Robium

Physical AI skills for your agents

An open-source, continuously evolving collection of field-tested robotics expertise. Install robium as a plugin to empower your favorite AI coding agent with the robotics skills it needs. Covers ROS 2, Nav2, Gazebo, MuJoCo, NVIDIA Isaac Sim, Isaac Lab, and LeRobot, for Claude Code, Codex, Gemini CLI, and Cursor.

Install

Quick: clone the repository and install the native plugin or skills for every supported agent detected on your machine:

npx robium-ai setup                  # auto-detects your agents
npx robium-ai setup --agent codex    # or target one

The default clone lives at ~/robium. Target one agent with --agent.

Native install: clone the repository, then use your agent's own package flow where one is available:

git clone https://github.com/robium-ai/robium ~/robium

# Claude Code: plugin with skills, architect agent, and capture hooks
claude plugin marketplace add ~/robium && claude plugin install robium@robium

# Codex: native plugin
codex plugin marketplace add ~/robium && codex plugin add robium@robium

# Gemini CLI: extension with automatic updates
gemini extensions install https://github.com/robium-ai/robium --auto-update

# Cursor: versioned skills
mkdir -p ~/.cursor/skills && ln -s ~/robium/skills/* ~/.cursor/skills/

The repository remains the source of truth. Pull it to update symlink-based installs; for native plugins, re-run npx robium-ai setup -y and start a new session so the host refreshes its installed cache.

Your application stays in its own repository; Robium lives beside it. Use the reference apps as starting points and contribute reusable fixes back.

How it fits

Robium provides robotics expertise. Your project provides the context. Your AI coding agent handles architecture, implementation, simulation, testing, and deployment.

Captured build learnings can improve future skill versions. See the workflow at robium.ai.

What's inside

robium/
├── skills/          the catalog: versioned, hand-crafted, validator-enforced
├── agents/          robium-architect: researches the stack, writes your brief
├── .claude-plugin/  Claude Code package
├── .codex-plugin/   Codex package manifest
├── .agents/plugins/ Codex-native repository marketplace
├── hooks/           shared Claude/Codex learning-capture hooks
├── AGENTS.md        canonical Codex-native maintainer guidance
├── gemini-extension.json  Gemini CLI extension
├── learnings/       field evidence from real builds, input to the learning loop
└── cli/             npx robium-ai: setup, doctor, skill search

The reference applications live in robium-ai/robium-apps and the robium.ai site + live-demo infrastructure in robium-ai/robium-website.

The catalog in one view: every skill is one folder under skills/, browsable on robium.ai:

Umbrella skills own decisions (which simulator, where data comes from, how to test); tool skills own the mechanics of one library. architect is the entry point and routes to everything else.

A catalog that maintains itself

Robotics guidance rots fast: APIs move, versions pair differently, commands change shape. robium is built to notice:

• Capture: hooks record what broke and what fixed it during real build sessions; learnings/ holds the evidence.
• Mine: the ecosystem's proven patterns are read out of real repos, with citations that must still hold at the pinned commit.
• Absorb: evidence-gated pull requests fold both back into the versioned skills. Agents draft; a human merges every change.
• Verify: version facts are checked against live upstream docs at authoring time, and each claim states how it was verified. Prior skill versions stay browsable under archive/.

Contributing

The contribution unit is small on purpose: one skill, no build system. Pick a robotics tool you know, copy the template, pass the validator:

uv run skills/skill-author/scripts/validate_skills.py

CONTRIBUTING.md has the five-step walkthrough; good-first-skill issues are the on-ramp. Questions: Discord or Discussions.

License

MIT. See CONTRIBUTING.md for the skill format, quality bar, and development workflow.