Corebrum OS — A Linux desktop for ROS 2, AI agents, and Physical AI

Hey ROS community!

I’ve been hacking on a side project called Corebrum OS (https://corebrum.com), and I figured it was time to share what I’ve been building!

It started with a simple question:

What if we had a Linux operating system built specifically for robotics and Physical AI, with ROS 2 and AI agents baked right in?

I’ve always liked the developer experience of Omarchy, so I started experimenting with a similar approach. But instead of Arch Linux, Corebrum runs on Ubuntu Server 24.04 and 26.04, with a lightweight Sway desktop.

The goal isn’t to reinvent ROS or create another robotics framework. It’s to make it easier to go from a fresh Linux installation to a working robot with perception, AI agents, and motion control.

What’s under the hood?

• ROS 2 — The foundation for robot communication, sensors, and control.

• RealSense — Integrated 3D stereo depth sensing.

• AI Agents — AI coding and reasoning agents that can interact with the operating system and robotics stack. The following AI Agents are built into the OS: Antigravity, Claude, Codex, Cursor, Grok, Hermes, OpenClaw, and Ollama

• Corebrum Synapse (formerly AgenticROS) — Connects the AI Agents to ROS 2 sensors, capabilities, and robot actions.

• Corebrum Skills — Reusable robotics capabilities that agents can discover and execute. (https://skills.corebrum.com)

• Corebrum Cloud — Remote connectivity, robot management, mapping, and teleoperation. (https://cloud.corebrum.com)

• Zenoh — A first-class communication layer for distributed robotics.

• Sway — A lightweight Wayland desktop inspired by the Omarchy developer experience.

• Developer Tools — A robotics-focused Linux environment for coding, visualization, and debugging.

The fun part: AI agents can control robots!

This is where things get interesting.

Corebrum Synapse, which started as my AgenticROS project, bridges AI agents with the ROS 2 ecosystem.

Instead of limiting agents to writing code or answering questions, Synapse lets them interact with real robots through ROS 2.

For example, an agent can:

• Start a RealSense camera and access sensor data.

• Find and follow a person.

• Navigate to a location.

• Execute robotic skills and chain them together into missions.

• Interact with robot hardware through reusable capabilities.

The AI agent handles higher-level reasoning while ROS 2 handles the underlying perception, navigation, and motion-control operations.

I’m also experimenting with multiple AI agents, rather than locking the system into one model or provider.

Corebrum Skills: Give your robot new abilities

One of my favorite parts of the project is Corebrum Skills.

Think of these as reusable building blocks for robot behavior.

Examples include:

• Follow Me

• Find Object

• Detect Humans

• Start/Stop SLAM

• Navigate to Pose

• Explore

• Dock to Charger

• MoveIt Pick

Skills can be combined into larger missions, allowing AI agents to accomplish more complex tasks without having to reinvent the underlying robotics code each time.

I’m also interested in making it easy for other developers to publish and share their own skills.

Corebrum Cloud: Robots don’t have to stay on your desk

I’ve also been hacking on Corebrum Cloud, which extends the experience beyond the local robot.

The idea is to make it easier to connect to robots remotely without having to manually configure a bunch of networking infrastructure.

Some of the capabilities I’m exploring include:

• Remote robot connectivity

• Browser-based teleoperation

• Remote terminal access

• Video and voice telepresence

• Connecting robots and AI agents across networks

It’s still evolving, but it’s been fun seeing how much of the robotics development experience can be made accessible remotely.

How small can we go?

I’ve been testing Corebrum on different hardware platforms, and so far I’ve gotten it running on:

• Raspberry Pi 5 (8GB)

• NVIDIA Jetson AGX Orin

• NVIDIA Jetson Thor

• x86-64 computers running Ubuntu Server 24.04 and 26.04

Having the same general environment running across ARM SBCs, NVIDIA edge AI hardware, and x86 machines has been a fun challenge.

What’s next?

I’m experimenting with a broader plugin architecture so developers can bring their own robotics capabilities into Corebrum and expose them to AI agents through Synapse.

I’m also interested in resource isolation, latency monitoring, and distributed robotics across multiple machines.

There’s plenty left to build, break, and improve!

Demo video

I’ve attached a short video showing Corebrum in action.

I’d love feedback from the ROS community, especially around the agent-to-ROS architecture, Zenoh, reusable skills, and making the environment useful across different robot platforms.

If you’re experimenting with ROS 2, AI agents, or Physical AI, I’d love to hear what you’re building too.

What would YOU want in a Linux OS built specifically for ROS developers and robots?

Happy hacking!