autonomous-ai/lampOS

The open-source operating system for physical AI.

autonomous-ai/lampOS – Physical AI Operating System

What it is – A full‑stack software platform that turns a robot (the “Lamp”, “Intern”, Reachy Mini, or any custom hardware) into a physical embodiment for large‑language‑model agents such as Hermes, Claude Code, Codex, OpenClaw, etc. The OS handles the low‑level drivers, safety checks, voice‑in/voice‑out pipelines, and a “skill” system that lets the agent control lights, motors, cameras, microphones and other sensors.

Key components

  • HAL (Hardware Abstraction Layer) – Go services that expose hardware via HTTP routes, enforce a safety gate, and run realtime voice pipelines (Gemini Live, OpenAI Realtime, Pipecat, etc.).
  • Agentic runtimes – Plug‑in implementations (Hermes, OpenClaw, PicoClaw, Codex, Claude Code, OpenCode) that present a uniform AgentGateway API to the OS.
  • Skills – Markdown‑defined behaviours (skills/servo‑tracking/, skills/harness‑use/, …) that the agent can invoke. Skills emit hardware markers ([HW:/path:{json}]) which the OS translates into driver calls.
  • OS server – Central Go daemon (os‑server) on port 5000 that routes voice/text requests, runs intent validation, delegates to Harness (paired desktop agents) or to the main LLM runtime, and strips hardware markers.
  • Apps & UI – A web UI (system/web/) and a mobile “Autonomous” app for robot provisioning, Wi‑Fi setup, skill store, and live monitoring.
  • Cross‑robot support – Built‑in support for three reference bodies (Lamp, Intern, Reachy Mini) and a “bring‑your‑own‑robot” guide that requires four markdown spec files (ROBOT.md, SOUL.md, SAFETY.md, SKILL.md).

How you use it

  1. Add a robot via the mobile app (or SSH for Reachy Mini). The app provisions Wi‑Fi and installs the OS on the device.
  2. Choose an LLM agent (Hermes, Claude Code, etc.) and configure voice, personality, and tools.
  3. Talk to the robot – voice or text commands are parsed locally; simple intents (e.g., “turn off the lights”) are executed directly, while more ambiguous requests are forwarded to the LLM.
  4. Install or create skills – pick from the Skill Store or write a new markdown skill; the OS distributes it to any compatible robot.
  5. Extend – add new hardware drivers, boards, or whole new robots by following the documented markdown contracts.

Why it matters – It provides a reusable, open‑source stack for “embodied AI”: turning conversational agents into agents that can see, hear, speak, move, and interact with the physical world. By separating hardware, safety, intent handling, and the LLM runtime, developers can experiment with new bodies or new agents without rewriting the whole system.

License – Apache‑2.0 for most of the repo; the HAL drivers are GPL‑3.0.

Related

  • Project
    ros/meta-ros

    A collection of OpenEmbedded layers that enable the Robot Operating System (ROS) to be built into custom embedded Linux releases using the Yocto Project.

  • Project
    GalaxyGeneralRobotics/Humanoid-GPT

    A GPT-style humanoid motion Transformer pre-trained on 2 billion frames to enable zero-shot whole-body motion tracking for humanoid robots like the Unitree G1.

  • Project
    OpenLegged/URDF-Studio

    URDF Studio is a browser‑based robot authoring tool that lets you edit kinematics, geometry, and hardware parameters, visualise robots with a React‑Three‑Fiber canvas, import/export many robot description formats, and optionally use an AI assistant to generate or inspect robots. It’s built with React, TypeScript, Vite, Three.js and includes a reusable `react‑robot‑canvas` package.

  • Project
    makerforgetech/modular-biped

    An open-source, modular bipedal robot platform that integrates Raspberry Pi and Arduino to support AI-driven features like computer vision, LLM interaction, and autonomous movement.