🦾 ROSClaw

The Universal OS Bridging Multimodal AI Agents with the Physical World.

🌐 English • 中文 • Architecture • Quick Start • Discord

"Teach Once, Embody Anywhere. Share Skills, Shape Reality."

⚠️ Project Status: Pre-Alpha (V0.1)

ROSClaw is currently in active development. We are NOT building another simple "LLM-to-ROS2 API wrapper" (there are plenty of those). We are building a full Embodied AI Operating System, featuring a MuJoCo-based Digital Twin Firewall, Reflex Handshake protocol for Multi-Agent coordination, and an Auto-EAP Data Flywheel.

Current release is a structural proof-of-concept. Production-grade ROS 2 Action servers, MoveIt 2 integration, and the full e-URDF firewall will be released in the upcoming v0.2.0 milestone. Stay tuned.

🌍 The Vision: Democratizing Physical AI

Phenomenal frameworks like Claude Code, OpenClaw, and WorkBuddy have democratized the digital world—empowering anyone to orchestrate AI teams to build software effortlessly.

ROSClaw brings this exact revolution to the Physical World.

We are not just building a bridge between LLMs and robots; we are building an Open Ecosystem for Physical Skills. If a developer in Tokyo teaches a robotic arm the "precision screwdriving" skill via ROSClaw, a factory worker in Berlin can instantly download that skill and deploy it on a completely different humanoid robot—no re-programming required.

By unifying heterogeneous hardware behind the universal Model Context Protocol (MCP) and abstracting physics through our OS kernel, we enable creators to share, iterate, and deploy Embodied AI across thousands of industries.

The future we imagine: A skill marketplace where physical intelligence flows as freely as software—teach once, embody everywhere.

💡 What does "ROS" in ROSClaw mean?

We transcend legacy middleware. In our architecture, the agent in the digital world is simply called an "Agent" (via OpenClaw). When that Agent enters the physical universe, it is called a "Robot".

ROSClaw is the Universal Embodied Operating System that facilitates this transition. We are NOT just another "LLM-to-ROS2 API wrapper." While traditional frameworks (like recent academic papers sharing the ROSClaw namesake) focus purely on translating text to ROS 2 topics, we treat legacy ROS 1, ROS 2, and IoT protocols merely as "Southbound Drivers."

Our core value lies in the OS Kernel: the Semantic-Physical Firewall, the Asynchronous Brain-Cerebellum Routing, and the Autonomous Data Flywheel. We are a "Super Plugin" that connects the entire physical ecosystem to any multimodal agent seamlessly.

---

✨ Core Innovations

ROSClaw is an Agent-Agnostic Embodied OS built on four pillars:

1. 🌐 Universal MCP Hub

Plug-and-play with ANY AI Agent framework. We translate complex ROS 2 topics and DDS streams into clean JSON schemas that Claude Code, OpenClaw, or any MCP-compatible agent can command natively.

2. 🧠 Asynchronous Brain-Cerebellum Routing

Decouples the Cognitive Brain (LLMs at ~1Hz) from the Physical Cerebellum (ROS 2/VLA at 1000Hz). Network latency or LLM delays never compromise physical stability.

3. 🛡️ Digital Twin Firewall (MuJoCo)

LLM hallucinations in the physical world are catastrophic. Before any command executes, it is fast-forwarded in a Headless Digital Twin (MuJoCo). If collision or torque overload is predicted, the action is blocked and the Agent self-corrects.

4. 🔄 Skill Flywheel (TODO)

Every execution feeds an Event-Driven Ring Buffer. Data is packaged into LeRobot formats to continuously fine-tune VLA models. Talk to Train—evolve your robot's physical intuition daily.

---

🗺️ Architecture: Agent-Agnostic by Design

graph TD
    subgraph Agents[Any MCP-Compatible Agent]
        CC[Claude Code]
        OC[OpenClaw]
        QB[QClaw / WorkBuddy]
        AC[AutoGen / Custom]
    end

    subgraph OS[ROSClaw OS Kernel: Layers 1-4]
        MCP[MCP Hub: JSON-RPC Bridge]
        DT[Digital Twin: MuJoCo Validation]
        RB[Ring Buffer: Data Capture]
    end

    subgraph Runtime[Physical Runtime: Layer 5]
        VLA[VLA Engine: OpenVLA/π0]
        ROS2[ROS 2 / DDS]
    end

    subgraph Hardware[Hardware Layer]
        G1[Unitree G1]
        UR5[UR5e Arm]
        Other[Your Robot]
    end

    CC & OC & QB & AC <-->|JSON-RPC| MCP
    MCP <-->|Physics Check| DT
    MCP <-->|Safe Execution| VLA
    VLA <-->|1000Hz Control| ROS2
    ROS2 <--> G1 & UR5 & Other
    ROS2 -.->|Events| RB

Loading

Key Insight: Layers 1-4 form the stable kernel. Any Agent (Layer 6+) can connect via MCP without hardware-specific knowledge.

---

🚀 Quick Start

Zero configuration. Native compatibility. Get your robot online in 30 seconds.

1. Install ROSClaw OS Kernel

curl -sSL https://rosclaw.io/get | bash

2. Plug into ANY Agent Framework

Claude Code:

claude mcp add rosclaw -- rosclaw-hub --auto-discover
# Then: "Claude, move the UR5 arm to home position and validate first"

OpenClaw / WorkBuddy (mcp_servers.json):

{
  "mcpServers": {
    "rosclaw-embodiment": {
      "command": "rosclaw-hub",
      "args": ["--enable-digital-twin"]
    }
  }
}

---

🎯 Roadmap: Where We're Going

Phase	Status	Key Deliverables
1	✅	Digital Twin Firewall, UR5 MCP Server, MuJoCo models
1.5	🚧	Testing (42 tests), CI/CD, PyPI release
2	📋	Data Flywheel, OpenVLA/π0 integration, Skill library
3	📋	G1/Panda support via sdk_to_mcp, ClawHub skill marketplace
4	🔮	Neural Twin, Multi-agent collaboration, TSN

Active TODOs

• Replace print with logging module
• YAML configuration for model paths
• PyPI release v0.1.0
• sdk_to_mcp integration docs

---

💎 Supported Embodiments & Ecosystem

We are actively unifying all hardware through our official south-bound drivers:

• Unitree G1 (via rosclaw-g1-dds-mcp)
• Universal Robots (UR5e) (via rosclaw-ur-ros2-mcp)
• General PTZ Gimbals (via rosclaw-gimbal-mcp)

🚀 sdk_to_mcp: Zero-Code Hardware Integration

Have a new robot with an SDK? Our sdk_to_mcp toolchain auto-generates MCP servers from official SDK documentation—no manual driver development required.

# Example: Generate MCP server from robot SDK docs
python -m sdk_to_mcp generate --sdk-doc robot_sdk.pdf --output rosclaw-newrobot-mcp

---

🛡️ Safety Architecture

Digital Twin Firewall

Every motion is validated in MuJoCo before physical execution:

from rosclaw.firewall import DigitalTwinFirewall, mujoco_firewall, SafetyLevel

# Method 1: Direct validation
firewall = DigitalTwinFirewall("src/rosclaw/specs/ur5e.xml")
result = firewall.validate_trajectory(trajectory_points)
if not result.is_safe:
    raise SafetyViolationError(f"Unsafe: {result.violation_details}")

# Method 2: Decorator
@mujoco_firewall(model_path="ur5e.xml", safety_level=SafetyLevel.STRICT)
def execute_motion(trajectory):
    # Only runs if validation passes
    ...

Validation Checks

• Collision Detection: Self-collision and environment collision
• Joint Limits: Position, velocity, and torque limits
• Workspace: TCP position within safe bounds
• Smoothness: Jerk and acceleration limits

---

🙏 Acknowledgements

ROSClaw stands on the shoulders of giants:

Academic Partners

Tongji University

Shanghai Research Institute of Autonomous Intelligent Unmanned Systems

This project is proudly supported by Tongji University and the Shanghai Research Institute of Autonomous Intelligent Unmanned Systems (SRIAS). Their commitment to advancing autonomous intelligent systems research provides the foundation for innovations like ROSClaw.

Open Source Community

• The Agent Ecosystem (OpenClaw, Claude Code, etc.): For pioneering the digital workflows that inspired our physical architecture.
• RoboClaw: For pioneering the Embodied closed-loop and Entangled Action Pairs (EAP).
• mjlab: For providing the blazingly fast MuJoCo backend that powers our Digital Twin Firewall.

---

Bridging AGI to the Physical Universe.
rosclaw.io