Skip to main content
This quickstart gets dimOS running on your laptop. You install it, then play back a recorded Unitree Go2 session and watch the robot map and navigate an office in a live visualization. You do not need a robot or a GPU for this. When you are ready for more, the same install works with physics simulation, a real robot, or an LLM agent you can talk to.

​
Before you begin

You need a machine running Ubuntu 22.04 or newer or macOS 12.6 or newer, with Python 3.12, about 10 GB of free disk, and 16 GB of RAM. A GPU is only required later for perception and AI features, so any reasonably modern laptop can run this quickstart. The full hardware matrix, including tested configurations and Jetson boards, is on the system requirements page. If you use a coding agent such as Claude Code or OpenClaw, point it at the repository’s AGENTS.md so it understands the codebase conventions.

​
Install dimOS

There are two ways to install, and you only need one of them. The installer script walks you through the whole setup interactively. It installs the system packages dimOS needs (such as git-lfs and portaudio), installs the uv Python package manager if you don’t have it, creates a virtual environment, and installs dimOS into it with the extras you choose.
If you prefer to read the script before running it, it lives at scripts/install.sh in the repository. When the installer finishes, activate the environment it created and skip ahead to Run your first replay. Everything in Option B has already been done for you.

​
Option B: manual install

First install the system dependencies for your OS by following the matching guide: Then create a Python 3.12 environment. The examples use uv, though plain python -m venv and pip work the same way:
Finally install dimOS with the extras this quickstart uses:
Extras keep the install lean. The base extra brings the runtime, modules, transports, and CLI, while unitree adds WebRTC support and the skills for the Go2 and G1 robots, whether real or replayed.

​
Run your first replay

dimOS applications are launched from blueprints. A blueprint is a ready to run bundle of modules that you start by name with dimos run. The --replay flag feeds it recorded sensor data instead of connecting to a robot:
What you should see: a Rerun visualization window opens, and after a short wait it fills with the robot’s camera feed, a LiDAR point cloud, and a map of an office being built up as the robot navigates through it, with its planned path drawn on top.
On the first run, roughly 75 MB of recorded session data is downloaded before anything appears, so the Rerun window may stay black for a minute or two. That is normal. If it stays black well beyond the download, check the terminal output for errors.
Congratulations, you have a full dimOS navigation stack running on recorded data. Everything below is optional and independent, so pick whichever branch matches your goal.

​
Next steps

​
Simulation (MuJoCo)

Instead of replaying recorded data, you can run the robot in a physics simulation. Install the sim extra and pass --simulation:

​
Real robot

With a Unitree Go2 on the same network, point dimOS at its IP address and drop the --replay flag:
Before driving real hardware, read the Unitree Go2 platform guide. It covers network setup, latency, time sync, and the safety habits that keep you and the robot out of trouble. Do not skip it.

​
LLM agent

The agentic blueprints add an LLM agent that controls the robot through natural language. The default agent uses OpenAI’s gpt-4o, so you need an OPENAI_API_KEY in your environment before starting it. Other providers and local models are covered in the agents guide.
Every robot skill is also exposed over MCP, which means external tools and coding agents can call them directly:
Manage the background run with dimos status, dimos log -f, and dimos stop. The full command reference is in the CLI guide.

​
More blueprints to try

To learn how blueprints are composed, or to write your own, see the blueprints guide.

​
What next?

Add an LLM agent

Natural language control, agent configuration, and MCP-exposed skills.

Pick your platform

Hardware support and bring-up guides for the Go2 quadruped and G1 humanoid.

Core concepts

Modules, streams, and blueprints, the building blocks behind every workflow.

Capabilities

Navigation, perception, spatial memory, and manipulation in depth.