDPOINT=https://hf-mirror.com

Q9: GR00T model loading error - shape mismatch

Reason: The old version of the N1-2B model was downloaded, but the code is from the N1.6 version. Solution: Download the nvidia/GR00T-N1.6-3B model.

Q10: Flash-Attn compilation issues

Symptom 1: SSH/VNC connection drops during compilation, but CPU usage is high.

This is normal! Compilation consumes a lot of resources, causing the system to respond slowly.

Monitor CPU usage in the Alibaba Cloud console.
Wait until CPU drops below 5% to consider compilation complete.
Do not force a server restart! Otherwise, the compilation will be futile.

Symptom 2: CPU utilization is close to 0%, compilation stalls

Indicated that the compilation process exited abnormally and needs to be retried:

bashexport CUDA_HOME=/usr/local/cuda
export PATH=$CUDA_HOME/bin:$PATH
MAX_JOBS=4 pip install flash-attn --no-build-isolation -v

Symptom 3: `flash-attn` disappears after restarting

Compilation is interrupted and needs to be recompiled. If you don’t want to wait, you can skip flash-attn:

bashcd ~/Isaac-GR00T
pip install -e . --no-deps
pip install transformers==4.51.3 safetensors einops peft diffusers tyro omegaconf pandas dm-tree termcolor av albumentations huggingface_hub deepspeed accelerate
pip install click datasets gymnasium lmdb matplotlib msgpack-numpy pyzmq wandb torchcodec

Symptom 4: Installation succeeds, but there are many version warnings.

gr00t 0.1.0 requires flash-attn==2.7.4.post1, but you have flash-attn 2.8.3 which is incompatible.
...

Can be ignored! This is because GR00T’s setup.py locks very strict version numbers, but actually a slightly newer version can also work properly. As long as python -c "import flash_attn" doesn’t report any errors, it’s fine.

💡 flash-attn is a performance optimization component and is not required. Skipping it does not affect the functionality of GR00T.

Q11: MuJoCo cannot display the window

Solution:

Ensure VNC is running: netstat -tlnp | grep 5901
Set export DISPLAY=:1
Replace MUJOCO_GL in the code with glx, and onscreen with True

Q12: Unable to connect after VNC starts

Symptoms: websockify logs show Connection refused

Solution:

bash# 杀掉旧进程
vncserver -kill :1
pkill -9 websockify

# 修复 xstartup 配置
cat > ~/.vnc/xstartup << 'EOF'
#!/bin/sh
unset SESSION_MANAGER
unset DBUS_SESSION_BUS_ADDRESS
exec startxfce4
EOF
chmod +x ~/.vnc/xstartup

# 重启 VNC（注意 -localhost no 参数）
vncserver :1 -geometry 1920x1080 -depth 24 -localhost no
nohup websockify --web=/usr/share/novnc/ 6080 localhost:5901 > /dev/null 2>&1 &

Q13: Preemptible instance released

Suggestions:

Set a reasonable maximum price
Back up important data promptly
Use a data disk to store models and code

Q14: Insufficient disk space

Solutions:

Clean up Docker cache: docker system prune -a
Clean up pip cache: pip cache purge
Clean up conda cache: conda clean -a
It is recommended to use a system disk of 150GB+ or mount a separate data drive

Q15: VNC screen lagging

Note: VNC transmission of 3D simulation images will experience delays and frame loss, which is normal. VNC is mainly used for “verifying operational status,” not for smooth manipulation.

Advanced Solutions:

Use NoMachine (better performance)
Record the video and download it locally for viewing

Q16: GR00T-WholeBodyControl clone failed

Symptoms:

fatal: repository 'https://github.com/NVIDIA/GR00T-WholeBodyControl.git/' not found
error: RPC failed; curl 16 Error in the HTTP2 framing layer

Reason: The sub-module address is under NVlabs, not NVIDIA; or there are network issues.

Review: Section 7.1

Solution:

bashcd ~/Isaac-GR00T/external_dependencies

# 方法1：正确地址
git clone --depth 1 https://github.com/NVlabs/GR00T-WholeBodyControl.git

# 方法2（推荐）：用 ghproxy 加速
curl -L -o wbc.tar.gz https://ghproxy.cn/https://github.com/NVlabs/GR00T-WholeBodyControl/archive/refs/heads/main.tar.gz
tar -xzf wbc.tar.gz
mv GR00T-WholeBodyControl-main GR00T-WholeBodyControl
rm wbc.tar.gz

Q17: Transformers version is incompatible

Symptoms: AttributeError: 'Eagle3_VLConfig' object has no attribute '_attn_implementation_autoset'

Reason: The transformers version is too new.

Review: Section 6.3, confirm that transformers==4.51.3 was used during installation

Solution:

bashpip install transformers==4.51.3

Appendix: Quick Reference of Common Commands

bash# 启动 VNC
vncserver :1 -geometry 1920x1080 -depth 24 -localhost no
nohup websockify --web=/usr/share/novnc/ 6080 localhost:5901 > /dev/null 2>&1 &

# 进入 Isaac Sim 容器
docker start isaac-sim-gui
docker exec -it isaac-sim-gui bash

# 在容器内运行 Isaac Lab demo
export DISPLAY=:1
cd /isaac-sim/IsaacLab
./isaaclab.sh -p source/standalone/demos/arms.py

# 激活 GR00T 环境
conda activate groot

# 启动 GR00T Server
python gr00t/eval/run_gr00t_server.py \
  --model-path /root/groot_g1_model \
  --embodiment-tag UNITREE_G1 \
  --use-sim-policy-wrapper

Reference Links

Tutorial completion time: January 2026 *
Test environment: Alibaba Cloud ecs.gn7i-c16g1.4xlarge (A10 24GB) *

Complete Tutorial on Deploying Isaac Lab + GR00T on Alibaba Cloud

Table of Contents

1. Environment Overview

Final Architecture

Why is it designed this way?

Key Design Decisions

Deployment order (must be executed in sequence)

Hardware Requirements

2. Creating Alibaba Cloud Instances

2.0 Preparations

Register a NGC account (required to pull the Isaac Sim image)

2.1 Cost Description

2.2 Select Instance Specifications

2.2 Image Selection

2.3 Network Configuration

2.4 Security Group Configuration

3. Basic Environment Configuration

3.1 SSH Connection

3.2 Installing NVIDIA drivers (if not pre-installed in the image)

3.3 Installing Docker

3.4 Configuring Docker Domestic Images

3.5 Installing NVIDIA Container Toolkit

4. VNC Remote Desktop Configuration

4.1 Installing the Desktop Environment and VNC

4.2 Configure VNC

4.3 Accessing VNC

Method 1: SSH Tunnel (Recommended, safer)

Method 2: Direct access (port must be opened)

4.4 VNC Self-starting (Optional)

5. Deployment of Isaac Sim + Isaac Lab

5.1 Log in to NGC and pull images

5.2 Starting Containers

5.3 Preparation: Download dependencies on the host machine

5.4 Installing dependencies within the container

5.5 Fixing Isaac Lab configuration files

5.6 Installing Isaac Lab

5.7 Testing Isaac Lab

6. GR00T Environment Configuration

6.1 Installing Miniconda

6.2 Creating the GR00T Environment

Install flash-attn (optional, recommended to skip)

6.3 Installing Isaac-GR00T

Install GR00T

6.4 Download GR00T Model

6.5 Testing GR00T Inference

7. GR00T + MuJoCo Evaluation

7.1 Preparation Work

7.2 Modify the setup script and run it

7.3 Modify the code to enable real-time visualization

7.4 Run Evaluation

8. Common Issues and Pitfalls Records

Q1: How to install NVIDIA Container Toolkit?

Q2: WebRTC Livestream Black Screen

Q3: Isaac Lab cloning failed

Q4: Robomimic installation failed

Q5: Setup.py configuration error

Q6: Error in the latest version of Isaac Lab

Q7: rsl-rl version conflict

Q8: HuggingFace cannot be accessed

Q9: GR00T model loading error - shape mismatch

Q10: Flash-Attn compilation issues

Symptom 1: SSH/VNC connection drops during compilation, but CPU usage is high.

Symptom 2: CPU utilization is close to 0%, compilation stalls

Symptom 3: flash-attn disappears after restarting

Symptom 4: Installation succeeds, but there are many version warnings.

Q11: MuJoCo cannot display the window

Q12: Unable to connect after VNC starts

Q13: Preemptible instance released

Q14: Insufficient disk space

Q15: VNC screen lagging

Q16: GR00T-WholeBodyControl clone failed

Q17: Transformers version is incompatible

Appendix: Quick Reference of Common Commands

Reference Links

Symptom 3: `flash-attn` disappears after restarting