Installation

Prerequisites

• Python 3.10 or higher
• PyTorch 2.1 or higher
• CUDA 11.8 or higher (for GPU support)

Install via pip

pip install vlarobot

Install with simulation support

pip install vlarobot[sim]

Install from source

git clone https://github.com/vlarobot/vlarobot.git
cd vlarobot
pip install -e ".[dev]"

Install with Docker

For a reproducible environment with all dependencies pre-configured:

# Pull the official image
docker pull vlarobot/vlarobot:latest

# Or build from source
docker build -t vlarobot .

# Run with GPU support
docker run --gpus all -it vlarobot bash

Docker Compose is also available for multi-service setups:

docker compose up -d

Platform-Specific Notes

Linux (Ubuntu 22.04+)

Recommended platform. Full GPU support with CUDA. Install NVIDIA drivers and CUDA toolkit before installing vlarobot.

macOS

CPU-only inference is supported. Training requires a Linux machine with GPU. Install via pip — MuJoCo simulation works on Apple Silicon.

Windows

WSL2 with Ubuntu is recommended. Native Windows support is experimental. GPU passthrough via WSL2 works with CUDA 12+.

Verify installation

vlarobot --version
# vlarobot, version 0.1.0

vlarobot presets
# Available model presets:
#   - openvla-7b
#   - smolvla-450m
#   - dream-vla-7b

Dependencies

Core dependencies are installed automatically:

Troubleshooting

CUDA not found — Ensure NVIDIA drivers and CUDA toolkit are installed. Run nvidia-smi to verify. PyTorch must be installed with CUDA support: pip install torch --index-url https://download.pytorch.org/whl/cu118

Import errors after install — Try reinstalling in a clean virtual environment: python -m venv .venv && source .venv/bin/activate && pip install vlarobot

Out of memory during inference — Use a smaller model (SmolVLA-450M runs on 8GB VRAM) or enable gradient checkpointing.

MuJoCo rendering issues — Install mujoco and set MUJOCO_GL=egl for headless rendering.