🛠️ Installation Guide#

Important

We are actively fixing mistakes in the document. If you find any errors in the documentation, please feel free to open an issue. Your help in improving the document is greatly appreciated 🥰.

😄 Don’t worry — both Quick Installation and Dataset Preparation are beginner-friendly.

Prerequisites#

InternManip works across most hardware setups. Just note the following exceptions:

  • GenManip Benchmark must run on NVIDIA RTX series GPUs (e.g., RTX 4090).

  • GR00T requires CUDA 12.4 installed system-wide (not via Conda).

Overall Requirements#

  • OS: Ubuntu 20.04/22.04

  • GPU Compatibility:

GPU Model Training & Inference Simulation
CALVIN Simpler-Env Genmanip
NVIDIA RTX Series
(Driver: 535.216.01+ ) 		✅ ✅ ✅ ✅
NVIDIA V/A/H100 ✅ ✅ ✅ ❌

Note

We provide a flexible installation tool for users who want to use InternManip for different purposes. Users can choose to install the training and inference environment, and the individual simulation environment independently.

Model-Specific Requirements#

Models Minimum GPU Requirement System RAM
(Train/Inference) 		CUDA
Training (Full) Training (LoRA) Inference
GR00T-N1/1.5 RTX 4090 / A100 - RTX 3090 / A100 24GB / 8GB ⚠️ 12.4
Pi-0 RTX 4090 (48G) / A100 RTX 4090 RTX 3090 / A100 70GB / 8GB -
Diffusion Policy (DP) RTX 2080 - RTX 2070 16GB / 8GB -

Quick Installation#

We provide a unified installation script that handles environment setup and dependency installation automatically.

Step 1: Clone the Repository and Install uv#

# Clone the main repository
git clone https://github.com/internrobotics/internmanip.git
cd internmanip

# Initialize and update submodules
git submodule update --init --recursive

# Skip the following commands if you have installed uv
# Install uv package manager
curl -LsSf https://astral.sh/uv/install.sh | sh

# Restart your shell or source the profile
source ~/.bashrc

Step 2: Run the Installation Script#

# Make the script executable
chmod +x install.sh

# View available options
./install.sh --help

🧠 Full Installation (Advanced users):#

./install.sh --all

Available Installation Options:#

Usage:
  --calvin [NAME]         Create Calvin benchmark virtual environment and install dependencies
  --simpler-env [NAME]    Create SimplerEnv benchmark virtual environment and install dependencies
  --genmanip [NAME]       Create GenManip benchmark virtual environment and install dependencies
  --model [NAME]          Create model virtual environment and install dependencies
  --all                   Create all virtual environments and install dependencies (recommended for advanced users)
  --beginner              Create beginner virtual environments and install dependencies (without genmanip, recommended for beginners)

Customization Options:
  --venv-dir DIR          Set custom virtual environment root directory (default: .venv)
  --python-version V      Set default Python version (recommended default: 3.10)

Examples:
  ./install.sh --venv-dir ./my_envs --model
  ./install.sh --calvin calvin-test --model model-test
  ./install.sh --python-version 3.10 --calvin calvin-dev --simpler-env simpler-dev
  ./install.sh --all
  ./install.sh --beginner
  --help                  Show help information

Selective Installation:#

# Install only specific components
./install.sh --gr00t --genmanip

Activate Virtual Environment#

After installation, virtual environments are created in the .venv directory by default.

# List available environments
ls .venv/

# Activate a specific environment
source .venv/{environment_name}/bin/activate

# Example: Activate GR00T environment
source .venv/gr00t/bin/activate

# Deactivate environment
deactivate

Optionally, users can customize the virtual environments directory path by passing the --venv-dir {path} option when executing install.sh.

./install.sh --venv-dir ./my_envs --model

Verification (WIP)#

To check your installation, you can evaluate the pretrained Pi-0 on the Simpler-Env benchmark using the following command:

python scripts/eval/start_evaluator.py --config run_configs/examples/internmanip_demo.py

If it installed properly, you can find the evaluation results and the visualization in the eval_results/bridgedata_v2/pi0 directory:

🎉 Congratulations! You have successfully installed InternManip.

Dataset Preparation (WIP)#

Automatic Download#

Datasets, model weights, and benchmark assets are automatically downloaded when running the code for the first time. The default download location is ${repo_root}/data. The system will prompt you to download required datasets.

Warning

Please ensure you have enough disk space available before starting the download. For cluster users, we recommend creating a symbolic link to a data storage for ${repo_root}/data to avoid disk space issues.

Manual Download#

If you prefer manual dataset preparation:

  1. Visit our platform: Dataset Platform

  2. Download datasets based on your needs: